Hurle, M. R. | bioRxiv Science

bioRxiv2018

Predicting clinically promising therapeutic hypotheses using tensor factorization

Determining which target to pursue is a challenging and error-prone first step in developing a therapeutic treatment for a disease, where missteps are potentially very costly given the long-time frames and high expenses of drug development. We identified examples of successes and failures of target-indication pairs in clinical trials across 875 targets and 574 disease indications to build a gold-standard data set of 6,140 known clinical outcomes. We used information from Open Targets and others databases that covered 17 different sources of evidence for target-indication association and represented the data as a matrix of 21,437x2,211x17 with over two million non-null values. We designed and executed three benchmarking strategies to examine the performance of multiple machine learning models: Logistic Regression, Elasticnet, Random Forest, Tensor Factorization and Gradient Boosting Machine. With ten-fold cross validation, tensor factorization achieved AUROC=0.82{+/-}0.02 and AUPRC=0.71{+/-}0.03. Across multiple validation schemes, this was comparable or better than other methods. Tensor factorization is a general form of matrix factorization that has been successfully exploited in recommendation systems that suggest items to users based on their existing preference on a small number of items. Our application, using Bayesian probabilistic modelling, extends the capacity of matrix factorization to model multiple relationships between and among targets and indications. We use the model to show that our predicted probabilities of success correlate with clinical phases, and within clinical phase we can predict which trials are most likely to succeed.

bioinformatics↗

bioRxiv2017

Systematic interrogation of diverse Omic data reveals interpretable, robust, and generalizable transcriptomic features of clinically successful therapeutic targets

Target selection is the first and pivotal step in drug discovery. An incorrect choice may not manifest itself for many years after hundreds of millions of research dollars have been spent. We collected a set of 332 targets that succeeded or failed in phase III clinical trials, and explored whether Omic features describing the target genes could predict clinical success. We obtained features from the recently published comprehensive resource: Harmonizome. Nineteen features appeared to be significantly correlated with phase III clinical trial outcomes, but only 4 passed validation schemes that used bootstrapping or modified permutation tests to assess feature robustness and generalizability while accounting for target class selection bias. We also used classifiers to perform multivariate feature selection and found that classifiers with a single feature performed as well in cross-validation as classifiers with more features (AUROC=0.57 and AUPR=0.81). The two predominantly selected features were mean mRNA expression across tissues and standard deviation of expression across tissues, where successful targets tended to have lower mean expression and higher expression variance than failed targets. This finding supports the conventional wisdom that it is favorable for a target to be present in the tissue(s) affected by a disease and absent from other tissues. Overall, our results suggest that it is feasible to construct a model integrating interpretable target features to inform target selection. We anticipate deeper insights and better models in the future, as researchers can reuse the data we have provided to improve methods for handling sample biases and learn more informative features. Code, documentation, and data for this study have been deposited on GitHub at https://github.com/arouillard/omic-features-successful-targets.\n\nAUTHOR SUMMARYDrug discovery often begins with a hypothesis that changing the abundance or activity of a target--a biological molecule, usually a protein--will cure a disease or ameliorate its symptoms. Whether a target hypothesis translates into a successful therapy depends in part on the characteristics of the target, but it is not completely understood which target characteristics are important for success. We sought to answer this question with a supervised machine learning approach. We obtained outcomes of target hypotheses tested in clinical trials, scoring targets as successful or failed, and then obtained thousands of features (i.e. properties or characteristics) of targets from dozens of biological datasets. We statistically tested which features differed between successful and failed targets, and built a computational model that used these features to predict success or failure of targets in clinical trials. We found that successful targets tended to have more variable mRNA abundance from tissue to tissue and lower average abundance across tissues than failed targets. Thus, it is probably favorable for a target to be present in the tissue(s) affected by a disease and absent from other tissues. Our work demonstrates the feasibility of predicting clinical trial outcomes from target features.

bioinformatics↗