We introduce Concept Bottleneck Protein Language Models (CB-pLM), a generative masked language model with a layer where each neuron corresponds to an interpretable concept. Our architecture offers three key benefits: i) Control: We can intervene on concept values to precisely control the properties of generated proteins, achieving a 3$\times$ larger change in desired concept values compared to baselines. ii) Interpretability: A linear mapping between concept values and predicted tokens allows transparent analysis of the model's decision-making process. iii) Debugging: This transparency facilitates easy debugging of trained models. Our models achieve pre-training perplexity and downstream task performance comparable to traditional masked protein language models, demonstrating that interpretability does not compromise performance. While adaptable to any language model, we focus on masked protein language models due to their importance in drug discovery and the ability to validate our model's capabilities through real-world experiments and expert knowledge. We scale our CB-pLM from 24 million to 3 billion parameters, making them the largest Concept Bottleneck Models trained and the first capable of generative language modeling.

We introduce a generative model with an intrinsically interpretable layer---a concept bottleneck layer---that constrains the model to encode human-understandable concepts. The concept bottleneck layer partitions the generative model into three parts: the pre-concept bottleneck portion, the CB layer, and the post-concept bottleneck portion. To train CB generative models, we complement the traditional task-based loss function for training generative models with a concept loss and an orthogonality loss. The CB layer and these loss terms are model agnostic, which we demonstrate by applying the CB layer to three different families of generative models: generative adversarial networks, variational autoencoders, and diffusion models. On multiple datasets across different types of generative models, steering a generative model, with the CB layer, outperforms all baselines---in some cases, it is \textit{10 times} more effective. In addition, we show how the CB layer can be used to interpret the output of the generative model and debug the model during or post training.

Saliency methods have been widely used to highlight important input features in model predictions. Most existing methods use backpropagation on a modified gradient function to generate saliency maps. Thus, noisy gradients can result in unfaithful feature attributions. In this paper, we tackle this issue and introduce a {\it saliency guided training} procedure for neural networks to reduce noisy gradients used in predictions while retaining the predictive performance of the model. Our saliency guided training procedure iteratively masks features with small and potentially noisy gradients while maximizing the similarity of model outputs for both masked and unmasked inputs. We apply the saliency guided training procedure to various synthetic and real data sets from computer vision, natural language processing, and time series across diverse neural architectures, including Recurrent Neural Networks, Convolutional Networks, and Transformers. Through qualitative and quantitative evaluations, we show that saliency guided training procedure significantly improves model interpretability across various domains while preserving its predictive performance.

It is challenging to associate features such as human health outcomes, diet, environmental conditions, or other metadata to microbial community measurements, due in part to their quantitative properties. Microbiome multi-omics are typically noisy, sparse (zero-inflated), high-dimensional, extremely non-normal, and often in the form of count or compositional measurements. Here we introduce an optimized combination of novel and established methodology to assess multivariable association of microbial community features with complex metadata in population-scale observational studies. Our approach, MaAsLin 2 (Microbiome Multivariable Associations with Linear Models), uses generalized linear and mixed models to accommodate a wide variety of modern epidemiological studies, including cross-sectional and longitudinal designs, as well as a variety of data types (e.g., counts and relative abundances) with or without covariates and repeated measurements. To construct this method, we conducted a large-scale evaluation of a broad range of scenarios under which straightforward identification of meta-omics associations can be challenging. These simulation studies reveal that MaAsLin 2's linear model preserves statistical power in the presence of repeated measures and multiple covariates, while accounting for the nuances of meta-omics features and controlling false discovery. We also applied MaAsLin 2 to a microbial multi-omics dataset from the Integrative Human Microbiome (HMP2) project which, in addition to reproducing established results, revealed a unique, integrated landscape of inflammatory bowel diseases (IBD) across multiple time points and omics profiles.

Saliency methods are used extensively to highlight the importance of input features in model predictions. These methods are mostly used in vision and language tasks, and their applications to time series data is relatively unexplored. In this paper, we set out to extensively compare the performance of various saliency-based interpretability methods across diverse neural architectures, including Recurrent Neural Network, Temporal Convolutional Networks, and Transformers in a new benchmark of synthetic time series data. We propose and report multiple metrics to empirically evaluate the performance of saliency methods for detecting feature importance over time using both precision (i.e., whether identified features contain meaningful signals) and recall (i.e., the number of features with signal identified as important). Through several experiments, we show that (i) in general, network architectures and saliency methods fail to reliably and accurately identify feature importance over time in time series data, (ii) this failure is mainly due to the conflation of time and feature domains, and (iii) the quality of saliency maps can be improved substantially by using our proposed two-step temporal saliency rescaling (TSR) approach that first calculates the importance of each time step before calculating the importance of each feature at a time step.

MOTIVATION: Integrative analysis of genomic data that includes statistical methods in combination with visual exploration has gained widespread adoption. Many existing methods involve a combination of tools and resources: user interfaces that provide visualization of large genomic datasets, and computational environments that focus on data analyses over various subsets of a given dataset. Over the last few years, we have developed Epiviz as an integrative and interactive genomic data analysis tool that incorporates visualization tightly with state-of-the-art statistical analysis framework. RESULTS: In this article, we present Epiviz Feed, a proactive and automatic visual analytics system integrated with Epiviz that alleviates the burden of manually executing data analysis required to test biologically meaningful hypotheses. Results of interest that are proactively identified by server-side computations are listed as notifications in a feed. The feed turns genomic data analysis into a collaborative work between the analyst and the computational environment, which shortens the analysis time and allows the analyst to explore results efficiently. We discuss three ways where the proposed system advances the field of genomic data analysis: (i) takes the first step of proactive data analysis by utilizing available CPU power from the server to automate the analysis process; (ii) summarizes hypothesis test results in a way that analysts can easily understand and investigate; (iii) enables filtering and grouping of analysis results for quick search. This effort provides initial work on systems that substantially expand how computational and visualization frameworks can be tightly integrated to facilitate interactive genomic data analysis. AVAILABILITY AND IMPLEMENTATION: The source code for Epiviz Feed application is available at http://github.com/epiviz/epiviz_feed_polymer. The Epiviz Computational Server is available at http://github.com/epiviz/epiviz-feed-computation. Please refer to Epiviz documentation site for details: http://epiviz.github.io/.

MOTIVATION: Genomic data repositories like The Cancer Genome Atlas, Encyclopedia of DNA Elements, Bioconductor's AnnotationHub and ExperimentHub etc., provide public access to large amounts of genomic data as flat files. Researchers often download a subset of data files from these repositories to perform exploratory data analysis. We developed Epiviz File Server, a Python library that implements an in situ data query system for local or remotely hosted indexed genomic files, not only for visualization but also data transformation. The File Server library decouples data retrieval and transformation from specific visualization and analysis tools and provides an abstract interface to define computations independent of the location, format or structure of the file. We demonstrate the File Server in two use cases: (i) integration with Galaxy workflows and (ii) using Epiviz to create a custom genome browser from the Epigenome Roadmap dataset. AVAILABILITY AND IMPLEMENTATION: Epiviz File Server is open source and is available on GitHub at http://github.com/epiviz/epivizFileServer. The documentation for the File Server library is available at http://epivizfileserver.rtfd.io.

MOTIVATION: Advances in sequencing technology, inference algorithms and differential testing methodology have enabled transcript-level analysis of RNA-seq data. Yet, the inherent inferential uncertainty in transcript-level abundance estimation, even among the most accurate approaches, means that robust transcript-level analysis often remains a challenge. Conversely, gene-level analysis remains a common and robust approach for understanding RNA-seq data, but it coarsens the resulting analysis to the level of genes, even if the data strongly support specific transcript-level effects. RESULTS: We introduce a new data-driven approach for grouping together transcripts in an experiment based on their inferential uncertainty. Transcripts that share large numbers of ambiguously-mapping fragments with other transcripts, in complex patterns, often cannot have their abundances confidently estimated. Yet, the total transcriptional output of that group of transcripts will have greatly reduced inferential uncertainty, thus allowing more robust and confident downstream analysis. Our approach, implemented in the tool terminus, groups together transcripts in a data-driven manner allowing transcript-level analysis where it can be confidently supported, and deriving transcriptional groups where the inferential uncertainty is too high to support a transcript-level result. AVAILABILITY AND IMPLEMENTATION: Terminus is implemented in Rust, and is freely available and open source. It can be obtained from https://github.com/COMBINE-lab/Terminus. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Recent efforts to improve the interpretability of deep neural networks use saliency to characterize the importance of input features to predictions made by models. Work on interpretability using saliency-based methods on Recurrent Neural Networks (RNNs) has mostly targeted language tasks, and their applicability to time series data is less understood. In this work we analyze saliency-based methods for RNNs, both classical and gated cell architectures. We show that RNN saliency vanishes over time, biasing detection of salient features only to later time steps and are, therefore, incapable of reliably detecting important features at arbitrary time intervals. To address this vanishing saliency problem, we propose a novel RNN cell structure (input-cell attention), which can extend any RNN cell architecture. At each time step, instead of only looking at the current input vector, input-cell attention uses a fixed-size matrix embedding, each row of the matrix attending to different inputs from current or previous time steps. Using synthetic data, we show that the saliency map produced by the input-cell attention RNN is able to faithfully detect important features regardless of their occurrence in time. We also apply the input-cell attention RNN on a neuroscience task analyzing functional Magnetic Resonance Imaging (fMRI) data for human subjects performing a variety of tasks. In this case, we use saliency to characterize brain regions (input features) for which activity is important to distinguish between tasks. We show that standard RNN architectures are only capable of detecting important brain regions in the last few time steps of the fMRI data, while the input-cell attention model is able to detect important brain region activity across time without latter time step biases.

SUMMARY: We developed the metagenomeFeatures R Bioconductor package along with annotation packages for three 16S rRNA databases (Greengenes, RDP and SILVA) to facilitate working with 16S rRNA databases and marker-gene survey feature data. The metagenomeFeatures package defines two classes, MgDb for working with 16S rRNA sequence databases, and mgFeatures for marker-gene survey feature data. The associated annotation packages provide a consistent interface to the different databases facilitating database comparison and exploration. The mgFeatures-class represents a crucial step in the development of a common data structure for working with 16S marker-gene survey data in R. AVAILABILITY AND IMPLEMENTATION: https://bioconductor.org/packages/release/bioc/html/metagenomeFeatures.html. SUPPLEMENTARY INFORMATION: Supplementary material is available at Bioinformatics online.

BACKGROUND: Ultra-fast pseudo-alignment approaches are the tool of choice in transcript-level RNA sequencing (RNA-seq) analyses. Unfortunately, these methods couple the tasks of pseudo-alignment and transcript quantification. This coupling precludes the direct usage of pseudo-alignment to other expression analyses, including alternative splicing or differential gene expression analysis, without including a non-essential transcript quantification step. RESULTS: In this paper, we introduce a transcriptome segmentation approach to decouple these two tasks. We propose an efficient algorithm to generate maximal disjoint segments given a transcriptome reference library on which ultra-fast pseudo-alignment can be used to produce per-sample segment counts. We show how to apply these maximally unambiguous count statistics in two specific expression analyses - alternative splicing and gene differential expression - without the need of a transcript quantification step. Our experiments based on simulated and experimental data showed that the use of segment counts, like other methods that rely on local coverage statistics, provides an advantage over approaches that rely on transcript quantification in detecting and correctly estimating local splicing in the case of incomplete transcript annotations. CONCLUSIONS: The transcriptome segmentation approach implemented in Yanagi exploits the computational and space efficiency of pseudo-alignment approaches. It significantly expands their applicability and interpretability in a variety of RNA-seq analyses by providing the means to model and capture local coverage variation in these analyses.

Analysis of differential alternative splicing from RNA-seq data is complicated by the fact that many RNA-seq reads map to multiple transcripts, and that annotated transcripts from a given gene are often a small subset of many possible complete transcripts for that gene. Here we describe Yanagi, a tool which segments a transcriptome into disjoint regions to create a segments library from a complete transcriptome annotation that preserves all of its consecutive regions of a given length L while distinguishing annotated alternative splicing events in the transcriptome. In this paper, we formalize this concept of transcriptome segmentation and propose an efficient algorithm for generating segment libraries based on a length parameter dependent on specific RNA-Seq library construction. The resulting segment sequences can be used with pseudo-alignment tools to quantify expression at the segment level. We characterize the segment libraries for the reference transcriptomes of Drosophila melanogaster and Homo sapiens. Finally, we demonstrate the utility of quantification using a segment library based on an analysis of differential exon skipping in Drosophila melanogaster and Homo sapiens. The notion of transcript segmentation as introduced here and implemented in Yanagi will open the door for the application of lightweight, ultra-fast pseudo-alignment algorithms in a wide variety of analyses of transcription variation.

This special issue introduces current developments in bioinformatics, the use of computational techniques in the service of the life sciences, in the specific context of the analysis and interpretation of DNA molecules, where DNA is the primary informational molecule of all life. To fully appreciate the papers in this issue, some background knowledge in biology is required. The DNA in a cell is organized into a genome consisting of one or more chromosomes, each of which is a DNA molecule, which can be thought of as a long string over the alphabet of nucleotides, A, C, G, and T. Within a DNA string or sequence are substrings that are genes that normally encode the recipe for the functional proteins in the cell. The human genome is organized into 23 pairs of chromosomes, one maternal and one paternal chromosome set, containing a total of approximately six billion nucleotides. Amazingly, all ~1 trillion cells in one’s body, of diverse form and function, contain the same “molecular blueprint” with a near perfect copy of their genome.

MOTIVATION: DNA methylation aberrations are now known to, almost universally, accompany the initiation and progression of cancers. In particular, the colon cancer epigenome contains specific genomic regions that, along with differences in methylation levels with respect to normal colon tissue, also show increased epigenetic and gene expression heterogeneity at the population level, i.e. across tumor samples, in comparison with other regions in the genome. Tumors are highly heterogeneous at the clonal level as well, and the relationship between clonal and population heterogeneity is poorly understood. RESULTS: We present an approach that uses sequencing reads from high-throughput sequencing of bisulfite-converted DNA to reconstruct heterogeneous cell populations by assembling cell-specific methylation patterns. Our methodology is based on the solution of a specific class of minimum cost network flow problems. We use our methods to analyze the relationship between clonal heterogeneity and population heterogeneity in high-coverage data from multiple samples of colon tumor and matched normal tissues. AVAILABILITY AND IMPLEMENTATION: http://github.com/hcorrada/methylFlow CONTACT: [email protected] SUPPLEMENTARY INFORMATION: SUPPLEMENTARY INFORMATION is available at Bioinformatics online.

Sequencing and microarray samples often are collected or processed in multiple batches or at different times. This often produces technical biases that can lead to incorrect results in the downstream analysis. There are several existing batch adjustment tools for '-omics' data, but they do not indicate a priori whether adjustment needs to be conducted or how correction should be applied. We present a software pipeline, BatchQC, which addresses these issues using interactive visualizations and statistics that evaluate the impact of batch effects in a genomic dataset. BatchQC can also apply existing adjustment tools and allow users to evaluate their benefits interactively. We used the BatchQC pipeline on both simulated and real data to demonstrate the effectiveness of this software toolkit. AVAILABILITY AND IMPLEMENTATION: BatchQC is available through Bioconductor: http://bioconductor.org/packages/BatchQC and GitHub: https://github.com/mani2012/BatchQC CONTACT: [email protected] information: Supplementary data are available at Bioinformatics online.

MOTIVATION: Developing targeted therapeutics and identifying biomarkers relies on large amounts of research participant data. Beyond human DNA, scientists now investigate the DNA of micro-organisms inhabiting the human body. Recent work shows that an individual's collection of microbial DNA consistently identifies that person and could be used to link a real-world identity to a sensitive attribute in a research dataset. Unfortunately, the current suite of DNA-specific privacy-preserving analysis tools does not meet the requirements for microbiome sequencing studies. RESULTS: To address privacy concerns around microbiome sequencing, we implement metagenomic analyses using secure computation. Our implementation allows comparative analysis over combined data without revealing the feature counts for any individual sample. We focus on three analyses and perform an evaluation on datasets currently used by the microbiome research community. We use our implementation to simulate sharing data between four policy-domains. Additionally, we describe an application of our implementation for patients to combine data that allows drug developers to query against and compensate patients for the analysis. AVAILABILITY AND IMPLEMENTATION: The software is freely available for download at: http://cbcb.umd.edu/∼hcorrada/projects/secureseq.html SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online. CONTACT: [email protected].

BACKGROUND: Computational and visual data analysis for genomics has traditionally involved a combination of tools and resources, of which the most ubiquitous consist of genome browsers, focused mainly on integrative visualization of large numbers of big datasets, and computational environments, focused on data modeling of a small number of moderately sized datasets. Workflows that involve the integration and exploration of multiple heterogeneous data sources, small and large, public and user specific have been poorly addressed by these tools. In our previous work, we introduced Epiviz, which bridges the gap between the two types of tools, simplifying these workflows. RESULTS: In this paper we expand on the design decisions behind Epiviz, and introduce a series of new advanced features that further support the type of interactive exploratory workflow we have targeted. We discuss three ways in which Epiviz advances the field of genomic data analysis: 1) it brings code to interactive visualizations at various different levels; 2) takes the first steps in the direction of collaborative data analysis by incorporating user plugins from source control providers, as well as by allowing analysis states to be shared among the scientific community; 3) combines established analysis features that have never before been available simultaneously in a genome browser. In our discussion section, we present security implications of the current design, as well as a series of limitations and future research steps. CONCLUSIONS: Since many of the design choices of Epiviz are novel in genomics data analysis, this paper serves both as a document of our own approaches with lessons learned, as well as a start point for future efforts in the same direction for the genomics community.

MOTIVATION: The recently released Infinium HumanMethylation450 array (the '450k' array) provides a high-throughput assay to quantify DNA methylation (DNAm) at ∼450 000 loci across a range of genomic features. Although less comprehensive than high-throughput sequencing-based techniques, this product is more cost-effective and promises to be the most widely used DNAm high-throughput measurement technology over the next several years. RESULTS: Here we describe a suite of computational tools that incorporate state-of-the-art statistical techniques for the analysis of DNAm data. The software is structured to easily adapt to future versions of the technology. We include methods for preprocessing, quality assessment and detection of differentially methylated regions from the kilobase to the megabase scale. We show how our software provides a powerful and flexible development platform for future methods. We also illustrate how our methods empower the technology to make discoveries previously thought to be possible only with sequencing-based methods. AVAILABILITY AND IMPLEMENTATION: http://bioconductor.org/packages/release/bioc/html/minfi.html. CONTACT: [email protected]; [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Abstract Motivation: Base-calling of sequencing data produced by high-throughput sequencing platforms is a fundamental process in current bioinformatics analysis. However, existing third-party probabilistic or machine-learning methods that significantly improve the accuracy of base-calls on these platforms are impractical for production use due to their computational inefficiency. Results: We directly formulate base-calling as a blind deconvolution problem and implemented BlindCall as an efficient solver to this inverse problem. BlindCall produced base-calls at accuracy comparable to state-of-the-art probabilistic methods while processing data at rates 10 times faster in most cases. The computational complexity of BlindCall scales linearly with read length making it better suited for new long-read sequencing technologies. Availability and Implementation: BlindCall is implemented as a set of Matlab scripts available for download at http://cbcb.umd.edu/∼hcorrada/secgen. Contact: [email protected]

BACKGROUND: Early screening for cancer is arguably one of the greatest public health advances over the last fifty years. However, many cancer screening tests are invasive (digital rectal exams), expensive (mammograms, imaging) or both (colonoscopies). This has spurred growing interest in developing genomic signatures that can be used for cancer diagnosis and prognosis. However, progress has been slowed by heterogeneity in cancer profiles and the lack of effective computational prediction tools for this type of data. RESULTS: We developed anti-profiles as a first step towards translating experimental findings suggesting that stochastic across-sample hyper-variability in the expression of specific genes is a stable and general property of cancer into predictive and diagnostic signatures. Using single-chip microarray normalization and quality assessment methods, we developed an anti-profile for colon cancer in tissue biopsy samples. To demonstrate the translational potential of our findings, we applied the signature developed in the tissue samples, without any further retraining or normalization, to screen patients for colon cancer based on genomic measurements from peripheral blood in an independent study (AUC of 0.89). This method achieved higher accuracy than the signature underlying commercially available peripheral blood screening tests for colon cancer (AUC of 0.81). We also confirmed the existence of hyper-variable genes across a range of cancer types and found that a significant proportion of tissue-specific genes are hyper-variable in cancer. Based on these observations, we developed a universal cancer anti-profile that accurately distinguishes cancer from normal regardless of tissue type (ten-fold cross-validation AUC > 0.92). CONCLUSIONS: We have introduced anti-profiles as a new approach for developing cancer genomic signatures that specifically takes advantage of gene expression heterogeneity. We have demonstrated that anti-profiles can be successfully applied to develop peripheral-blood based diagnostics for cancer and used anti-profiles to develop a highly accurate universal cancer signature. By using single-chip normalization and quality assessment methods, no further retraining of signatures developed by the anti-profile approach would be required before their application in clinical settings. Our results suggest that anti-profiles may be used to develop inexpensive and non-invasive universal cancer screening tests.

BACKGROUND: In systems biology, the task of reverse engineering gene pathways from data has been limited not just by the curse of dimensionality (the interaction space is huge) but also by systematic error in the data. The gene expression barcode reduces spurious association driven by batch effects and probe effects. The binary nature of the resulting expression calls lends itself perfectly to modern regularization approaches that thrive in high-dimensional settings. RESULTS: The Partitioned LASSO-Patternsearch algorithm is proposed to identify patterns of multiple dichotomous risk factors for outcomes of interest in genomic studies. A partitioning scheme is used to identify promising patterns by solving many LASSO-Patternsearch subproblems in parallel. All variables that survive this stage proceed to an aggregation stage where the most significant patterns are identified by solving a reduced LASSO-Patternsearch problem in just these variables. This approach was applied to genetic data sets with expression levels dichotomized by gene expression bar code. Most of the genes and second-order interactions thus selected and are known to be related to the outcomes. CONCLUSIONS: We demonstrate with simulations and data analyses that the proposed method not only selects variables and patterns more accurately, but also provides smaller models with better prediction accuracy, in comparison to several alternative methodologies.

Managing uncertain data using probabilistic frameworks has attracted much interest lately in the database literature, and a central computational challenge is probabilistic inference. This paper presents a broad class of aggregate queries, called MPF queries, inspired by the literature on probabilistic inference in statistics and machine learning. An MPF (Marginalize a Product Function) query is an aggregate query over a stylized join of several relations. In probabilistic inference, this join corresponds to taking the product of several probability distributions, while the aggregate operation corresponds to marginalization. Probabilistic inference can be expressed directly as MPF queries in a relational setting, and therefore, by optimizing evaluation of MPF queries, we provide scalable support for probabilistic inference in database systems. To optimize MPF queries, we build on ideas from database query optimization as well as traditional algorithms such as Variable Elimination and Belief Propagation from the probabilistic inference literature.