Jiahao Tang, Henry Hengyuan Zhao, Lijian Wu, Zijian Zhang, Yifei Tao, Dongxing Mao, Yang Wan, Jingru Tan, Min Zeng, Min Li, Alex Jinpeng Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Despite continuous advancements in deep learning for understanding human motion, existing models often struggle to accurately identify action timing and specific body parts, typically supporting only single-round interaction. This limitation is particularly pronounced in home exercise monitoring, neurological disorder assessment, and rehabilitation, where precise motion analysis is crucial for ensuring exercise efficacy, detecting early signs of neurological conditions, and guiding personalized recovery programs. In this paper, we propose MoChat, a multimodal large language model capable of spatio-temporal grounding of human motion and multi-turn dialogue understanding. To achieve this, we first group spatial features in skeleton frames according to human anatomical structures and process them through a Joints-Grouped Skeleton Encoder. The encoder's outputs are fused with large language model embeddings to generate spatio-aware representations. A cross-attention-based Regression Head module is then designed to align hidden-layer embeddings and skeletal sequence embeddings, enabling precise temporal grounding. Furthermore, we develop a pipeline for temporal grounding task to extract timestamps from skeleton-text pairs and construct a multi-turn instruction dialogues for spatial grounding task. Finally, various task instructions are generated for jointly training. Experimental results demonstrate that MoChat achieves state-of-the-art performance across multiple metrics in motion understanding tasks, making it as the first model capable of fine-grained spatio-temporal grounding of human motion.

Predicting drug-target interactions (DTIs) plays an essential role in drug discovery and drug repurposing. Although significant performance improvements have been achieved in DTI prediction, existing methods have not fully explored the properties of protein and drug molecular structure to make results interpretable. In this study, we propose MotifGT-DTI, a novel motif-based model with a graph transformer (GT) for DTI prediction. Specifically, MotifGT-DTI captures complex molecular patterns of drug molecular graph motifs and protein 3-D pocket subgraphs with GT. To attain protein characteristics more comprehensively, MotifGT-DTI fuses 1-D sequence and 3-D structure features with cross-attention from two views. Then, the structural-level association patterns of drug molecules and proteins are connected via a bilinear attention network. Experimental results show that MotifGT-DTI achieves the best accuracy compared to state-of-the-art baselines on four public datasets. In the three cold-start scenarios, the prediction results provided by our method are competitive in accuracy, generalization ability, and stability, highlighting its promising potential for practical applications. Furthermore, the visualization study demonstrates that MotifGT-DTI finds functional molecular motifs and provides interpretability for predicted results. The datasets and codes are publicly available at https://github.com/Dimpleney/MotifGT-DTI.

Accurate diagnosis of Parkinson's disease (PD) from MRI remains challenging due to symptom variability and pathological heterogeneity. Most existing methods rely on conventional magnitude-based MRI modalities, such as T1weighted images (T1w), which are less sensitive to PD pathology than Quantitative Susceptibility Mapping (QSM), a phasebased MRI technique that quantifies iron deposition in deep gray matter nuclei. In this study, we propose GateFuseNet, an adaptive 3D multimodal fusion network that integrates QSM and T1w images for PD diagnosis. The core innovation lies in a gated fusion module that learns modality-specific attention weights and channel-wise gating vectors for selective feature modulation. This hierarchical gating mechanism enhances ROIaware features while suppressing irrelevant signals. Experimental results show that our method outperforms three existing state-of-the-art approaches, achieving 85.00 % accuracy and 92.06% AUC. Ablation studies further validate the contributions of ROI guidance, multimodal integration, and fusion positioning. Grad-CAM visualizations confirm the model's focus on clinically relevant pathological regions. The source codes and pretrained models can be found at https://github.com/YangGaoUQ/GateFuseNet

The tertiary structure of RNA plays a crucial role in determining its biological functions, stability, and interactions with other molecules. Accurate prediction of RNA tertiary structure is essential for understanding RNA's functional roles in cellular processes. Although the accuracy of RNA secondary structure prediction is currently considered acceptable, using these predictions as explicit geometric constraints in tertiary structure modeling remains challenges. In the study, we propose RNA3D-SSCL, an end-to-end deep learning framework for RNA tertiary structure prediction. RNA3D-SSCL leverages deep learning techniques to predict the three-dimensional folding of RNAs, incorporating a novel loss function that integrates secondary structure constraints to improve prediction performance. By utilizing a combination of sequence and secondary structure features, RNA3D-SSCL is capable of generating more accurate RNA tertiary structures. RNA3D-SSCL was evaluated on an independent test set, revealing significant improvements compared to existing methods. Ablation studies confirm that the secondary structure-constrained loss function led to a notable reduction in RMSD and an improvement in TM-score, indicating higher prediction performance. The source code can be obtained at https://github.com/CSUBioGroup/RNA3D-SSCL.

Ion channels are critical targets in drug development and play an important role in the treatment of various diseases. However, existing prediction methods either generalize across all drug targets or focus narrowly on specific ion channel proteins, leading to limitations in accuracy and versatility. In this study, we propose DeepDICI, a novel deep learning framework that integrates topology-aware protein embeddings with spatial-channel interaction modeling to overcome these challenges. Our approach leverages a LLaMA3-style language model to capture longrange structural dependencies between transmembrane segments and cytoplasmic domains, effectively simulating the grammar of channel folding. In addition, a Cross-Attention network dynamically aligns drug substructures with channel functional domains, enabling the precise identification of state-dependent binding interfaces. Comprehensive evaluations in multiple scenarios demonstrate that DeepDICI achieves the highest precision of prediction$(P R C>0.98)$and maintains exceptional robustness under cold start conditions, showing a 28.9% improvement in terms of recall for novel chemical scaffolds and a 14.3% increase in terms of accuracy for unseen channel targets. All the results indicate that DeepDICI represents a unified and interpretable paradigm for ion channel-targeted drug discovery, bridging computational predictions with biophysical insights into structuredynamic protein systems. The code and datasets for DeepDICI are freely available at https://github.com/CSUBioGroup/DeepDICI.

In recent years, with the improvement of living standards, the prevalence of depression has been steadily increasing, making it a growing public health concern. Gene expression data can reveal links between genes and diseases. Studies have shown that gene expression in depression patients differs significantly from healthy individuals, offering potential for early detection. However, existing methods often depend on selecting differentially expressed genes, which may overlook important signals from other genes and are vulnerable to batch effects, limiting model generalization. To address these limitations, we propose DP-GPT, a GPT-driven gene text feature embedding framework fused with gene expression data for depression prediction. DP-GPT integrates gene expression data with features extracted by GPT. Specifically, gene names and summaries are obtained from the NCBI Gene database, then embedded using GPT to generate feature vectors. These are fused with sample gene expression data and fed into a classifier for prediction. Extensive experiments show that DP-GPT achieves superior performance in depression prediction. The source code can be obtained from https://github.com/CSUBioGroup/DP-GPT.

Intrinsically disordered regions (IDRs) play essential roles in cellular signaling and regulation, primarily mediating interactions through short peptide motifs. Linear Interacting Peptides (LIPs) are a recently defined class of bindingassociated IDRs that undergo disorder-to-order transitions upon binding. LIPs encompass several well-characterized subclasses, including molecular recognition features (MoRFs) and short linear motifs (SLiMs). However, many current predictors focus exclusively on MoRFs prediction and exhibit limited effectiveness in identifying the broader class of LIPs. To address this limitation, we propose LipPredictor, the first deep learning framework specifically designed for LIP prediction. The model incorporates protein language model embeddings as input to a shared feature extraction module, which is composed of convolutional neural networks and a multi-head attention layer. It further adopts a dual-branch architecture that couples residue-level classification with auxiliary multi-scale anchor detection, enabling accurate identification of LIPs across diverse segment lengths through joint training. These results demonstrate its robustness and effectiveness in predicting LIPs. The source code can be obtained from https://github.com/Chenxi-Xia/LipPredictor.

Whole Slide Image (WSI) representation is critical for cancer subtyping, cancer recognition and mutation prediction.Training an end-to-end WSI representation model poses significant challenges, as a standard gigapixel slide can contain tens of thousands of image tiles, making it difficult to compute gradients of all tiles in a single mini-batch due to current GPU limitations. To address this challenge, we propose a method of dynamic residual encoding with slide-level contrastive learning (DRE-SLCL) for end-to-end WSI representation. Our approach utilizes a memory bank to store the features of tiles across all WSIs in the dataset. During training, a mini-batch usually contains multiple WSIs. For each WSI in the batch, a subset of tiles is randomly sampled and their features are computed using a tile encoder. Then, additional tile features from the same WSI are selected from the memory bank. The representation of each individual WSI is generated using a residual encoding technique that incorporates both the sampled features and those retrieved from the memory bank. Finally, the slide-level contrastive loss is computed based on the representations and histopathology reports ofthe WSIs within the mini-batch. Experiments conducted over cancer subtyping, cancer recognition, and mutation prediction tasks proved the effectiveness of the proposed DRE-SLCL method.

Multimodal Biomedical Image Incremental Learning (MBIIL) is essential for handling diverse tasks and modalities in the biomedical domain, as training separate models for each modality or task significantly increases inference costs. Existing incremental learning methods focus on task expansion within a single modality, whereas MBIIL seeks to train a unified model incrementally across modalities. The MBIIL faces two challenges: I) How to preserve previously learned knowledge during incremental updates? II) How to effectively leverage knowledge acquired from existing modalities to support new modalities? To address these challenges, we propose MSLoRA-CR, a method that fine-tunes Modality-Specific LoRA modules while incorporating Contrastive Regularization to enhance intra-modality knowledge sharing and promote inter-modality knowledge differentiation. Our approach builds upon a large vision-language model (LVLM), keeping the pretrained model frozen while incrementally adapting new LoRA modules for each modality or task. Experiments on the incremental learning of biomedical images demonstrate that MSLoRA-CR outperforms both the state-of-the-art (SOTA) approach of training separate models for each modality and the general incremental learning method (incrementally fine-tuning LoRA). Specifically, MSLoRA-CR achieves a 1.88% improvement in overall performance compared to unconstrained incremental learning methods while maintaining computational efficiency. Our code is publicly available at https://github.com/VentusAislant/MSLoRA_CR.

MOTIVATION: Accurately predicting RNA subcellular localization is crucial for understanding the cellular functions and regulatory mechanisms of RNAs. Although many computational methods have been developed to predict the subcellular localization of lncRNAs, miRNAs, and circRNAs, very few of them are designed to simultaneously predict the subcellular localization of multiple types of RNAs. In addition, the emergence of pre-trained RNA language model has shown remarkable performance in various bioinformatics tasks, such as structure prediction and functional annotation. Despite these advancements, there remains a significant gap in applying pre-trained RNA language models specifically for predicting RNA subcellular localization. RESULTS: In this study, we proposed RNALoc-LM, the first interpretable deep-learning framework that leverages a pre-trained RNA language model for predicting RNA subcellular localization. RNALoc-LM uses a pre-trained RNA language model to encode RNA sequences, then captures local patterns and long-range dependencies through TextCNN and BiLSTM modules. A multi-head attention mechanism is used to focus on important regions within the RNA sequences. The results demonstrate that RNALoc-LM significantly outperforms both deep-learning baselines and existing state-of-the-art predictors. Additionally, motif analysis highlights RNALoc-LM's potential for discovering important motifs, while an ablation study confirms the effectiveness of the RNA sequence embeddings generated by the pre-trained RNA language model. AVAILABILITY AND IMPLEMENTATION: The RNALoc-LM web server is available at http://csuligroup.com:8000/RNALoc-LM. The source code can be obtained from https://github.com/CSUBioGroup/RNALoc-LM.

MOTIVATION: Accurate prediction of single-guide RNA (sgRNA) activity is crucial for optimizing the CRISPR/Cas9 gene-editing system, as it directly influences the efficiency and accuracy of genome modifications. However, existing prediction methods mainly rely on large-scale experimental data of a single Cas9 variant to construct Cas9 protein (variants)-specific sgRNA activity prediction models, which limits their generalization ability and prediction performance across different Cas9 protein (variants), as well as their scalability to the continuously discovered new variants. RESULTS: In this study, we proposed PLM-CRISPR, a novel deep learning-based model that leverages protein language models to capture Cas9 protein (variants) representations for cross-variant sgRNA activity prediction. PLM-CRISPR uses tailored feature extraction modules for both sgRNA and protein sequences, incorporating a cross-variant training strategy and a dynamic feature fusion mechanism to effectively model their interactions. Extensive experiments demonstrate that PLM-CRISPR outperforms existing methods across datasets spanning seven Cas9 protein (variants) in three real-world scenarios, demonstrating its superior performance in handling data-scarce situations, including cases with few or no samples for novel variants. Comparative analyses with traditional machine learning and deep learning models further confirm the effectiveness of PLM-CRISPR. Additionally, motif analysis reveals that PLM-CRISPR accurately identifies high-activity sgRNA sequence patterns across diverse Cas9 protein (variants). Overall, PLM-CRISPR provides a robust, scalable, and generalizable solution for sgRNA activity prediction across diverse Cas9 protein (variants). AVAILABILITY AND IMPLEMENTATION: The source code can be obtained from https://github.com/CSUBioGroup/PLM-CRISPR.

MOTIVATION: 2'-O-methylation (2OMe) is a common post-transcriptional modification in RNA that plays a crucial role in regulating gene expression and is implicated in various biological processes and diseases. Computational methods offer an efficient alternative to the time-consuming and costly experimental identification of 2OMe sites. Recent advancements in RNA pre-trained language models have revolutionized RNA bioinformatics. However, there remains a gap in their application specifically for predicting 2OMe sites. RESULTS: In the study, we propose a novel deep learning framework, 2OMe-LM, for predicting 2OMe sites in RNA. 2OMe-LM integrates RNA sequence features derived from RNA pre-trained language models with those obtained from the word2vec technique. Then, 2OMe-LM employs fully connected layers and a bidirectional long short-term memory network to process the two types of features separately, followed by a feature fusion module for the final prediction. Additionally, an attention block is incorporated to provide the interpretability of the prediction results. The results demonstrate that 2OMe-LM significantly outperforms existing state-of-the-art predictors, with features from RNA pre-trained language models proving to be critical. Motif analysis further demonstrates 2OMe-LM's potential for discovering 2OMe-related motifs. AVAILABILITY AND IMPLEMENTATION: The 2OMe-LM web server is available at https://csuligroup.com:9200/2OMe-LM. The source code can be obtained from https://github.com/CSUBioGroup/2OMe-LM.

Drug repositioning, a promising strategy in drug development, aims to identify new indications for existing drugs while reducing costs and safety risks. Leveraging their unique advantages in modeling higher-order relations among nodes, hypergraphs and hypergraph neural networks (HGNN) have become increasingly popular in drug repositioning. However, most HGNN-based methods overlook the diverse relations generated during the convolution and do not explicitly model high-order interactions, limiting their ability to capture high-order interaction information adequately. To address these limitations, we propose HGCNDR, a hypergraph convolutional network with explicit high-order interaction extraction for drug repositioning. HGCNDR introduces a relation-aware hypergraph convolution operation to handle distinct relation types and a Hadamard product-based strategy to effectively model high-order interactions among drugs and diseases, efficiently extracting the resulting high-order interaction information. Specifically, HGCNDR constructs two feature graphs and a hypergraph based on drug similarity features, disease similarity features, and drug-disease association networks. HGCNDR then employs graph convolutional networks to extract embeddings from the feature graphs, while using the relation-aware hypergraph convolution operation and the strategy to extract structural and high-order interaction information embeddings from the hypergraph. Additionally, to preserve the common semantics between the embeddings extracted from the feature graphs and the hypergraph, HGCNDR introduces a consistency constraint. The experimental results demonstrate that HGCNDR has competitive performance compared to several baseline methods. Moreover, case studies on Alzheimer's disease and Breast carcinoma confirm that HGCNDR can retrieve more actual drug-disease associations in the top prediction results.

Proteins play crucial roles in diverse biological functions. Accurately annotating their functions is essential for understanding cellular mechanisms and developing therapies for complex diseases. Computational methods have been proposed as alternatives to labor-intensive and expensive experimental approaches. Existing computational methods have demonstrated that protein evolution information and Protein-Protein Interactions (PPIs) are essential for protein function prediction. However, traditional computational approaches for generating evolution information are time-consuming. On the other hand, proteins lacking interactions are ignored in previous studies. To address these limitations, we propose a novel deep learning framework, named DeepFMB, which incorporates multi-type biological knowledge. DeepFMB leverages a pre-trained protein language model to extract evolution information. Moreover, DeepFMB generates PPI-related features and orthology-related features using graph neural networks on the constructed PPI and orthology networks. Then, these multi-type features are fused adaptively for protein function prediction. Compared to eight state-of-the-art methods, DeepFMB outperforms all of them in terms of F-max and AUPR. Additionally, with the combination of sequence similarity-based inference, our predicted model predicts protein functions more accurately. Experimental results also validate the superior performance of our methods in predicting low-frequency GO terms. Ablation studies demonstrate that the multi-type biological knowledge we use is highly relevant to protein functions.

The subcellular localization of circular RNAs (circRNAs) is crucial for understanding their functional relevance and regulatory mechanisms. CircRNA subcellular localization exhibits variations across different cell lines, demonstrating the diversity and complexity of circRNA regulation within distinct cellular contexts. However, existing computational methods for predicting circRNA subcellular localization often ignore the importance of cell line specificity and instead train a general model on aggregated data from all cell lines. Considering the diversity and context-dependent behavior of circRNAs across different cell lines, it is imperative to develop cell line-specific models to accurately predict circRNA subcellular localization. In the study, we proposed CellCircLoc, a sequence-based deep learning model for circRNA subcellular localization prediction, which is trained for different cell lines. CellCircLoc utilizes a combination of convolutional neural networks, Transformer blocks, and bidirectional long short-term memory to capture both sequence local features and long-range dependencies within the sequences. In the Transformer blocks, CellCircLoc uses an attentive convolution mechanism to capture the importance of individual nucleotides. Extensive experiments demonstrate the effectiveness of CellCircLoc in accurately predicting circRNA subcellular localization across different cell lines, outperforming other computational models that do not consider cell line specificity. Moreover, the interpretability of CellCircLoc facilitates the discovery of important motifs associated with circRNA subcellular localization.

In recent years, the increasing number of individuals diagnosed with depression and the growing awareness of its impact on modern society have highlighted the significance of accurate depression diagnosis. Microarray data has played a crucial role in uncovering the genetic mechanisms underlying depression. However, existing methods for depression prediction using microarray data often rely on the selection of differentially expressed genes. This approach disregards important information from other genes and is susceptible to batch effects, thereby limiting generalizability and model stability. To address these limitations, we propose DP-BERT, a depression prediction model based on Bidirectional Encoder Representations from Transformers (BERT). DP-BERT follows a pre-training and fine-tuning paradigm, leveraging a large amount of unlabeled microarray data from diverse sequencing platforms for pretraining to extract comprehensive genetic-level representations of psychiatric disorders. Subsequently, supervised fine-tuning is performed for depression prediction. Experimental results demonstrate that the pre-trained model achieves superior performance in depression prediction. The source code can be obtained from https://github.com/CSUBioGroup/DP-BERT.

Sleep posture identification is crucial for accurately assessing sleep quality and diagnosing related diseases. In the realm of non-intrusive sleep monitoring, non-contact technologies are becoming increasingly mainstream. Millimeter-wave radar is frequently utilized in sleep posture identification due to its high resolution, strong penetration, and excellent sensitivity. However, traditional radar-based methods for sleep posture identification often struggle with reliability when dealing with diverse individuals and complex sleep environments. To address these challenges, we propose UniSleepPos, which designs a novel dual-view fusion mechanism to integrate depression and elevation angle signals obtained from radar, thus accurately capturing the posture information of the monitored subject in three-dimensional space. Furthermore, we combine sleep posture identification with individual characteristics, utilizing existing individual labels as prior knowledge to assist in sleep posture identification. The integration of prior knowledge provides a valuable information source for the model, helping to enhance its understanding of the data and improve its performance. We collected sleep posture data from eight volunteers using millimeter-wave radar devices under various environmental conditions. Leave-one-subject-out experiments were conducted to validate the effectiveness of UniSleepPos. The results indicated that UniSleepPos significantly outperforms existing methods, demonstrating its potential for practical applications.

Accurately predicting essential proteins is vital for comprehending organism survival, aiding in drug discovery, and informing strategies for treating diseases. While previous computational methods for essential protein prediction have predominantly focused on network-based approaches, recent advancements have seen rapid development in sequence-based prediction methods. However, existing sequence-based prediction methods tend to focus only on sequence-level features, ignoring other biological information at diverse levels. To make use of the diverse information across various biological levels, in this study, we introduce ComLMEss, a novel deep learning framework that combines three protein language models. ComLMEss integrates ProtTrans, ESMFold and OntoProtein, which contain different levels of biological information include protein sequence, conservation, structural, and functional information. ComLMEss employs convolutional neural networks and transformer structure to refine and contextualize the representations from three language models, enabling accurate and robust predictions. Experimental results demonstrate that ComLMEss consistently outperforms existing methods. Ablation studies confirm that the effectiveness of combining different language models focus on different biological information. All results underscore the potential of ComLMEss in essential protein prediction. The source code can be obtained at https://github.com/CSUBioGroup/ComLMEss.

Large Vision-Language Models (LVLMs) have garnered substantial attention in the biomedical image analysis domain due to their robust vision understanding capabilities. However, current methods rely heavily on dataset- and modality-specific fine-tuning. This involves tuning separate models for each dataset and biomedical modality. In this paper, we introduce a method for aligning multimodal biomedical images and language using a single LVLM, dubbed UniMed-LVLM. Specifically, we devise a General Projection Module (GPM) by integrating multiple image projection branches and implementing dynamic routing between the vision encoder and language decoder within the LLaVA-Med framework. Subsequently, we progressively align multiple biomedical modalities using a Parameter-Efficient Fine-Tuning (PEFT) technique known as Low-Rank Adaptation (LoRA). The model is initially trained on the LLaVA-Med dataset and then fine-tuned on four biomedical image analysis datasets: PathVqa, Slake, VqaRad, and Fitzpatrick17k, enabling the simultaneous analysis of radiology, pathology, and dermatology images. A single model is fine-tuned on three modalities across these datasets and evaluated on all test sets. Experimental results show that UniMed-LVLM improves the average evaluation score by 1.88% across the four datasets, validating its effectiveness in handling multimodal biomedical images.

Proteins play an important role in life activities and are the basic units for performing functions. Accurately annotating functions to proteins is crucial for understanding the intricate mechanisms of life and developing effective treatments for complex diseases. Traditional biological experiments struggle to keep pace with the growing number of known proteins. With the development of high-throughput sequencing technology, a wide variety of biological data provides the possibility to accurately predict protein functions by computational methods. Consequently, many computational methods have been proposed. Due to the diversity of application scenarios, it is necessary to conduct a comprehensive evaluation of these computational methods to determine the suitability of each algorithm for specific cases. In this study, we present a comprehensive benchmark, BeProf, to process data and evaluate representative computational methods. We first collect the latest datasets and analyze the data characteristics. Then, we investigate and summarize 17 state-of-the-art computational methods. Finally, we propose a novel comprehensive evaluation metric, design eight application scenarios and evaluate the performance of existing methods on these scenarios. Based on the evaluation, we provide practical recommendations for different scenarios, enabling users to select the most suitable method for their specific needs. All of these servers can be obtained from https://csuligroup.com/BEPROF and https://github.com/CSUBioGroup/BEPROF.

MOTIVATION: Transcription factors are pivotal in the regulation of gene expression, and accurate identification of transcription factor binding sites (TFBSs) at high resolution is crucial for understanding the mechanisms underlying gene regulation. The task of identifying TFBSs from DNA sequences is a significant challenge in the field of computational biology today. To address this challenge, a variety of computational approaches have been developed. However, these methods face limitations in their ability to achieve high-resolution identification and often lack interpretability. RESULTS: We propose BertSNR, an interpretable deep learning framework for identifying TFBSs at single-nucleotide resolution. BertSNR integrates sequence-level and token-level information by multi-task learning based on pre-trained DNA language models. Benchmarking comparisons show that our BertSNR outperforms the existing state-of-the-art methods in TFBS predictions. Importantly, we enhanced the interpretability of the model through attentional weight visualization and motif analysis, and discovered the subtle relationship between attention weight and motif. Moreover, BertSNR effectively identifies TFBSs in promoter regions, facilitating the study of intricate gene regulation. AVAILABILITY AND IMPLEMENTATION: The BertSNR source code can be found at https://github.com/lhy0322/BertSNR.

Automatic International Classification of Disease (ICD) coding plays a crucial role in assigning ICD codes to electronic medical records. This task presents a challenging multi-label text classification problem due to the vast number of ICD codes and the imbalanced label distribution. However, accurately predicting all labels simultaneously is extremely difficult for such the large label space. In this paper, we propose a novel model called Deep Iterative Learning Model (DILM-ICD), which uses an iterative learning framework to perform automatic ICD coding task. The iterative learning framework can refine the prediction results by repeating the iteration modules, which simulates the human-like coding process. In addition, we propose a multi-head text-label matching mechanism, which combines the embedding ICD description information to better match the relationship between text and label. The combination of the iterative learning framework with the multi-head text-label matching mechanism enables the model pay attention to lowfrequency ICD codes. DILM-ICD is evaluated on the MIMIC-III-full dataset and MIMIC-III-50 dataset. The experimental results show that DILM-ICD achieves state-of-the-art results across multiple evaluation metrics, which demonstrates the effectiveness of our proposed model.

Proteins play crucial roles in diverse biological functions, and accurately annotating their functions is essential for understanding cellular mechanisms and developing therapies for complex diseases. Computational methods have been proposed as alternatives to laborious experimental approaches. However, existing network-based methods focus on the protein-protein interaction (PPI) networks, while the proteins without interactions are ignored. To address this limitation, we propose a novel deep learning framework for protein function prediction, named PFP-GMB, which incorporates multi-type biological knowledge to consider the proteins not present in the PPI networks. PFP-GMB leverages a pre-trained protein language model to extract sequence representations. Moreover, PPIs and orthology relationships are used to generate functional related features via graph neural networks and attention mechanisms. Finally, these multi-type features are fused for protein function prediction. Compared to eight state-of-the-art methods, PFP-GMB outperforms all of them in terms of F-max and AUPR. The ablation studies further confirm the relevance and significance of the multi-type biological knowledge incorporated into PFP-GMB for protein function prediction.

Transformers achieve excellent performance in a variety of domains since they can capture long-distance dependencies through the self-attention mechanism. However, self-attention is computationally costly due to its quadratic complexity and high memory consumption. In this paper, we propose a novel Transformer variant (Singularformer) that uses neural networks to learn the singular value decomposition process of the attention matrix to design a linear-complexity and memory-efficient global self-attention mechanism. Specifically, we decompose the attention matrix into the product of three matrix factors based on singular value decomposition and design neural networks to learn these matrix factors, then the associative law of matrix multiplication is used to linearize the calculation of self-attention. The above procedure allows us to compute self-attention as two-dimensional reduction processes in the first and second token dimensional spaces, followed by a multi-head self-attention computational process on the first dimensional reduced token features. Experimental results on 8 real-world datasets demonstrate that Singularformer performs favorably against the other Transformer variants with lower time and space complexity. Our source code is publicly available at https://github.com/CSUBioGroup/Singularformer.

The subcellular localization of long non-coding RNAs (lncRNAs) is crucial for understanding lncRNA functions. Most of existing lncRNA subcellular localization prediction methods use k-mer frequency features to encode lncRNA sequences. However, k-mer frequency features lose sequence order information and fail to capture sequence patterns and motifs of different lengths. In this paper, we proposed GraphLncLoc, a graph convolutional network-based deep learning model, for predicting lncRNA subcellular localization. Unlike previous studies encoding lncRNA sequences by using k-mer frequency features, GraphLncLoc transforms lncRNA sequences into de Bruijn graphs, which transforms the sequence classification problem into a graph classification problem. To extract the high-level features from the de Bruijn graph, GraphLncLoc employs graph convolutional networks to learn latent representations. Then, the high-level feature vectors derived from de Bruijn graph are fed into a fully connected layer to perform the prediction task. Extensive experiments show that GraphLncLoc achieves better performance than traditional machine learning models and existing predictors. In addition, our analyses show that transforming sequences into graphs has more distinguishable features and is more robust than k-mer frequency features. The case study shows that GraphLncLoc can uncover important motifs for nucleus subcellular localization. GraphLncLoc web server is available at http://csuligroup.com:8000/GraphLncLoc/.

Emerging evidence has proved that circular RNAs (circRNAs) are implicated in pathogenic processes. They are regarded as promising biomarkers for diagnosis due to covalently closed loop structures. As opposed to traditional experiments, computational approaches can identify circRNA-disease associations at a lower cost. Aggregating multi-source pathogenesis data helps to alleviate data sparsity and infer potential associations at the system level. The majority of computational approaches construct a homologous network using multi-source data, but they lose the heterogeneity of the data. Effective methods that use the features of multi-source data are considered as a matter of urgency. In this paper, we propose a model (CDHGNN) based on edge-weighted graph attention and heterogeneous graph neural networks for potential circRNA-disease association prediction. The circRNA network, micro RNA network, disease network and heterogeneous network are constructed based on multi-source data. To reflect association probabilities between nodes, an edge-weighted graph attention network model is designed for node features. To assign attention weights to different types of edges and learn contextual meta-path, CDHGNN infers potential circRNA-disease association based on heterogeneous neural networks. CDHGNN outperforms state-of-the-art algorithms in terms of accuracy. Edge-weighted graph attention networks and heterogeneous graph networks have both improved performance significantly. Furthermore, case studies suggest that CDHGNN is capable of identifying specific molecular associations and investigating biomolecular regulatory relationships in pathogenesis. The code of CDHGNN is freely available at https://github.com/BioinformaticsCSU/CDHGNN.

The design of enzyme catalytic stability is of great significance in medicine and industry. However, traditional methods are time-consuming and costly. Hence, a growing number of complementary computational tools have been developed, e.g. ESMFold, AlphaFold2, Rosetta, RosettaFold, FireProt, ProteinMPNN. They are proposed for algorithm-driven and data-driven enzyme design through artificial intelligence (AI) algorithms including natural language processing, machine learning, deep learning, variational autoencoder/generative adversarial network, message passing neural network (MPNN). In addition, the challenges of design of enzyme catalytic stability include insufficient structured data, large sequence search space, inaccurate quantitative prediction, low efficiency in experimental validation and a cumbersome design process. The first principle of the enzyme catalytic stability design is to treat amino acids as the basic element. By designing the sequence of an enzyme, the flexibility and stability of the structure are adjusted, thus controlling the catalytic stability of the enzyme in a specific industrial environment or in an organism. Common indicators of design goals include the change in denaturation energy (ΔΔG), melting temperature (ΔTm), optimal temperature (Topt), optimal pH (pHopt), etc. In this review, we summarized and evaluated the enzyme design in catalytic stability by AI in terms of mechanism, strategy, data, labeling, coding, prediction, testing, unit, integration and prospect.

Circular RNAs (circRNAs) are reverse-spliced and covalently closed RNAs. Their interactions with RNA-binding proteins (RBPs) have multiple effects on the progress of many diseases. Some computational methods are proposed to identify RBP binding sites on circRNAs but suffer from insufficient accuracy, robustness and explanation. In this study, we first take the characteristics of both RNA and RBP into consideration. We propose a method for discriminating circRNA-RBP binding sites based on multi-scale characterizing sequence and structure features, called CRMSS. For circRNAs, we use sequence ${k}\hbox{-}{mer}$ embedding and the forming probabilities of local secondary structures as features. For RBPs, we combine sequence and structure frequencies of RNA-binding domain regions to generate features. We capture binding patterns with multi-scale residual blocks. With BiLSTM and attention mechanism, we obtain the contextual information of high-level representation for circRNA-RBP binding. To validate the effectiveness of CRMSS, we compare its predictive performance with other methods on 37 RBPs. Taking the properties of both circRNAs and RBPs into account, CRMSS achieves superior performance over state-of-the-art methods. In the case study, our model provides reliable predictions and correctly identifies experimentally verified circRNA-RBP pairs. The code of CRMSS is freely available at https://github.com/BioinformaticsCSU/CRMSS.

MOTIVATION: Protein essentiality is usually accepted to be a conditional trait and strongly affected by cellular environments. However, existing computational methods often do not take such characteristics into account, preferring to incorporate all available data and train a general model for all cell lines. In addition, the lack of model interpretability limits further exploration and analysis of essential protein predictions. RESULTS: In this study, we proposed DeepCellEss, a sequence-based interpretable deep learning framework for cell line-specific essential protein predictions. DeepCellEss utilizes a convolutional neural network and bidirectional long short-term memory to learn short- and long-range latent information from protein sequences. Further, a multi-head self-attention mechanism is used to provide residue-level model interpretability. For model construction, we collected extremely large-scale benchmark datasets across 323 cell lines. Extensive computational experiments demonstrate that DeepCellEss yields effective prediction performance for different cell lines and outperforms existing sequence-based methods as well as network-based centrality measures. Finally, we conducted some case studies to illustrate the necessity of considering specific cell lines and the superiority of DeepCellEss. We believe that DeepCellEss can serve as a useful tool for predicting essential proteins across different cell lines. AVAILABILITY AND IMPLEMENTATION: The DeepCellEss web server is available at http://csuligroup.com:8000/DeepCellEss. The source code and data underlying this study can be obtained from https://github.com/CSUBioGroup/DeepCellEss. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: There is mounting evidence that the subcellular localization of lncRNAs can provide valuable insights into their biological functions. In the real world of transcriptomes, lncRNAs are usually localized in multiple subcellular localizations. Furthermore, lncRNAs have specific localization patterns for different subcellular localizations. Although several computational methods have been developed to predict the subcellular localization of lncRNAs, few of them are designed for lncRNAs that have multiple subcellular localizations, and none of them take motif specificity into consideration. RESULTS: In this study, we proposed a novel deep learning model, called LncLocFormer, which uses only lncRNA sequences to predict multi-label lncRNA subcellular localization. LncLocFormer utilizes eight Transformer blocks to model long-range dependencies within the lncRNA sequence and shares information across the lncRNA sequence. To exploit the relationship between different subcellular localizations and find distinct localization patterns for different subcellular localizations, LncLocFormer employs a localization-specific attention mechanism. The results demonstrate that LncLocFormer outperforms existing state-of-the-art predictors on the hold-out test set. Furthermore, we conducted a motif analysis and found LncLocFormer can capture known motifs. Ablation studies confirmed the contribution of the localization-specific attention mechanism in improving the prediction performance. AVAILABILITY AND IMPLEMENTATION: The LncLocFormer web server is available at http://csuligroup.com:9000/LncLocFormer. The source code can be obtained from https://github.com/CSUBioGroup/LncLocFormer.

The understanding of protein functions is critical to many biological problems such as the development of new drugs and new crops. To reduce the huge gap between the increase of protein sequences and annotations of protein functions, many methods have been proposed to deal with this problem. These methods use Gene Ontology (GO) to classify the functions of proteins and consider one GO term as a class label. However, they ignore the co-occurrence of GO terms that is helpful for protein function prediction. We propose a new deep learning model, named DeepPFP-CO, which uses Graph Convolutional Network (GCN) to explore and capture the co-occurrence of GO terms to improve the protein function prediction performance. In this way, we can further deduce the protein functions by fusing the predicted propensity of the center function and its co-occurrence functions. We use Fmax and AUPR to evaluate the performance of DeepPFP-CO and compare DeepPFP-CO with state-of-the-art methods such as DeepGOPlus and DeepGOA. The computational results show that DeepPFP-CO outperforms DeepGOPlus and other methods. Moreover, we further analyze our model at the protein level. The results have demonstrated that DeepPFP-CO improves the performance of protein function prediction. DeepPFP-CO is available at https://csuligroup.com/DeepPFP/.

Recent studies have confirmed the significant effects of long non-coding RNAs (IncRNAs) in understanding the mechanism of diseases. Because of the relatively small number of validated associations between IncRNAs and diseases, and previous computational methods have limited performance without capturing important features of sequences and ontology information, we developed LDAGSO, a novel deep learning framework to predict IncRNA and disease associations from IncRNA sequences and disease ontology. For IncRNA sequences, we converted them into graph structure based on k-mer technique and de Bruijn graph, and captured high-level features of the graph using graph convolutional networks. For diseases, we extracted ontology term paths from the disease ontology tree, and treated them as sentences to obtain their feature representation using Bidirectional Encoder Representations from Transformers (BERT) technique. Finally, these two kinds of features were fed into a fully connected layer to perform the task of association prediction between 1ncRNAs and diseases. According to the results, our approach provides state-of-the-art results when evaluated by leave-one-out cross-validation.

Biomedical named entity recognition (BioNER) is to extract entities, such as genes and proteins, from biomedical texts, where there is often a lack of high-quality training data. Recent work addresses this issue by multi-task learning with multiple datasets. However, these methods are usually over-parameterized, and some even suffer from negative transfer issues. To address above problems, we propose adversarial sparse sharing mechanism, which trains a sparsely shared encoder on multiple tasks with both task-agnostic and task-specific subnetworks. With adversarial training, we guide the task-agnostic subnetwork to learn shared task-invariant features and the task-specific subnetwork to learn task-dependent features. For a particular task, only the shared and its subnetwork are activated, which greatly reduces the number of parameters and avoids interference among tasks. Experimental results on 15 benchmark BioNER datasets show that our proposed method outperforms or is competitive with baseline methods with fewer parameters. Our code is released at: https://github.com/CSU-NLP-Group/ASNet

The coronavirus disease 2019 pandemic has alerted people of the threat caused by viruses. Vaccine is the most effective way to prevent the disease from spreading. The interaction between antibodies and antigens will clear the infectious organisms from the host. Identifying B-cell epitopes is critical in vaccine design, development of disease diagnostics and antibody production. However, traditional experimental methods to determine epitopes are time-consuming and expensive, and the predictive performance using the existing in silico methods is not satisfactory. This paper develops a general framework to predict variable-length linear B-cell epitopes specific for human-adapted viruses with machine learning approaches based on Protvec representation of peptides and physicochemical properties of amino acids. QR decomposition is incorporated during the embedding process that enables our models to handle variable-length sequences. Experimental results on large immune epitope datasets validate that our proposed model's performance is superior to the state-of-the-art methods in terms of AUROC (0.827) and AUPR (0.831) on the testing set. Moreover, sequence analysis also provides the results of the viral category for the corresponding predicted epitopes with high precision. Therefore, this framework is shown to reliably identify linear B-cell epitopes of human-adapted viruses given protein sequences and could provide assistance for potential future pandemics and epidemics.

Long non-coding RNAs (lncRNAs) are a class of RNA molecules with more than 200 nucleotides. A growing amount of evidence reveals that subcellular localization of lncRNAs can provide valuable insights into their biological functions. Existing computational methods for predicting lncRNA subcellular localization use k-mer features to encode lncRNA sequences. However, the sequence order information is lost by using only k-mer features. We proposed a deep learning framework, DeepLncLoc, to predict lncRNA subcellular localization. In DeepLncLoc, we introduced a new subsequence embedding method that keeps the order information of lncRNA sequences. The subsequence embedding method first divides a sequence into some consecutive subsequences and then extracts the patterns of each subsequence, last combines these patterns to obtain a complete representation of the lncRNA sequence. After that, a text convolutional neural network is employed to learn high-level features and perform the prediction task. Compared with traditional machine learning models, popular representation methods and existing predictors, DeepLncLoc achieved better performance, which shows that DeepLncLoc could effectively predict lncRNA subcellular localization. Our study not only presented a novel computational model for predicting lncRNA subcellular localization but also introduced a new subsequence embedding method which is expected to be applied in other sequence-based prediction tasks. The DeepLncLoc web server is freely accessible at http://bioinformatics.csu.edu.cn/DeepLncLoc/, and source code and datasets can be downloaded from https://github.com/CSUBioGroup/DeepLncLoc.

MOTIVATION: Exploring drug-protein interactions (DPIs) provides a rapid and precise approach to assist in laboratory experiments for discovering new drugs. Network-based methods usually utilize a drug-protein association network and predict DPIs by the information of its associated proteins or drugs, called 'guilt-by-association' principle. However, the 'guilt-by-association' principle is not always true because sometimes similar proteins cannot interact with similar drugs. Recently, learning-based methods learn molecule properties underlying DPIs by utilizing existing databases of characterized interactions but neglect the network-level information. RESULTS: We propose a novel method, namely BridgeDPI. We devise a class of virtual nodes to bridge the gap between drugs and proteins and construct a learnable drug-protein association network. The network is optimized based on the supervised signals from the downstream task-the DPI prediction. Through information passing on this drug-protein association network, a Graph Neural Network can capture the network-level information among diverse drugs and proteins. By combining the network-level information and the learning-based method, BridgeDPI achieves significant improvement in three real-world DPI datasets. Moreover, the case study further verifies the effectiveness and reliability of BridgeDPI. AVAILABILITY AND IMPLEMENTATION: The source code of BridgeDPI can be accessed at https://github.com/SenseTime-Knowledge-Mining/BridgeDPI. The source data used in this study is available on the https://github.com/IBM/InterpretableDTIP (for the BindingDB dataset), https://github.com/masashitsubaki/CPI_prediction (for the C.ELEGANS and HUMAN) datasets, http://dude.docking.org/ (for the DUD-E dataset), repectively.

Essential proteins are considered the foundation of life as they are indispensable for the survival of living organisms. Computational methods for essential protein discovery provide a fast way to identify essential proteins. But most of them heavily rely on various biological information, especially protein-protein interaction networks, which limits their practical applications. With the rapid development of high-throughput sequencing technology, sequencing data has become the most accessible biological data. However, using only protein sequence information to predict essential proteins has limited accuracy. In this paper, we propose EP-EDL, an ensemble deep learning model using only protein sequence information to predict human essential proteins. EP-EDL integrates multiple classifiers to alleviate the class imbalance problem and to improve prediction accuracy and robustness. In each base classifier, we employ multi-scale text convolutional neural networks to extract useful features from protein sequence feature matrices with evolutionary information. Our computational results show that EP-EDL outperforms the state-of-the-art sequence-based methods. Furthermore, EP-EDL provides a more practical and flexible way for biologists to accurately predict essential proteins. The source code and datasets can be downloaded from https://github.com/CSUBioGroup/EP-EDL.

Automatic International Classification of Diseases (ICD) coding is defined as a kind of text multi-label classification problem, which is difficult because the number of labels is very large and the distribution of labels is unbalanced. The label-wise attention mechanism is widely used in automatic ICD coding because it can assign weights to every word in full Electronic Medical Records (EMR) for different ICD codes. However, the label-wise attention mechanism is redundant and costly in computing. In this paper, we propose a pseudo label-wise attention mechanism to tackle the problem. Instead of computing different attention modes for different ICD codes, the pseudo label-wise attention mechanism automatically merges similar ICD codes and computes only one attention mode for the similar ICD codes, which greatly compresses the number of attention modes and improves the predicted accuracy. In addition, we apply a more convenient and effective way to obtain the ICD vectors, and thus our model can predict new ICD codes by calculating the similarities between EMR vectors and ICD vectors. Our model demonstrates effectiveness in extensive computational experiments. On the public MIMIC-III dataset and private Xiangya dataset, our model achieves the best performance on micro F1 (0.583 and 0.806), micro AUC (0.986 and 0.994), P@8 (0.756 and 0.413), and costs much smaller GPU memory (about 26.1% of the models with label-wise attention). Furthermore, we verify the ability of our model in predicting new ICD codes. The interpretablility analysis and case study show the effectiveness and reliability of the patterns obtained by the pseudo label-wise attention mechanism.

Accurate prediction of essential proteins by using computational methods can effectively reduce the cost of wet-lab experiments. Existing computational methods usually rely on constructed protein-protein interaction (PPI) networks with different kinds of biological data. However, high-quality PPI networks and other biological data are not available for all proteins. Thus, it is very necessary and valuable to develop accurate methods for fast and effective prediction of essential proteins by using only protein sequences. We propose EPGBDT, a machine learning ensemble model, to improve the performance of essential protein prediction by using only protein sequences. EP-GBDT has an ensemble structure that combines multiple Gradient Boosting Decision Tree (GBDT) base classifiers. In addition, to reduce the effects of imbalanced dataset, EP-GBDT uses a sampling technique. The results show that EP-GBDT outperforms state-of-the-art sequence-based methods and network-based centrality measures. The source code and datasets can be downloaded from https://github.com/CSUBioGroup/EP-GBDT.

MOTIVATION: Emerging studies indicate that circular RNAs (circRNAs) are widely involved in the progression of human diseases. Due to its special structure which is stable, circRNAs are promising diagnostic and prognostic biomarkers for diseases. However, the experimental verification of circRNA-disease associations is expensive and limited to small-scale. Effective computational methods for predicting potential circRNA-disease associations are regarded as a matter of urgency. Although several models have been proposed, over-reliance on known associations and the absence of characteristics of biological functions make precise predictions are still challenging. RESULTS: In this study, we propose a method for predicting CircRNA-disease associations based on sequence and ontology representations, named CDASOR, with convolutional and recurrent neural networks. For sequences of circRNAs, we encode them with continuous k-mers, get low-dimensional vectors of k-mers, extract their local feature vectors with 1D CNN and learn their long-term dependencies with bi-directional long short-term memory. For diseases, we serialize disease ontology into sentences containing the hierarchy of ontology, obtain low-dimensional vectors for disease ontology terms and get terms' dependencies. Furthermore, we get association patterns of circRNAs and diseases from known circRNA-disease associations with neural networks. After the above steps, we get circRNAs' and diseases' high-level representations, which are informative to improve the prediction. The experimental results show that CDASOR provides an accurate prediction. Importing the characteristics of biological functions, CDASOR achieves impressive predictions in the de novo test. In addition, 6 of the top-10 predicted results are verified by the published literature in the case studies. AVAILABILITY AND IMPLEMENTATION: The code and data of CDASOR are freely available at https://github.com/BioinformaticsCSU/CDASOR.

Computational methods including centrality and machine learning-based methods have been proposed to identify essential proteins for understanding the minimum requirements of the survival and evolution of a cell. In centrality methods, researchers are required to design a score function which is based on prior knowledge, yet is usually not sufficient to capture the complexity of biological information. In machine learning-based methods, some selected biological features cannot represent the complete properties of biological information as they lack a computational framework to automatically select features. To tackle these problems, we propose a deep learning framework to automatically learn biological features without prior knowledge. We use node2vec technique to automatically learn a richer representation of protein-protein interaction (PPI) network topologies than a score function. Bidirectional long short term memory cells are applied to capture non-local relationships in gene expression data. For subcellular localization information, we exploit a high dimensional indicator vector to characterize their feature. To evaluate the performance of our method, we tested it on PPI network of S. cerevisiae. Our experimental results demonstrate that the performance of our method is better than traditional centrality methods and is superior to existing machine learning-based methods. To explore which of the three types of biological information is the most vital element, we conduct an ablation study by removing each component in turn. Our results show that the PPI network embedding contributes most to the improvement. In addition, gene expression profiles and subcellular localization information are also helpful to improve the performance in identification of essential proteins.

A growing amount of evidence suggests that long non-coding RNAs (lncRNAs) play important roles in the regulation of biological processes in many human diseases. However, the number of experimentally verified lncRNA-disease associations is very limited. Thus, various computational approaches are proposed to predict lncRNA-disease associations. Current matrix factorization-based methods cannot capture the complex non-linear relationship between lncRNAs and diseases, and traditional machine learning-based methods are not sufficiently powerful to learn the representation of lncRNAs and diseases. Considering these limitations in existing computational methods, we propose a deep matrix factorization model to predict lncRNA-disease associations (DMFLDA in short). DMFLDA uses a cascade of non-linear hidden layers to learn latent representation to represent lncRNAs and diseases. By using non-linear hidden layers, DMFLDA captures the more complex non-linear relationship between lncRNAs and diseases than traditional matrix factorization-based methods. In addition, DMFLDA learns features directly from the lncRNA-disease interaction matrix and thus can obtain more accurate representation learning for lncRNAs and diseases than traditional machine learning methods. The low dimensional representations of the lncRNAs and diseases are fused to estimate the new interaction value. To evaluate the performance of DMFLDA, we perform leave-one-out cross-validation and 5-fold cross-validation on known experimentally verified lncRNA-disease associations. The experimental results show that DMFLDA performs better than the existing methods. The case studies show that many predicted interactions of colorectal cancer, prostate cancer, and renal cancer have been verified by recent biomedical literature. The source code and datasets can be obtained from https://github.com/CSUBioGroup/DMFLDA.

Knowledge of protein functions plays an important role in biology and medicine. With the rapid development of high-throughput technologies, a huge number of proteins have been discovered. However, there are a great number of proteins without functional annotations. A protein usually has multiple functions and some functions or biological processes require interactions of a plurality of proteins. Additionally, Gene Ontology provides a useful classification for protein functions and contains more than 40,000 terms. We propose a deep learning framework called DeepGOA to predict protein functions with protein sequences and protein-protein interaction (PPI) networks. For protein sequences, we extract two types of information: sequence semantic information and subsequence-based features. We use the word2vec technique to numerically represent protein sequences, and utilize a Bi-directional Long and Short Time Memory (Bi-LSTM) and multi-scale convolutional neural network (multi-scale CNN) to obtain the global and local semantic features of protein sequences, respectively. Additionally, we use the InterPro tool to scan protein sequences for extracting subsequence-based information, such as domains and motifs. Then, the information is plugged into a neural network to generate high-quality features. For the PPI network, the Deepwalk algorithm is applied to generate its embedding information of PPI. Then the two types of features are concatenated together to predict protein functions. To evaluate the performance of DeepGOA, several different evaluation methods and metrics are utilized. The experimental results show that DeepGOA outperforms DeepGO and BLAST.

In recent years, more and more evidence indicates that circular RNAs (circRNAs) with covalently closed loop play various roles in biological processes. Dysregulation and mutation of circRNAs may be implicated in diseases. Due to its stable structure and resistance to degradation, circRNAs provide great potential to be diagnostic biomarkers. Therefore, predicting circRNA-disease associations is helpful in disease diagnosis. However, there are few experimentally validated associations between circRNAs and diseases. Although several computational methods have been proposed, precisely representing underlying features and grasping the complex structures of data are still challenging. In this paper, we design a new method, called DMFCDA (Deep Matrix Factorization CircRNA-Disease Association), to infer potential circRNA-disease associations. DMFCDA takes both explicit and implicit feedback into account. Then, it uses a projection layer to automatically learn latent representations of circRNAs and diseases. With multi-layer neural networks, DMFCDA can model the non-linear associations to grasp the complex structure of data. We assess the performance of DMFCDA using leave-one cross-validation and 5-fold cross-validation on two datasets. Computational results show that DMFCDA efficiently infers circRNA-disease associations according to AUC values, the percentage of precisely retrieved associations in various top ranks, and statistical comparison. We also conduct case studies to evaluate DMFCDA. All results show that DMFCDA provides accurate predictions.

Genes that are thought to be critical for the survival of organisms or cells are called essential genes. The prediction of essential genes and their products (essential proteins) is of great value in exploring the mechanism of complex diseases, the study of the minimal required genome for living cells and the development of new drug targets. As laboratory methods are often complicated, costly and time-consuming, a great many of computational methods have been proposed to identify essential genes/proteins from the perspective of the network level with the in-depth understanding of network biology and the rapid development of biotechnologies. Through analyzing the topological characteristics of essential genes/proteins in protein-protein interaction networks (PINs), integrating biological information and considering the dynamic features of PINs, network-based methods have been proved to be effective in the identification of essential genes/proteins. In this paper, we survey the advanced methods for network-based prediction of essential genes/proteins and present the challenges and directions for future research.

MOTIVATION: Protein-protein interactions (PPIs) play important roles in many biological processes. Conventional biological experiments for identifying PPI sites are costly and time-consuming. Thus, many computational approaches have been proposed to predict PPI sites. Existing computational methods usually use local contextual features to predict PPI sites. Actually, global features of protein sequences are critical for PPI site prediction. RESULTS: A new end-to-end deep learning framework, named DeepPPISP, through combining local contextual and global sequence features, is proposed for PPI site prediction. For local contextual features, we use a sliding window to capture features of neighbors of a target amino acid as in previous studies. For global sequence features, a text convolutional neural network is applied to extract features from the whole protein sequence. Then the local contextual and global sequence features are combined to predict PPI sites. By integrating local contextual and global sequence features, DeepPPISP achieves the state-of-the-art performance, which is better than the other competing methods. In order to investigate if global sequence features are helpful in our deep learning model, we remove or change some components in DeepPPISP. Detailed analyses show that global sequence features play important roles in DeepPPISP. AVAILABILITY AND IMPLEMENTATION: The DeepPPISP web server is available at http://bioinformatics.csu.edu.cn/PPISP/. The source code can be obtained from https://github.com/CSUBioGroup/DeepPPISP. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Knowledge of protein-binding residues (PBRs) improves our understanding of protein-protein interactions, contributes to the prediction of protein functions and facilitates protein-protein docking calculations. While many sequence-based predictors of PBRs were published, they offer modest levels of predictive performance and most of them cross-predict residues that interact with other partners. One unexplored option to improve the predictive quality is to design consensus predictors that combine results produced by multiple methods. RESULTS: We empirically investigate predictive performance of a representative set of nine predictors of PBRs. We report substantial differences in predictive quality when these methods are used to predict individual proteins, which contrast with the dataset-level benchmarks that are currently used to assess and compare these methods. Our analysis provides new insights for the cross-prediction concern, dissects complementarity between predictors and demonstrates that predictive performance of the top methods depends on unique characteristics of the input protein sequence. Using these insights, we developed PROBselect, first-of-its-kind consensus predictor of PBRs. Our design is based on the dynamic predictor selection at the protein level, where the selection relies on regression-based models that accurately estimate predictive performance of selected predictors directly from the sequence. Empirical assessment using a low-similarity test dataset shows that PROBselect provides significantly improved predictive quality when compared with the current predictors and conventional consensuses that combine residue-level predictions. Moreover, PROBselect informs the users about the expected predictive quality for the prediction generated from a given input protein. AVAILABILITY AND IMPLEMENTATION: PROBselect is available at http://bioinformatics.csu.edu.cn/PROBselect/home/index. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Drug discovery is known for the large amount of money and time it consumes and the high risk it takes. Drug repositioning has, therefore, become a popular approach to save time and cost by finding novel indications for approved drugs. In order to distinguish these novel indications accurately in a great many of latent associations between drugs and diseases, it is necessary to exploit abundant heterogeneous information about drugs and diseases. RESULTS: In this article, we propose a meta-path-based computational method called NEDD to predict novel associations between drugs and diseases using heterogeneous information. First, we construct a heterogeneous network as an undirected graph by integrating drug-drug similarity, disease-disease similarity, and known drug-disease associations. NEDD uses meta paths of different lengths to explicitly capture the indirect relationships, or high order proximity, within drugs and diseases, by which the low dimensional representation vectors of drugs and diseases are obtained. NEDD then uses a random forest classifier to predict novel associations between drugs and diseases. CONCLUSIONS: The experiments on a gold standard dataset which contains 1933 validated drug-disease associations show that NEDD produces superior prediction results compared with the state-of-the-art approaches.

Identifying drug-target interactions (DTIs) is an important task in drug discovery. Various computational models have been proposed to predict potential association between drugs and targets. However, it is still a great challenge to accurately predict the potential drug-target interactions with rare known drug-target interactions. In this work, we propose a heterogeneous network embedding model to predict drug-target interactions, called HNEDTI. Based on the assumption that similar drugs share similar patterns of relationships with target proteins, we integrate the drug-drug similarity network, target-target similarity network and known drug-target interactions into a heterogeneous network. HNEDTI can learn more accurate feature representation of drugs and targets by extract both local and global information of the heterogeneous network from different lengths of meta-paths. The low dimensional feature representation vectors of drugs and targets are applied to random forest model to predict whether the given drug-target pair has an interaction. The evaluation on four benchmark datasets (Enzyme, Ion Channel, GPCR and Nuclear Receptor) shows that our method HNEDTI outperforms the previous methods.

Long non-coding RNAs (lncRNAs) are the foundation for understanding mechanisms of many human diseases. Considering the limited number of known experimentally verified associations between lncRNAs and diseases, it is appealing to develop accurate and effective computational methods to identify lncRNA-disease associations. Conventional matrix factorization-based methods cannot model complicated associations between lncRNAs and diseases. In this study, we propose a novel computational framework, through combining linear and non-linear features, which is used for lncRNA-disease association prediction. In our model, a conventional matrix factorization method is applied to extract linear features between lncRNAs and diseases. Deep learning techniques (fully connected layers) are applied to extract nonlinear features between lncRNAs and diseases. Finally, linear and non-linear features are fused to improve predictive performance. Compared to previous studies, our model can take advantages of the combination of linear and non-linear features between lncRNAs and diseases, and thus can effectively identify potential lncRNA-disease associations. The results show that our method achieves state-of-the-art performance in the leave-one-out cross-validation. The source codes of our method can be found at https://github.com/CSUBioGroup/DMFLDA2.

BACKGROUND: Essential proteins are crucial for cellular life and thus, identification of essential proteins is an important topic and a challenging problem for researchers. Recently lots of computational approaches have been proposed to handle this problem. However, traditional centrality methods cannot fully represent the topological features of biological networks. In addition, identifying essential proteins is an imbalanced learning problem; but few current shallow machine learning-based methods are designed to handle the imbalanced characteristics. RESULTS: We develop DeepEP based on a deep learning framework that uses the node2vec technique, multi-scale convolutional neural networks and a sampling technique to identify essential proteins. In DeepEP, the node2vec technique is applied to automatically learn topological and semantic features for each protein in protein-protein interaction (PPI) network. Gene expression profiles are treated as images and multi-scale convolutional neural networks are applied to extract their patterns. In addition, DeepEP uses a sampling method to alleviate the imbalanced characteristics. The sampling method samples the same number of the majority and minority samples in a training epoch, which is not biased to any class in training process. The experimental results show that DeepEP outperforms traditional centrality methods. Moreover, DeepEP is better than shallow machine learning-based methods. Detailed analyses show that the dense vectors which are generated by node2vec technique contribute a lot to the improved performance. It is clear that the node2vec technique effectively captures the topological and semantic properties of PPI network. The sampling method also improves the performance of identifying essential proteins. CONCLUSION: We demonstrate that DeepEP improves the prediction performance by integrating multiple deep learning techniques and a sampling method. DeepEP is more effective than existing methods.

ICD-9 (the Ninth Revision of International Classification of Diseases) is widely used to describe a patient's diagnosis. Accurate automated ICD-9 coding is important because manual coding is expensive, time-consuming, and inefficient. Inspired by the recent successes of deep learning, in this study, we present a deep learning framework called DeepLabeler to automatically assign ICD-9 codes. DeepLabeler combines the convolutional neural network with the 'Document to Vector' technique to extract and encode local and global features. Our proposed DeepLabeler demonstrates its effectiveness by achieving state-of-the-art performance, i.e., 0.335 micro F-measure on MIMIC-II dataset and 0.408 micro F-measure on MIMIC-III dataset. It outperforms classical hierarchy-based SVM and flat-SVM both on these two datasets by at least 14 percent. Furthermore, we analyze the deep neural network structure to discover the vital elements in the success of DeepLabeler. We find that the convolutional neural network is the most effective component in our network and the 'Document to Vector' technique is also necessary for enhancing classification performance since it extracts well-recognized global features. Extensive experimental results demonstrate that the great promise of deep learning techniques in the field of text multi-label classification and automated medical coding.