About

Michael K. Gilson, Ph.D., M.D., is a Professor at the Skaggs School of Pharmacy and Pharmaceutical Sciences, where he also serves as Chair in Computer-Aided Drug Design and Co-Director of the UCSD Center for Drug Discovery Innovation. His research focuses on computer simulations of molecules to facilitate drug discovery, aiming to improve the realism, speed, and accuracy of these simulations through methods and software development. He manages BindingDB, an open database containing over 2 million measured protein-small molecule binding data points for more than 1 million compounds, and actively engages in molecular design and synthesis projects related to cancer and anesthesia. Gilson's academic background includes an A.B. in Bioengineering from Harvard College, a Ph.D. in Biochemistry and Molecular Biophysics from Columbia University, and an M.D. from Columbia University College of Physicians and Surgeons. His notable contributions include the development of statistical thermodynamics of protein-drug binding, structure-based development of mutation-resistant HIV-protease inhibitors, and the invention of the Mining Minima technology for computer-aided drug design. He has held leadership roles such as serving on the Executive Committee of the UC Drug Discovery Consortium and co-founding VeraChem LLC. His work has been recognized through awards including the Howard Hughes Medical Institute Physician Research Fellowship and the Endowed Chair in Computer-Aided Drug Design at UC San Diego.

Research topics

• Computer Science
• Physics
• Statistical physics
• Mathematics
• Quantum mechanics
• Chemistry
• Classical mechanics
• Physical chemistry
• Materials science
• Reliability engineering
• Engineering
• Geometry
• Statistics
• Telecommunications
• Thermodynamics
• Computational chemistry
• Chemical physics
• Environmental science

Selected publications

• ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

  arXiv (Cornell University) · 2026-05-12

  preprintOpen access

  Advances in large language models (LLMs) have recently opened new and promising avenues for small-molecule drug discovery. Yet existing LLM-based approaches for molecular generation often suffer from high rates of invalid and low-quality ligand candidates, a result of the syntactic limitations of current models with regard to molecular strings. In this paper, we introduce $\texttt{ToolMol}$, an evolutionary agentic framework for de novo drug design. $\texttt{ToolMol}$ combines a multi-objective genetic algorithm with an agentic LLM operator that iteratively updates the ligand population. We build a comprehensive toolbox of RDKit-backed functions that allows our agentic operator to consisently make precise ligand modifications. $\texttt{ToolMol}$ achieves state-of-the-art performance on multi-objective property optimization tasks, discovering drug-like and synthesizable ligands that have $>10\%$ stronger predicted binding affinity compared to existing methods, evaluated on three protein targets. $\texttt{ToolMol}$ ligands additionally achieve state-of-the-art results in gold-standard Absolute Binding Free Energy scores, gaining over existing methods by over $35\%$. By studying chain-of-thought reasoning traces, we observe that tool-calling enables the model to more faithfully execute its planned modifications, efficiently exploiting the strong chemical prior knowledge in LLMs.

  PublisherDOI

• ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

  ArXiv.org · 2026-05-12

  articleOpen access

  Advances in large language models (LLMs) have recently opened new and promising avenues for small-molecule drug discovery. Yet existing LLM-based approaches for molecular generation often suffer from high rates of invalid and low-quality ligand candidates, a result of the syntactic limitations of current models with regard to molecular strings. In this paper, we introduce $\texttt{ToolMol}$, an evolutionary agentic framework for de novo drug design. $\texttt{ToolMol}$ combines a multi-objective genetic algorithm with an agentic LLM operator that iteratively updates the ligand population. We build a comprehensive toolbox of RDKit-backed functions that allows our agentic operator to consisently make precise ligand modifications. $\texttt{ToolMol}$ achieves state-of-the-art performance on multi-objective property optimization tasks, discovering drug-like and synthesizable ligands that have $>10\%$ stronger predicted binding affinity compared to existing methods, evaluated on three protein targets. $\texttt{ToolMol}$ ligands additionally achieve state-of-the-art results in gold-standard Absolute Binding Free Energy scores, gaining over existing methods by over $35\%$. By studying chain-of-thought reasoning traces, we observe that tool-calling enables the model to more faithfully execute its planned modifications, efficiently exploiting the strong chemical prior knowledge in LLMs.

  PublisherOA PDF

• BPS2026 – Development and benchmarking of the open force field initiative self-consistent force field for proteins and small molecules

  Biophysical Journal · 2026-02-01

  articleSenior author
  PublisherDOI

• A Generalized Theory for the Structural and Spatial Mapping of Energy, Entropy, and Free Energy

  ChemRxiv · 2026-03-17

  articleOpen access1st authorCorresponding

  Systems in which the free energy density is nonuniform in space are familiar: the surface tension of a water droplet and the surface energy of a solid are good examples. Some such cases can be treated with prior theory, notably inhomogeneous solvation theory (IST), but IST is applicable only to liquids. Despite this limitation, IST has proven useful as a guide to the design of ligands to bind a targeted protein, based on the idea that ligands which displace water at high free energy will tend to bind more tightly, {\em ceteris paribus}. Here, we present Generalized Thermodynamic Mapping Theory (GTMT), a more general theory that is applicable to the entirety of a biomolecular or other chemical system, and which thus may provide additional guidance for molecular design. For example, it might highlight parts of a ligand whose local free energy rises on binding, thus suggesting where modifications could improve affinity. Starting from classical statistical thermodynamics, we derive structural decompositions that assign energy and entropy to individual atoms or to larger chemical components such as amino acid residues, and spatial decompositions that define continuously varying thermodynamic densities throughout the system. The potential energy is decomposed via the multibody expansion, and the entropy via the mutual information expansion. The resulting thermodynamic densities satisfy key desiderata: their spatial integrals yield the correct total thermodynamic quantities, the densities vanish where the atomic number density is zero, and all entropy terms above first order go to zero in the absence of correlation. The entropy decomposition in GTMT is closely related to that of IST, but it is simpler, largely because it expresses the entropy in terms of normalized probability density functions instead of non-normalized correlation functions. We show that, while GTMT is formally exact for any number of particles, N, IST is formally inexact for finite N, although it becomes exact in the thermodynamic limit. The unified framework presented here enables thermodynamic mapping of solute and solvent alike and is expected to support a range of applications, including in structure-based drug design, protein design, the analysis of allostery, and materials science.

  PublisherOA PDFDOI

• BindingDB Dataset, January 1, 2026. In BindingDB: Measured Binding Data for Protein-Ligand and Other Molecular Systems

  California Digital Library · 2026-01-01

  datasetOpen access1st authorCorresponding

  This object is the BindingDB dataset as of January 1, 2026. The core data for a given binding measurement comprises the identity of the two molecules involved, usually a protein and a small organic molecule; a measured indicator of their affinity, such as a dissociation constant or IC50; and the identity of the document from which the measurement was curated. Proteins are identified primarily by UniProt ID and small organic molecules by SMILES strings. This BindingDB release contains about 3.18 million experimental binding data.

  OA PDFDOI

• A Generalized Theory for the Structural and Spatial Mapping of Energy, Entropy, and Free Energy

  ChemRxiv · 2026-04-16

  article1st authorCorresponding
  PublisherDOI

• Assessment of Pharmaceutical Protein-Ligand Pose and Affinity Predictions in CASP16

  2025-04-25 · 8 citations

  preprintOpen access1st authorCorresponding

  The protein-ligand component of the 16th Critical Assessment of Structure Prediction (CASP16) challenged participants to predict both binding poses and affinities of small molecules to protein targets, with a focus on drug-like compounds from pharmaceutical discovery projects. Thirty research groups submitted predictions for 229 protein-ligand pose targets and 140 affinity targets across five protein systems. Template-based pose-prediction methods did particularly well, with the best groups achieving mean LDDT-PLI values of 0.69 (scale of 0-1 with 1 best). For comparison, we also ran a set of automated baseline pose-prediction methods, including ones using deep neural networks. Of these, AlphaFold 3 did particularly well, with a mean LDDT-PLI of 0.8, thus outscoring the best CASP16 predictor. The CASP affinity predictions showed modest correlation with experimental data (maximum Kendall’s τ = 0.42), well below the theoretical maximum possible given experimental uncertainty. As seen in prior challenges, providing experimental structures did not improve affinity predictions in the second stage of the challenge, suggesting that the scoring functions used here are a key limiting factor. Overall, the accuracy achieved by CASP participants is similar to that observed in the prior Drug Design Data Resource (D3R) blinded prediction challenges. The present results highlight the progress and persistent challenges in computational protein-ligand modeling and provide valuable benchmarks for the field of computer-aided drug design.

  PublisherOA PDFDOI

• Thermodynamics of Water Displacement from Binding Sites and its Contributions to Supramolecular and Biomolecular Affinity

  Angewandte Chemie International Edition · 2025-06-04 · 4 citations

  articleOpen accessSenior author

  Abstract The role of water displacement in noncovalent binding has been debated in the fields of supramolecular chemistry and drug design. We use molecular dynamics simulations of idealized host‐guest systems to address the long‐standing controversy of whether water is merely a bystander or an actual driver of noncovalent binding in aqueous solution. To isolate hydration effects, we consider a pseudo‐hard‐sphere guest binding to a series of cucurbit[8]uril‐based macrocyclic host models whose energetic interactions with water vary widely. The computed free energy cost of displacing water from binding sites ranges from 0 to +37 kcal mol −1 , strongly influencing binding affinities. However, neither water density nor excess chemical potential reliably indicates the thermodynamic favorability of cavity water. These results support the concept that “unfavorable” binding site water contributes to high‐affinity binding and resolve the paradox of stable but thermodynamically unfavorable cavity water. This work provides insights into the nature of the hydrophobic effect in molecular recognition and offers a framework for understanding the role of water in binding across various host‐guest and protein‐ligand systems.

  PublisherOA PDFDOI

• A Simple, Polarizable, Rigid, 3-Point Water Model Using the Direct Polarization Approximation

  Journal of Chemical Theory and Computation · 2025-07-07

  articleOpen accessSenior authorCorresponding

  We present dPol, a 3-point, rigid, polarizable water model that uses the direct approximation of polarization. We show that, with a moderate computational cost (∼3× slower than TIP3P), dPol achieves additional accuracy over widely used nonpolarizable 3-point rigid water models. Unlike most polarizable force fields, dPol allows the use of a 2 fs time-step with a conventional molecular dynamics integrator. The partial charges and polarizabilities used in dPol are derived from quantum chemistry calculations, while the Lennard-Jones parameters and geometry are adjusted to reproduce liquid properties under ambient conditions. The final dPol water model reproduces key room-temperature physical properties used in training, with a heat of vaporization of 10.43 kcal/mol, a dielectric constant of 80.7, a high-frequency dielectric constant of 1.60, a molecular polarizability of 1.41 Å3, a gas-phase dipole moment of 1.89 D, and a mean liquid-phase dipole moment of 2.55 D. Importantly, dPol also closely reproduces properties outside the training set, including the oxygen–oxygen radial distribution function of liquid water, as well as the self-diffusion coefficient (2.3×10–5 cm2 s–1) and shear viscosity (0.87 mPa s). Predicted temperature-dependent properties are also largely reproduced; although dPol does not correctly place the density maximum, this is not expected to impede successful application of the model to biomolecular systems near room temperature. The dPol water model is, by design, compatible with our AM1-BCC-dPol polarizable electrostatic model for small organic molecules [J. Chem. Theory Comput., 2024, 20, 1293–1305]. These models in combination establish a foundation for the integration of electronic polarizability into efficient force fields for heterogeneous systems of biological and pharmaceutical interest.

  PublisherOA PDFDOI

• The <scp>CASP</scp> 16 Experimental Protein–Ligand Datasets

  Proteins Structure Function and Bioinformatics · 2025-10-03 · 7 citations

  articleSenior authorCorresponding

  This paper presents the experimental protein-ligand datasets used as benchmarks in the CASP 16 blind prediction experiment-the first CASP round to incorporate targets from pharmaceutical discovery projects. We have assembled and characterized protein-ligand complexes for four proteins that are known or candidate drug targets: human chymase, human cathepsin G, human autotaxin, and the SARS-CoV-2 main protease. The collection encompasses over 200 co-crystal structures at resolutions better than 2.7 Å, paired with binding affinity measurements for approximately 160 compounds covering a broad affinity range (nanomolar to high micromolar). These data enabled the CASP16 pose-prediction and affinity-prediction challenges. Many systems feature potentially challenging characteristics, including chymase's electropositive surface and acidic ligands, which require proper handling of titratable ligand groups; autotaxin complexes with and without zinc coordination; and a SARS-CoV-2 protease crystal form exhibiting an unusually open active site conformation. We describe the experimental approaches-from protein production and crystallization to binding assay development-that yielded these reference data. Contributed by scientists at F. Hoffmann-La Roche and Idorsia Pharmaceuticals, these datasets represent actual drug discovery projects and therefore provide a realistic testbed for assessing how computational methods perform on pharmaceutically relevant targets. An accompanying paper in the present special journal issue provides a comprehensive assessment of the pose and affinity predictions for these pharmaceutical protein-ligand systems.

  PublisherDOI

Recent grants

• NIH Grant R41GM062050

  NIH · $100k · 2002

• AN OPEN RESOURCE TO ADVANCE COMPUTER-AIDED DRUG DESIGN

  NIH · $3.6M · 2014–2020

• BindingDB: Data and Tools for Drug Discovery, Chemical Biology and Systems Pharmacology

  NIH · $6.9M · 2004–2021

• NIH Grant R42GM062050

  NIH · $2.1M · 2006

• NIH Grant R29GM054053

  NIH · $432k · 2001

Frequent coauthors

• David L. Mobley

  University of California, Irvine

  64 shared

• Niel M. Henriksen

  56 shared

• Tom Kurtzman

  The Graduate Center, CUNY

  45 shared

• John D. Chodera

  42 shared

• Ray Luo

  University of California, Irvine

  41 shared

• Jian Yin

  40 shared

• Andrew T. Fenley

  Virginia Tech

  39 shared

• Michael R. Shirts

  University of Colorado Boulder

  38 shared

Education

• M.D.

  Columbia University Vagelos College of Physicians and Surgeons

  1989

• Ph.D., Biochemistry and Molecular Biophysics

  Columbia University

  1988

• A.B., Bioengineering

  Harvard College

  1981

Awards & honors

• Howard Hughes Medical Institute Physician Research Fellowshi…
• Distinguished Lecture in Computational and Mathematical Biol…
• Endowed Chair in Computer-Aided Drug Design, UC San Diego, 2…
• 5th Annual Kollman Lectureship, UC San Francisco, 2016