NIH grant: Integrating physics-based and machine learning methods to accelerate computational drug design (1K99GM164832-01)

retained text

NIH grant 1K99GM164832-01: Integrating physics-based and machine learning methods to accelerate computational drug design
Institute: {'code': 'GM', 'abbreviation': 'NIGMS', 'name': 'National Institute of General Medical Sciences', 'admin_org_id': '12500024', 'admin_funding_url': ''}. PI: Alzbeta  Kubincova. Award: $126489.
Start: 2026-08-01T00:00:00.
Project Abstract: Small-molecule drug discovery is expensive and time consuming. Currently, each new drug brought to market costs around $1.4-$2.6 billion, and well over $400 million of this cost goes into preclinical studies. Yet more than 90% of discovery programs fail, often because of failing to identify a “lead” molecule binding to its target with sufficient potency, or later due to other liabilities, such as bad selectivity or toxicity. Computational methods are applied throughout early-stage drug discovery to prioritize compounds for experiments, and being able to virtually screen a large number of ligands increases the chances of finding leads, as well as reducing costs by relying less on synthesis. Particularly, the combination of accurate binding free energy calculations using molecular dynamics with the high throughput of active learning methods using surrogate machine learning models attracted interest in the pharmaceutical industry for retrieving potential lead compounds from large chemical libraries. This application focuses on new innovations to provide utility in prospective drug discovery applications, where active learning and free-energy methods are failing so far. Aim 1 involves the development of an improved active learning workflow which optimizes for multiple objectives (potency, selectivity, drug-likeness) to mitigate potential liabilities, and can leverage data at varying accuracy at the same time (docking scores and free energy) to strike a good balance between throughput and accuracy. Aim 2 extends the domain of applicability of relative binding free energy calculations to larger modifications of the ligand structure by accounting for multiple potential binding modes of a given ligand. The candidate has experience in developing molecular dynamics and active learning methodologies, as well as their application to drug discovery. The candidate’s long-term goal is broadening the scope and improving the throughput of computational chemistry methods to accurately score large numbers of small-molecule drug candidates, in order to boost the success rate of discovery programs. Short term research goals include the validation of new algorithms for active learning and ligand preparation for molecular dynamics, and their robust software implementation which can be built upon in the future. The training plan involves regular meetings with the mentoring committee led by Dr. David Mobley (binding free energy calculations), with the co-mentors Dr. Michael Gilson (computer-aided drug design), Dr. Brian Paegel (combinatorial chemistry), Dr. Stefano Forli (docking) and Dr. Pierre Baldi (machine learning). The practical training will be complemented by formal coursework, seminars, journal club meetings and presentations at scientific conferences. The University of California, Irvine will provide the candidate with an excellent research environment and compute resources on the HPC3 computer cluster, as well as ample learning opportunities in the form of courses and seminars (machine learning, drug discovery, rigor of research).

document fetched 2026-08-09 06:52