In WP1, we published two open-source Python packages, stanscofi, and benchscofi, to enable easier access to drug-repurposing data sets and develop machine-learning approaches for drug repurposing. stanscofi automates data processing, visualization, training, and validation of methods, and it standardizes the implementation of drug repurposing algorithms (see Figure 1). benchscofi implements 21 drug repurposing algorithms from the state-of-the-art to enforce a quick and robust assessment of the performance of drug repurposing.
In WP2, we designed a drug repurposing approach called JELI (Joint Embedding-classifier Learning for improved Interpretability), which retrieves connections between diseases, drugs, and relevant biological features (e.g. gene expression) in a partially completed graph (e.g. with gene-to-gene interactions) to make recommendations on drug candidates with interpretable outputs. See Figure 2. We released the method in an open-source Python package. We demonstrated the prediction performance over several validation metrics and the potential explainability of JELI on synthetic and drug-repurposing data sets.
In WP3, we crafted an imputation method called F3I (Fast Iterative Improvement for Imputation) to deal with missing data in sparse drug repurposing data sets. This method combines some of the nearest data points to a sample with missing values via an automatically learned parameter that optimizes the preservation of the initial data distribution. This imputation method can be seamlessly chained with a drug repurposing algorithm to optimize for reasonable imputation and prediction. We have applied F3I to synthetic, drug repurposing, and well-known computer vision benchmarks and showed that F3I is robust to different mechanisms accounting for missing data and large sparsity (see Figure 3). We also proved theoretical guarantees on the quality of imputation by F3I.
In WP4, we applied JELI (from WP2) to predict new therapeutic candidates for melanoma and successfully retrieved a significantly perturbed biological pathway connected to melanin biosynthesis. This result allowed us to illustrate the explainability potential of JELI, which can be leveraged to prioritize and assess the relevance of predicted drug candidates.
In WP5, we considered designing a method to obtain theoretical guarantees on the rate of false positive candidates (predicted as promising therapies by a recommender system but unsuccessful in practice). So far, we have implemented a first approach based on bootstrapped Neyman-Pearson classification to extend the state-of-the-art to cases with unknown and adverse drug-disease annotations. This approach allows us to return a subset of candidates given a desired upper bound on the false positive rate.
In WP6, we considered the list of the Top 8 candidates predicted by JELI against melanoma and aimed to validate them independently by molecular protein docking. So far, we have identified protein targets of interest for melanoma by studying the pharmacophores (set of features describing the molecular interaction site) of known, successful drugs against melanoma. The objective is then to assess the affinity between those protein targets and the pharmacophores of the predicted candidates.
