Identifying a small molecule from its mass spectrum is the primary open problem in computational metabolomics. This is typically cast as information retrieval: an unknown spectrum is matched against spectra predicted computationally from a large database of chemical structures. However, current approaches to spectrum prediction model the output space in ways that force a tradeoff between capturing high resolution mass information and tractable learning. We resolve this tradeoff by casting spectrum prediction as a mapping from an input molecular graph to a probability distribution over chemical formulas. We further discover that a large corpus of mass spectra can be closely approximated using a fixed vocabulary constituting only 2% of all observed formulas. This enables efficient spectrum prediction using an architecture similar to graph classification - GrAFF-MS - achieving significantly lower prediction error and greater retrieval accuracy than previous approaches.

Better understanding the transcriptomic response produced by a compound perturbing its targets can shed light on the underlying biological processes regulated by the compound. However, establishing the relationship between the induced transcriptomic response and the target of a compound is non-trivial, partly because targets are rarely differentially expressed. Therefore, connecting both modalities requires orthogonal information (e.g., pathway or functional information). Here, we present a comprehensive study aimed at exploring this relationship by leveraging thousands of transcriptomic experiments and target data for over 2000 compounds. Firstly, we confirm that compound-target information does not correlate as expected with the transcriptomic signatures induced by a compound. However, we reveal how the concordance between both modalities increases by connecting pathway and target information. Additionally, we investigate whether compounds that target the same proteins induce a similar transcriptomic response and conversely, whether compounds with similar transcriptomic responses share the same target proteins. While our findings suggest that this is generally not the case, we did observe that compounds with similar transcriptomic profiles are more likely to share at least one protein target and common therapeutic applications. Finally, we demonstrate how to exploit the relationship between both modalities for mechanism of action deconvolution by presenting a case scenario involving a few compound pairs with high similarity.

Recent advances in Knowledge Graphs (KGs) and Knowledge Graph Embedding Models (KGEMs) have led to their adoption in a broad range of fields and applications. The current publishing system in machine learning requires newly introduced KGEMs to achieve state-of-the-art performance, surpassing at least one benchmark in order to be published. Despite this, dozens of novel architectures are published every year, making it challenging for users, even within the field, to deduce the most suitable configuration for a given application. A typical biomedical application of KGEMs is drug-disease prediction in the context of drug discovery, in which a KGEM is trained to predict triples linking drugs and diseases. These predictions can be later tested in clinical trials following extensive experimental validation. However, given the infeasibility of evaluating each of these predictions and that only a minimal number of candidates can be experimentally tested, models that yield higher precision on the top prioritized triples are preferred. In this paper, we apply the concept of ensemble learning on KGEMs for drug discovery to assess whether combining the predictions of several models can lead to an overall improvement in predictive performance. First, we trained and benchmarked 10 KGEMs to predict drug-disease triples on two independent biomedical KGs designed for drug discovery. Following, we applied different ensemble methods that aggregate the predictions of these models by leveraging the distribution or the position of the predicted triple scores. We then demonstrate how the ensemble models can achieve better results than the original KGEMs by benchmarking the precision (i.e., number of true positives prioritized) of their top predictions. Lastly, we released the source code presented in this work at https://github.com/enveda/kgem-ensembles-in-drug-discovery.

Network-based approaches are becoming increasingly popular for drug discovery as they provide a systems-level overview of the mechanisms underlying disease pathophysiology. They have demonstrated significant early promise over other methods of biological data representation, such as in target discovery, side effect prediction and drug repurposing. In parallel, an explosion of -omics data for the deep characterization of biological systems routinely uncovers molecular signatures of disease for similar applications. Here, we present RPath, a novel algorithm that prioritizes drugs for a given disease by reasoning over causal paths in a knowledge graph (KG), guided by both drug-perturbed as well as disease-specific transcriptomic signatures. First, our approach identifies the causal paths that connect a drug to a particular disease. Next, it reasons over these paths to identify those that correlate with the transcriptional signatures observed in a drug-perturbation experiment, and anti-correlate to signatures observed in the disease of interest. The paths which match this signature profile are then proposed to represent the mechanism of action of the drug. We demonstrate how RPath consistently prioritizes clinically investigated drug-disease pairs on multiple datasets and KGs, achieving better performance over other similar methodologies. Furthermore, we present two case studies showing how one can deconvolute the predictions made by RPath as well as predict novel targets.