Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks

An analytical framework for decoding cell type-specific genetic variation of gene regulation

A deeper understanding of genetic regulation and functional mechanisms underlying genetic associations with complex traits and diseases is impeded by cellular heterogeneity and linkage disequilibrium. To address these limits, we introduce Huatuo, a framework to decode genetic variation of gene regulation at cell type and single-nucleotide resolutions by integrating deep-learning-based variant predictions with population-based association analyses. We apply Huatuo to generate a comprehensive cell type-specific genetic variation landscape across human tissues and further evaluate their potential roles in complex diseases and traits. Finally, we show that Huatuo's inferences permit prioritizations of driver cell types associated with complex traits and diseases and allow for systematic insights into the mechanisms of phenotype-causal genetic variation.

An intrinsically interpretable neural network architecture for sequence-to-function learning

MOTIVATION

Sequence-based deep learning approaches have been shown to predict a multitude of functional genomic readouts, including regions of open chromatin and RNA expression of genes. However, a major limitation of current methods is that model interpretation relies on computationally demanding post hoc analyses, and even then, one can often not explain the internal mechanics of highly parameterized models. Here, we introduce a deep learning architecture called totally interpretable sequence-to-function model (tiSFM). tiSFM improves upon the performance of standard multilayer convolutional models while using fewer parameters. Additionally, while tiSFM is itself technically a multilayer neural network, internal model parameters are intrinsically interpretable in terms of relevant sequence motifs.

RESULTS

We analyze published open chromatin measurements across hematopoietic lineage cell-types and demonstrate that tiSFM outperforms a state-of-the-art convolutional neural network model custom-tailored to this dataset. We also show that it correctly identifies context-specific activities of transcription factors with known roles in hematopoietic differentiation, including Pax5 and Ebf1 for B-cells, and Rorc for innate lymphoid cells. tiSFM's model parameters have biologically meaningful interpretations, and we show the utility of our approach on a complex task of predicting the change in epigenetic state as a function of developmental transition.

AVAILABILITY AND IMPLEMENTATION

The source code, including scripts for the analysis of key findings, can be found at https://github.com/boooooogey/ATAConv, implemented in Python.

The evolution and mutational robustness of chromatin accessibility in Drosophila

The evolution of regulatory regions in the genome plays a critical role in shaping the diversity of life. While this process is primarily sequence-dependent, the enormous complexity of biological systems has made it difficult to understand the factors underlying regulation and its evolution. Here, we apply deep neural networks as a tool to investigate the sequence determinants underlying chromatin accessibility in different tissues of Drosophila. We train hybrid convolution-attention neural networks to accurately predict ATAC-seq peaks using only local DNA sequences as input. We show that a model trained in one species has nearly identical performance when tested in another species, implying that the sequence determinants of accessibility are highly conserved. Indeed, model performance remains excellent even in distantly-related species. By using our model to examine species-specific gains in chromatin accessibility, we find that their orthologous inaccessible regions in other species have surprisingly similar model outputs, suggesting that these regions may be ancestrally poised for evolution. We then use in silico saturation mutagenesis to reveal evidence of selective constraint acting specifically on inaccessible chromatin regions. We further show that chromatin accessibility can be accurately predicted from short subsequences in each example. However, in silico knock-out of these sequences does not qualitatively impair classification, implying that chromatin accessibility is mutationally robust. Subsequently, we demonstrate that chromatin accessibility is predicted to be robust to large-scale random mutation even in the absence of selection. We also perform in silico evolution experiments under the regime of strong selection and weak mutation (SSWM) and show that chromatin accessibility can be extremely malleable despite its mutational robustness. However, selection acting in different directions in a tissue-specific manner can substantially slow adaptation. Finally, we identify motifs predictive of chromatin accessibility and recover motifs corresponding to known chromatin accessibility activators and repressors. These results demonstrate the conservation of the sequence determinants of accessibility and the general robustness of chromatin accessibility, as well as the power of deep neural networks as tools to answer fundamental questions in regulatory genomics and evolution.

ExplaiNN: interpretable and transparent neural networks for genomics

Deep learning models such as convolutional neural networks (CNNs) excel in genomic tasks but lack interpretability. We introduce ExplaiNN, which combines the expressiveness of CNNs with the interpretability of linear models. ExplaiNN can predict TF binding, chromatin accessibility, and de novo motifs, achieving performance comparable to state-of-the-art methods. Its predictions are transparent, providing global (cell state level) as well as local (individual sequence level) biological insights into the data. ExplaiNN can serve as a plug-and-play platform for pretrained models and annotated position weight matrices. ExplaiNN aims to accelerate the adoption of deep learning in genomic sequence analysis by domain experts.

Finding motifs using DNA images derived from sparse representations

MOTIVATION

Motifs play a crucial role in computational biology, as they provide valuable information about the binding specificity of proteins. However, conventional motif discovery methods typically rely on simple combinatoric or probabilistic approaches, which can be biased by heuristics such as substring-masking for multiple motif discovery. In recent years, deep neural networks have become increasingly popular for motif discovery, as they are capable of capturing complex patterns in data. Nonetheless, inferring motifs from neural networks remains a challenging problem, both from a modeling and computational standpoint, despite the success of these networks in supervised learning tasks.

RESULTS

We present a principled representation learning approach based on a hierarchical sparse representation for motif discovery. Our method effectively discovers gapped, long, or overlapping motifs that we show to commonly exist in next-generation sequencing datasets, in addition to the short and enriched primary binding sites. Our model is fully interpretable, fast, and capable of capturing motifs in a large number of DNA strings. A key concept emerged from our approach-enumerating at the image level-effectively overcomes the k-mers paradigm, enabling modest computational resources for capturing the long and varied but conserved patterns, in addition to capturing the primary binding sites.

AVAILABILITY AND IMPLEMENTATION

Our method is available as a Julia package under the MIT license at https://github.com/kchu25/MOTIFs.jl, and the results on experimental data can be found at https://zenodo.org/record/7783033.

Transfer learning identifies sequence determinants of cell-type specific regulatory element accessibility

Dysfunction of regulatory elements through genetic variants is a central mechanism in the pathogenesis of disease. To better understand disease etiology, there is consequently a need to understand how DNA encodes regulatory activity. Deep learning methods show great promise for modeling of biomolecular data from DNA sequence but are limited to large input data for training. Here, we develop ChromTransfer, a transfer learning method that uses a pre-trained, cell-type agnostic model of open chromatin regions as a basis for fine-tuning on regulatory sequences. We demonstrate superior performances with ChromTransfer for learning cell-type specific chromatin accessibility from sequence compared to models not informed by a pre-trained model. Importantly, ChromTransfer enables fine-tuning on small input data with minimal decrease in accuracy. We show that ChromTransfer uses sequence features matching binding site sequences of key transcription factors for prediction. Together, these results demonstrate ChromTransfer as a promising tool for learning the regulatory code.

Computational Assessment of the Expression-modulating Potential for Non-coding Variants

Large-scale genome-wide association studies (GWAS) and expression quantitative trait locus (eQTL) studies have identified multiple non-coding variants associated with genetic diseases by affecting gene expression. However, pinpointing causal variants effectively and efficiently remains a serious challenge. Here, we developed CARMEN, a novel algorithm to identify functional non-coding expression-modulating variants. Multiple evaluations demonstrated CARMEN's superior performance over state-of-the-art tools. Applying CARMEN to GWAS and eQTL datasets further pinpointed several causal variants other than the reported lead single-nucleotide polymorphisms (SNPs). CARMEN scales well with the massive datasets, and is available online as a web server at http://carmen.gao-lab.org.

A survey on algorithms to characterize transcription factor binding sites

Transcription factors (TFs) are key regulatory proteins that control the transcriptional rate of cells by binding short DNA sequences called transcription factor binding sites (TFBS) or motifs. Identifying and characterizing TFBS is fundamental to understanding the regulatory mechanisms governing the transcriptional state of cells. During the last decades, several experimental methods have been developed to recover DNA sequences containing TFBS. In parallel, computational methods have been proposed to discover and identify TFBS motifs based on these DNA sequences. This is one of the most widely investigated problems in bioinformatics and is referred to as the motif discovery problem. In this manuscript, we review classical and novel experimental and computational methods developed to discover and characterize TFBS motifs in DNA sequences, highlighting their advantages and drawbacks. We also discuss open challenges and future perspectives that could fill the remaining gaps in the field.

DiMo: discovery of microRNA motifs using deep learning and motif embedding

MicroRNAs are small regulatory RNAs that decrease gene expression after transcription in various biological disciplines. In bioinformatics, identifying microRNAs and predicting their functionalities is critical. Finding motifs is one of the most well-known and important methods for identifying the functionalities of microRNAs. Several motif discovery techniques have been proposed, some of which rely on artificial intelligence-based techniques. However, in the case of few or no training data, their accuracy is low. In this research, we propose a new computational approach, called DiMo, for identifying motifs in microRNAs and generally macromolecules of small length. We employ word embedding techniques and deep learning models to improve the accuracy of motif discovery results. Also, we rely on transfer learning models to pre-train a model and use it in cases of a lack of (enough) training data. We compare our approach with five state-of-the-art works using three real-world datasets. DiMo outperforms the selected related works in terms of precision, recall, accuracy and f1-score.

Decoding enhancer complexity with machine learning and high-throughput discovery

Enhancers are genomic DNA elements controlling spatiotemporal gene expression. Their flexible organization and functional redundancies make deciphering their sequence-function relationships challenging. This article provides an overview of the current understanding of enhancer organization and evolution, with an emphasis on factors that influence these relationships. Technological advancements, particularly in machine learning and synthetic biology, are discussed in light of how they provide new ways to understand this complexity. Exciting opportunities lie ahead as we continue to unravel the intricacies of enhancer function.

TSPTFBS 2.0: trans-species prediction of transcription factor binding sites and identification of their core motifs in plants

Introduction

An emerging approach using promoter tiling deletion via genome editing is beginning to become popular in plants. Identifying the precise positions of core motifs within plant gene promoter is of great demand but they are still largely unknown. We previously developed TSPTFBS of 265 Arabidopsis transcription factor binding sites (TFBSs) prediction models, which now cannot meet the above demand of identifying the core motif.

Methods

Here, we additionally introduced 104 maize and 20 rice TFBS datasets and utilized DenseNet for model construction on a large-scale dataset of a total of 389 plant TFs. More importantly, we combined three biological interpretability methods including DeepLIFT, in-silico tiling deletion, and in-silico mutagenesis to identify the potential core motifs of any given genomic region.

Results

For the results, DenseNet not only has achieved greater predictability than baseline methods such as LS-GKM and MEME for above 389 TFs from Arabidopsis, maize and rice, but also has greater performance on trans-species prediction of a total of 15 TFs from other six plant species. A motif analysis based on TF-MoDISco and global importance analysis (GIA) further provide the biological implication of the core motif identified by three interpretability methods. Finally, we developed a pipeline of TSPTFBS 2.0, which integrates 389 DenseNet-based models of TF binding and the above three interpretability methods.

Discussion

TSPTFBS 2.0 was implemented as a user-friendly web-server (http://www.hzau-hulab.com/TSPTFBS/), which can support important references for editing targets of any given plant promoters and it has great potentials to provide reliable editing target of genetic screen experiments in plants.

Correcting gradient-based interpretations of deep neural networks for genomics

Post hoc attribution methods can provide insights into the learned patterns from deep neural networks (DNNs) trained on high-throughput functional genomics data. However, in practice, their resultant attribution maps can be challenging to interpret due to spurious importance scores for seemingly arbitrary nucleotides. Here, we identify a previously overlooked attribution noise source that arises from how DNNs handle one-hot encoded DNA. We demonstrate this noise is pervasive across various genomic DNNs and introduce a statistical correction that effectively reduces it, leading to more reliable attribution maps. Our approach represents a promising step towards gaining meaningful insights from DNNs in regulatory genomics.

EvoAug: improving generalization and interpretability of genomic deep neural networks with evolution-inspired data augmentations

Deep neural networks (DNNs) hold promise for functional genomics prediction, but their generalization capability may be limited by the amount of available data. To address this, we propose EvoAug, a suite of evolution-inspired augmentations that enhance the training of genomic DNNs by increasing genetic variation. Random transformation of DNA sequences can potentially alter their function in unknown ways, so we employ a fine-tuning procedure using the original non-transformed data to preserve functional integrity. Our results demonstrate that EvoAug substantially improves the generalization and interpretability of established DNNs across prominent regulatory genomics prediction tasks, offering a robust solution for genomic DNNs.

A non-coding variant linked to metabolic obesity with normal weight affects actin remodelling in subcutaneous adipocytes

Recent large-scale genomic association studies found evidence for a genetic link between increased risk of type 2 diabetes and decreased risk for adiposity-related traits, reminiscent of metabolically obese normal weight (MONW) association signatures. However, the target genes and cellular mechanisms driving such MONW associations remain to be identified. Here, we systematically identify the cellular programmes of one of the top-scoring MONW risk loci, the 2q24.3 risk locus, in subcutaneous adipocytes. We identify a causal genetic variant, rs6712203, an intronic single-nucleotide polymorphism in the COBLL1 gene, which changes the conserved transcription factor motif of POU domain, class 2, transcription factor 2, and leads to differential COBLL1 gene expression by altering the enhancer activity at the locus in subcutaneous adipocytes. We then establish the cellular programme under the genetic control of the 2q24.3 MONW risk locus and the effector gene COBLL1, which is characterized by impaired actin cytoskeleton remodelling in differentiating subcutaneous adipocytes and subsequent failure of these cells to accumulate lipids and develop into metabolically active and insulin-sensitive adipocytes. Finally, we show that perturbations of the effector gene Cobll1 in a mouse model result in organismal phenotypes matching the MONW association signature, including decreased subcutaneous body fat mass and body weight along with impaired glucose tolerance. Taken together, our results provide a mechanistic link between the genetic risk for insulin resistance and low adiposity, providing a potential therapeutic hypothesis and a framework for future identification of causal relationships between genome associations and cellular programmes in other disorders.

Validation of genetic variants from NGS data using deep convolutional neural networks

Accurate somatic variant calling from next-generation sequencing data is one most important tasks in personalised cancer therapy. The sophistication of the available technologies is ever-increasing, yet, manual candidate refinement is still a necessary step in state-of-the-art processing pipelines. This limits reproducibility and introduces a bottleneck with respect to scalability. We demonstrate that the validation of genetic variants can be improved using a machine learning approach resting on a Convolutional Neural Network, trained using existing human annotation. In contrast to existing approaches, we introduce a way in which contextual data from sequencing tracks can be included into the automated assessment. A rigorous evaluation shows that the resulting model is robust and performs on par with trained researchers following published standard operating procedure.

Improving deep models of protein-coding potential with a Fourier-transform architecture and machine translation task

Ribosomes are information-processing macromolecular machines that integrate complex sequence patterns in messenger RNA (mRNA) transcripts to synthesize proteins. Studies of the sequence features that distinguish mRNAs from long noncoding RNAs (lncRNAs) may yield insight into the information that directs and regulates translation. Computational methods for calculating protein-coding potential are important for distinguishing mRNAs from lncRNAs during genome annotation, but most machine learning methods for this task rely on previously known rules to define features. Sequence-to-sequence (seq2seq) models, particularly ones using transformer networks, have proven capable of learning complex grammatical relationships between words to perform natural language translation. Seeking to leverage these advancements in the biological domain, we present a seq2seq formulation for predicting protein-coding potential with deep neural networks and demonstrate that simultaneously learning translation from RNA to protein improves classification performance relative to a classification-only training objective. Inspired by classical signal processing methods for gene discovery and Fourier-based image-processing neural networks, we introduce LocalFilterNet (LFNet). LFNet is a network architecture with an inductive bias for modeling the three-nucleotide periodicity apparent in coding sequences. We incorporate LFNet within an encoder-decoder framework to test whether the translation task improves the classification of transcripts and the interpretation of their sequence features. We use the resulting model to compute nucleotide-resolution importance scores, revealing sequence patterns that could assist the cellular machinery in distinguishing mRNAs and lncRNAs. Finally, we develop a novel approach for estimating mutation effects from Integrated Gradients, a backpropagation-based feature attribution, and characterize the difficulty of efficient approximations in this setting.

AdaLiftOver: high-resolution identification of orthologous regulatory elements with Adaptive liftOver

MOTIVATION

Elucidating functionally similar orthologous regulatory regions for human and model organism genomes is critical for exploiting model organism research and advancing our understanding of results from genome-wide association studies (GWAS). Sequence conservation is the de facto approach for finding orthologous non-coding regions between human and model organism genomes. However, existing methods for mapping non-coding genomic regions across species are challenged by the multi-mapping, low precision, and low mapping rate issues.

RESULTS

We develop Adaptive liftOver (AdaLiftOver), a large-scale computational tool for identifying functionally similar orthologous non-coding regions across species. AdaLiftOver builds on the UCSC liftOver framework to extend the query regions and prioritizes the resulting candidate target regions based on the conservation of the epigenomic and the sequence grammar features. Evaluations of AdaLiftOver with multiple case studies, spanning both genomic intervals from epigenome datasets across a wide range of model organisms and GWAS SNPs, yield AdaLiftOver as a versatile method for deriving hard-to-obtain human epigenome datasets as well as reliably identifying orthologous loci for GWAS SNPs.

AVAILABILITY AND IMPLEMENTATION

The R package and the data for AdaLiftOver is available from https://github.com/keleslab/AdaLiftOver.

An intrinsically interpretable neural network architecture for sequence to function learning

MOTIVATION

Sequence-based deep learning approaches have been shown to predict a multitude of functional genomic readouts, including regions of open chromatin and RNA expression of genes. However, a major limitation of current methods is that model interpretation relies on computationally demanding post hoc analyses, and even then, one can often not explain the internal mechanics of highly parameterized models. Here, we introduce a deep learning architecture called tiSFM (totally interpretable sequence to function model). tiSFM improves upon the performance of standard multi-layer convolutional models while using fewer parameters. Additionally, while tiSFM is itself technically a multi-layer neural network, internal model parameters are intrinsically interpretable in terms of relevant sequence motifs.

RESULTS

We analyze published open chromatin measurements across hematopoietic lineage cell-types and demonstrate that tiSFM outperforms a state-of-the-art convolutional neural network model custom-tailored to this dataset. We also show that it correctly identifies context specific activities of transcription factors with known roles in hematopoietic differentiation, including Pax5 and Ebf1 for B-cells, and Rorc for innate lymphoid cells. tiSFM's model parameters have biologically meaningful interpretations, and we show the utility of our approach on a complex task of predicting the change in epigenetic state as a function of developmental transition.

AVAILABILITY AND IMPLEMENTATION

The source code, including scripts for the analysis of key findings, can be found at https://github.com/boooooogey/ATAConv, implemented in Python.

Evidence for the role of transcription factors in the co-transcriptional regulation of intron retention

BACKGROUND

Alternative splicing is a widespread regulatory phenomenon that enables a single gene to produce multiple transcripts. Among the different types of alternative splicing, intron retention is one of the least explored despite its high prevalence in both plants and animals. The recent discovery that the majority of splicing is co-transcriptional has led to the finding that chromatin state affects alternative splicing. Therefore, it is plausible that transcription factors can regulate splicing outcomes.

RESULTS

We provide evidence for the hypothesis that transcription factors are involved in the regulation of intron retention by studying regions of open chromatin in retained and excised introns. Using deep learning models designed to distinguish between regions of open chromatin in retained introns and non-retained introns, we identified motifs enriched in IR events with significant hits to known human transcription factors. Our model predicts that the majority of transcription factors that affect intron retention come from the zinc finger family. We demonstrate the validity of these predictions using ChIP-seq data for multiple zinc finger transcription factors and find strong over-representation for their peaks in intron retention events.

CONCLUSIONS

This work opens up opportunities for further studies that elucidate the mechanisms by which transcription factors affect intron retention and other forms of splicing.

AVAILABILITY

Source code available at https://github.com/fahadahaf/chromir.

A Unified Deep Learning Framework for Single-Cell ATAC-Seq Analysis Based on ProdDep Transformer Encoder

Recent advances in single-cell sequencing assays for the transposase-accessibility chromatin (scATAC-seq) technique have provided cell-specific chromatin accessibility landscapes of cis-regulatory elements, providing deeper insights into cellular states and dynamics. However, few research efforts have been dedicated to modeling the relationship between regulatory grammars and single-cell chromatin accessibility and incorporating different analysis scenarios of scATAC-seq data into the general framework. To this end, we propose a unified deep learning framework based on the ProdDep Transformer Encoder, dubbed PROTRAIT, for scATAC-seq data analysis. Specifically motivated by the deep language model, PROTRAIT leverages the ProdDep Transformer Encoder to capture the syntax of transcription factor (TF)-DNA binding motifs from scATAC-seq peaks for predicting single-cell chromatin accessibility and learning single-cell embedding. Based on cell embedding, PROTRAIT annotates cell types using the Louvain algorithm. Furthermore, according to the identified likely noises of raw scATAC-seq data, PROTRAIT denoises these values based on predated chromatin accessibility. In addition, PROTRAIT employs differential accessibility analysis to infer TF activity at single-cell and single-nucleotide resolution. Extensive experiments based on the Buenrostro2018 dataset validate the effeteness of PROTRAIT for chromatin accessibility prediction, cell type annotation, and scATAC-seq data denoising, therein outperforming current approaches in terms of different evaluation metrics. Besides, we confirm the consistency between the inferred TF activity and the literature review. We also demonstrate the scalability of PROTRAIT to analyze datasets containing over one million cells.

Using Attribution Sequence Alignment to Interpret Deep Learning Models for miRNA Binding Site Prediction

MicroRNAs (miRNAs) are small non-coding RNAs that play a central role in the post-transcriptional regulation of biological processes. miRNAs regulate transcripts through direct binding involving the Argonaute protein family. The exact rules of binding are not known, and several in silico miRNA target prediction methods have been developed to date. Deep learning has recently revolutionized miRNA target prediction. However, the higher predictive power comes with a decreased ability to interpret increasingly complex models. Here, we present a novel interpretation technique, called attribution sequence alignment, for miRNA target site prediction models that can interpret such deep learning models on a two-dimensional representation of miRNA and putative target sequence. Our method produces a human readable visual representation of miRNA:target interactions and can be used as a proxy for the further interpretation of biological concepts learned by the neural network. We demonstrate applications of this method in the clustering of experimental data into binding classes, as well as using the method to narrow down predicted miRNA binding sites on long transcript sequences. Importantly, the presented method works with any neural network model trained on a two-dimensional representation of interactions and can be easily extended to further domains such as protein–protein interactions.

Massively parallel characterization of psychiatric disorder-associated and cell-type-specific regulatory elements in the developing human cortex

Nucleotide changes in gene regulatory elements are important determinants of neuronal development and disease. Using massively parallel reporter assays in primary human cells from mid-gestation cortex and cerebral organoids, we interrogated the cis-regulatory activity of 102,767 sequences, including differentially accessible cell-type specific regions in the developing cortex and single-nucleotide variants associated with psychiatric disorders. In primary cells, we identified 46,802 active enhancer sequences and 164 disorder-associated variants that significantly alter enhancer activity. Activity was comparable in organoids and primary cells, suggesting that organoids provide an adequate model for the developing cortex. Using deep learning, we decoded the sequence basis and upstream regulators of enhancer activity. This work establishes a comprehensive catalog of functional gene regulatory elements and variants in human neuronal development.

Deep learning for detecting and elucidating human T-cell leukemia virus type 1 integration in the human genome

Human T-cell leukemia virus type 1 (HTLV-1), a retrovirus, is the causative agent for adult T cell leukemia/lymphoma and many other human diseases. Accurate and high throughput detection of HTLV-1 virus integration sites (VISs) across the host genomes plays a crucial role in the prevention and treatment of HTLV-1-associated diseases. Here, we developed DeepHTLV, the first deep learning framework for VIS prediction de novo from genome sequence, motif discovery, and cis-regulatory factor identification. We demonstrated the high accuracy of DeepHTLV with more efficient and interpretive feature representations. Decoding the informative features captured by DeepHTLV resulted in eight representative clusters with consensus motifs for potential HTLV-1 integration. Furthermore, DeepHTLV revealed interesting cis-regulatory elements in regulation of VISs that have significant association with the detected motifs. Literature evidence demonstrated nearly half (34) of the predicted transcription factors enriched with VISs were involved in HTLV-1-associated diseases. DeepHTLV is freely available at https://github.com/bsml320/DeepHTLV.

Obtaining genetics insights from deep learning via explainable artificial intelligence

Artificial intelligence (AI) models based on deep learning now represent the state of the art for making functional predictions in genomics research. However, the underlying basis on which predictive models make such predictions is often unknown. For genomics researchers, this missing explanatory information would frequently be of greater value than the predictions themselves, as it can enable new insights into genetic processes. We review progress in the emerging area of explainable AI (xAI), a field with the potential to empower life science researchers to gain mechanistic insights into complex deep learning models. We discuss and categorize approaches for model interpretation, including an intuitive understanding of how each approach works and their underlying assumptions and limitations in the context of typical high-throughput biological datasets.

Deep learning in regulatory genomics: from identification to design

Genomics and deep learning are a natural match since both are data-driven fields. Regulatory genomics refers to functional noncoding DNA regulating gene expression. In recent years, deep learning applications on regulatory genomics have achieved remarkable advances so-much-so that it has revolutionized the rules of the game of the computational methods in this field. Here, we review two emerging trends: (i) the modeling of very long input sequence (up to 200 kb), which requires self-matched modularization of model architecture; (ii) on the balance of model predictability and model interpretability because the latter is more able to meet biological demands. Finally, we discuss how to employ these two routes to design synthetic regulatory DNA, as a promising strategy for optimizing crop agronomic properties.

ResidualBind: Uncovering Sequence-Structure Preferences of RNA-Binding Proteins with Deep Neural Networks

Deep neural networks have demonstrated improved performance at predicting sequence specificities of DNA- and RNA-binding proteins. However, it remains unclear why they perform better than previous methods that rely on k-mers and position weight matrices. Here, we highlight a recent deep learning-based software package, called ResidualBind, that analyzes RNA-protein interactions using only RNA sequence as an input feature and performs global importance analysis for model interpretability. We discuss practical considerations for model interpretability to uncover learned sequence motifs and their secondary structure preferences.

Applications of deep learning in understanding gene regulation

Gene regulation is a central topic in cell biology. Advances in omics technologies and the accumulation of omics data have provided better opportunities for gene regulation studies than ever before. For this reason deep learning, as a data-driven predictive modeling approach, has been successfully applied to this field during the past decade. In this article, we aim to give a brief yet comprehensive overview of representative deep-learning methods for gene regulation. Specifically, we discuss and compare the design principles and datasets used by each method, creating a reference for researchers who wish to replicate or improve existing methods. We also discuss the common problems of existing approaches and prospectively introduce the emerging deep-learning paradigms that will potentially alleviate them. We hope that this article will provide a rich and up-to-date resource and shed light on future research directions in this area.

Learning single-cell chromatin accessibility profiles using meta-analytic marker genes

MOTIVATION

Single-cell assay for transposase accessible chromatin using sequencing (scATAC-seq) is a valuable resource to learn cis-regulatory elements such as cell-type specific enhancers and transcription factor binding sites. However, cell-type identification of scATAC-seq data is known to be challenging due to the heterogeneity derived from different protocols and the high dropout rate.

RESULTS

In this study, we perform a systematic comparison of seven scATAC-seq datasets of mouse brain to benchmark the efficacy of neuronal cell-type annotation from gene sets. We find that redundant marker genes give a dramatic improvement for a sparse scATAC-seq annotation across the data collected from different studies. Interestingly, simple aggregation of such marker genes achieves performance comparable or higher than that of machine-learning classifiers, suggesting its potential for downstream applications. Based on our results, we reannotated all scATAC-seq data for detailed cell types using robust marker genes. Their meta scATAC-seq profiles are publicly available at https://gillisweb.cshl.edu/Meta_scATAC. Furthermore, we trained a deep neural network to predict chromatin accessibility from only DNA sequence and identified key motifs enriched for each neuronal subtype. Those predicted profiles are visualized together in our database as a valuable resource to explore cell-type specific epigenetic regulation in a sequence-dependent and -independent manner.

Deep learning predicts the impact of regulatory variants on cell-type-specific enhancers in the brain

Motivation

Previous studies have shown that the heritability of multiple brain-related traits and disorders is highly enriched in transcriptional enhancer regions. However, these regions often contain many individual variants, while only a subset of them are likely to causally contribute to a trait. Statistical fine-mapping techniques can identify putative causal variants, but their resolution is often limited, especially in regions with multiple variants in high linkage disequilibrium. In these cases, alternative computational methods to estimate the impact of individual variants can aid in variant prioritization.

Results

Here, we develop a deep learning pipeline to predict cell-type-specific enhancer activity directly from genomic sequences and quantify the impact of individual genetic variants in these regions. We show that the variants highlighted by our deep learning models are targeted by purifying selection in the human population, likely indicating a functional role. We integrate our deep learning predictions with statistical fine-mapping results for 8 brain-related traits, identifying 63 distinct candidate causal variants predicted to contribute to these traits by modulating enhancer activity, representing 6% of all genome-wide association study signals analyzed. Overall, our study provides a valuable computational method that can prioritize individual variants based on their estimated regulatory impact, but also highlights the limitations of existing methods for variant prioritization and fine-mapping.

Availability and implementation

The data underlying this article, nucleotide-level importance scores, and code for running the deep learning pipeline are available at https://github.com/Pandaman-Ryan/AgentBind-brain.

Contact

mgymrek@ucsd.edu.

Supplementary information

Supplementary data are available at Bioinformatics Advances online.

Predicting Chromatin Interactions from DNA Sequence Using DeepC

The genome 3D structure is central to understanding how disease-associated genetic variants in the noncoding genome regulate their target genes. Genome architecture spans large-scale structures determined by fine-grained regulatory elements, making it challenging to predict the effects of sequence and structural variants. Experimental approaches for chromatin interaction mapping remain costly and time-consuming, limiting their use for interrogating changes of chromatin architecture associated with genomic variation at scale. Computational models to predict chromatin interactions have either interpreted chromatin at coarse resolution or failed to capture the long-range dependencies of larger sequence contexts. To bridge this gap, we previously developed deepC, a deep neural network approach to predict chromatin interactions from DNA sequence at megabase scale. deepC employs dilated convolutional layers to achieve simultaneously a large sequence context while interpreting the DNA sequence at single base pair resolution. Using transfer learning of convolutional weights trained to predict a compendium of chromatin features across cell types allows deepC to predict cell type-specific chromatin interactions from DNA sequence alone. Here, we present a detailed workflow to predict chromatin interactions with deepC. We detail the necessary data pre-processing steps, guide through deepC model training, and demonstrate how to employ trained models to predict chromatin interactions and the effect of sequence variations on genome architecture.

Predicting mutational function using machine learning

Genetic variations are one of the major causes of phenotypic variations between human individuals. Although beneficial as being the substrate of evolution, germline mutations may cause diseases, including Mendelian diseases and complex diseases such as diabetes and heart diseases. Mutations occurring in somatic cells are a main cause of cancer and likely cause age-related phenotypes and other age-related diseases. Because of the high abundance of genetic variations in the human genome, i.e., millions of germline variations per human subject and thousands of additional somatic mutations per cell, it is technically challenging to experimentally verify the function of every possible mutation and their interactions. Significant progress has been made to solve this problem using computational approaches, especially machine learning (ML). Here, we review the progress and achievements made in recent years in this field of research. We classify the computational models in two ways: one according to their prediction goals including protein structural alterations, gene expression changes, and disease risks, and the other according to their methodologies, including non-machine learning methods, classical machine learning methods, and deep neural network methods. For models in each category, we discuss their architecture, prediction accuracy, and potential limitations. This review provides new insights into the applications and future directions of computational approaches in understanding the role of mutations in aging and disease.

SENIES: DNA Shape Enhanced Two-Layer Deep Learning Predictor for the Identification of Enhancers and Their Strength

Identifying enhancers is a critical task in bioinformatics due to their primary role in regulating gene expression. For this reason, various computational algorithms devoted to enhancer identification have been put forward over the years. More features are extracted from the single DNA sequences to boost the performance. Nevertheless, DNA structural information is neglected, which is an essential factor affecting the binding preferences of transcription factors to regulatory elements like enhancers. Here, we propose SENIES, a DNA shape enhanced deep learning predictor, to identify enhancers and their strength. The predictor consists of two layers where the first layer is for enhancer and non-enhancer identification, and the second layer is for predicting the strength of enhancers. Apart from two common sequence-derived features (i.e., one-hot and k-mer), DNA shape is introduced to describe the 3D structures of DNA sequences. Performance comparison with state-of-the-art methods conducted on public datasets demonstrates the effectiveness and robustness of our predictor. The code implementation of SENIES is publicly available at https://github.com/hlju-liye/SENIES.

DeepPHiC: predicting promoter-centered chromatin interactions using a novel deep learning approach

MOTIVATION

Promoter-centered chromatin interactions, which include promoter-enhancer (PE) and promoter-promoter (PP) interactions, are important to decipher gene regulation and disease mechanisms. The development of next-generation sequencing technologies such as promoter capture Hi-C (pcHi-C) leads to the discovery of promoter-centered chromatin interactions. However, pcHi-C experiments are expensive and thus may be unavailable for tissues/cell types of interest. In addition, these experiments may be underpowered due to insufficient sequencing depth or various artifacts, which results in a limited finding of interactions. Most existing computational methods for predicting chromatin interactions are based on in situ Hi-C and can detect chromatin interactions across the entire genome. However, they may not be optimal for predicting promoter-centered chromatin interactions.

RESULTS

We develop a supervised multi-modal deep learning model, which utilizes a comprehensive set of features such as genomic sequence, epigenetic signal, anchor distance, evolutionary features and DNA structural features to predict tissue/cell type-specific PE and PP interactions. We further extend the deep learning model in a multi-task learning and a transfer learning framework and demonstrate that the proposed approach outperforms state-of-the-art deep learning methods. Moreover, the proposed approach can achieve comparable prediction performance using predefined biologically relevant tissues/cell types compared to using all tissues/cell types in the pretraining especially for predicting PE interactions. The prediction performance can be further improved by using computationally inferred biologically relevant tissues/cell types in the pretraining, which are defined based on the common genes in the proximity of two anchors in the chromatin interactions.

AVAILABILITY AND IMPLEMENTATION

https://github.com/lichen-lab/DeepPHiC.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

maxATAC: Genome-scale transcription-factor binding prediction from ATAC-seq with deep neural networks

Transcription factors read the genome, fundamentally connecting DNA sequence to gene expression across diverse cell types. Determining how, where, and when TFs bind chromatin will advance our understanding of gene regulatory networks and cellular behavior. The 2017 ENCODE-DREAM in vivo Transcription-Factor Binding Site (TFBS) Prediction Challenge highlighted the value of chromatin accessibility data to TFBS prediction, establishing state-of-the-art methods for TFBS prediction from DNase-seq. However, the more recent Assay-for-Transposase-Accessible-Chromatin (ATAC)-seq has surpassed DNase-seq as the most widely-used chromatin accessibility profiling method. Furthermore, ATAC-seq is the only such technique available at single-cell resolution from standard commercial platforms. While ATAC-seq datasets grow exponentially, suboptimal motif scanning is unfortunately the most common method for TFBS prediction from ATAC-seq. To enable community access to state-of-the-art TFBS prediction from ATAC-seq, we (1) curated an extensive benchmark dataset (127 TFs) for ATAC-seq model training and (2) built "maxATAC", a suite of user-friendly, deep neural network models for genome-wide TFBS prediction from ATAC-seq in any cell type. With models available for 127 human TFs, maxATAC is the largest collection of high-performance TFBS prediction models for ATAC-seq. maxATAC performance extends to primary cells and single-cell ATAC-seq, enabling improved TFBS prediction in vivo. We demonstrate maxATAC's capabilities by identifying TFBS associated with allele-dependent chromatin accessibility at atopic dermatitis genetic risk loci.

Accuracy and data efficiency in deep learning models of protein expression

Synthetic biology often involves engineering microbial strains to express high-value proteins. Thanks to progress in rapid DNA synthesis and sequencing, deep learning has emerged as a promising approach to build sequence-to-expression models for strain optimization. But such models need large and costly training data that create steep entry barriers for many laboratories. Here we study the relation between accuracy and data efficiency in an atlas of machine learning models trained on datasets of varied size and sequence diversity. We show that deep learning can achieve good prediction accuracy with much smaller datasets than previously thought. We demonstrate that controlled sequence diversity leads to substantial gains in data efficiency and employed Explainable AI to show that convolutional neural networks can finely discriminate between input DNA sequences. Our results provide guidelines for designing genotype-phenotype screens that balance cost and quality of training data, thus helping promote the wider adoption of deep learning in the biotechnology sector.

Predicting RNA secondary structure by a neural network: what features may be learned?

Deep learning is a class of machine learning techniques capable of creating internal representation of data without explicit preprogramming. Hence, in addition to practical applications, it is of interest to analyze what features of biological data may be learned by such models. Here, we describe PredPair, a deep learning neural network trained to predict base pairs in RNA structure from sequence alone, without any incorporated prior knowledge, such as the stacking energies or possible spatial structures. PredPair learned the Watson-Crick and wobble base-pairing rules and created an internal representation of the stacking energies and helices. Application to independent experimental (DMS-Seq) data on nucleotide accessibility in mRNA showed that the nucleotides predicted as paired indeed tend to be involved in the RNA structure. The performance of the constructed model was comparable with the state-of-the-art method based on the thermodynamic approach, but with a higher false positives rate. On the other hand, it successfully predicted pseudoknots. t-SNE clusters of embeddings of RNA sequences created by PredPair tend to contain embeddings from particular Rfam families, supporting the predictions of PredPair being in line with biological classification.

RNAlight: a machine learning model to identify nucleotide features determining RNA subcellular localization

Different RNAs have distinct subcellular localizations. However, nucleotide features that determine these distinct distributions of lncRNAs and mRNAs have yet to be fully addressed. Here, we develop RNAlight, a machine learning model based on LightGBM, to identify nucleotide k-mers contributing to the subcellular localizations of mRNAs and lncRNAs. With the Tree SHAP algorithm, RNAlight extracts nucleotide features for cytoplasmic or nuclear localization of RNAs, indicating the sequence basis for distinct RNA subcellular localizations. By assembling k-mers to sequence features and subsequently mapping to known RBP-associated motifs, different types of sequence features and their associated RBPs were additionally uncovered for lncRNAs and mRNAs with distinct subcellular localizations. Finally, we extended RNAlight to precisely predict the subcellular localizations of other types of RNAs, including snRNAs, snoRNAs and different circular RNA transcripts, suggesting the generality of using RNAlight for RNA subcellular localization prediction.

Scalable Functional Assays for the Interpretation of Human Genetic Variation

Scalable sequence-function studies have enabled the systematic analysis and cataloging of hundreds of thousands of coding and noncoding genetic variants in the human genome. This has improved clinical variant interpretation and provided insights into the molecular, biophysical, and cellular effects of genetic variants at an astonishing scale and resolution across the spectrum of allele frequencies. In this review, we explore current applications and prospects for the field and outline the principles underlying scalable functional assay design, with a focus on the study of single-nucleotide coding and noncoding variants.

The genetic and biochemical determinants of mRNA degradation rates in mammals

BACKGROUND

Degradation rate is a fundamental aspect of mRNA metabolism, and the factors governing it remain poorly characterized. Understanding the genetic and biochemical determinants of mRNA half-life would enable more precise identification of variants that perturb gene expression through post-transcriptional gene regulatory mechanisms.

RESULTS

We establish a compendium of 39 human and 27 mouse transcriptome-wide mRNA decay rate datasets. A meta-analysis of these data identified a prevalence of technical noise and measurement bias, induced partially by the underlying experimental strategy. Correcting for these biases allowed us to derive more precise, consensus measurements of half-life which exhibit enhanced consistency between species. We trained substantially improved statistical models based upon genetic and biochemical features to better predict half-life and characterize the factors molding it. Our state-of-the-art model, Saluki, is a hybrid convolutional and recurrent deep neural network which relies only upon an mRNA sequence annotated with coding frame and splice sites to predict half-life (r=0.77). The key novel principle learned by Saluki is that the spatial positioning of splice sites, codons, and RNA-binding motifs within an mRNA is strongly associated with mRNA half-life. Saluki predicts the impact of RNA sequences and genetic mutations therein on mRNA stability, in agreement with functional measurements derived from massively parallel reporter assays.

CONCLUSIONS

Our work produces a more robust ground truth for transcriptome-wide mRNA half-lives in mammalian cells. Using these revised measurements, we trained Saluki, a model that is over 50% more accurate in predicting half-life from sequence than existing models. Saluki succinctly captures many of the known determinants of mRNA half-life and can be rapidly deployed to predict the functional consequences of arbitrary mutations in the transcriptome.

Deep learning approaches for noncoding variant prioritization in neurodegenerative diseases

Determining how noncoding genetic variants contribute to neurodegenerative dementias is fundamental to understanding disease pathogenesis, improving patient prognostication, and developing new clinical treatments. Next generation sequencing technologies have produced vast amounts of genomic data on cell type-specific transcription factor binding, gene expression, and three-dimensional chromatin interactions, with the promise of providing key insights into the biological mechanisms underlying disease. However, this data is highly complex, making it challenging for researchers to interpret, assimilate, and dissect. To this end, deep learning has emerged as a powerful tool for genome analysis that can capture the intricate patterns and dependencies within these large datasets. In this review, we organize and discuss the many unique model architectures, development philosophies, and interpretation methods that have emerged in the last few years with a focus on using deep learning to predict the impact of genetic variants on disease pathogenesis. We highlight both broadly-applicable genomic deep learning methods that can be fine-tuned to disease-specific contexts as well as existing neurodegenerative disease research, with an emphasis on Alzheimer's-specific literature. We conclude with an overview of the future of the field at the intersection of neurodegeneration, genomics, and deep learning.

A cis-regulatory lexicon of DNA motif combinations mediating cell-type-specific gene regulation

Gene expression is controlled by transcription factors (TFs) that bind cognate DNA motif sequences in cis-regulatory elements (CREs). The combinations of DNA motifs acting within homeostasis and disease, however, are unclear. Gene expression, chromatin accessibility, TF footprinting, and H3K27ac-dependent DNA looping data were generated and a random-forest-based model was applied to identify 7,531 cell-type-specific cis-regulatory modules (CRMs) across 15 diploid human cell types. A co-enrichment framework within CRMs nominated 838 cell-type-specific, recurrent heterotypic DNA motif combinations (DMCs), which were functionally validated using massively parallel reporter assays. Cancer cells engaged DMCs linked to neoplasia-enabling processes operative in normal cells while also activating new DMCs only seen in the neoplastic state. This integrative approach identifies cell-type-specific cis-regulatory combinatorial DNA motifs in diverse normal and diseased human cells and represents a general framework for deciphering cis-regulatory sequence logic in gene regulation.

Deciphering the impact of genetic variation on human polyadenylation using APARENT2

BACKGROUND

3'-end processing by cleavage and polyadenylation is an important and finely tuned regulatory process during mRNA maturation. Numerous genetic variants are known to cause or contribute to human disorders by disrupting the cis-regulatory code of polyadenylation signals. Yet, due to the complexity of this code, variant interpretation remains challenging.

RESULTS

We introduce a residual neural network model, APARENT2, that can infer 3'-cleavage and polyadenylation from DNA sequence more accurately than any previous model. This model generalizes to the case of alternative polyadenylation (APA) for a variable number of polyadenylation signals. We demonstrate APARENT2's performance on several variant datasets, including functional reporter data and human 3' aQTLs from GTEx. We apply neural network interpretation methods to gain insights into disrupted or protective higher-order features of polyadenylation. We fine-tune APARENT2 on human tissue-resolved transcriptomic data to elucidate tissue-specific variant effects. By combining APARENT2 with models of mRNA stability, we extend aQTL effect size predictions to the entire 3' untranslated region. Finally, we perform in silico saturation mutagenesis of all human polyadenylation signals and compare the predicted effects of [Formula: see text] million variants against gnomAD. While loss-of-function variants were generally selected against, we also find specific clinical conditions linked to gain-of-function mutations. For example, we detect an association between gain-of-function mutations in the 3'-end and autism spectrum disorder. To experimentally validate APARENT2's predictions, we assayed clinically relevant variants in multiple cell lines, including microglia-derived cells.

CONCLUSIONS

A sequence-to-function model based on deep residual learning enables accurate functional interpretation of genetic variants in polyadenylation signals and, when coupled with large human variation databases, elucidates the link between functional 3'-end mutations and human health.

Deep learning of cross-species single-cell landscapes identifies conserved regulatory programs underlying cell types

Despite extensive efforts to generate and analyze reference genomes, genetic models to predict gene regulation and cell fate decisions are lacking for most species. Here, we generated whole-body single-cell transcriptomic landscapes of zebrafish, Drosophila and earthworm. We then integrated cell landscapes from eight representative metazoan species to study gene regulation across evolution. Using these uniformly constructed cross-species landscapes, we developed a deep-learning-based strategy, Nvwa, to predict gene expression and identify regulatory sequences at the single-cell level. We systematically compared cell-type-specific transcription factors to reveal conserved genetic regulation in vertebrates and invertebrates. Our work provides a valuable resource and offers a new strategy for studying regulatory grammar in diverse biological systems.

DeepNup: Prediction of Nucleosome Positioning from DNA Sequences Using Deep Neural Network

Nucleosome positioning is involved in diverse cellular biological processes by regulating the accessibility of DNA sequences to DNA-binding proteins and plays a vital role. Previous studies have manifested that the intrinsic preference of nucleosomes for DNA sequences may play a dominant role in nucleosome positioning. As a consequence, it is nontrivial to develop computational methods only based on DNA sequence information to accurately identify nucleosome positioning, and thus intend to verify the contribution of DNA sequences responsible for nucleosome positioning. In this work, we propose a new deep learning-based method, named DeepNup, which enables us to improve the prediction of nucleosome positioning only from DNA sequences. Specifically, we first use a hybrid feature encoding scheme that combines One-hot encoding and Trinucleotide composition encoding to encode raw DNA sequences; afterwards, we employ multiscale convolutional neural network modules that consist of two parallel convolution kernels with different sizes and gated recurrent units to effectively learn the local and global correlation feature representations; lastly, we use a fully connected layer and a sigmoid unit serving as a classifier to integrate these learned high-order feature representations and generate the final prediction outcomes. By comparing the experimental evaluation metrics on two benchmark nucleosome positioning datasets, DeepNup achieves a better performance for nucleosome positioning prediction than that of several state-of-the-art methods. These results demonstrate that DeepNup is a powerful deep learning-based tool that enables one to accurately identify potential nucleosome sequences.

Deep mendelian randomization: Investigating the causal knowledge of genomic deep learning models

Multi-task deep learning (DL) models can accurately predict diverse genomic marks from sequence, but whether these models learn the causal relationships between genomic marks is unknown. Here, we describe Deep Mendelian Randomization (DeepMR), a method for estimating causal relationships between genomic marks learned by genomic DL models. By combining Mendelian randomization with in silico mutagenesis, DeepMR obtains local (locus specific) and global estimates of (an assumed) linear causal relationship between marks. In a simulation designed to test recovery of pairwise causal relations between transcription factors (TFs), DeepMR gives accurate and unbiased estimates of the 'true' global causal effect, but its coverage decays in the presence of sequence-dependent confounding. We then apply DeepMR to examine the global relationships learned by a state-of-the-art DL model, BPNet, between TFs involved in reprogramming. DeepMR's causal effect estimates validate previously hypothesized relationships between TFs and suggest new relationships for future investigation.

Assessment and Optimization of Explainable Machine Learning Models Applied to Transcriptomic Data

Explainable artificial intelligence aims to interpret how machine learning models make decisions, and many model explainers have been developed in the computer vision field. However, understanding of the applicability of these model explainers to biological data is still lacking. In this study, we comprehensively evaluated multiple explainers by interpreting pre-trained models for predicting tissue types from transcriptomic data and by identifying the top contributing genes from each sample with the greatest impacts on model prediction. To improve the reproducibility and interpretability of results generated by model explainers, we proposed a series of optimization strategies for each explainer on two different model architectures of multilayer perceptron (MLP) and convolutional neural network (CNN). We observed three groups of explainer and model architecture combinations with high reproducibility. Group II, which contains three model explainers on aggregated MLP models, identified top contributing genes in different tissues that exhibited tissue-specific manifestation and were potential cancer biomarkers. In summary, our work provides novel insights and guidance for exploring biological mechanisms using explainable machine learning models.

A review of brain imaging biomarker genomics in Alzheimer’s disease: implementation and perspectives

Alzheimer’s disease (AD) is a progressive neurodegenerative disease with phenotypic changes closely associated with both genetic variants and imaging pathology. Brain imaging biomarker genomics has been developed in recent years to reveal potential AD pathological mechanisms and provide early diagnoses. This technique integrates multimodal imaging phenotypes with genetic data in a noninvasive and high-throughput manner. In this review, we summarize the basic analytical framework of brain imaging biomarker genomics and elucidate two main implementation scenarios of this technique in AD studies: (1) exploring novel biomarkers and seeking mutual interpretability and (2) providing a diagnosis and prognosis for AD with combined use of machine learning methods and brain imaging biomarker genomics. Importantly, we highlight the necessity of brain imaging biomarker genomics, discuss the strengths and limitations of current methods, and propose directions for development of this research field.

LanceOtron: a deep learning peak caller for genome sequencing experiments

MOTIVATION

Genome sequencing experiments have revolutionized molecular biology by allowing researchers to identify important DNA-encoded elements genome wide. Regions where these elements are found appear as peaks in the analog signal of an assay's coverage track, and despite the ease with which humans can visually categorize these patterns, the size of many genomes necessitates algorithmic implementations. Commonly used methods focus on statistical tests to classify peaks, discounting that the background signal does not completely follow any known probability distribution and reducing the information-dense peak shapes to simply maximum height. Deep learning has been shown to be highly accurate for many pattern recognition tasks, on par or even exceeding human capabilities, providing an opportunity to reimagine and improve peak calling.

RESULTS

We present the peak calling framework LanceOtron, which combines deep learning for recognizing peak shape with multifaceted enrichment calculations for assessing significance. In benchmarking ATAC-seq, ChIP-seq and DNase-seq, LanceOtron outperforms long-standing, gold-standard peak callers through its improved selectivity and near-perfect sensitivity.

AVAILABILITY AND IMPLEMENTATION

A fully featured web application is freely available from LanceOtron.molbiol.ox.ac.uk, command line interface via python is pip installable from PyPI at https://pypi.org/project/lanceotron/, and source code and benchmarking tests are available at https://github.com/LHentges/LanceOtron.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.