Research map / Probabilistic and scientific modeling / Molecules and proteins

Looking Under the Streetlight: Evaluation in Generative Molecular Dynamics

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Diffusion for molecular conformations

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Position paper arguing that generative molecular dynamics evaluation relies on insufficient references and misaligned observables, identifying six patterns (P1-P6) and proposing a tiered checklist and benchmark suite design.

Problem

Evaluation in generative molecular dynamics (GenMD) compares models against short or convenient references that cannot contain the rare events and slow kinetics the field claims to model, risking over-claimed advantage over classical baselines.

Approach

Frames GenMD evaluation failures as a community-level "streetlight effect" and identifies six recurring misaligned practices, pairing each with concrete observables (e.g., reverse ESS, multi-basin statistics, kinetic observables), classical baselines, cost accounting (N_break, amortized data cost), and a tiered checklist plus proposed reference-benchmark suites.

Key results

• RMSF is nearly identical across three metastable states of a CATH domain simulation, so it cannot distinguish ensembles differing in state populations.
• A flow trained by maximum likelihood on a small system had good test log-likelihood and Ramachandran agreement yet only about 3% ESS while mis-modeling a low-population metastable state.
• A model flagged as hallucinating against a ~ns reference had its predicted states later confirmed as physical by ultra-long unbiased MD and enhanced sampling.

Surprising findings

• RMSF agreement can be near-perfect even when the ensemble misrepresents multi-basin populations.
• Energy histogram agreement and good test log-likelihood can coexist with severe missing-mode failures detectable only by ESS.

Limitations

• Explicitly does not estimate prevalence of misaligned evaluation patterns in the literature.
• Some proposed checks (long-timescale references, enhanced sampling, experimental kinetics) are costly and may slow exploratory work.
• Proposed benchmark suite is described conceptually and not yet built.

Experimental setup

Datasets: MDQM9 mdCATH ATLAS MISATO CATH-derived MD

Metrics: RMSF effective sample size folding free energy mean first-passage time relaxation timescales native contact fraction

Baselines: AMBER CHARMM replica exchange MD metadynamics transition interface sampling forward flux sampling

Builds on: Boltzmann Generators time-lagged independent component analysis Markov state models effective sample size reweighting

Closest papers by meaning

• Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.88)
• Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.85)
• FlashMD: long-stride, universal prediction of molecular dynamics (NeurIPS 2025 · similarity 0.85)
• Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.84)
• Geometric Representation Condition Improves Equivariant Molecule Generation (ICML 2025 · similarity 0.84)
• UniGEM: A Unified Approach to Generation and Property Prediction for Molecules (ICLR 2025 · similarity 0.84)