Matthew James O'Meara, PhD

My computational pharmacology research program aims to develop and apply molecular simulation and artificial intelligence approaches to drive the discovery of small molecule biological probes and drugs. In the age of breakthrough biotechnologies, small molecules remain highly valuable: they have diverse physiochemical properties to target diverse biological processes, they can be multiplexed into a wide range of assays with precise spatial and temporal control, and they can be low-cost to produce and deliver, thus setting the standard for accessible therapeutics. However, they remain expensive to develop. Fundamentally, useful small molecules lie at the intersection of two complex spaces: the combinatorial space of chemical synthesis, and the topologically rich space of molecular interactions. State of the art drug discovery pipelines involve complex screening, lead optimization, and pre-clinical and clinical evaluation of safety and efficacy, while simultaneously aiming to minimize the overall cost, time, and attrition. An exciting strategy to accelerate discovery is to center the decision-making process on computational modeling, leveraging molecular simulations, machine learning, and chemoinformatic data analysis to actively drive the design-make-test cycle. Building on my background in computer science and pharmaceutical chemistry, my computational lab will drive the revolution in medicinal chemistry to leverage the broad advances in bioinformatics and AI to drive small molecule discovery.

Core expertise: virtual screening. We use molecular simulations and machine learning to model how small molecules dock into macromolecular binding sites. Through my PhD as a core-developer in the Rosetta Commons and as a postdoc working with UCSF DOCK, I have developed deep experience in applying molecular simulation to virtual screening in drug discovery campaigns. A guiding principle of my lab is to pragmatically balance modeling physical realism against computational cost during prospective discovery. In my previous work, I leveraged Rosetta to flexibly model receptors and ligands to answer “how does this compound bind” type of questions (Hernandez, et al., BMCL, 2022); and I leveraged DOCK with pre-computed scoring grids and ligand databases to rapidly screen billions of commercially available make-on-demand molecules to answer “what new molecules should we experimentally test” type of questions (Lyu, et al., Nature, 2019; Alon, et al., Nature, 2021). As the accessibility of compute and make-on-demand chemistry grows, allowing molecular simulation to be used more broadly, an emerging challenge is to clarify how to effectively apply them. Building on my experience with statistics and data modeling, my lab develops benchmarking and exploratory analysis tools to enable robust application of molecular simulations for drug discovery. Further my lab develops, applies, and publishes emerging deep learning models to virtual screening. Whereas previous generation machine learning models trained on sparse and biased data curated from the literature, and high-throughput screens often fail to generalize, the emerging high-capacity deep learning models can learn high-quality representations for molecules. These foundational models can then be fine-tuned using precious experimental data to efficiently identify novel drug-like bioactive molecules. Crucially, deep learning models require high-quality large-scale training data. An innovation in our group is to leverage simulation-based virtual screening to train and evaluate deep-learning molecule representation models.

Core expertise: computational and statistical models to navigate drug discovery. The keys to facilitating pharmacology decision making are to model for each experimental system, the potential outcomes, the information we aim to learn, and the real-world costs and constraints. Together, these can help to select informative experiments, optimize assay designs, and plan the discovery strategy. To do this we apply Bayesian and causal inference methods for high-content screening to control for experimental biases and isolate responsive sub-populations, and characterize model uncertainty and collaborate with high-content screening experimental labs at UMich and beyond to design and analyze experiments.

Core expertise: Chemoinformatic data science. Effective collaborative drug discovery campaigns require aggregation and synthesis of experimental and modeling data to facilitate decision making. Pharmacology data encompass diverse concepts at varying levels of specificity that can be difficult harmonize into common, accessible databases. Ideally, computational workflows can be developed to capture, curate, and analyze the data that work for all members of the team. As part of the UCSF Coronavirus Research Group, I developed a real-time platform for prediction nomination, triage, testing of drugs, chemical probes, and tool compounds to target potential host factors enabling SARS-CoV-2 infection (Gordon*, ..., O'Meara*, et al., Nature 2020). My group uses chemoinformatics to organize and drive drug-discovery collaborations.