Rachel Dania Melamed

Machine learning models for causal inference aim to adjust for confounding factors that are associated with both an exposure and an outcome, creating a spurious biased association. But, these methods are rarely empirically evaluated to assess their success in mitigating such bias. Recent advances in knowledge representation, including both foundation models and knowledge graphs, could enrich these models, but rigorous evaluations are needed in order to assess their potential. Here, we ask whether enriching existing causal inference models with knowledge representations from foundation models can improve confounding control. Rather than using semi-simulated data to address this question, we focus on examples of real confounding: we emulate target randomized active comparator trials that are subject to confounding by indication. Our results can guide researchers aiming to develop or apply methods for discovering causal effects from observational data.

Understanding which disease genes are altered by a drug can provide insight into the biology of effect, help us understand adverse drug effects, and suggest new drug uses. Here, we build on our model Draphnet in a new formulation with a similar goal. Draphnet was designed to explain drug therapeutic and side effects by learning a network connecting drugs to the disease genes they alter. Our new model, DraPhormer, has a similar goal but instead of relying on a linear model, learning of drug to gene connections uses a transformer model. DraPhormer integrates drug molecular data, disease genetics, and known drug effects on diseases, along with language models representing all of these entities. We show in simulations that DraPhormer can explain the genetic mechanisms of drug effects. Then, we present our design for incorporating drug and disease biology into the model. Finally, we benchmark the model’s ability to learn drug indications and side effects in real data.

Most clinical trials fail due to either lack of efficacy or safety concerns. Human genetics can address both failure reasons: disease-associated genes are not only promising therapeutic targets but also predict drug side effects. However, because the same genetic signal underlies both outcomes, we need methods that disentangle which disease genes mediate therapeutic benefit versus adverse side effects. We use DraphNet, our previously developed model that maps drug molecular effects onto disease genes to generate two gene sets per drug: one linked to its therapeutic effects (IND genes) and one linked to its side effects (SE genes). We show that IND and SE genes overlap for 76% of the tested drugs (compared to a null model). We also show that drugs sharing greater IND similarity also have greater SE similarity (ρ=0.57, p<1e-300). To show how our approach enables insights into drug biology, we construct groupings of drugs based on their IND and SE genes. We find that drugs in the same IND grouping are enriched for co-occurrence in the same SE grouping (OR=212.37). We present two examples to illustrate the kind of insights this network enables: identification of drugs with shared IND but distinct SE genes as repurposing candidates, and identification of drugs with shared SE but distinct IND genes to assist treatment selection in patients with comorbidities. Finally, we develop a neural network that directly links drug molecular effects onto disease genes and learns a gene-level score that quantifies each gene’s relative contribution to drug therapeutic versus side effects on disease.

In order to discover drugs that could be repurposed for a public health emergency like the COVID-19 pandemic, the National COVID Cohort Collaborative (N3C) database compiles health records including 22 millions individuals and 8.9 million cases of COVID-19. Here, we sought to use this data to systematically investigate whether antidepressants could impact COVID-19 outcomes, adjusting for known risk factors for severe outcome. We conducted large scale target trial emulation, comparing all pairs of 18 antidepressants to each other. Because the best approach for discovering such drug effects from observational data is not known, we applied a series of methods for identifying drug effects by estimating the counterfactual outcome that would be observed in a randomized trial. We found that all methods for counterfactual outcome estimation were prone to bias due to poorly controlled unmeasured confounding. We describe this bias, which appears to be induced partly by conditioning on treatment exposure, opening a back door path, via collider effects, between treatment and exposure via unmeasured confounders. Via empirical simulations, we show that our approach is able to detect this bias. In result, we state that we are not able to confidently identify any antidepressant impacting COVID-19 outcomes.

Using large data to propose new uses of drugs has potential to rapidly prioritize new treatments for major diseases of public health importance. One comprehensive data set, LINCS L1000 Connectivity Map, profiles gene expression associated with thousands of compounds, including many with known clinical uses. But, some recent studies have questioned the reliability of this data, and the best approach to use this resource for drug repositioning is not well established. Here, we develop a novel generalizable approach by hypothesizing that new treatments for a disease should induce similar gene expression to existing treatments for a disease. Using the Drug Repurposing Hub compendium of known treatments, we formulate a combined logistic regression model to predict new drug indications, and we assess generalizability of our findings using independent clinical trials on experimental drug uses. We support the hypothesis that drugs sharing an indication induce more similar gene expression, additionally demonstrating that the simpler Spearman correlation (p = 7.71e-38), outperforms the popular Connectivity Score (p = 5.2e-6). Our final model, combining predicted drug indications across three diverse cell lines, generalizes to predict experimental clinical trials with AUC of 0.708. By developing a new approach to using LINCS L1000 data for drug repositioning, we both propose plausible new disease treatments and provide an interpretable rationale for predictions. Our findings not only put forward new drug repositioning candidates, browseable at https://bsultan.shinyapps.io/web-app , but they also provide guidelines for future researchers employing L1000 data for drug repurposing.

Background Combinations of common drugs may, when taken together, have unexpected effects on incidence of diseases like cancer. It is not feasible to test for all combination drug effects in clinical trials, but in the real world, drugs are frequently taken in combination. Then, undiscovered effects may protect users of drug combinations from cancer—or increase their risk. By analyzing massive health data containing numerous people exposed to drug combinations, we have an opportunity to discover these associations. Method We describe, apply, and evaluate an approach for discovering drug combination associations with cancer using health data. Our approach builds on marginal structural model methods to emulate a randomized trial where one arm is assigned to take a drug alone, while the other arm takes that drug in combination with a second drug. Results Here, we perform drug combination-wide analysis to estimate effects of over 9000 drug combinations on incidence of all common cancer types, using claims data covering more than 100 million people. But, because the discovery of associations from observational data is always prone to confounding, we develop a number of strategies to distinguish confounding from biomedically relevant findings. We describe a robustly supported beneficial drug combination that may synergistically impact lipid levels to reduce the risk of cancer. Conclusions These findings can suggest new clinical uses for drug combinations to prevent or treat cancer. Our approach can be adapted to mine electronic health records for interactive effects on other late-onset common diseases.

Uncovering genes dysregulated in neuropsychiatric diseases promises to put forward therapeutic targets. By linking disease risk variants to tissue-specific gene regulation, the Genotype-Tissue Expression (GTEx) project has enabled discovery of disease genes and the brain tissues where these pathogenic effects occur. But, GTEx requires huge investment in obtaining human brain tissues, limiting the sample size and tissue diversity available. Crucially, this resource does not contain all key tissues where brain diseases are known to originate. Here, we propose a novel graph convolutional model that integrates the GTEx brain expression data with the Allen Brain Atlas, a resource spanning over 100 fine grained brain tissues that are annotated with an ontology of tissue organization. Our model, Brain Ontology Expression (BRONTE), allows completion of the GTEx data for 221 individuals to impute expanded expression in 103 uncollected detailed brain tissues. We show that expression data imputed by BRONTE recovers measured variation in expression across subjects, genes, and brain tissues. Most importantly, our method both recapitulates known relationships between genetic variations and gene expression, and simultaneously discovers both new genes impacted by genetic variants, and the tissues impacted. This allows us to put forward novel mechanisms of effect for genetic variation associated with 30 brain-related traits, supplying therapeutic targets in overlooked tissues of the brain.

Combining two clinically approved drugs has potential to improve treatment for common disease. But, with many thousands of combinations possible, clinically testing all pairs of drugs, with all common diseases, is not feasible. Here, we propose DRACO, a new machine learning method for discovering therapeutic drug combinations by leveraging knowledge about which health conditions each drug has been tested for. Our graph neural network model is able to generalize from a curated set of drug combinations to discover new combination therapies. We showcase our approach’s power to answer the question: given a drug and a health condition, what second drug would create an effective combination? Of our top three predicted candidate combinations, 74% have been previously reported. In the harder task of distinguishing the small number of reported combinations from millions of possible combination-condition triplets, our model ranks 94% of unseen triplets at the highest 0.1%. We expect DRACO to be a useful tool for proposing new therapies across thousands of diseases.

Human genetics demonstrates great potential for drug discovery, but challenges in identifying causal genes limit its clinical translation. Pleiotropy, the phenomenon where genetic variants or genes influence multiple traits, has been previously used to identify drug targets shared between phenotypically similar diseases. Here, we expand the use of pleiotropy to develop and evaluate a gene-agnostic method that predicts novel drug therapeutic and side effects across the phenome. We hypothesize that diseases with high genetic similarity to a drug’s known indications can point to new drug uses. To test this, we develop five metrics to quantify the genetic similarity between pairs of 178 diseases integrating genome-wide genetic correlation, gene-level associations and tissue-specific gene regulation. Comparing these metrics with data on indications and side effects of 1,711 common drugs, we find that more genetically similar diseases tend to share more drugs. We then use genetic similarity to predict drug therapeutic effects: our predictions with probability >0.1 show a 2.03-fold increased likelihood of progressing from Phase I clinical trials to regulatory approval. As well, predictions with probability >0.2 are 1.42 times more likely to correspond to true adverse effects. Notably, the indications model predicts side effects better than expected by chance, and vice-versa, implying a shared genetic basis for therapeutic and adverse drug effects. Together, our results suggest that genetic similarity can reveal new drug-disease links, putting forward a new use of genetics that bypasses the need for disease and drug target identification.