Sam Abujudeh, Marika Catapano, Paidi Creed, Jaime Domingues, Craig Glastonbury, Josep Montserrat, Francesca Mulas, Povilas Norvaisas, Delphine Rolando, Aaron Sim, Amparo Toboso-Navasa, Hamish Tomlinson, Arpad Vezer.
Cancer, sarcopenia, diabetes, and ALS, are but a few diseases that present notable heterogeneity between patients in both symptoms and aetiology. This heterogeneity extends to the patients’ response to experimental treatments, therefore presenting a significant challenge for drug discovery.
Embracing this challenge, at BenevolentAI we have developed an approach using omics data to identify patient subgroups. Here we present this approach, illustrated with a case study from our patient stratification drug discovery programs.
Unsupervised machine learning methods can be used to identify subgroup-specific patterns; the biological interpretation of these patterns is key for the identification of endotype-specific disease-modifying targets. We have developed a systematic evaluation of the identified patterns by assessing confounding variables, clinical covariates and biological mechanisms. Therefore, biologically meaningful subgroup-specific patterns are defined by the lack of confounding effects, their correlation with clinical variables and their implication in biological pathways.
Our pipeline tackles the challenges of interpreting subgroup-specific patterns derived from high-dimensional omics data, while uncovering pathobiological aspects of the disease specific to a group of patients. The identification of biological mechanisms that explain a given disease in different patient subgroups will facilitate drug discovery by providing targets regulating those mechanisms.
At BenevolentAI, Drug Discovery Scientists with Biology, Pharmacology and Chemistry backgrounds, work together with Bioinformaticians and AI Scientists. With our joint effort, we focus on identifying heterogeneous diseases that will benefit from our Precision Medicine approach. We look for omics datasets that reflect this heterogeneity while capturing clinical information; this will allow our models to find patient subgroups. Then, Drug Discovery Scientists, as myself, interpret the underlying biology to find disease drivers in each endotype. Target identification is followed by target validation in experimental settings that mimic the characteristics of each patient subpopulation so we find the right target for the right patient.