Thesis Defense | Filipp Nikitin | August 18, 2026 | 9am
CPCB and CBD are proud to announce:
Title: Generative Modeling of Molecular Geometry Across Equilibrium and Reactive Regimes
Filipp Nikitin
August 18, 20269am
Mellon Institute Room 355
For zoom details, please contact Nicole Stenger
Committee:
Olexandr Isayev, Ph.D. (Advisor, Committee Chair)
Jose Lugo-Martinez, Ph.D.
David R. Koes, Ph.D.
Dylan M. Anstine, Ph.D.
Abstract:
Generative molecular geometry modeling is a pivotal task in molecular discovery, with applications ranging from virtual screening to mechanistic modeling. Recent diffusion and flow-based generative models have demonstrated strong performance in generating molecular graphs and approximate geometries; however, many existing approaches struggle to produce high-quality 3D structures, fail to capture energetically meaningful configurations, and rely on benchmarks that miss key geometric or energetic aspects. As a result, it remains challenging to determine how architectural choices, training objectives, and datasets contribute to chemically relevant generative performance.
This dissertation presents a unified diffusion and flow-based generative modeling framework for molecular data, designed to jointly model continuous 3D coordinates and discrete chemical features. The first contribution develops scalable generative models for unconditional 3D molecule generation using both diffusion and flow matching objectives within a shared architectural framework. Building on recent advances in joint continuous and discrete denoising objectives, this work also introduces geometry-aware benchmarking protocols that explicitly evaluate molecular structure precision and energetic consistency, addressing key limitations of existing evaluation methodologies.
The second contribution targets low-energy conformer generation by developing a large-scale, quantum-mechanically informed conformer dataset and a corresponding generative model. Leveraging extensive conformational sampling enabled by recent advances in machine-learned interatomic potentials, this work enables generative models that preferentially sample physically meaningful, low-energy molecular geometries while maintaining chemical diversity. Particular emphasis is placed on explicitly modeling stereochemistry, including chiral centers and E/Z isomerism, together with molecular protonation states, which are frequently ignored or mishandled by existing generative approaches. By incorporating stereochemistry-aware graph augmentation and training directly on a large-scale conformer dataset optimized to near quantum-mechanical accuracy under implicit solvation, the resulting model learns implicit representations of the molecular conformational energy landscape.
The third contribution extends the generative modeling framework to chemical reactivity through direct transition state generation and dataset construction. An automated dataset-generation pipeline combines reaction enumeration, quantum-chemical validation, and model-guided augmentation to yield approximately 1.62 million transition states. Trained on this corpus, the model generates transition state geometries from reactant-product bond connectivity without requiring pre-optimized structures, expert-curated initial guesses, or expensive pathway search methods. Stereochemistry-augmented reactant-product graphs preserve and control stereochemical outcomes, enabling selective exo/endo Diels-Alder transition states and E/Z elimination pathways. The model also generalizes across diverse reaction classes, including click cycloadditions, amide hydrolysis, carbonate formation, epoxidation, esterification, Diels-Alder cycloadditions, and the Hajos-Parrish-Eder-Sauer-Wiechert organocatalytic cycle.
Together, these contributions establish a unified generative modeling approach for molecular geometry generation, conformational search, and reactivity prediction. This work advances diffusion and flow-based models toward practical use in large-scale molecular discovery and mechanistic analysis, while providing standardized benchmarks, datasets, and implementations to support reproducible and chemically meaningful evaluation.