Physics-Informed Generative Models for Molecular Structures

This direction is split into other directions: steric clash prevention, geometric validity constraints. Its last version stays readable and citable here.

Version 1 of 1194 papers233 conceptsfrozen 2026-10-04

Deep generative models have revolutionized our ability to sample the vast space of chemical compounds. By treating molecular coordinates as points in a high-dimensional distribution, architectures like Diffusion Models and Flow Matching can "dream up" new drug candidates or protein folds. However, a persistent "reality gap" remains: a model can achieve a perfect statistical score on geometric benchmarks while proposing molecules that are physically impossible.

Recent research has shifted from purely data-driven generation toward physics-informed generative modeling. This direction seeks to bake the "hard" laws of chemistry—such as bond valency, steric exclusion, and energy minimization—directly into the neural architecture.

The Reality Gap: Statistical vs. Physical Validity

In traditional computer vision, a "hallucinated" pixel in a generated image might go unnoticed. In chemistry, a hallucinated atom is a catastrophe. As Mayk Caldas Ramos notes, Large Language Models often propose "Texas Carbons"—carbon atoms with five bonds—violating fundamental valency.

Even when models respect basic graph connectivity, their 3D realizations often fail. Charles B. Harris and Pietro Lio introduced POSECHECK, a benchmark revealing that state-of-the-art models frequently generate "clashing" structures where atoms overlap in space or rings are under extreme internal strain. These models prioritize geometric similarity (low RMSD) to training data, but as Alex Morehead and colleagues argue in PoseBench, geometric accuracy does not imply chemical realism.

Force-Guided Diffusion and Energy Landscapes

To bridge this gap, researchers are integrating physical energy functions into the generative process. The intuition is simple: if we know the "energy" of a molecular configuration, we can use the gradient of that energy to push the model toward more stable shapes.

In CONFDIFF, Yan Wang and others use "force-guided" diffusion to sample protein conformations. Instead of just denoising random coordinates, the model is guided by an internal force field that prevents unphysical artifacts like fragmented backbones. Similarly, Ziyang Yu developed Force-Guided Bridge Matching, which learns to "jump" across time intervals in molecular dynamics while remaining grounded by an intermediate force field.

By embedding real-world potential energy calculations into the loss function—an approach pioneered by John Kevin Cava and John Vant—models learn to generate "pathways of least resistance." This ensures that the generated structures aren't just static snapshots, but represent states that could actually exist in a thermal environment.

Beyond Points: Manifolds and Priors

Standard models often treat atoms as dimensionless points. This makes it easy for the model to accidentally place two atoms in the same location. Liang Yan and Divin Yan addressed this in NucleusDiff by explicitly modeling the "electron cloud" manifold around nuclei. By treating atoms as volumes with repulsive boundaries, they eliminated "separation violations" and significantly improved binding affinity.

Other researchers focus on the starting point of the generation. Most diffusion models start from unstructured Gaussian noise, which has no concept of chemical connectivity. Hannes Stark and Bowen Jing proposed FLOWSITE, which uses a "harmonic prior." This prior acts like a set of virtual springs that maintain the basic topology of the molecule even when it is "noisy," making it much easier for the model to converge on a valid 3D structure.

The Hybrid Future: AI + Physics Relaxation

Even the best generative models, like AlphaFold 3 by Josh Abramson and the DeepMind team, can still produce stereochemical errors like chirality violations (mirror-image errors) in complex environments.

The emerging consensus, highlighted in the PoseX benchmark, is that while AI has overtaken traditional physics in speed and diversity, it still benefits from a final "physics-based energy relaxation." Models like Uni-Mol Docking V2 by Zhifeng Gao show that by prioritizing physical plausibility alongside geometric accuracy, we can produce results where 95% of generated poses are free from the "hallucinations" that previously plagued the field.

The goal of this research direction is not to replace physics with AI, but to use AI to navigate the vast chemical landscape while using physics as the "guardrails" that ensure every step taken is a real one.

Papers in this version

Showing 60 of 194 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physics_informed_generative_models_for_m_6c515aa4,
  title = {Physics-Informed Generative Models for Molecular Structures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-generative-models-for-molecular-structures/rdr_6c515aa48fcd475a8fcf1153632a1907},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_physics_informed_generative_models_for_m_6c515aa4_v1,
  title = {Physics-Informed Generative Models for Molecular Structures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-generative-models-for-molecular-structures/rdr_6c515aa48fcd475a8fcf1153632a1907/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}