Using Interpretable Deep Learning to Model Cancer Dependencies

Advancing Computational Toxicology by Interpretable Machine Learning

Chemical toxicity evaluations for drugs, consumer products, and environmental chemicals have a critical impact on human health. Traditional animal models to evaluate chemical toxicity are expensive, time-consuming, and often fail to detect toxicants in humans. Computational toxicology is a promising alternative approach that utilizes machine learning (ML) and deep learning (DL) techniques to predict the toxicity potentials of chemicals. Although the applications of ML- and DL-based computational models in chemical toxicity predictions are attractive, many toxicity models are "black boxes" in nature and difficult to interpret by toxicologists, which hampers the chemical risk assessments using these models. The recent progress of interpretable ML (IML) in the computer science field meets this urgent need to unveil the underlying toxicity mechanisms and elucidate the domain knowledge of toxicity models. In this review, we focused on the applications of IML in computational toxicology, including toxicity feature data, model interpretation methods, use of knowledge base frameworks in IML development, and recent applications. The challenges and future directions of IML modeling in toxicology are also discussed. We hope this review can encourage efforts in developing interpretable models with new IML algorithms that can assist new chemical assessments by illustrating toxicity mechanisms in humans.

SL-scan identifies synthetic lethal interactions in cancer using metabolic networks

Exploiting synthetic lethality is a promising strategy for developing targeted cancer therapies. However, identifying clinically significant synthetic lethal (SL) interactions among a large number of gene combinations is a challenging computational task. In this study, we developed the SL-scan pipeline based on metabolic network modeling to discover SL interaction. The SL-scan pipeline identifies the association between simulated Flux Balance Analysis knockout scores and mutation data across cancer cell lines and predicts putative SL interactions. We assessed the concordance of the SL pairs predicted by SL-scan with those of obtained from analysis of the CRISPR, shRNA, and PRISM datasets. Our results demonstrate that the SL-scan pipeline outperformed existing SL prediction approaches based on metabolic networks in identifying SL pairs in various cancers. This study emphasizes the importance of integrating multiple data sources, particularly mutation data, when identifying SL pairs for targeted cancer therapies. The findings of this study may lead to the development of novel targeted cancer therapies.

Knowledge graph aids comprehensive explanation of drug and chemical toxicity

In computational toxicology, prediction of complex endpoints has always been challenging, as they often involve multiple distinct mechanisms. State-of-the-art models are either limited by low accuracy, or lack of interpretability due to their black-box nature. Here, we introduce AIDTox, an interpretable deep learning model which incorporates curated knowledge of chemical-gene connections, gene-pathway annotations, and pathway hierarchy. AIDTox accurately predicts cytotoxicity outcomes in HepG2 and HEK293 cells. It also provides comprehensive explanations of cytotoxicity covering multiple aspects of drug activity, including target interaction, metabolism, and elimination. In summary, AIDTox provides a computational framework for unveiling cellular mechanisms for complex toxicity endpoints.

Network Biology-Inspired Machine Learning Features Predict Cancer Gene Targets and Reveal Target Coordinating Mechanisms

Anticipating and understanding cancers' need for specific gene activities is key for novel therapeutic development. Here we utilized DepMap, a cancer gene dependency screen, to demonstrate that machine learning combined with network biology can produce robust algorithms that both predict what genes a cancer is dependent on and what network features coordinate such gene dependencies. Using network topology and biological annotations, we constructed four groups of novel engineered machine learning features that produced high accuracies when predicting binary gene dependencies. We found that in all examined cancer types, F1 scores were greater than 0.90, and model accuracy remained robust under multiple hyperparameter tests. We then deconstructed these models to identify tumor type-specific coordinators of gene dependency and identified that in certain cancers, such as thyroid and kidney, tumors' dependencies are highly predicted by gene connectivity. In contrast, other histologies relied on pathway-based features such as lung, where gene dependencies were highly predictive by associations with cell death pathway genes. In sum, we show that biologically informed network features can be a valuable and robust addition to predictive pharmacology models while simultaneously providing mechanistic insights.

Predicting genetic interactions, cell line dependencies and drug sensitivities with variational graph auto-encoder

Large scale cancer genomics data provide crucial information about the disease and reveal points of intervention. However, systematic data have been collected in specific cell lines and their collection is laborious and costly. Hence, there is a need to develop computational models that can predict such data for any genomic context of interest. Here we develop novel models that build on variational graph auto-encoders and can integrate diverse types of data to provide high quality predictions of genetic interactions, cell line dependencies and drug sensitivities, outperforming previous methods. Our models, data and implementation are available at: https://github.com/aijag/drugGraphNet.

Opportunities and challenges in interpretable deep learning for drug sensitivity prediction of cancer cells

In precision oncology, therapy stratification is done based on the patients' tumor molecular profile. Modeling and prediction of the drug response for a given tumor molecular type will further improve therapeutic decision-making for cancer patients. Indeed, deep learning methods hold great potential for drug sensitivity prediction, but a major problem is that these models are black box algorithms and do not clarify the mechanisms of action. This puts a limitation on their clinical implementation. To address this concern, many recent studies attempt to overcome these issues by developing interpretable deep learning methods that facilitate the understanding of the logic behind the drug response prediction. In this review, we discuss strengths and limitations of recent approaches, and suggest future directions that could guide further improvement of interpretable deep learning in drug sensitivity prediction in cancer research.

Knowledge-guided deep learning models of drug toxicity improve interpretation

In drug development, a major reason for attrition is the lack of understanding of cellular mechanisms governing drug toxicity. The black-box nature of conventional classification models has limited their utility in identifying toxicity pathways. Here we developed DTox (deep learning for toxicology), an interpretation framework for knowledge-guided neural networks, which can predict compound response to toxicity assays and infer toxicity pathways of individual compounds. We demonstrate that DTox can achieve the same level of predictive performance as conventional models with a significant improvement in interpretability. Using DTox, we were able to rediscover mechanisms of transcription activation by three nuclear receptors, recapitulate cellular activities induced by aromatase inhibitors and pregnane X receptor (PXR) agonists, and differentiate distinctive mechanisms leading to HepG2 cytotoxicity. Virtual screening by DTox revealed that compounds with predicted cytotoxicity are at higher risk for clinical hepatic phenotypes. In summary, DTox provides a framework for deciphering cellular mechanisms of toxicity in silico.