Accurate identification of protein particles in cryo-electron microscopy (cryo-EM) micrographs is crucial for high-resolution structure determination, but remains challenging due to the heavy reliance on extensive annotated datasets and the difficulty of ensuring robustness under low signal-to-noise ratio (SNR) conditions. Current approaches require large annotations and exhibit poor generalization to new protein targets. We present CryoFSL (Cryo-EM Few Shot-Learning), a novel few-shot learning framework built on Segment Anything Model 2 with lightweight adapters, enabling robust particle picking with as few as five labeled micrographs and significantly reducing the annotation burden. The framework's hierarchical adapter design supports dynamic feature modulation for low-SNR and heterogeneous conditions, resolving the trade-off between annotation burden and performance. CryoFSL surpasses both traditional template-based methods and state-of-the-art deep learning models across diverse proteins in the few-shot learning setting, achieving superior recall, precision, and 3D reconstruction resolution with minimal supervision. It maintains stability across heterogeneous micrographs and consistently detects high-quality particles with fewer false-positives. Notably, CryoFSL achieves competitive resolution in density map reconstruction with just a fraction of the particles picked by other methods, redefining efficiency and quality in cryo-EM analysis. This work paves the way for scalable, generalizable, and annotation-efficient particle-picking pipelines. The code is available at https://github.com/biplabpoudel25/CryoFSL.

MOTIVATION: Predicting protein function is a fundamental and challenging task that requires integrating diverse biological data modalities to capture complex functional relationships. Traditional machine learning methods often rely on single modalities or combine only a limited number (typically two), without aligning them in a unified representation, thereby constraining predictive accuracy. Moreover, most existing machine learning approaches are limited to preselected subsets of Gene Ontology (GO) function terms with sufficient annotations, making the prediction of novel function terms a persistent challenge. RESULTS: Here, we present FunBind, a multimodal AI model that jointly learns from five modalities, i.e., protein sequences, textual descriptions, domain annotations, structures, and GO terms, to enhance prediction accuracy and infer previously unseen functions. FunBind operates in two modes: (1) self-supervised pretraining using contrastive learning to align the sequence modality with other heterogeneous modalities in a unified latent space, enabling unsupervised zero-shot function prediction, and (2) supervised fine-tuning of the pretrained model to leverage all non-function modalities for comprehensive and accurate function classification. Our results show that FunBind's zero-shot capabilities allow it to generalize effectively to novel function terms never encountered before, while its joint multimodal fine-tuning strategy outperforms single-modality models and current state-of-the-art deep learning methods in typical function prediction settings. AVAILABILITY: https://github.com/jianlin-cheng/FunBind.

MOTIVATION: Deciphering gene regulatory networks (GRNs) from single-cell transcriptomics data remains a fundamental challenge in computational biology. It is hindered by data sparsity, high dimensionality, and the lack of scalable, generalizable inference models. To address this, we present GRNFormer, a generalizable graph transformer framework for accurate GRN inference from transcriptomics data across species, cell types, and platforms without requiring cell-type annotations or prior regulatory information. RESULTS: GRNFormer integrates a transformer-based gene expression encoder (Gene-Transcoder) with a variational graph autoencoder (GraViTAE) employing pairwise attention to jointly learn the representations of genes (nodes) and their co-expression relationships (edges). Leveraging TF-Walker, a transcription factor-anchored subgraph sampling strategy, it effectively captures gene regulatory interactions from either single-cell or bulk RNA-seq datasets. Benchmarking on standard datasets demonstrates that GRNFormer outperforms existing traditional and deep learning state-of-the-art methods in blind evaluations, achieving average sampled area under the receiver operating characteristic curve (Sampled_AUROC) and sampled area under the precision-recall curve (Sampled_AUPRC) values between 0.90 and 0.98 as well as 0.87-0.98 average sampled F1 score. The model robustly recovers both known and novel regulatory networks, including pluripotency circuits in human embryonic stem cells (hESCs) and immune cell modules in peripheral blood mononuclear cells (PBMCs). The architecture enables scalable, biologically interpretable GRN inference across various datasets, cell types, and species, establishing GRNFormer as a robust and transferable tool for network biology. AVAILABILITY AND IMPLEMENTATION: GRNFormer is available on GitHub (https://github.com/BioinfoMachineLearning/GRNformer); the version used in this work is archived on Zenodo (https://doi.org/10.5281/zenodo.18868395), with evaluation resources for reproducibility.

Predicting protein complex structures is essential for protein function analysis, protein design, and drug discovery. While AI methods like AlphaFold can predict accurate structural models for many protein complexes, reliably estimating the quality of these predicted models (estimation of model accuracy, or EMA) for model ranking and selection remains a major challenge. A key barrier to developing effective machine learning-based EMA methods is the lack of large, diverse, and well-annotated datasets for training and evaluation. To address this gap, we introduce PSBench, a benchmark suite comprising five large-scale, labeled datasets, four of which were generated during the 15th and 16th community-wide Critical Assessment of Protein Structure Prediction (CASP15 and CASP16), and one curated for new Protein Data Bank (PDB) entries deposited between July 2024 and August 2025. PSBench includes over 1.4 million structural models covering a wide range of protein sequence lengths, complex stoichiometries, functional classes, and modeling difficulties. Each model is annotated with multiple complementary quality scores at the global, local, and interface levels. PSBench also provides multiple evaluation metrics and baseline EMA methods to facilitate rigorous comparisons. To demonstrate PSBench’s utility, we trained and evaluated GATE, a graph transformer-based EMA method, on the CASP15 data. GATE was blindly tested in CASP16 (2024), where it ranked among the top-performing EMA methods. These results highlight PSBench as a valuable resource for advancing EMA research in protein complex modeling. PSBench is publicly available at: https://github.com/BioinfoMachineLearning/PSBench.

Cryo-electron microscopy (cryo-EM) has revolutionized structural biology by enabling the determination of high-resolution 3-Dimensional (3D) structures of large biological macromolecules. Protein particle picking, the process of identifying individual protein particles in cryo-EM micrographs for building protein structures, has progressed from manual and template-based methods to sophisticated artificial intelligence (AI)-driven approaches in recent years. This review critically examines the evolution and current state of cryo-EM particle picking methods, with an emphasis on the impact of AI. We conducted a comparative evaluation of popular AI-based particle picking methods, using both general machine learning metrics and specific cryo-EM structure determination metrics. This analysis involved constructing the 3D density map from the picked protein particles and assessing the obtained resolution and particle orientation diversity, underscoring the significant impact of AI on cryo-EM particle picking. Despite the advancements, we also identified key obstacles, such as handling complex micrographs with small proteins. The analysis provides insights into the future development of more sophisticated and fully automated AI methods in cryo-EM particle recognition.

Gene Regulatory Networks (GRNs) are intricate biological systems that control gene expression and regulation in response to environmental and developmental cues. Advances in computational biology, coupled with high-throughput sequencing technologies, have significantly improved the accuracy of GRN inference and modeling. Modern approaches increasingly leverage artificial intelligence (AI), particularly machine learning techniques-including supervised, unsupervised, semi-supervised, and contrastive learning-to analyze large-scale omics data and uncover regulatory gene interactions. To support both the application of GRN inference in studying gene regulation and the development of novel machine learning methods, we present a comprehensive review of machine learning-based GRN inference methodologies, along with the datasets and evaluation metrics commonly used. Special emphasis is placed on the emerging role of cutting-edge deep learning techniques in enhancing inference performance. The major challenges and potential future directions for improving GRN inference are also discussed.

MOTIVATION: Powerful generative AI models of protein-ligand structure have recently been proposed, but few of these methods support both flexible protein-ligand docking and affinity estimation. Of those that do, none can directly model multiple binding ligands concurrently or have been rigorously benchmarked on pharmacologically relevant drug targets, hindering their widespread adoption in drug discovery efforts. RESULTS: In this work, we propose FlowDock, the first deep geometric generative model based on conditional flow matching (CFM) that learns to directly map unbound (apo) structures to their bound (holo) counterparts for an arbitrary number of binding ligands. Furthermore, FlowDock provides predicted structural confidence scores and binding affinity values with each of its generated protein-ligand complex structures, enabling fast virtual screening of new (multi-ligand) drug targets. For the well-known PoseBusters Benchmark dataset, FlowDock outperforms single-sequence AlphaFold 3 (AF3) with a 51% blind docking success rate using unbound (apo) protein input structures and without any information derived from multiple sequence alignments, and for the challenging new DockGen-E dataset, FlowDock outperforms single-sequence AF3 and matches single-sequence Chai-1 for binding pocket generalization. Additionally, in the ligand category of the 16th community-wide Critical Assessment of Techniques for Structure Prediction, FlowDock ranked among the top-5 methods for pharmacological binding affinity estimation across 140 protein-ligand complexes, demonstrating the efficacy of its learned representations in virtual screening. AVAILABILITY AND IMPLEMENTATION: Source code, data, and pre-trained models are available at https://github.com/BioinfoMachineLearning/FlowDock.

MOTIVATION: Cryogenic electron microscopy (cryo-EM) is a core experimental technique used to determine the structure of macromolecules such as proteins. However, the effectiveness of cryo-EM is often hindered by the noise and missing density values in cryo-EM density maps caused by experimental conditions such as low contrast and conformational heterogeneity. Although various global and local map-sharpening techniques are widely employed to improve cryo-EM density maps, it is still challenging to efficiently improve their quality for building better protein structures from them. RESULTS: In this study, we introduce CryoTEN-a 3D UNETR++ style transformer to improve cryo-EM maps effectively. CryoTEN is trained using a diverse set of 1295 cryo-EM maps as inputs and their corresponding simulated maps generated from known protein structures as targets. An independent test set containing 150 maps is used to evaluate CryoTEN, and the results demonstrate that it can robustly enhance the quality of cryo-EM density maps. In addition, automatic de novo protein structure modeling shows that protein structures built from the density maps processed by CryoTEN have substantially better quality than those built from the original maps. Compared to the existing state-of-the-art deep learning methods for enhancing cryo-EM density maps, CryoTEN ranks second in improving the quality of density maps, while running >10 times faster and requiring much less GPU memory than them. AVAILABILITY AND IMPLEMENTATION: The source code and data are freely available at https://github.com/jianlin-cheng/cryoten.

MOTIVATION: Proteins interact with a variety of molecules, including other proteins, DNAs, RNAs, ligands, ions, and lipids. These interactions play a crucial role in cellular communication, metabolic regulation, gene regulation, and structural integrity, making proteins fundamental to nearly all biological functions. Accurately predicting protein interaction (binding) sites is essential for understanding protein interaction and function. RESULTS: In this work, we introduce MPBind, a multitask protein binding site prediction method, which integrates protein language models (PLMs) that can extract structural and functional information from sequences and equivariant graph neural networks (EGNNs) that can effectively capture geometric features of 3D protein structures. Through multitask learning, it can predict binding sites on proteins that interact with five key categories of binding partners: proteins, DNA/RNA, ligands, lipids, and ions. MPBind generalizes across the five molecular classes with state-of-the-art accuracy, achieving AUROC scores of 0.83 and 0.81 for protein-protein and protein-DNA/RNA-binding site prediction, respectively. Moreover, MPBind outperforms both general and task-specific binding site prediction methods, making it a useful, versatile tool for protein binding site prediction. AVAILABILITY AND IMPLEMENTATION: The source code of MPBind is available at the GitHub repository: https://github.com/jianlin-cheng/MPBind.

We introduce ProteinWorkshop, a comprehensive benchmark suite for representation learning on protein structures with Geometric Graph Neural Networks. We consider large-scale pre-training and downstream tasks on both experimental and predicted structures to enable the systematic evaluation of the quality of the learned structural representation and their usefulness in capturing functional relationships for downstream tasks. We find that: (1) large-scale pretraining on AlphaFold structures and auxiliary tasks consistently improve the performance of both rotation-invariant and equivariant GNNs, and (2) more expressive equivariant GNNs benefit from pretraining to a greater extent compared to invariant models. We aim to establish a common ground for the machine learning and computational biology communities to rigorously compare and advance protein structure representation learning. Our open-source codebase reduces the barrier to entry for working with large protein structure datasets by providing: (1) storage-efficient dataloaders for large-scale structural databases including AlphaFoldDB and ESM Atlas, as well as (2) utilities for constructing new tasks from the entire PDB. ProteinWorkshop is available at: github.com/a-r-j/ProteinWorkshop.

Picking protein particles in cryo-electron microscopy (cryo-EM) micrographs is a crucial step in the cryo-EM-based structure determination. However, existing methods trained on a limited amount of cryo-EM data still cannot accurately pick protein particles from noisy cryo-EM images. The general foundational artificial intelligence-based image segmentation model such as Meta's Segment Anything Model (SAM) cannot segment protein particles well because their training data do not include cryo-EM images. Here, we present a novel approach (CryoSegNet) of integrating an attention-gated U-shape network (U-Net) specially designed and trained for cryo-EM particle picking and the SAM. The U-Net is first trained on a large cryo-EM image dataset and then used to generate input from original cryo-EM images for SAM to make particle pickings. CryoSegNet shows both high precision and recall in segmenting protein particles from cryo-EM micrographs, irrespective of protein type, shape and size. On several independent datasets of various protein types, CryoSegNet outperforms two top machine learning particle pickers crYOLO and Topaz as well as SAM itself. The average resolution of density maps reconstructed from the particles picked by CryoSegNet is 3.33 Å, 7% better than 3.58 Å of Topaz and 14% better than 3.87 Å of crYOLO. It is publicly available at https://github.com/jianlin-cheng/CryoSegNet.

The genome-wide single-cell chromosome conformation capture technique, i.e. single-cell Hi-C (ScHi-C), was recently developed to interrogate the conformation of the genome of individual cells. However, single-cell Hi-C data are much sparser than bulk Hi-C data of a population of cells, and noise in single-cell Hi-C makes it difficult to apply and analyze them in biological research. Here, we developed the first generative diffusion models (HiCDiff) to denoise single-cell Hi-C data in the form of chromosomal contact matrices. HiCDiff uses a deep residual network to remove the noise in the reverse process of diffusion and can be trained in both unsupervised and supervised learning modes. Benchmarked on several single-cell Hi-C test datasets, the diffusion models substantially remove the noise in single-cell Hi-C data. The unsupervised HiCDiff outperforms most supervised non-diffusion deep learning methods and achieves the performance comparable to the state-of-the-art supervised deep learning method in terms of multiple metrics, demonstrating that diffusion models are a useful approach to denoising single-cell Hi-C data. Moreover, its good performance holds on denoising bulk Hi-C data.

MOTIVATION: Cryo-electron microscopy (cryo-EM) is a powerful technique for determining the structures of large protein complexes. Picking single protein particles from cryo-EM micrographs (images) is a crucial step in reconstructing protein structures from them. However, the widely used template-based particle picking process requires some manual particle picking and is labor-intensive and time-consuming. Though machine learning and artificial intelligence (AI) can potentially automate particle picking, the current AI methods pick particles with low precision or low recall. The erroneously picked particles can severely reduce the quality of reconstructed protein structures, especially for the micrographs with low signal-to-noise ratio. RESULTS: To address these shortcomings, we devised CryoTransformer based on transformers, residual networks, and image processing techniques to accurately pick protein particles from cryo-EM micrographs. CryoTransformer was trained and tested on the largest labeled cryo-EM protein particle dataset-CryoPPP. It outperforms the current state-of-the-art machine learning methods of particle picking in terms of the resolution of 3D density maps reconstructed from the picked particles as well as F1-score, and is poised to facilitate the automation of the cryo-EM protein particle picking. AVAILABILITY AND IMPLEMENTATION: The source code and data for CryoTransformer are openly available at: https://github.com/jianlin-cheng/CryoTransformer.

MOTIVATION: The field of geometric deep learning has recently had a profound impact on several scientific domains such as protein structure prediction and design, leading to methodological advancements within and outside of the realm of traditional machine learning. Within this spirit, in this work, we introduce GCPNet, a new chirality-aware SE(3)-equivariant graph neural network designed for representation learning of 3D biomolecular graphs. We show that GCPNet, unlike previous representation learning methods for 3D biomolecules, is widely applicable to a variety of invariant or equivariant node-level, edge-level, and graph-level tasks on biomolecular structures while being able to (1) learn important chiral properties of 3D molecules and (2) detect external force fields. RESULTS: Across four distinct molecular-geometric tasks, we demonstrate that GCPNet's predictions (1) for protein-ligand binding affinity achieve a statistically significant correlation of 0.608, more than 5%, greater than current state-of-the-art methods; (2) for protein structure ranking achieve statistically significant target-local and dataset-global correlations of 0.616 and 0.871, respectively; (3) for Newtownian many-body systems modeling achieve a task-averaged mean squared error less than 0.01, more than 15% better than current methods; and (4) for molecular chirality recognition achieve a state-of-the-art prediction accuracy of 98.7%, better than any other machine learning method to date. AVAILABILITY AND IMPLEMENTATION: The source code, data, and instructions to train new models or reproduce our results are freely available at https://github.com/BioinfoMachineLearning/GCPNet.

Semi-supervised learning (SSL) has recently received increased attention from machine learning researchers. By enabling effective propagation of known labels in graph-based deep learning (GDL) algorithms, SSL is poised to become an increasingly used technique in GDL in the coming years. However, there are currently few explorations in the graph-based SSL literature on exploiting classical dimensionality reduction techniques for improved label propagation. In this work, we investigate the use of dimensionality reduction techniques such as PCA, t-SNE, and UMAP to see their effect on the performance of graph neural networks (GNNs) designed for semi-supervised propagation of node labels. Our study makes use of benchmark semi-supervised GDL datasets such as the Cora and Citeseer datasets to allow meaningful comparisons of the representations learned by each algorithm when paired with a dimensionality reduction technique. Our comprehensive benchmarks and clus-tering visualizations quantitatively and qualitatively demonstrate that, under certain conditions, employing a priori and a posteriori dimensionality reduction to GNN inputs and outputs, respectively, can simultaneously improve the effectiveness of semi-supervised node label propagation and node clustering. Our source code is freely available on GitHub.

MOTIVATION: Millions of protein sequences have been generated by numerous genome and transcriptome sequencing projects. However, experimentally determining the function of the proteins is still a time consuming, low-throughput, and expensive process, leading to a large protein sequence-function gap. Therefore, it is important to develop computational methods to accurately predict protein function to fill the gap. Even though many methods have been developed to use protein sequences as input to predict function, much fewer methods leverage protein structures in protein function prediction because there was lack of accurate protein structures for most proteins until recently. RESULTS: We developed TransFun-a method using a transformer-based protein language model and 3D-equivariant graph neural networks to distill information from both protein sequences and structures to predict protein function. It extracts feature embeddings from protein sequences using a pre-trained protein language model (ESM) via transfer learning and combines them with 3D structures of proteins predicted by AlphaFold2 through equivariant graph neural networks. Benchmarked on the CAFA3 test dataset and a new test dataset, TransFun outperforms several state-of-the-art methods, indicating that the language model and 3D-equivariant graph neural networks are effective methods to leverage protein sequences and structures to improve protein function prediction. Combining TransFun predictions and sequence similarity-based predictions can further increase prediction accuracy. AVAILABILITY AND IMPLEMENTATION: The source code of TransFun is available at https://github.com/jianlin-cheng/TransFun.

MOTIVATION: Quality assessment (QA) of predicted protein tertiary structure models plays an important role in ranking and using them. With the recent development of deep learning end-to-end protein structure prediction techniques for generating highly confident tertiary structures for most proteins, it is important to explore corresponding QA strategies to evaluate and select the structural models predicted by them since these models have better quality and different properties than the models predicted by traditional tertiary structure prediction methods. RESULTS: We develop EnQA, a novel graph-based 3D-equivariant neural network method that is equivariant to rotation and translation of 3D objects to estimate the accuracy of protein structural models by leveraging the structural features acquired from the state-of-the-art tertiary structure prediction method-AlphaFold2. We train and test the method on both traditional model datasets (e.g. the datasets of the Critical Assessment of Techniques for Protein Structure Prediction) and a new dataset of high-quality structural models predicted only by AlphaFold2 for the proteins whose experimental structures were released recently. Our approach achieves state-of-the-art performance on protein structural models predicted by both traditional protein structure prediction methods and the latest end-to-end deep learning method-AlphaFold2. It performs even better than the model QA scores provided by AlphaFold2 itself. The results illustrate that the 3D-equivariant graph neural network is a promising approach to the evaluation of protein structural models. Integrating AlphaFold2 features with other complementary sequence and structural features is important for improving protein model QA. AVAILABILITY AND IMPLEMENTATION: The source code is available at https://github.com/BioinfoMachineLearning/EnQA. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Proteins interact to form complexes to carry out essential biological functions. Computational methods such as AlphaFold-multimer have been developed to predict the quaternary structures of protein complexes. An important yet largely unsolved challenge in protein complex structure prediction is to accurately estimate the quality of predicted protein complex structures without any knowledge of the corresponding native structures. Such estimations can then be used to select high-quality predicted complex structures to facilitate biomedical research such as protein function analysis and drug discovery. RESULTS: In this work, we introduce a new gated neighborhood-modulating graph transformer to predict the quality of 3D protein complex structures. It incorporates node and edge gates within a graph transformer framework to control information flow during graph message passing. We trained, evaluated and tested the method (called DProQA) on newly-curated protein complex datasets before the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) and then blindly tested it in the 2022 CASP15 experiment. The method was ranked 3rd among the single-model quality assessment methods in CASP15 in terms of the ranking loss of TM-score on 36 complex targets. The rigorous internal and external experiments demonstrate that DProQA is effective in ranking protein complex structures. AVAILABILITY AND IMPLEMENTATION: The source code, data, and pre-trained models are available at https://github.com/jianlin-cheng/DProQA.

MOTIVATION: Sphagnum-dominated peatlands store a substantial amount of terrestrial carbon. The genus is undersampled and under-studied. No experimental crystal structure from any Sphagnum species exists in the Protein Data Bank and fewer than 200 Sphagnum-related genes have structural models available in the AlphaFold Protein Structure Database. Tools and resources are needed to help bridge these gaps, and to enable the analysis of other structural proteomes now made possible by accurate structure prediction. RESULTS: We present the predicted structural proteome (25 134 primary transcripts) of Sphagnum divinum computed using AlphaFold, structural alignment results of all high-confidence models against an annotated nonredundant crystallographic database of over 90,000 structures, a structure-based classification of putative Enzyme Commission (EC) numbers across this proteome, and the computational method to perform this proteome-scale structure-based annotation. AVAILABILITY AND IMPLEMENTATION: All data and code are available in public repositories, detailed at https://github.com/BSDExabio/SAFA. The structural models of the S. divinum proteome have been deposited in the ModelArchive repository at https://modelarchive.org/doi/10.5452/ma-ornl-sphdiv.

MOTIVATION: The spatial genome organization of a eukaryotic cell is important for its function. The development of single-cell technologies for probing the 3D genome conformation, especially single-cell chromosome conformation capture techniques, has enabled us to understand genome function better than before. However, due to extreme sparsity and high noise associated with single-cell Hi-C data, it is still difficult to study genome structure and function using the HiC-data of one single cell. RESULTS: In this work, we developed a deep learning method ScHiCEDRN based on deep residual networks and generative adversarial networks for the imputation and enhancement of Hi-C data of a single cell. In terms of both image evaluation and Hi-C reproducibility metrics, ScHiCEDRN outperforms the four deep learning methods (DeepHiC, HiCPlus, HiCSR, and Loopenhance) on enhancing the raw single-cell Hi-C data of human and Drosophila. The experiments also show that it can generate single-cell Hi-C data more suitable for identifying topologically associating domain boundaries and reconstructing 3D chromosome structures than the existing methods. Moreover, ScHiCEDRN's performance generalizes well across different single cells and cell types, and it can be applied to improving population Hi-C data. AVAILABILITY AND IMPLEMENTATION: The source code of ScHiCEDRN is available at the GitHub repository: https://github.com/BioinfoMachineLearning/ScHiCEDRN.

MOTIVATION: The state-of-art protein structure prediction methods such as AlphaFold are being widely used to predict structures of uncharacterized proteins in biomedical research. There is a significant need to further improve the quality and nativeness of the predicted structures to enhance their usability. In this work, we develop ATOMRefine, a deep learning-based, end-to-end, all-atom protein structural model refinement method. It uses a SE(3)-equivariant graph transformer network to directly refine protein atomic coordinates in a predicted tertiary structure represented as a molecular graph. RESULTS: The method is first trained and tested on the structural models in AlphaFoldDB whose experimental structures are known, and then blindly tested on 69 CASP14 regular targets and 7 CASP14 refinement targets. ATOMRefine improves the quality of both backbone atoms and all-atom conformation of the initial structural models generated by AlphaFold. It also performs better than two state-of-the-art refinement methods in multiple evaluation metrics including an all-atom model quality score-the MolProbity score based on the analysis of all-atom contacts, bond length, atom clashes, torsion angles, and side-chain rotamers. As ATOMRefine can refine a protein structure quickly, it provides a viable, fast solution for improving protein geometry and fixing structural errors of predicted structures through direct coordinate refinement. AVAILABILITY AND IMPLEMENTATION: The source code of ATOMRefine is available in the GitHub repository (https://github.com/BioinfoMachineLearning/ATOMRefine). All the required data for training and testing are available at https://doi.org/10.5281/zenodo.6944368.

Successful integration of deep neural networks (DNNs) or deep learning (DL) has resulted in breakthroughs in many areas. However, deploying these highly accurate models for data-driven, learned, automatic, and practical machine learning (ML) solutions to end-user applications remains challenging. DL algorithms are often computationally expensive, power-hungry, and require large memory to process complex and iterative operations of millions of parameters. Hence, training and inference of DL models are typically performed on high-performance computing (HPC) clusters in the cloud. Data transmission to the cloud results in high latency, round-trip delay, security and privacy concerns, and the inability of real-time decisions. Thus, processing on edge devices can significantly reduce cloud transmission cost. Edge devices are end devices closest to the user, such as mobile phones, cyber–physical systems (CPSs), wearables, the Internet of Things (IoT), embedded and autonomous systems, and intelligent sensors. These devices have limited memory, computing resources, and power-handling capability. Therefore, optimization techniques at both the hardware and software levels have been developed to handle the DL deployment efficiently on the edge. Understanding the existing research, challenges, and opportunities is fundamental to leveraging the next generation of edge devices with artificial intelligence (AI) capability. Mainly, four research directions have been pursued for efficient DL inference on edge devices: 1) novel DL architecture and algorithm design; 2) optimization of existing DL methods; 3) development of algorithm–hardware codesign; and 4) efficient accelerator design for DL deployment. This article focuses on surveying each of the four research directions, providing a comprehensive review of the state-of-the-art tools and techniques for efficient edge inference.

New drug production, from target identification to marketing approval, takes over 12 years and can cost around $2.6 billion. Furthermore, the COVID-19 pandemic has unveiled the urgent need for more powerful computational methods for drug discovery. Here, we review the computational approaches to predicting protein-ligand interactions in the context of drug discovery, focusing on methods using artificial intelligence (AI). We begin with a brief introduction to proteins (targets), ligands (e.g. drugs) and their interactions for nonexperts. Next, we review databases that are commonly used in the domain of protein-ligand interactions. Finally, we survey and analyze the machine learning (ML) approaches implemented to predict protein-ligand binding sites, ligand-binding affinity and binding pose (conformation) including both classical ML algorithms and recent deep learning methods. After exploring the correlation between these three aspects of protein-ligand interaction, it has been proposed that they should be studied in unison. We anticipate that our review will aid exploration and development of more accurate ML-based prediction strategies for studying protein-ligand interactions.

MOTIVATION: Deep learning has revolutionized protein tertiary structure prediction recently. The cutting-edge deep learning methods such as AlphaFold can predict high-accuracy tertiary structures for most individual protein chains. However, the accuracy of predicting quaternary structures of protein complexes consisting of multiple chains is still relatively low due to lack of advanced deep learning methods in the field. Because interchain residue-residue contacts can be used as distance restraints to guide quaternary structure modeling, here we develop a deep dilated convolutional residual network method (DRCon) to predict interchain residue-residue contacts in homodimers from residue-residue co-evolutionary signals derived from multiple sequence alignments of monomers, intrachain residue-residue contacts of monomers extracted from true/predicted tertiary structures or predicted by deep learning, and other sequence and structural features. RESULTS: Tested on three homodimer test datasets (Homo_std dataset, DeepHomo dataset and CASP-CAPRI dataset), the precision of DRCon for top L/5 interchain contact predictions (L: length of monomer in a homodimer) is 43.46%, 47.10% and 33.50% respectively at 6 Å contact threshold, which is substantially better than DeepHomo and DNCON2_inter and similar to Glinter. Moreover, our experiments demonstrate that using predicted tertiary structure or intrachain contacts of monomers in the unbound state as input, DRCon still performs well, even though its accuracy is lower than using true tertiary structures in the bound state are used as input. Finally, our case study shows that good interchain contact predictions can be used to build high-accuracy quaternary structure models of homodimers. AVAILABILITY AND IMPLEMENTATION: The source code of DRCon is available at https://github.com/jianlin-cheng/DRCon. The datasets are available at https://zenodo.org/record/5998532#.YgF70vXMKsB. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

received valuable feedback from readers in the field regarding our paper that was published in a special issue of BMC Bioinformatics in 2020 (BMC Bioinformatics, 21(S21):534, 2020), we discovered some visualization errors and typos in two Figures (Figs. 17 and 18) in the published manuscript.The Fourier Shell Correlation (FSC) values in Figs.17b and 18b in the published paper were plotted in the incorrect order by mistake.The error only occurred in visualizing the data, while the original raw data are still the same as before.Therefore, we replot the data to create a new version of Figs. 17 and 18 and fix some typos (e.g., the title of the x-axis of Figs.17b and 18b) according to the feedback.Figures 17 and 18 are given below.

BACKGROUND: Estimation of the accuracy (quality) of protein structural models is important for both prediction and use of protein structural models. Deep learning methods have been used to integrate protein structure features to predict the quality of protein models. Inter-residue distances are key information for predicting protein's tertiary structures and therefore have good potentials to predict the quality of protein structural models. However, few methods have been developed to fully take advantage of predicted inter-residue distance maps to estimate the accuracy of a single protein structural model. RESULT: We developed an attentive 2D convolutional neural network (CNN) with channel-wise attention to take only a raw difference map between the inter-residue distance map calculated from a single protein model and the distance map predicted from the protein sequence as input to predict the quality of the model. The network comprises multiple convolutional layers, batch normalization layers, dense layers, and Squeeze-and-Excitation blocks with attention to automatically extract features relevant to protein model quality from the raw input without using any expert-curated features. We evaluated DISTEMA's capability of selecting the best models for CASP13 targets in terms of ranking loss of GDT-TS score. The ranking loss of DISTEMA is 0.079, lower than several state-of-the-art single-model quality assessment methods. CONCLUSION: This work demonstrates that using raw inter-residue distance information with deep learning can predict the quality of protein structural models reasonably well. DISTEMA is freely at https://github.com/jianlin-cheng/DISTEMA.

The information about the domain architecture of proteins is useful for studying protein structure and function. However, accurate prediction of protein domain boundaries (i.e., sequence regions separating two domains) from sequence remains a significant challenge. In this work, we develop a deep learning method based on multi-head U-Nets (called DistDom) to predict protein domain boundaries utilizing 1D sequence features and predicted 2D inter-residue distance map as input. The 1D features contain the evolutionary and physicochemical information of protein sequences, whereas the 2D distance map includes the structural information of proteins that was rarely used in domain boundary prediction before. The 1D and 2D features are processed by the 1D and 2D U-Nets respectively to generate hidden features. The hidden features are then used by the multi-head attention to predict the probability of each residue of a protein being in a domain boundary, leveraging both local and global information in the features. The residue-level domain boundary predictions can be used to classify proteins as single-domain or multi-domain proteins. It classifies the CASP14 single-domain and multi-domain targets at the accuracy of 75.9%, 13.28% more accurate than the state-of-the-art method. Tested on the CASP14 multi-domain protein targets with expert annotated domain boundaries, the average per-target F1 measure score of the domain boundary prediction by DistDom is 0.263, 29.56% higher than the state-of-the-art method.

MOTIVATION: The Gene Network Estimation Tool (GNET) is designed to build gene regulatory networks (GRNs) from transcriptomic gene expression data with a probabilistic graphical model. The data preprocessing, model construction and visualization modules of the original GNET software were developed on different programming platforms, which were inconvenient for users to deploy and use. RESULTS: Here, we present GNET2, an improved implementation of GNET as an integrated R package. GNET2 provides more flexibility for parameter initialization and regulatory module construction based on the core iterative modeling process of the original algorithm. The data exchange interface of GNET2 is handled within an R session automatically. Given the growing demand for regulatory network reconstruction from transcriptomic data, GNET2 offers a convenient option for GRN inference on large datasets. AVAILABILITY AND IMPLEMENTATION: The source code of GNET2 is available at https://github.com/jianlin-cheng/GNET2. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Accurate prediction of residue-residue distances is important for protein structure prediction. We developed several protein distance predictors based on a deep learning distance prediction method and blindly tested them in the 14th Critical Assessment of Protein Structure Prediction (CASP14). The prediction method uses deep residual neural networks with the channel-wise attention mechanism to classify the distance between every two residues into multiple distance intervals. The input features for the deep learning method include co-evolutionary features as well as other sequence-based features derived from multiple sequence alignments (MSAs). Three alignment methods are used with multiple protein sequence/profile databases to generate MSAs for input feature generation. Based on different configurations and training strategies of the deep learning method, five MULTICOM distance predictors were created to participate in the CASP14 experiment. RESULTS: Benchmarked on 37 hard CASP14 domains, the best performing MULTICOM predictor is ranked 5th out of 30 automated CASP14 distance prediction servers in terms of precision of top L/5 long-range contact predictions [i.e. classifying distances between two residues into two categories: in contact (<8 Angstrom) and not in contact otherwise] and performs better than the best CASP13 distance prediction method. The best performing MULTICOM predictor is also ranked 6th among automated server predictors in classifying inter-residue distances into 10 distance intervals defined by CASP14 according to the precision of distance classification. The results show that the quality and depth of MSAs depend on alignment methods and sequence databases and have a significant impact on the accuracy of distance prediction. Using larger training datasets and multiple complementary features improves prediction accuracy. However, the number of effective sequences in MSAs is only a weak indicator of the quality of MSAs and the accuracy of predicted distance maps. In contrast, there is a strong correlation between the accuracy of contact/distance predictions and the average probability of the predicted contacts, which can therefore be more effectively used to estimate the confidence of distance predictions and select predicted distance maps. AVAILABILITY AND IMPLEMENTATION: The software package, source code and data of DeepDist2 are freely available at https://github.com/multicom-toolbox/deepdist and https://zenodo.org/record/4712084#.YIIM13VKhQM. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Due to the complexity of the biological systems, the prediction of the potential DNA binding sites for transcription factors remains a difficult problem in computational biology. Genomic DNA sequences and experimental results from parallel sequencing provide available information about the affinity and accessibility of genome and are commonly used features in binding sites prediction. The attention mechanism in deep learning has shown its capability to learn long-range dependencies from sequential data, such as sentences and voices. Until now, no study has applied this approach in binding site inference from massively parallel sequencing data. The successful applications of attention mechanism in similar input contexts motivate us to build and test new methods that can accurately determine the binding sites of transcription factors. RESULTS: In this study, we propose a novel tool (named DeepGRN) for transcription factors binding site prediction based on the combination of two components: single attention module and pairwise attention module. The performance of our methods is evaluated on the ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge datasets. The results show that DeepGRN achieves higher unified scores in 6 of 13 targets than any of the top four methods in the DREAM challenge. We also demonstrate that the attention weights learned by the model are correlated with potential informative inputs, such as DNase-Seq coverage and motifs, which provide possible explanations for the predictive improvements in DeepGRN. CONCLUSIONS: DeepGRN can automatically and effectively predict transcription factor binding sites from DNA sequences and DNase-Seq coverage. Furthermore, the visualization techniques we developed for the attention modules help to interpret how critical patterns from different types of input features are recognized by our model.

Abstract Background Driven by deep learning, inter-residue contact/distance prediction has been significantly improved and substantially enhanced ab initio protein structure prediction. Currently, most of the distance prediction methods classify inter-residue distances into multiple distance intervals instead of directly predicting real-value distances. The output of the former has to be converted into real-value distances to be used in tertiary structure prediction. Results To explore the potentials of predicting real-value inter-residue distances, we develop a multi-task deep learning distance predictor (DeepDist) based on new residual convolutional network architectures to simultaneously predict real-value inter-residue distances and classify them into multiple distance intervals. Tested on 43 CASP13 hard domains, DeepDist achieves comparable performance in real-value distance prediction and multi-class distance prediction. The average mean square error (MSE) of DeepDist’s real-value distance prediction is 0.896 Å 2 when filtering out the predicted distance ≥ 16 Å, which is lower than 1.003 Å 2 of DeepDist’s multi-class distance prediction. When distance predictions are converted into contact predictions at 8 Å threshold (the standard threshold in the field), the precision of top L/5 and L/2 contact predictions of DeepDist’s multi-class distance prediction is 79.3% and 66.1%, respectively, higher than 78.6% and 64.5% of its real-value distance prediction and the best results in the CASP13 experiment. Conclusions DeepDist can predict inter-residue distances well and improve binary contact prediction over the existing state-of-the-art methods. Moreover, the predicted real-value distances can be directly used to reconstruct protein tertiary structures better than multi-class distance predictions due to the lower MSE. Finally, we demonstrate that predicting the real-value distance map and multi-class distance map at the same time performs better than predicting real-value distances alone.

MOTIVATION: Deep learning has become the dominant technology for protein contact prediction. However, the factors that affect the performance of deep learning in contact prediction have not been systematically investigated. RESULTS: We analyzed the results of our three deep learning-based contact prediction methods (MULTICOM-CLUSTER, MULTICOM-CONSTRUCT and MULTICOM-NOVEL) in the CASP13 experiment and identified several key factors [i.e. deep learning technique, multiple sequence alignment (MSA), distance distribution prediction and domain-based contact integration] that influenced the contact prediction accuracy. We compared our convolutional neural network (CNN)-based contact prediction methods with three coevolution-based methods on 75 CASP13 targets consisting of 108 domains. We demonstrated that the CNN-based multi-distance approach was able to leverage global coevolutionary coupling patterns comprised of multiple correlated contacts for more accurate contact prediction than the local coevolution-based methods, leading to a substantial increase of precision by 19.2 percentage points. We also tested different alignment methods and domain-based contact prediction with the deep learning contact predictors. The comparison of the three methods showed deeper sequence alignments and the integration of domain-based contact prediction with the full-length contact prediction improved the performance of contact prediction. Moreover, we demonstrated that the domain-based contact prediction based on a novel ab initio approach of parsing domains from MSAs alone without using known protein structures was a simple, fast approach to improve contact prediction. Finally, we showed that predicting the distribution of inter-residue distances in multiple distance intervals could capture more structural information and improve binary contact prediction. AVAILABILITY AND IMPLEMENTATION: https://github.com/multicom-toolbox/DNCON2/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Cryo-EM data generated by electron tomography (ET) contains images for individual protein particles in different orientations and tilted angles. Individual cryo-EM particles can be aligned to reconstruct a 3D density map of a protein structure. However, low contrast and high noise in particle images make it challenging to build 3D density maps at intermediate to high resolution (1-3 Å). To overcome this problem, we propose a fully automated cryo-EM 3D density map reconstruction approach based on deep learning particle picking. RESULTS: A perfect 2D particle mask is fully automatically generated for every single particle. Then, it uses a computer vision image alignment algorithm (image registration) to fully automatically align the particle masks. It calculates the difference of the particle image orientation angles to align the original particle image. Finally, it reconstructs a localized 3D density map between every two single-particle images that have the largest number of corresponding features. The localized 3D density maps are then averaged to reconstruct a final 3D density map. The constructed 3D density map results illustrate the potential to determine the structures of the molecules using a few samples of good particles. Also, using the localized particle samples (with no background) to generate the localized 3D density maps can improve the process of the resolution evaluation in experimental maps of cryo-EM. Tested on two widely used datasets, Auto3DCryoMap is able to reconstruct good 3D density maps using only a few thousand protein particle images, which is much smaller than hundreds of thousands of particles required by the existing methods. CONCLUSIONS: We design a fully automated approach for cryo-EM 3D density maps reconstruction (Auto3DCryoMap). Instead of increasing the signal-to-noise ratio by using 2D class averaging, our approach uses 2D particle masks to produce locally aligned particle images. Auto3DCryoMap is able to accurately align structural particle shapes. Also, it is able to construct a decent 3D density map from only a few thousand aligned particle images while the existing tools require hundreds of thousands of particle images. Finally, by using the pre-processed particle images, Auto3DCryoMap reconstructs a better 3D density map than using the original particle images.

BACKGROUND: Cryo-electron microscopy (Cryo-EM) is widely used in the determination of the three-dimensional (3D) structures of macromolecules. Particle picking from 2D micrographs remains a challenging early step in the Cryo-EM pipeline due to the diversity of particle shapes and the extremely low signal-to-noise ratio of micrographs. Because of these issues, significant human intervention is often required to generate a high-quality set of particles for input to the downstream structure determination steps. RESULTS: Here we propose a fully automated approach (DeepCryoPicker) for single particle picking based on deep learning. It first uses automated unsupervised learning to generate particle training datasets. Then it trains a deep neural network to classify particles automatically. Results indicate that the DeepCryoPicker compares favorably with semi-automated methods such as DeepEM, DeepPicker, and RELION, with the significant advantage of not requiring human intervention. CONCLUSIONS: Our framework combing supervised deep learning classification with automated un-supervised clustering for generating training data provides an effective approach to pick particles in cryo-EM images automatically and accurately.

Matrix Factorization (MF) is a successful collaborative filtering approach used in recommendation systems. However, its performance decreases significantly when users of the system have limited, inadequate feedback data. This problem is also known as the data sparsity problem. To handle this problem, hybrid approaches were proposed recently to integrate items' contextual information with MF-based approaches, which improved the performance of recommendations. Nevertheless, learning better representation of the items' contents is still a challenge that needs to be further enhanced. In this paper, we propose a Collaborative Attentive Autoencoder (CATA) that learns latent factors of items through an attention mechanism that can capture the most pertinent part of information for making better recommendations. Comprehensive experiments on two real-world datasets have shown our method performs better than the state-of-the-art models according to various evaluation metrics.

MOTIVATION: Several methods have been proposed for the paired RNA-seq analysis. However, many of them do not consider the heterogeneity in treatment effect among pairs that can naturally arise in real data. In addition, it has been reported in literature that the false discovery rate (FDR) control of some popular methods has been problematic. In this paper, we present a full hierarchical Bayesian model for the paired RNA-seq count data that accounts for variation of treatment effects among pairs and controls the FDR through the posterior expected FDR. RESULTS: Our simulation studies show that most competing methods can have highly inflated FDR for small to moderate sample sizes while PairedFB is able to control FDR close to the nominal levels. Furthermore, PairedFB has overall better performance in ranking true differentially expressed genes (DEGs) on the top than others, especially when the sample size gets bigger or when the heterogeneity level of treatment effects is high. In addition, PairedFB can be applied to identify the biologically significant DEGs with controlled FDR. The real data analysis also indicates PairedFB tends to find more biologically relevant genes even when the sample size is small. PairedFB is also shown to be robust with respect to the model misspecification in terms of its relative performance compared to others. AVAILABILITY AND IMPLEMENTATION: Software to implement this method (PairedFB) can be downloaded at: https://sites.google.com/a/udel.edu/qiujing/publication. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Three-dimensional (3D) genome organization plays important functional roles in cells. User-friendly tools for reconstructing 3D genome models from chromosomal conformation capturing data and analyzing them are needed for the study of 3D genome organization. RESULTS: We built a comprehensive graphical tool (GenomeFlow) to facilitate the entire process of modeling and analysis of 3D genome organization. This process includes the mapping of Hi-C data to one-dimensional (1D) reference genomes, the generation, normalization and visualization of two-dimensional (2D) chromosomal contact maps, the reconstruction and the visualization of the 3D models of chromosome and genome, the analysis of 3D models and the integration of these models with functional genomics data. This graphical tool is the first of its kind in reconstructing, storing, analyzing and annotating 3D genome models. It can reconstruct 3D genome models from Hi-C data and visualize them in real-time. This tool also allows users to overlay gene annotation, gene expression data and genome methylation data on top of 3D genome models. AVAILABILITY AND IMPLEMENTATION: The source code and user manual: https://github.com/jianlin-cheng/GenomeFlow. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: An important task of macromolecular structure determination by cryo-electron microscopy (cryo-EM) is the identification of single particles in micrographs (particle picking). Due to the necessity of human involvement in the process, current particle picking techniques are time consuming and often result in many false positives and negatives. Adjusting the parameters to eliminate false positives often excludes true particles in certain orientations. The supervised machine learning (e.g. deep learning) methods for particle picking often need a large training dataset, which requires extensive manual annotation. Other reference-dependent methods rely on low-resolution templates for particle detection, matching and picking, and therefore, are not fully automated. These issues motivate us to develop a fully automated, unbiased framework for particle picking. RESULTS: We design a fully automated, unsupervised approach for single particle picking in cryo-EM micrographs. Our approach consists of three stages: image preprocessing, particle clustering, and particle picking. The image preprocessing is based on multiple techniques including: image averaging, normalization, cryo-EM image contrast enhancement correction (CEC), histogram equalization, restoration, adaptive histogram equalization, guided image filtering, and morphological operations. Image preprocessing significantly improves the quality of original cryo-EM images. Our particle clustering method is based on an intensity distribution model which is much faster and more accurate than traditional K-means and Fuzzy C-Means (FCM) algorithms for single particle clustering. Our particle picking method, based on image cleaning and shape detection with a modified Circular Hough Transform algorithm, effectively detects the shape and the center of each particle and creates a bounding box encapsulating the particles. CONCLUSIONS: AutoCryoPicker can automatically and effectively recognize particle-like objects from noisy cryo-EM micrographs without the need of labeled training data or human intervention making it a useful tool for cryo-EM protein structure determination.

In face recognition, there currently exists a significant challenge, most prominent when complex conditions such as large expression, pose, illumination, and even low resolution are introduced. Therefore, there exists a demand to address this issue, particularly concerning the key challenge of efficient feature representation, appreciating the extensive feature space that can be utilized to improve efficiency. This paper proposes a new approach in which a localized Deep-CNN structure is applied to demonstrate its' effectiveness and efficiency. The main contribution of this model is to directly learn the localized visual features by splitting the learning face into four local blocks. Intuitively, the Localized Deep-CNN model mimics the primary visual context to joint feature representations by extraction to produce local relational visualized features for the learning face. Our Deep-CNN model achieves 97.13% accuracy rate on the Labeled Faces in the Wild (LFW) dataset, compared to the humanlevel recognition rate which is 98.76%.

Face verification is still a big challenging problem due to the different image conditions such as expression, pose, and illumination. To address these challenges, we propose a new Deep Leaning structure called Localized Deep-Norm CNN. Our model focuses on finding the correlations of features inside the sub region of each learning face by adding a localized feature normalization layer. The model can recover all the important correlated features of face images. Intuitively, the Localized Deep-Norm CNN model mimics the primary visual context of the learned face image by combining the localized extracted features representations. The local relational face features are extracted and normalized by assigning each sub-block to a local CNN model. Then, the global features are constructed by combining these localized high-level features to one fully connected layer to produce the final feature space of 4608d dimensions. In our model, two different optimization techniques are proposed to optimize the loss functions. The first optimization modifies the SoftMax loss function by using a cosine similarity metric instead of Euclidean inner-product layer. The second optimization is done by combining different loss functions with different metric learning. Our model achieves robustness accuracy of 99.19% in characterizing the similarity of multiple faces which is 0.16% improving on the LFW performance results.

The data sparsity is a significant challenge for collaborative filtering methods in recommendation systems for making accurate recommendations. Several approaches have been published to address this issue. Most of them usually use only one source of data to train their model, and other approaches still have lower performance especially when the sparsity of data is very high. In this paper, we use a deep learning based model, DeepHCF, to tackle this problem. DeepHCF uses two sources of data, ratings matrix and item reviews, to train two deep models via joint training. The user-item ratings matrix data is trained using a Multi-Layer Perceptron (MLP), while other side information is trained using a Convolutional Neural Network (CNN). The users and items' latent features learned by each model are utilized by factorization machines for our model prediction. Extensive experimental results on four different realworld datasets show that DeepHCF achieves on average a 7.42% improvement over the second most accurate method, when the dataset has over 99.9% of sparsity.

Motivation: Significant improvements in the prediction of protein residue-residue contacts are observed in the recent years. These contacts, predicted using a variety of coevolution-based and machine learning methods, are the key contributors to the recent progress in ab initio protein structure prediction, as demonstrated in the recent CASP experiments. Continuing the development of new methods to reliably predict contact maps is essential to further improve ab initio structure prediction. Results: In this paper we discuss DNCON2, an improved protein contact map predictor based on two-level deep convolutional neural networks. It consists of six convolutional neural networks-the first five predict contacts at 6, 7.5, 8, 8.5 and 10 Å distance thresholds, and the last one uses these five predictions as additional features to predict final contact maps. On the free-modeling datasets in CASP10, 11 and 12 experiments, DNCON2 achieves mean precisions of 35, 50 and 53.4%, respectively, higher than 30.6% by MetaPSICOV on CASP10 dataset, 34% by MetaPSICOV on CASP11 dataset and 46.3% by Raptor-X on CASP12 dataset, when top L/5 long-range contacts are evaluated. We attribute the improved performance of DNCON2 to the inclusion of short- and medium-range contacts into training, two-level approach to prediction, use of the state-of-the-art optimization and activation functions, and a novel deep learning architecture that allows each filter in a convolutional layer to access all the input features of a protein of arbitrary length. Availability and implementation: The web server of DNCON2 is at http://sysbio.rnet.missouri.edu/dncon2/ where training and testing datasets as well as the predictions for CASP10, 11 and 12 free-modeling datasets can also be downloaded. Its source code is available at https://github.com/multicom-toolbox/DNCON2/. Contact: [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Protein fold recognition is an important problem in structural bioinformatics. Almost all traditional fold recognition methods use sequence (homology) comparison to indirectly predict the fold of a target protein based on the fold of a template protein with known structure, which cannot explain the relationship between sequence and fold. Only a few methods had been developed to classify protein sequences into a small number of folds due to methodological limitations, which are not generally useful in practice. Results: We develop a deep 1D-convolution neural network (DeepSF) to directly classify any protein sequence into one of 1195 known folds, which is useful for both fold recognition and the study of sequence-structure relationship. Different from traditional sequence alignment (comparison) based methods, our method automatically extracts fold-related features from a protein sequence of any length and maps it to the fold space. We train and test our method on the datasets curated from SCOP1.75, yielding an average classification accuracy of 75.3%. On the independent testing dataset curated from SCOP2.06, the classification accuracy is 73.0%. We compare our method with a top profile-profile alignment method-HHSearch on hard template-based and template-free modeling targets of CASP9-12 in terms of fold recognition accuracy. The accuracy of our method is 12.63-26.32% higher than HHSearch on template-free modeling targets and 3.39-17.09% higher on hard template-based modeling targets for top 1, 5 and 10 predicted folds. The hidden features extracted from sequence by our method is robust against sequence mutation, insertion, deletion and truncation, and can be used for other protein pattern recognition problems such as protein clustering, comparison and ranking. Availability and implementation: The DeepSF server is publicly available at: http://iris.rnet.missouri.edu/DeepSF/. Contact: [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

BACKGROUND: Contact-guided protein structure prediction methods are becoming more and more successful because of the latest advances in residue-residue contact prediction. To support contact-driven structure prediction, effective tools that can quickly build tertiary structural models of good quality from predicted contacts need to be developed. RESULTS: We develop an improved contact-driven protein modelling method, CONFOLD2, and study how it may be effectively used for ab initio protein structure prediction with predicted contacts as input. It builds models using various subsets of input contacts to explore the fold space under the guidance of a soft square energy function, and then clusters the models to obtain the top five models. CONFOLD2 obtains an average reconstruction accuracy of 0.57 TM-score for the 150 proteins in the PSICOV contact prediction dataset. When benchmarked on the CASP11 contacts predicted using CONSIP2 and CASP12 contacts predicted using Raptor-X, CONFOLD2 achieves a mean TM-score of 0.41 on both datasets. CONCLUSION: CONFOLD2 allows to quickly generate top five structural models for a protein sequence when its secondary structures and contacts predictions at hand. The source code of CONFOLD2 is publicly available at https://github.com/multicom-toolbox/CONFOLD2/ .

Motivation: Protein model quality assessment (QA) plays a very important role in protein structure prediction. It can be divided into two groups of methods: single model and consensus QA method. The consensus QA methods may fail when there is a large portion of low quality models in the model pool. Results: In this paper, we develop a novel single-model quality assessment method QAcon utilizing structural features, physicochemical properties, and residue contact predictions. We apply residue-residue contact information predicted by two protein contact prediction methods PSICOV and DNcon to generate a new score as feature for quality assessment. This novel feature and other 11 features are used as input to train a two-layer neural network on CASP9 datasets to predict the quality of a single protein model. We blindly benchmarked our method QAcon on CASP11 dataset as the MULTICOM-CLUSTER server. Based on the evaluation, our method is ranked as one of the top single model QA methods. The good performance of the features based on contact prediction illustrates the value of using contact information in protein quality assessment. Availability and Implementation: The web server and the source code of QAcon are freely available at: http://cactus.rnet.missouri.edu/QAcon. Contact: [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

BACKGROUND: Residue-residue contacts are key features for accurate de novo protein structure prediction. For the optimal utilization of these predicted contacts in folding proteins accurately, it is important to study the challenges of reconstructing protein structures using true contacts. Because contact-guided protein modeling approach is valuable for predicting the folds of proteins that do not have structural templates, it is necessary for reconstruction studies to focus on hard-to-predict protein structures. RESULTS: Using a data set consisting of 496 structural domains released in recent CASP experiments and a dataset of 150 representative protein structures, in this work, we discuss three techniques to improve the reconstruction accuracy using true contacts - adding secondary structures, increasing contact distance thresholds, and adding non-contacts. We find that reconstruction using secondary structures and contacts can deliver accuracy higher than using full contact maps. Similarly, we demonstrate that non-contacts can improve reconstruction accuracy not only when the used non-contacts are true but also when they are predicted. On the dataset consisting of 150 proteins, we find that by simply using low ranked predicted contacts as non-contacts and adding them as additional restraints, can increase the reconstruction accuracy by 5% when the reconstructed models are evaluated using TM-score. CONCLUSIONS: Our findings suggest that secondary structures are invaluable companions of contacts for accurate reconstruction. Confirming some earlier findings, we also find that larger distance thresholds are useful for folding many protein structures which cannot be folded using the standard definition of contacts. Our findings also suggest that for more accurate reconstruction using predicted contacts it is useful to predict contacts at higher distance thresholds (beyond 8 Å) and predict non-contacts.

BACKGROUND: Deep learning is one of the most powerful machine learning methods that has achieved the state-of-the-art performance in many domains. Since deep learning was introduced to the field of bioinformatics in 2012, it has achieved success in a number of areas such as protein residue-residue contact prediction, secondary structure prediction, and fold recognition. In this work, we developed deep learning methods to improve the prediction of torsion (dihedral) angles of proteins. RESULTS: We design four different deep learning architectures to predict protein torsion angles. The architectures including deep neural network (DNN) and deep restricted Boltzmann machine (DRBN), deep recurrent neural network (DRNN) and deep recurrent restricted Boltzmann machine (DReRBM) since the protein torsion angle prediction is a sequence related problem. In addition to existing protein features, two new features (predicted residue contact number and the error distribution of torsion angles extracted from sequence fragments) are used as input to each of the four deep learning architectures to predict phi and psi angles of protein backbone. The mean absolute error (MAE) of phi and psi angles predicted by DRNN, DReRBM, DRBM and DNN is about 20-21° and 29-30° on an independent dataset. The MAE of phi angle is comparable to the existing methods, but the MAE of psi angle is 29°, 2° lower than the existing methods. On the latest CASP12 targets, our methods also achieved the performance better than or comparable to a state-of-the art method. CONCLUSIONS: Our experiment demonstrates that deep learning is a valuable method for predicting protein torsion angles. The deep recurrent network architecture performs slightly better than deep feed-forward architecture, and the predicted residue contact number and the error distribution of torsion angles extracted from sequence fragments are useful features for improving prediction accuracy.

BACKGROUND: With the development of chromosomal conformation capturing techniques, particularly, the Hi-C technique, the study of the spatial conformation of a genome is becoming an important topic in bioinformatics and computational biology. The Hi-C technique can generate genome-wide chromosomal interaction (contact) data, which can be used to investigate the higher-level organization of chromosomes, such as Topologically Associated Domains (TAD), i.e., locally packed chromosome regions bounded together by intra chromosomal contacts. The identification of the TADs for a genome is useful for studying gene regulation, genomic interaction, and genome function. RESULTS: Here, we formulate the TAD identification problem as an unsupervised machine learning (clustering) problem, and develop a new TAD identification method called ClusterTAD. We introduce a novel method to represent chromosomal contacts as features to be used by the clustering algorithm. Our results show that ClusterTAD can accurately predict the TADs on a simulated Hi-C data. Our method is also largely complementary and consistent with existing methods on the real Hi-C datasets of two mouse cells. The validation with the chromatin immunoprecipitation (ChIP) sequencing (ChIP-Seq) data shows that the domain boundaries identified by ClusterTAD have a high enrichment of CTCF binding sites, promoter-related marks, and enhancer-related histone modifications. CONCLUSIONS: As ClusterTAD is based on a proven clustering approach, it opens a new avenue to apply a large array of clustering methods developed in the machine learning field to the TAD identification problem. The source code, the results, and the TADs generated for the simulated and real Hi-C datasets are available here: https://github.com/BDM-Lab/ClusterTAD .

With an unprecedented amount of data available, it is important to explore new methods for developing predictive models to mine this data for scientific discoveries. In this study, we propose a deep learning regression model based on MultiLayer Perceptron and Stacked Denoising Auto-encoder (MLP-SAE) to predict gene expression from genotypes of genetic variation. Specifically, we use a stacked denoising auto-encoder to train our regression model in order to extract useful features, and utilize the multilayer perceptron for backpropagation. We further improve our model by adding a dropout technique to prevent overfitting. Our results on a real genomic dataset show that our MLP-SAE model with dropout outperform Lasso, Random Forests, and MLP-SAE without dropout. Our study provides a new application of deep learning in mining genomics data, and demonstrates that deep learning has great potentials in building predictive models to help understand biological systems.

MOTIVATION: Speed, accuracy and robustness of building protein fragment library have important implications in de novo protein structure prediction since fragment-based methods are one of the most successful approaches in template-free modeling (FM). Majority of the existing fragment detection methods rely on database-driven search strategies to identify candidate fragments, which are inherently time-consuming and often hinder the possibility to locate longer fragments due to the limited sizes of databases. Also, it is difficult to alleviate the effect of noisy sequence-based predicted features such as secondary structures on the quality of fragment. RESULTS: Here, we present FRAGSION, a database-free method to efficiently generate protein fragment library by sampling from an Input-Output Hidden Markov Model. FRAGSION offers some unique features compared to existing approaches in that it (i) is lightning-fast, consuming only few seconds of CPU time to generate fragment library for a protein of typical length (300 residues); (ii) can generate dynamic-size fragments of any length (even for the whole protein sequence) and (iii) offers ways to handle noise in predicted secondary structure during fragment sampling. On a FM dataset from the most recent Critical Assessment of Structure Prediction, we demonstrate that FGRAGSION provides advantages over the state-of-the-art fragment picking protocol of ROSETTA suite by speeding up computation by several orders of magnitude while achieving comparable performance in fragment quality. AVAILABILITY AND IMPLEMENTATION: Source code and executable versions of FRAGSION for Linux and MacOS is freely available to non-commercial users at http://sysbio.rnet.missouri.edu/FRAGSION/ It is bundled with a manual and example data. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Recent experimental studies have suggested that proteins fold via stepwise assembly of structural units named 'foldons' through the process of sequential stabilization. Alongside, latest developments on computational side based on probabilistic modeling have shown promising direction to perform de novo protein conformational sampling from continuous space. However, existing computational approaches for de novo protein structure prediction often randomly sample protein conformational space as opposed to experimentally suggested stepwise sampling. RESULTS: Here, we develop a novel generative, probabilistic model that simultaneously captures local structural preferences of backbone and side chain conformational space of polypeptide chains in a united-residue representation and performs experimentally motivated conditional conformational sampling via stepwise synthesis and assembly of foldon units that minimizes a composite physics and knowledge-based energy function for de novo protein structure prediction. The proposed method, UniCon3D, has been found to (i) sample lower energy conformations with higher accuracy than traditional random sampling in a small benchmark of 6 proteins; (ii) perform comparably with the top five automated methods on 30 difficult target domains from the 11th Critical Assessment of Protein Structure Prediction (CASP) experiment and on 15 difficult target domains from the 10th CASP experiment; and (iii) outperform two state-of-the-art approaches and a baseline counterpart of UniCon3D that performs traditional random sampling for protein modeling aided by predicted residue-residue contacts on 45 targets from the 10th edition of CASP. AVAILABILITY AND IMPLEMENTATION: Source code, executable versions, manuals and example data of UniCon3D for Linux and OSX are freely available to non-commercial users at http://sysbio.rnet.missouri.edu/UniCon3D/ CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: The three-dimensional (3D) conformation of chromosomes and genomes play an important role in cellular processes such as gene regulation, DNA replication and genome methylation. Several methods have been developed to reconstruct 3D structures of individual chromosomes from chromosomal conformation capturing data such as Hi-C data. However, few methods can effectively reconstruct the 3D structures of an entire genome due to the difficulty of handling noisy and inconsistent inter-chromosomal contact data. RESULTS: We generalized a 3D chromosome reconstruction method to make it capable of reconstructing 3D models of genomes from both intra- and inter-chromosomal Hi-C contact data and implemented it as a software tool called MOGEN. We validated MOGEN on synthetic datasets of a polymer worm-like chain model and a yeast genome at first, and then applied it to generate an ensemble of 3D structural models of the genome of human B-cells from a Hi-C dataset. These genome models not only were validated by some known structural patterns of the human genome, such as chromosome compartmentalization, chromosome territories, co-localization of small chromosomes in the nucleus center with the exception of chromosome 18, enriched center-toward inter-chromosomal interactions between elongated or telomere regions of chromosomes, but also demonstrated the intrinsically dynamic orientations between chromosomes. Therefore, MOGEN is a useful tool for converting chromosomal contact data into 3D genome models to provide a better view into the spatial organization of genomes. AVAILABILITY AND IMPLEMENTATION: The software of MOGEN is available at: http://calla.rnet.missouri.edu/mogen/ CONTACT: : [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Abstract Background In recent years, successful contact prediction methods and contact-guided ab initio protein structure prediction methods have highlighted the importance of incorporating contact information into protein structure prediction methods. It is also observed that for almost all globular proteins, the quality of contact prediction dictates the accuracy of structure prediction. Hence, like many existing evaluation measures for evaluating 3D protein models, various measures are currently used to evaluate predicted contacts, with the most popular ones being precision, coverage and distance distribution score (Xd). Results We have built a web application and a downloadable tool, ConEVA, for comprehensive assessment and detailed comparison of predicted contacts. Besides implementing existing measures for contact evaluation we have implemented new and useful methods of contact visualization using chord diagrams and comparison using Jaccard similarity computations. For a set (or sets) of predicted contacts, the web application runs even when a native structure is not available, visualizing the contact coverage and similarity between predicted contacts. We applied the tool on various contact prediction data sets and present our findings and insights we obtained from the evaluation of effective contact assessments. ConEVA is publicly available at http://cactus.rnet.missouri.edu/coneva/ . Conclusion ConEVA is useful for a range of contact related analysis and evaluations including predicted contact comparison, investigation of individual protein folding using predicted contacts, and analysis of contacts in a structure of interest.

BACKGROUND: Protein quality assessment (QA) useful for ranking and selecting protein models has long been viewed as one of the major challenges for protein tertiary structure prediction. Especially, estimating the quality of a single protein model, which is important for selecting a few good models out of a large model pool consisting of mostly low-quality models, is still a largely unsolved problem. RESULTS: We introduce a novel single-model quality assessment method DeepQA based on deep belief network that utilizes a number of selected features describing the quality of a model from different perspectives, such as energy, physio-chemical characteristics, and structural information. The deep belief network is trained on several large datasets consisting of models from the Critical Assessment of Protein Structure Prediction (CASP) experiments, several publicly available datasets, and models generated by our in-house ab initio method. Our experiments demonstrate that deep belief network has better performance compared to Support Vector Machines and Neural Networks on the protein model quality assessment problem, and our method DeepQA achieves the state-of-the-art performance on CASP11 dataset. It also outperformed two well-established methods in selecting good outlier models from a large set of models of mostly low quality generated by ab initio modeling methods. CONCLUSION: DeepQA is a useful deep learning tool for protein single model quality assessment and protein structure prediction. The source code, executable, document and training/test datasets of DeepQA for Linux is freely available to non-commercial users at http://cactus.rnet.missouri.edu/DeepQA/ .

MOTIVATION: Sampling structural models and ranking them are the two major challenges of protein structure prediction. Traditional protein structure prediction methods generally use one or a few quality assessment (QA) methods to select the best-predicted models, which cannot consistently select relatively better models and rank a large number of models well. RESULTS: Here, we develop a novel large-scale model QA method in conjunction with model clustering to rank and select protein structural models. It unprecedentedly applied 14 model QA methods to generate consensus model rankings, followed by model refinement based on model combination (i.e. averaging). Our experiment demonstrates that the large-scale model QA approach is more consistent and robust in selecting models of better quality than any individual QA method. Our method was blindly tested during the 11th Critical Assessment of Techniques for Protein Structure Prediction (CASP11) as MULTICOM group. It was officially ranked third out of all 143 human and server predictors according to the total scores of the first models predicted for 78 CASP11 protein domains and second according to the total scores of the best of the five models predicted for these domains. MULTICOM's outstanding performance in the extremely competitive 2014 CASP11 experiment proves that our large-scale QA approach together with model clustering is a promising solution to one of the two major problems in protein structure modeling. AVAILABILITY AND IMPLEMENTATION: The web server is available at: http://sysbio.rnet.missouri.edu/multicom_cluster/human/.

BACKGROUND: With more and more protein sequences produced in the genomic era, predicting protein structures from sequences becomes very important for elucidating the molecular details and functions of these proteins for biomedical research. Traditional template-based protein structure prediction methods tend to focus on identifying the best templates, generating the best alignments, and applying the best energy function to rank models, which often cannot achieve the best performance because of the difficulty of obtaining best templates, alignments, and models. METHODS: We developed a large-scale conformation sampling and evaluation method and its servers to improve the reliability and robustness of protein structure prediction. In the first step, our method used a variety of alignment methods to sample relevant and complementary templates and to generate alternative and diverse target-template alignments, used a template and alignment combination protocol to combine alignments, and used template-based and template-free modeling methods to generate a pool of conformations for a target protein. In the second step, it used a large number of protein model quality assessment methods to evaluate and rank the models in the protein model pool, in conjunction with an exception handling strategy to deal with any additional failure in model ranking. RESULTS: The method was implemented as two protein structure prediction servers: MULTICOM-CONSTRUCT and MULTICOM-CLUSTER that participated in the 11th Critical Assessment of Techniques for Protein Structure Prediction (CASP11) in 2014. The two servers were ranked among the best 10 server predictors. CONCLUSIONS: The good performance of our servers in CASP11 demonstrates the effectiveness and robustness of the large-scale conformation sampling and evaluation. The MULTICOM server is available at: http://sysbio.rnet.missouri.edu/multicom_cluster/.

BACKGROUND: The entire collection of genetic information resides within the chromosomes, which themselves reside within almost every cell nucleus of eukaryotic organisms. Each individual chromosome is found to have its own preferred three-dimensional (3D) structure independent of the other chromosomes. The structure of each chromosome plays vital roles in controlling certain genome operations, including gene interaction and gene regulation. As a result, knowing the structure of chromosomes assists in the understanding of how the genome functions. Fortunately, the 3D structure of chromosomes proves possible to construct through computational methods via contact data recorded from the chromosome. We developed a unique computational approach based on optimization procedures known as adaptation, simulated annealing, and genetic algorithm to construct 3D models of human chromosomes, using chromosomal contact data. RESULTS: Our models were evaluated using a percentage-based scoring function. Analysis of the scores of the final 3D models demonstrated their effective construction from our computational approach. Specifically, the models resulting from our approach yielded an average score of 80.41%, with a high of 91%, across models for all chromosomes of a normal human B-cell. Comparisons made with other methods affirmed the effectiveness of our strategy. Particularly, juxtaposition with models generated through the publicly available method Markov chain Monte Carlo 5C (MCMC5C) illustrated the outperformance of our approach, as seen through a higher average score for all chromosomes. Our methodology was further validated using two consistency checking techniques known as convergence testing and robustness checking, which both proved successful. CONCLUSIONS: The pursuit of constructing accurate 3D chromosomal structures is fueled by the benefits revealed by the findings as well as any possible future areas of study that arise. This motivation has led to the development of our computational methodology. The implementation of our approach proved effective in constructing 3D chromosome models and proved consistent with, and more effective than, some other methods thereby achieving our goal of creating a tool to help advance certain research efforts. The source code, test data, test results, and documentation of our method, Gen3D, are available at our sourceforge site at: http://sourceforge.net/projects/gen3d/.

Ab initio protein secondary structure (SS) predictions are utilized to generate tertiary structure predictions, which are increasingly demanded due to the rapid discovery of proteins. Although recent developments have slightly exceeded previous methods of SS prediction, accuracy has stagnated around 80 percent and many wonder if prediction cannot be advanced beyond this ceiling. Disciplines that have traditionally employed neural networks are experimenting with novel deep learning techniques in attempts to stimulate progress. Since neural networks have historically played an important role in SS prediction, we wanted to determine whether deep learning could contribute to the advancement of this field as well. We developed an SS predictor that makes use of the position-specific scoring matrix generated by PSI-BLAST and deep learning network architectures, which we call DNSS. Graphical processing units and CUDA software optimize the deep network architecture and efficiently train the deep networks. Optimal parameters for the training process were determined, and a workflow comprising three separately trained deep networks was constructed in order to make refined predictions. This deep learning network approach was used to predict SS for a fully independent test dataset of 198 proteins, achieving a Q3 accuracy of 80.7 percent and a Sov accuracy of 74.2 percent.

BACKGROUND: It is important to predict the quality of a protein structural model before its native structure is known. The method that can predict the absolute local quality of individual residues in a single protein model is rare, yet particularly needed for using, ranking and refining protein models. RESULTS: We developed a machine learning tool (SMOQ) that can predict the distance deviation of each residue in a single protein model. SMOQ uses support vector machines (SVM) with protein sequence and structural features (i.e. basic feature set), including amino acid sequence, secondary structures, solvent accessibilities, and residue-residue contacts to make predictions. We also trained a SVM model with two new additional features (profiles and SOV scores) on 20 CASP8 targets and found that including them can only improve the performance when real deviations between native and model are higher than 5Å. The SMOQ tool finally released uses the basic feature set trained on 85 CASP8 targets. Moreover, SMOQ implemented a way to convert predicted local quality scores into a global quality score. SMOQ was tested on the 84 CASP9 single-domain targets. The average difference between the residue-specific distance deviation predicted by our method and the actual distance deviation on the test data is 2.637Å. The global quality prediction accuracy of the tool is comparable to other good tools on the same benchmark. CONCLUSION: SMOQ is a useful tool for protein single model quality assessment. Its source code and executable are available at: http://sysbio.rnet.missouri.edu/multicom_toolbox/.

BACKGROUND: Protein sequence profile-profile alignment is an important approach to recognizing remote homologs and generating accurate pairwise alignments. It plays an important role in protein sequence database search, protein structure prediction, protein function prediction, and phylogenetic analysis. RESULTS: In this work, we integrate predicted solvent accessibility, torsion angles and evolutionary residue coupling information with the pairwise Hidden Markov Model (HMM) based profile alignment method to improve profile-profile alignments. The evaluation results demonstrate that adding predicted relative solvent accessibility and torsion angle information improves the accuracy of profile-profile alignments. The evolutionary residue coupling information is helpful in some cases, but its contribution to the improvement is not consistent. CONCLUSION: Incorporating the new structural information such as predicted solvent accessibility and torsion angles into the profile-profile alignment is a useful way to improve pairwise profile-profile alignment methods.

BACKGROUND: Recognizing the correct structural fold among known template protein structures for a target protein (i.e. fold recognition) is essential for template-based protein structure modeling. Since the fold recognition problem can be defined as a binary classification problem of predicting whether or not the unknown fold of a target protein is similar to an already known template protein structure in a library, machine learning methods have been effectively applied to tackle this problem. In our work, we developed RF-Fold that uses random forest - one of the most powerful and scalable machine learning classification methods - to recognize protein folds. RESULTS: RF-Fold consists of hundreds of decision trees that can be trained efficiently on very large datasets to make accurate predictions on a highly imbalanced dataset. We evaluated RF-Fold on the standard Lindahl's benchmark dataset comprised of 976 × 975 target-template protein pairs through cross-validation. Compared with 17 different fold recognition methods, the performance of RF-Fold is generally comparable to the best performance in fold recognition of different difficulty ranging from the easiest family level, the medium-hard superfamily level, and to the hardest fold level. Based on the top-one template protein ranked by RF-Fold, the correct recognition rate is 84.5%, 63.4%, and 40.8% at family, superfamily, and fold levels, respectively. Based on the top-five template protein folds ranked by RF-Fold, the correct recognition rate increases to 91.5%, 79.3% and 58.3% at family, superfamily, and fold levels. CONCLUSIONS: The good performance achieved by the RF-Fold demonstrates the random forest's effectiveness for protein fold recognition.

BACKGROUND: A number of proteins contain regions which do not adopt a stable tertiary structure in their native state. Such regions known as disordered regions have been shown to participate in many vital cell functions and are increasingly being examined as drug targets. RESULTS: This work presents a new sequence based approach for the prediction of protein disorder. The method uses boosted ensembles of deep networks to make predictions and participated in the CASP10 experiment. In a 10 fold cross validation procedure on a dataset of 723 proteins, the method achieved an average balanced accuracy of 0.82 and an area under the ROC curve of 0.90. These results are achieved in part by a boosting procedure which is able to steadily increase balanced accuracy and the area under the ROC curve over several rounds. The method also compared competitively when evaluated against a number of state-of-the-art disorder predictors on CASP9 and CASP10 benchmark datasets. CONCLUSIONS: DNdisorder is available as a web service at http://iris.rnet.missouri.edu/dndisorder/.

BACKGROUND: In recent years, the use and importance of predicted protein residue-residue contacts has grown considerably with demonstrated applications such as drug design, protein tertiary structure prediction and model quality assessment. Nevertheless, reported accuracies in the range of 25-35% stubbornly remain the norm for sequence based, long range contact predictions on hard targets. This is in spite of a prolonged effort on behalf of the community to improve the performance of residue-residue contact prediction. A thorough study of the quality of current residue-residue contact predictions and the evaluation metrics used as well as an analysis of current methods is needed to stimulate further advancement in contact prediction and its application. Such a study will better explain the quality and nature of residue-residue contact predictions generated by current methods and as a result lead to better use of this contact information. RESULTS: We evaluated several sequence based residue-residue contact predictors that participated in the tenth Critical Assessment of protein Structure Prediction (CASP) experiment. The evaluation was performed using standard assessment techniques such as those used by the official CASP assessors as well as two novel evaluation metrics (i.e., cluster accuracy and cluster count). An in-depth analysis revealed that while most residue-residue contact predictions generated are not accurate at the residue level, there is quite a strong contact signal present when allowing for less than residue level precision. Our residue-residue contact predictor, DNcon, performed particularly well achieving an accuracy of 66% for the top L/10 long range contacts when evaluated in a neighbourhood of size 2. The coverage of residue-residue contact areas was also greater with DNcon when compared to other methods. We also provide an analysis of DNcon with respect to its underlying architecture and features used for classification. CONCLUSIONS: Our novel evaluation metrics demonstrate that current residue-residue contact predictions do contain a strong contact signal and are of better quality than standard evaluation metrics indicate. Our method, DNcon, is a robust, state-of-the-art residue-residue sequence based contact predictor and excelled under a number of evaluation schemes. It is available as a web service at http://iris.rnet.missouri.edu/dncon/.

Predicting protein function from sequence is useful for biochemical experiment design, mutagenesis analysis, protein engineering, protein design, biological pathway analysis, drug design, disease diagnosis, and genome annotation as a vast number of protein sequences with unknown function are routinely being generated by DNA, RNA and protein sequencing in the genomic era. However, despite significant progresses in the last several years, the accuracy of protein function prediction still needs to be improved in order to be used effectively in practice, particularly when little or no homology exists between a target protein and proteins with annotated function. Here, we developed a method that integrated profile-sequence alignment, profile-profile alignment, and Domain Co-Occurrence Networks (DCN) to predict protein function at different levels of complexity, ranging from obvious homology, to remote homology, to no homology. We tested the method blindingly in the 2011 Critical Assessment of Function Annotation (CAFA). Our experiments demonstrated that our three-level prediction method effectively increased the recall of function prediction while maintaining a reasonable precision. Particularly, our method can predict function terms defined by the Gene Ontology more accurately than three standard baseline methods in most situations, handle multi-domain proteins naturally, and make ab initio function prediction when no homology exists. These results show that our approach can combine complementary strengths of most widely used BLAST-based function prediction methods, rarely used in function prediction but more sensitive profile-profile comparison-based homology detection methods, and non-homology-based domain co-occurrence networks, to effectively extend the power of function prediction from high homology, to low homology, to no homology (ab initio cases).

BACKGROUND: High-throughput RNA sequencing (RNA-Seq) is a revolutionary technique to study the transcriptome of a cell under various conditions at a systems level. Despite the wide application of RNA-Seq techniques to generate experimental data in the last few years, few computational methods are available to analyze this huge amount of transcription data. The computational methods for constructing gene regulatory networks from RNA-Seq expression data of hundreds or even thousands of genes are particularly lacking and urgently needed. RESULTS: We developed an automated bioinformatics method to predict gene regulatory networks from the quantitative expression values of differentially expressed genes based on RNA-Seq transcriptome data of a cell in different stages and conditions, integrating transcriptional, genomic and gene function data. We applied the method to the RNA-Seq transcriptome data generated for soybean root hair cells in three different development stages of nodulation after rhizobium infection. The method predicted a soybean nodulation-related gene regulatory network consisting of 10 regulatory modules common for all three stages, and 24, 49 and 70 modules separately for the first, second and third stage, each containing both a group of co-expressed genes and several transcription factors collaboratively controlling their expression under different conditions. 8 of 10 common regulatory modules were validated by at least two kinds of validations, such as independent DNA binding motif analysis, gene function enrichment test, and previous experimental data in the literature. CONCLUSIONS: We developed a computational method to reliably reconstruct gene regulatory networks from RNA-Seq transcriptome data. The method can generate valuable hypotheses for interpreting biological data and designing biological experiments such as ChIP-Seq, RNA interference, and yeast two hybrid experiments.

MOTIVATION: Protein residue-residue contacts continue to play a larger and larger role in protein tertiary structure modeling and evaluation. Yet, while the importance of contact information increases, the performance of sequence-based contact predictors has improved slowly. New approaches and methods are needed to spur further development and progress in the field. RESULTS: Here we present DNCON, a new sequence-based residue-residue contact predictor using deep networks and boosting techniques. Making use of graphical processing units and CUDA parallel computing technology, we are able to train large boosted ensembles of residue-residue contact predictors achieving state-of-the-art performance. AVAILABILITY: The web server of the prediction method (DNCON) is available at http://iris.rnet.missouri.edu/dncon/. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: As genome sequencing is becoming routine in biomedical research, the total number of protein sequences is increasing exponentially, recently reaching over 108 million. However, only a tiny portion of these proteins (i.e. ~75,000 or < 0.07%) have solved tertiary structures determined by experimental techniques. The gap between protein sequence and structure continues to enlarge rapidly as the throughput of genome sequencing techniques is much higher than that of protein structure determination techniques. Computational software tools for predicting protein structure and structural features from protein sequences are crucial to make use of this vast repository of protein resources. RESULTS: To meet the need, we have developed a comprehensive MULTICOM toolbox consisting of a set of protein structure and structural feature prediction tools. These tools include secondary structure prediction, solvent accessibility prediction, disorder region prediction, domain boundary prediction, contact map prediction, disulfide bond prediction, beta-sheet topology prediction, fold recognition, multiple template combination and alignment, template-based tertiary structure modeling, protein model quality assessment, and mutation stability prediction. CONCLUSIONS: These tools have been rigorously tested by many users in the last several years and/or during the last three rounds of the Critical Assessment of Techniques for Protein Structure Prediction (CASP7-9) from 2006 to 2010, achieving state-of-the-art or near performance. In order to facilitate bioinformatics research and technological development in the field, we have made the MULTICOM toolbox freely available as web services and/or software packages for academic use and scientific research. It is available at http://sysbio.rnet.missouri.edu/multicom_toolbox/.

SUMMARY: We built a web server named APOLLO, which can evaluate the absolute global and local qualities of a single protein model using machine learning methods or the global and local qualities of a pool of models using a pair-wise comparison approach. Based on our evaluations on 107 CASP9 (Critical Assessment of Techniques for Protein Structure Prediction) targets, the predicted quality scores generated from our machine learning and pair-wise methods have an average per-target correlation of 0.671 and 0.917, respectively, with the true model quality scores. Based on our test on 92 CASP9 targets, our predicted absolute local qualities have an average difference of 2.60 Å with the actual distances to native structure. AVAILABILITY: http://sysbio.rnet.missouri.edu/apollo/. Single and pair-wise global quality assessment software is also available at the site.

BACKGROUND: Multiple Sequence Alignment (MSA) is a basic tool for bioinformatics research and analysis. It has been used essentially in almost all bioinformatics tasks such as protein structure modeling, gene and protein function prediction, DNA motif recognition, and phylogenetic analysis. Therefore, improving the accuracy of multiple sequence alignment is important for advancing many bioinformatics fields. RESULTS: We designed and developed a new method, MSACompro, to synergistically incorporate predicted secondary structure, relative solvent accessibility, and residue-residue contact information into the currently most accurate posterior probability-based MSA methods to improve the accuracy of multiple sequence alignments. The method is different from the multiple sequence alignment methods (e.g. 3D-Coffee) that use the tertiary structure information of some sequences since the structural information of our method is fully predicted from sequences. To the best of our knowledge, applying predicted relative solvent accessibility and contact map to multiple sequence alignment is novel. The rigorous benchmarking of our method to the standard benchmarks (i.e. BAliBASE, SABmark and OXBENCH) clearly demonstrated that incorporating predicted protein structural information improves the multiple sequence alignment accuracy over the leading multiple protein sequence alignment tools without using this information, such as MSAProbs, ProbCons, Probalign, T-coffee, MAFFT and MUSCLE. And the performance of the method is comparable to the state-of-the-art method PROMALS of using structural features and additional homologous sequences by slightly lower scores. CONCLUSION: MSACompro is an efficient and reliable multiple protein sequence alignment tool that can effectively incorporate predicted protein structural information into multiple sequence alignment. The software is available at http://sysbio.rnet.missouri.edu/multicom_toolbox/.

BACKGROUND: Accurate identification of protein domain boundaries is useful for protein structure determination and prediction. However, predicting protein domain boundaries from a sequence is still very challenging and largely unsolved. RESULTS: We developed a new method to integrate the classification power of machine learning with evolutionary signals embedded in protein families in order to improve protein domain boundary prediction. The method first extracts putative domain boundary signals from a multiple sequence alignment between a query sequence and its homologs. The putative sites are then classified and scored by support vector machines in conjunction with input features such as sequence profiles, secondary structures, solvent accessibilities around the sites and their positions. The method was evaluated on a domain benchmark by 10-fold cross-validation and 60% of true domain boundaries can be recalled at a precision of 60%. The trade-off between the precision and recall can be adjusted according to specific needs by using different decision thresholds on the domain boundary scores assigned by the support vector machines. CONCLUSIONS: The good prediction accuracy and the flexibility of selecting domain boundary sites at different precision and recall values make our method a useful tool for protein structure determination and modelling. The method is available at http://sysbio.rnet.missouri.edu/dobo/.

MOTIVATION: Protein structure prediction is one of the most important problems in structural bioinformatics. Here we describe MULTICOM, a multi-level combination approach to improve the various steps in protein structure prediction. In contrast to those methods which look for the best templates, alignments and models, our approach tries to combine complementary and alternative templates, alignments and models to achieve on average better accuracy. RESULTS: The multi-level combination approach was implemented via five automated protein structure prediction servers and one human predictor which participated in the eighth Critical Assessment of Techniques for Protein Structure Prediction (CASP8), 2008. The MULTICOM servers and human predictor were consistently ranked among the top predictors on the CASP8 benchmark. The methods can predict moderate- to high-resolution models for most template-based targets and low-resolution models for some template-free targets. The results show that the multi-level combination of complementary templates, alternative alignments and similar models aided by model quality assessment can systematically improve both template-based and template-free protein modeling. AVAILABILITY: The MULTICOM server is freely available at http://casp.rnet.missouri.edu/multicom_3d.html .

BACKGROUND: Protein folding rate is an important property of a protein. Predicting protein folding rate is useful for understanding protein folding process and guiding protein design. Most previous methods of predicting protein folding rate require the tertiary structure of a protein as an input. And most methods do not distinguish the different kinetic nature (two-state folding or multi-state folding) of the proteins. Here we developed a method, SeqRate, to predict both protein folding kinetic type (two-state versus multi-state) and real-value folding rate using sequence length, amino acid composition, contact order, contact number, and secondary structure information predicted from only protein sequence with support vector machines. RESULTS: We systematically studied the contributions of individual features to folding rate prediction. On a standard benchmark dataset, the accuracy of folding kinetic type classification is 80%. The Pearson correlation coefficient and the mean absolute difference between predicted and experimental folding rates (sec-1) in the base-10 logarithmic scale are 0.81 and 0.79 for two-state protein folders, and 0.80 and 0.68 for three-state protein folders. SeqRate is the first sequence-based method for protein folding type classification and its accuracy of fold rate prediction is improved over previous sequence-based methods. Its performance can be further enhanced with additional information, such as structure-based geometric contacts, as inputs. CONCLUSIONS: Both the web server and software of predicting folding rate are publicly available at http://casp.rnet.missouri.edu/fold_rate/index.html.

Predicting protein folding rate is useful for understanding protein folding process and guiding protein design. Most previous methods of predicting folding rate require the tertiary structure of a protein as an input. And most methods do not distinguish the different kinetic natures (two-state folding and multi-state folding) of the proteins. Here we developed a method, SeqRate, to predict both protein folding kinetic type (two-state versus multi-state) and real-value folding rate using features extracted from only protein sequence with support vector machines. On a standard benchmark dataset, the accuracy of folding kinetic type classification is 80%. The Pearson correlation coefficient and the mean absolute difference between predicted and experimental folding rates (sec-1) in the base-10 logarithmic scale are 0.81 and 0.79 for two-state protein folders, and 0.80 and 0.68 for three-state protein folders. SeqRate is the first sequence-based method for protein folding type classification and its accuracy of fold rate prediction is improved over previous sequence-based methods. Both the Web server and software of predicting folding rate are publicly available at http://casp.rnet.missouri.edu/fold_rate/index.html.

BACKGROUND: Disordered regions are segments of the protein chain which do not adopt stable structures. Such segments are often of interest because they have a close relationship with protein expression and functionality. As such, protein disorder prediction is important for protein structure prediction, structure determination and function annotation. RESULTS: This paper presents our protein disorder prediction server, PreDisorder. It is based on our ab initio prediction method (MULTICOM-CMFR) which, along with our meta (or consensus) prediction method (MULTICOM), was recently ranked among the top disorder predictors in the eighth edition of the Critical Assessment of Techniques for Protein Structure Prediction (CASP8). We systematically benchmarked PreDisorder along with 26 other protein disorder predictors on the CASP8 data set and assessed its accuracy using a number of measures. The results show that it compared favourably with other ab initio methods and its performance is comparable to that of the best meta and clustering methods. CONCLUSION: PreDisorder is a fast and reliable server which can be used to predict protein disordered regions on genomic scale. It is available at http://casp.rnet.missouri.edu/predisorder.html.

Ordinal regression is an important type of learning, which has properties of both classification and regression. Here we describe an effective approach to adapt a traditional neural network to learn ordinal categories. Our approach is a generalization of the perceptron method for ordinal regression. On several benchmark datasets, our method (NNRank) outperforms a neural network classification method. Compared with the ordinal regression methods using Gaussian processes and support vector machines, NNRank achieves comparable performance. Moreover, NNRank has the advantages of traditional neural networks: learning in both online and batch modes, handling very large training datasets, and making rapid predictions. These features make NNRank a useful and complementary tool for large-scale data mining tasks such as information retrieval, Web page ranking, collaborative filtering, and protein ranking in bioinformatics. The neural network software is available at: http://www.cs.missouri.edu/~chengji/cheng software.html.

MOTIVATION: Transmembrane beta-barrel (TMB) proteins are embedded in the outer membranes of mitochondria, Gram-negative bacteria and chloroplasts. These proteins perform critical functions, including active ion-transport and passive nutrient intake. Therefore, there is a need for accurate prediction of secondary and tertiary structure of TMB proteins. Traditional homology modeling methods, however, fail on most TMB proteins since very few non-homologous TMB structures have been determined. Yet, because TMB structures conform to specific construction rules that restrict the conformational space drastically, it should be possible for methods that do not depend on target-template homology to be applied successfully. RESULTS: We develop a suite (TMBpro) of specialized predictors for predicting secondary structure (TMBpro-SS), beta-contacts (TMBpro-CON) and tertiary structure (TMBpro-3D) of transmembrane beta-barrel proteins. We compare our results to the recent state-of-the-art predictors transFold and PRED-TMBB using their respective benchmark datasets, and leave-one-out cross-validation. Using the transFold dataset TMBpro predicts secondary structure with per-residue accuracy (Q(2)) of 77.8%, a correlation coefficient of 0.54, and TMBpro predicts beta-contacts with precision of 0.65 and recall of 0.67. Using the PRED-TMBB dataset, TMBpro predicts secondary structure with Q(2) of 88.3% and a correlation coefficient of 0.75. All of these performance results exceed previously published results by 4% or more. Working with the PRED-TMBB dataset, TMBpro predicts the tertiary structure of transmembrane segments with RMSD <6.0 A for 9 of 14 proteins. For 6 of 14 predictions, the RMSD is <5.0 A, with a GDT_TS score greater than 60.0. AVAILABILITY: http://www.igb.uci.edu/servers/psss.html.

BACKGROUND: Predicting protein residue-residue contacts is an important 2D prediction task. It is useful for ab initio structure prediction and understanding protein folding. In spite of steady progress over the past decade, contact prediction remains still largely unsolved. RESULTS: Here we develop a new contact map predictor (SVMcon) that uses support vector machines to predict medium- and long-range contacts. SVMcon integrates profiles, secondary structure, relative solvent accessibility, contact potentials, and other useful features. On the same test data set, SVMcon's accuracy is 4% higher than the latest version of the CMAPpro contact map predictor. SVMcon recently participated in the seventh edition of the Critical Assessment of Techniques for Protein Structure Prediction (CASP7) experiment and was evaluated along with seven other contact map predictors. SVMcon was ranked as one of the top predictors, yielding the second best coverage and accuracy for contacts with sequence separation > or = 12 on 13 de novo domains. CONCLUSION: We describe SVMcon, a new contact map predictor that uses SVMs and a large set of informative features. SVMcon yields good performance on medium- to long-range contact predictions and can be modularly incorporated into a structure prediction pipeline.

MOTIVATION: Recognizing proteins that have similar tertiary structure is the key step of template-based protein structure prediction methods. Traditionally, a variety of alignment methods are used to identify similar folds, based on sequence similarity and sequence-structure compatibility. Although these methods are complementary, their integration has not been thoroughly exploited. Statistical machine learning methods provide tools for integrating multiple features, but so far these methods have been used primarily for protein and fold classification, rather than addressing the retrieval problem of fold recognition-finding a proper template for a given query protein. RESULTS: Here we present a two-stage machine learning, information retrieval, approach to fold recognition. First, we use alignment methods to derive pairwise similarity features for query-template protein pairs. We also use global profile-profile alignments in combination with predicted secondary structure, relative solvent accessibility, contact map and beta-strand pairing to extract pairwise structural compatibility features. Second, we apply support vector machines to these features to predict the structural relevance (i.e. in the same fold or not) of the query-template pairs. For each query, the continuous relevance scores are used to rank the templates. The FOLDpro approach is modular, scalable and effective. Compared with 11 other fold recognition methods, FOLDpro yields the best results in almost all standard categories on a comprehensive benchmark dataset. Using predictions of the top-ranked template, the sensitivity is approximately 85, 56, and 27% at the family, superfamily and fold levels respectively. Using the 5 top-ranked templates, the sensitivity increases to 90, 70, and 48%.

Many biomedical problems relate to mutant functional properties across a sequence space of interest, e.g., flu, cancer, and HIV. Detailed knowledge of mutant properties and function improves medical treatment and prevention. A functional census of p53 cancer rescue mutants would aid the search for cancer treatments from p53 mutant rescue. We devised a general methodology for conducting a functional census of a mutation sequence space by choosing informative mutants early. The methodology was tested in a double-blind predictive test on the functional rescue property of 71 novel putative p53 cancer rescue mutants iteratively predicted in sets of three (24 iterations). The first double-blind 15-point moving accuracy was 47 percent and the last was 86 percent; r = 0.01 before an epiphanic 16th iteration and r = 0.92 afterward. Useful mutants were chosen early (overall r = 0.80). Code and data are freely available (http://www.igb.uci.edu/research/research.html, corresponding authors: R.H.L. for computation and R.K.B. for biology).

The formation of disulphide bridges among cysteines is an important fea- ture of protein structures. Here we develop new methods for the predic- tion of disulphide bond connectivity. We first build a large curated data set of proteins containing disulphide bridges and then use 2-Dimensional Recursive Neural Networks to predict bonding probabilities between cys- teine pairs. These probabilities in turn lead to a weighted graph matching problem that can be addressed efficiently. We show how the method con- sistently achieves better results than previous approaches on the same validation data. In addition, the method can easily cope with chains with arbitrary numbers of bonded cysteines. Therefore, it overcomes one of the major limitations of previous approaches restricting predictions to chains containing no more than 10 oxidized cysteines. The method can be applied both to situations where the bonded state of each cysteine is known or unknown, in which case bonded state can be predicted with 85% precision and 90% recall. The method also yields an estimate for the total number of disulphide bridges in each chain.