Explainable AI for computational pathology identifies model limitations and tissue biomarkers.

Deep learning models have shown promise in histopathology image analysis, but their opaque decision-making process poses challenges in high-risk medical scenarios. Here we introduce HIPPO, an explainable AI method that interrogates attention-based multiple instance learning (ABMIL) models in computational pathology by generating counterfactual examples through tissue patch modifications in whole slide images. Applying HIPPO to ABMIL models trained to detect breast cancer metastasis reveals that they may overlook small tumors and can be misled by non-tumor tissue, while attention maps-widely used for interpretation-often highlight regions that do not directly influence predictions. By interpreting ABMIL models trained on a prognostic prediction task, HIPPO identified tissue areas with stronger prognostic effects than high-attention regions, which sometimes showed counterintuitive influences on risk scores. These findings demonstrate HIPPO's capacity for comprehensive model evaluation, bias detection, and quantitative hypothesis testing. HIPPO greatly expands the capabilities of explainable AI tools to assess the trustworthy and reliable development, deployment, and regulation of weakly-supervised models in computational pathology.

Interpreting cis-regulatory interactions from large-scale deep neural networks.

The rise of large-scale, sequence-based deep neural networks (DNNs) for predicting gene expression has introduced challenges in their evaluation and interpretation. Current evaluations align DNN predictions with orthogonal experimental data, providing insights into generalization but offering limited insights into their decision-making process. Existing model explainability tools focus mainly on motif analysis, which becomes complex when interpreting longer sequences. Here we present cis-regulatory element model explanations (CREME), an in silico perturbation toolkit that interprets the rules of gene regulation learned by a genomic DNN. Applying CREME to Enformer, a state-of-the-art DNN, we identify cis-regulatory elements that enhance or silence gene expression and characterize their complex interactions. CREME can provide interpretations across multiple scales of genomic organization, from cis-regulatory elements to fine-mapped functional sequence elements within them, offering high-resolution insights into the regulatory architecture of the genome. CREME provides a powerful toolkit for translating the predictions of genomic DNNs into mechanistic insights of gene regulation.

Massive experimental quantification of amyloid nucleation allows interpretable deep learning of protein aggregation.

Protein aggregation is a pathological hallmark of more than fifty human diseases and a major problem for biotechnology. Methods have been proposed to predict aggregation from sequence, but these have been trained and evaluated on small and biased experimental datasets. Here we directly address this data shortage by experimentally quantifying the amyloid nucleation of >100,000 protein sequences. This unprecedented dataset reveals the limited performance of existing computational methods and allows us to train CANYA, a convolution-attention hybrid neural network that accurately predicts amyloid nucleation from sequence. We adapt genomic neural network interpretability analyses to reveal CANYA's decision-making process and learned grammar. Our results illustrate the power of massive experimental analysis of random sequence-spaces and provide an interpretable and robust neural network model to predict amyloid nucleation.

Evaluating the representational power of pre-trained DNA language models for regulatory genomics.

The emergence of genomic language models (gLMs) offers an unsupervised approach to learning a wide diversity of cis-regulatory patterns in the non-coding genome without requiring labels of functional activity generated by wet-lab experiments. Previous evaluations have shown that pre-trained gLMs can be leveraged to improve predictive performance across a broad range of regulatory genomics tasks, albeit using relatively simple benchmark datasets and baseline models. Since the gLMs in these studies were tested upon fine-tuning their weights for each downstream task, determining whether gLM representations embody a foundational understanding of cis-regulatory biology remains an open question. Here we evaluate the representational power of pre-trained gLMs to predict and interpret cell-type-specific functional genomics data that span DNA and RNA regulation. Our findings suggest that probing the representations of pre-trained gLMs do not offer substantial advantages over conventional machine learning approaches that use one-hot encoded sequences. This work highlights a major gap with current gLMs, raising potential issues in conventional pre-training strategies for the non-coding genome.

EvoAug-TF: extending evolution-inspired data augmentations for genomic deep learning to TensorFlow.

SUMMARY: Deep neural networks (DNNs) have been widely applied to predict the molecular functions of the non-coding genome. DNNs are data hungry and thus require many training examples to fit data well. However, functional genomics experiments typically generate limited amounts of data, constrained by the activity levels of the molecular function under study inside the cell. Recently, EvoAug was introduced to train a genomic DNN with evolution-inspired augmentations. EvoAug-trained DNNs have demonstrated improved generalization and interpretability with attribution analysis. However, EvoAug only supports PyTorch-based models, which limits its applications to a broad class of genomic DNNs based in TensorFlow. Here, we extend EvoAug's functionality to TensorFlow in a new package, we call EvoAug-TF. Through a systematic benchmark, we find that EvoAug-TF yields comparable performance with the original EvoAug package. AVAILABILITY AND IMPLEMENTATION: EvoAug-TF is freely available for users and is distributed under an open-source MIT license. Researchers can access the open-source code on GitHub (https://github.com/p-koo/evoaug-tf). The pre-compiled package is provided via PyPI (https://pypi.org/project/evoaug-tf) with in-depth documentation on ReadTheDocs (https://evoaug-tf.readthedocs.io). The scripts for reproducing the results are available at (https://github.com/p-koo/evoaug-tf_analysis).

EvoAug-TF: Extending evolution-inspired data augmentations for genomic deep learning to TensorFlow.

Deep neural networks (DNNs) have been widely applied to predict the molecular functions of regulatory regions in the non-coding genome. DNNs are data hungry and thus require many training examples to fit data well. However, functional genomics experiments typically generate limited amounts of data, constrained by the activity levels of the molecular function under study inside the cell. Recently, EvoAug was introduced to train a genomic DNN with evolution-inspired augmentations. EvoAug-trained DNNs have demonstrated improved generalization and interpretability with attribution analysis. However, EvoAug only supports PyTorch-based models, which limits its applications to a broad class of genomic DNNs based in TensorFlow. Here, we extend EvoAug's functionality to TensorFlow in a new package we call EvoAug-TF. Through a systematic benchmark, we find that EvoAug-TF yields comparable performance with the original EvoAug package. Availability: EvoAug-TF is freely available for users and is distributed under an open-source MIT license. Researchers can access the open-source code on GitHub (https://github.com/p-koo/evoaug-tf). The pre-compiled package is provided via PyPI (https://pypi.org/project/evoaug-tf) with in-depth documentation on ReadTheDocs (https://evoaug-tf.readthedocs.io). The scripts for reproducing the results are available at (https://github.com/p-koo/evoaug-tf_analysis).

Interpreting Cis-Regulatory Interactions from Large-Scale Deep Neural Networks for Genomics.

The rise of large-scale, sequence-based deep neural networks (DNNs) for predicting gene expression has introduced challenges in their evaluation and interpretation. Current evaluations align DNN predictions with experimental perturbation assays, which provides insights into the generalization capabilities within the studied loci but offers a limited perspective of what drives their predictions. Moreover, existing model explainability tools focus mainly on motif analysis, which becomes complex when interpreting longer sequences. Here we introduce CREME, an in silico perturbation toolkit that interrogates large-scale DNNs to uncover rules of gene regulation that it learns. Using CREME, we investigate Enformer, a prominent DNN in gene expression prediction, revealing cis-regulatory elements (CREs) that directly enhance or silence target genes. We explore the intricate complexity of higher-order CRE interactions, the relationship between CRE distance from transcription start sites on gene expression, as well as the biochemical features of enhancers and silencers learned by Enformer. Moreover, we demonstrate the flexibility of CREME to efficiently uncover a higher-resolution view of functional sequence elements within CREs. This work demonstrates how CREME can be employed to translate the powerful predictions of large-scale DNNs to study open questions in gene regulation.

Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models.

Deep neural networks (DNNs) have greatly advanced the ability to predict genome function from sequence. Interpreting genomic DNNs in terms of biological mechanisms, however, remains difficult. Here we introduce SQUID, a genomic DNN interpretability framework based on surrogate modeling. SQUID approximates genomic DNNs in user-specified regions of sequence space using surrogate models, i.e., simpler models that are mechanistically interpretable. Importantly, SQUID removes the confounding effects that nonlinearities and heteroscedastic noise in functional genomics data can have on model interpretation. Benchmarking analysis on multiple genomic DNNs shows that SQUID, when compared to established interpretability methods, identifies motifs that are more consistent across genomic loci and yields improved single-nucleotide variant-effect predictions. SQUID also supports surrogate models that quantify epistatic interactions within and between cis-regulatory elements. SQUID thus advances the ability to mechanistically interpret genomic DNNs.

ChampKit: A framework for rapid evaluation of deep neural networks for patch-based histopathology classification.

BACKGROUND AND OBJECTIVE: Histopathology is the gold standard for diagnosis of many cancers. Recent advances in computer vision, specifically deep learning, have facilitated the analysis of histopathology images for many tasks, including the detection of immune cells and microsatellite instability. However, it remains difficult to identify optimal models and training configurations for different histopathology classification tasks due to the abundance of available architectures and the lack of systematic evaluations. Our objective in this work is to present a software tool that addresses this need and enables robust, systematic evaluation of neural network models for patch classification in histology in a light-weight, easy-to-use package for both algorithm developers and biomedical researchers. METHODS: Here we present ChampKit (Comprehensive Histopathology Assessment of Model Predictions toolKit): an extensible, fully reproducible evaluation toolkit that is a one-stop-shop to train and evaluate deep neural networks for patch classification. ChampKit curates a broad range of public datasets. It enables training and evaluation of models supported by timm directly from the command line, without the need for users to write any code. External models are enabled through a straightforward API and minimal coding. As a result, Champkit facilitates the evaluation of existing and new models and deep learning architectures on pathology datasets, making it more accessible to the broader scientific community. To demonstrate the utility of ChampKit, we establish baseline performance for a subset of possible models that could be employed with ChampKit, focusing on several popular deep learning models, namely ResNet18, ResNet50, and R26-ViT, a hybrid vision transformer. In addition, we compare each model trained either from random weight initialization or with transfer learning from ImageNet pretrained models. For ResNet18, we also consider transfer learning from a self-supervised pretrained model. RESULTS: The main result of this paper is the ChampKit software. Using ChampKit, we were able to systemically evaluate multiple neural networks across six datasets. We observed mixed results when evaluating the benefits of pretraining versus random intialization, with no clear benefit except in the low data regime, where transfer learning was found to be beneficial. Surprisingly, we found that transfer learning from self-supervised weights rarely improved performance, which is counter to other areas of computer vision. CONCLUSIONS: Choosing the right model for a given digital pathology dataset is nontrivial. ChampKit provides a valuable tool to fill this gap by enabling the evaluation of hundreds of existing (or user-defined) deep learning models across a variety of pathology tasks. Source code and data for the tool are freely accessible at https://github.com/SBU-BMI/champkit.

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.

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.

ETV6 dependency in Ewing sarcoma by antagonism of EWS-FLI1-mediated enhancer activation.

The EWS-FLI1 fusion oncoprotein deregulates transcription to initiate the paediatric cancer Ewing sarcoma. Here we used a domain-focused CRISPR screen to implicate the transcriptional repressor ETV6 as a unique dependency in this tumour. Using biochemical assays and epigenomics, we show that ETV6 competes with EWS-FLI1 for binding to select DNA elements enriched for short GGAA repeat sequences. Upon inactivating ETV6, EWS-FLI1 overtakes and hyper-activates these cis-elements to promote mesenchymal differentiation, with SOX11 being a key downstream target. We show that squelching of ETV6 with a dominant-interfering peptide phenocopies these effects and suppresses Ewing sarcoma growth in vivo. These findings reveal targeting of ETV6 as a strategy for neutralizing the EWS-FLI1 oncoprotein by reprogramming of genomic occupancy.

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.

Light and temperature regulate m6A-RNA modification to regulate growth in plants.

N6-methyladenosine is a highly dynamic, abundant mRNA modification which is an excellent potential mechanism for fine tuning gene expression. Plants adapt to their surrounding light and temperature environment using complex gene regulatory networks. The role of m6A in controlling gene expression in response to variable environmental conditions has so far been unexplored. Here, we map the transcriptome-wide m6A landscape under various light and temperature environments. Identified m6A-modifications show a highly specific spatial distribution along transcripts with enrichment occurring in 5'UTR regions and around transcriptional end sites. We show that the position of m6A modifications on transcripts might influence cellular transcript localization and the presence of m6A-modifications is associated with alternative polyadenylation, a process which results in multiple RNA isoforms with varying 3'UTR lengths. RNA with m6A-modifications exhibit a higher preference for shorter 3'UTRs. These shorter 3'UTR regions might directly influence transcript abundance and localization by including or excluding cis-regulatory elements. We propose that environmental stimuli might change the m6A landscape of plants as one possible way of fine tuning gene regulation through alternative polyadenylation and transcript localization.

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.

Interpreting Potts and Transformer Protein Models Through the Lens of Simplified Attention.

The established approach to unsupervised protein contact prediction estimates coevolving positions using undirected graphical models. This approach trains a Potts model on a Multiple Sequence Alignment. Increasingly large Transformers are being pretrained on unlabeled, unaligned protein sequence databases and showing competitive performance on protein contact prediction. We argue that attention is a principled model of protein interactions, grounded in real properties of protein family data. We introduce an energy-based attention layer, factored attention, which, in a certain limit, recovers a Potts model, and use it to contrast Potts and Transformers. We show that the Transformer leverages hierarchical signal in protein family databases not captured by single-layer models. This raises the exciting possibility for the development of powerful structured models of protein family databases.

Selecting deep neural networks that yield consistent attribution-based interpretations for genomics.

Deep neural networks (DNNs) have advanced our ability to take DNA primary sequence as input and predict a myriad of molecular activities measured via high-throughput functional genomic assays. Post hoc attribution analysis has been employed to provide insights into the importance of features learned by DNNs, often revealing patterns such as sequence motifs. However, attribution maps typically harbor spurious importance scores to an extent that varies from model to model, even for DNNs whose predictions generalize well. Thus, the standard approach for model selection, which relies on performance of a held-out validation set, does not guarantee that a high-performing DNN will provide reliable explanations. Here we introduce two approaches that quantify the consistency of important features across a population of attribution maps; consistency reflects a qualitative property of human interpretable attribution maps. We employ the consistency metrics as part of a multivariate model selection framework to identify models that yield high generalization performance and interpretable attribution analysis. We demonstrate the efficacy of this approach across various DNNs quantitatively with synthetic data and qualitatively with chromatin accessibility data.

Evaluating deep learning for predicting epigenomic profiles.

Deep learning has been successful at predicting epigenomic profiles from DNA sequences. Most approaches frame this task as a binary classification relying on peak callers to define functional activity. Recently, quantitative models have emerged to directly predict the experimental coverage values as a regression. As new models continue to emerge with different architectures and training configurations, a major bottleneck is forming due to the lack of ability to fairly assess the novelty of proposed models and their utility for downstream biological discovery. Here we introduce a unified evaluation framework and use it to compare various binary and quantitative models trained to predict chromatin accessibility data. We highlight various modeling choices that affect generalization performance, including a downstream application of predicting variant effects. In addition, we introduce a robustness metric that can be used to enhance model selection and improve variant effect predictions. Our empirical study largely supports that quantitative modeling of epigenomic profiles leads to better generalizability and interpretability.

Improving representations of genomic sequence motifs in convolutional networks with exponential activations.

Deep convolutional neural networks (CNNs) trained on regulatory genomic sequences tend to build representations in a distributed manner, making it a challenge to extract learned features that are biologically meaningful, such as sequence motifs. Here we perform a comprehensive analysis on synthetic sequences to investigate the role that CNN activations have on model interpretability. We show that employing an exponential activation to first layer filters consistently leads to interpretable and robust representations of motifs compared to other commonly used activations. Strikingly, we demonstrate that CNNs with better test performance do not necessarily imply more interpretable representations with attribution methods. We find that CNNs with exponential activations significantly improve the efficacy of recovering biologically meaningful representations with attribution methods. We demonstrate these results generalise to real DNA sequences across several in vivo datasets. Together, this work demonstrates how a small modification to existing CNNs, i.e. setting exponential activations in the first layer, can significantly improve the robustness and interpretabilty of learned representations directly in convolutional filters and indirectly with attribution methods.