Gen-COMPAS (Tang, Pandey, Chen, Megías, Dehez & Chipot, Nature 2026) is an iterative loop that pairs a denoising diffusion generative model with a learned committor function to find the transition states of rare biomolecular conformational changes — without the pre-chosen collective variables (CVs) that metadynamics, umbrella sampling, and adaptive biasing force depend on [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl] arxiv.org. Across systems from a miniprotein to a pentameric ligand-gated ion channel, it recovers transition states, committor maps, and free-energy landscapes from just the two known end-point structures, at nanosecond-to-submicrosecond aggregate sampling where conventional methods need orders of magnitude more [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl].
The core idea, for a general reader
A large conformational change (folding, gating, transport) is a rare event: the molecule spends almost all its time sitting in one of two stable "valleys" and only occasionally hops over the mountain pass between them. Ordinary molecular dynamics (MD) has to wait out that hop in real simulated time, which can take microseconds to milliseconds — computationally enormous. Enhanced-sampling methods speed this up by artificially pushing the system along a hand-picked "reaction coordinate," but if that coordinate misses a slow motion, the pushed pathway is wrong arxiv.org.
Gen-COMPAS attacks the pass directly. Two ideas do the work:
- A diffusion model guesses what the intermediate looks like. The same class of generative model used for AI images is trained here to produce physically plausible structures that bridge the two end states — candidate "mountain-pass" shapes [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl] arxiv.org.
- The committor tells you which guesses are actually transition states. The committor q(x) is the probability that, starting from structure x, the molecule reaches the product basin before falling back to the reactant. A true transition state sits on the q = ½ surface (the "separatrix") — equally likely to roll down either side. Gen-COMPAS learns q directly in full conformational (Cartesian) space, so it never needs a predefined CV arxiv.org.
The iterative workflow
- Run very short (1–2 ns) unbiased MD of the two known metastable states — reactant A and product B — to seed a dataset arxiv.org.
- Train a denoising diffusion probabilistic model (Ho, Jain & Abbeel–style) on that data; it generates intermediate conformations connecting A and B arxiv.org.
- Learn the committor q in conformational space and select structures near the q = ½ separatrix as transition-state candidates arxiv.org.
- Run two targeted-MD simulations, one launched from A and one from B, to relax the system onto physically realistic structures matching each generative target arxiv.org.
- Shoot short unbiased MD from those near-separatrix points; the new trajectories retrain both the diffusion model and the committor predictor — a feedback loop that focuses sampling on the transition region arxiv.org.
On convergence, the accumulated unbiased sampling yields (i) the transition-state ensemble, (ii) committor maps projected onto any interpretable CV chosen after the fact, (iii) committor-consistent transition pathways, and (iv) free-energy landscapes (FELs) arxiv.org.
How much sampling it saves
The savings come from not biasing and not waiting: unbiased sampling is concentrated near the pass using generative targets, so correct thermodynamics and kinetics are preserved while the barrier crossing is bypassed arxiv.org.
- Trp-cage folding: required sampling fell from 208 µs to 594 ns — the authors call this "approximately hundreds times more efficient." That is a fold factor of ≈ 350× (computed: 208 µs ÷ 594 ns = 208,000 ns ÷ 594 ns ≈ 350; inputs 208 µs and 594 ns both from arxiv.org).
- General claim: transition-region ensembles are obtained at nanosecond-to-submicrosecond aggregate sampling where conventional MD or enhanced sampling need orders of magnitude more [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl] arxiv.org.
A second efficiency axis matters versus dataset-hungry emulators (MDGen, BioEmu), which need "hundreds of milliseconds of MD or large-scale experimental measurements" to pretrain: Gen-COMPAS needs no pre-generated dataset and operates on the system's real Hamiltonian/force field, so its FELs are directly comparable to MD rather than reconstructed from a learned latent space arxiv.org.
Caveat on the comparison: the head-to-head number is against brute-force reference MD (the DESRES Trp-cage trajectories). The paper frames metadynamics/hand-picked-CV methods qualitatively — they are "computationally demanding" and can bias outcomes — rather than giving a single matched speed-up figure against a specific metadynamics run [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl] arxiv.org.
Systems it was tested on
| System | Environment | What Gen-COMPAS recovered | Source |
|---|
| NANMA (alanine dipeptide), trialanine | Vacuum | Proof-of-concept: FEL, committor map, TSE, pathways matching prior work | arxiv.org |
| Trp-cage miniprotein | Explicit solvent | Bifurcated two-pathway folding (helix-nucleation-first vs hydrophobic-collapse-first); FEL matching µs DESRES reference | arxiv.org |
| Ribose-binding protein (RBP) | Explicit solvent | Binding-upon-folding via two cooperative open→closed routes (stepwise induced-fit vs simultaneous) | arxiv.org |
| Mitochondrial ADP/ATP carrier (AAC) | Membrane | Confirms an obligatory occluded (O-) intermediate on the C→O→M import path (holo, ADP-bound); O-state vanishes with a ~10 kcal/mol barrier in apo | arxiv.org |
| nAChR — pentameric ligand-gated ion channel | Membrane | Top-complexity system (Fig. 4 of the Nature paper); committors, transition states, and FELs from end-point structures alone | [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl] |
One reconciliation point worth flagging: the arXiv preprint (2510.24979) headlines the ladder "miniprotein → RBP → mitochondrial carrier," while the published Nature abstract and figures state the top system as a "pentameric, ligand-gated ion channel" (nAChR, Fig. 4) arxiv.org [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl]. The AAC results appear in the preprint text in full; the nAChR result is the added top-complexity case in the journal version.
Evidence-depth note: the mechanistic detail (workflow, Trp-cage/RBP/AAC results, kcal/mol barriers, the 208 µs → 594 ns figure) comes from the full text of the openly posted arXiv preprint, which I read in full. The published Nature article's methods/results body is paywalled; I read its full abstract, figure titles, data/code-availability statements, and complete reference list but not its main-text results. The nAChR-specific numbers therefore rest on the Nature abstract and figure captions, not on a full-text read of that section [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl].
Limitations the authors acknowledge
- It is an exploration ecosystem, not a final quantitative oracle. The authors state that for "applications demanding the highest quantitative precision," the generated FELs "should serve as a well-informed starting point" — the transition-state ensemble can then seed traditional enhanced sampling or transition path sampling to reach rigorously converged results arxiv.org.
- Accuracy is bounded by the force field. It can be systematically improved with higher-fidelity methods (ab initio MD, machine-learning force fields), which implies the baseline results inherit force-field error arxiv.org.
- It requires known end states and appropriate force-field parameters to start arxiv.org.
- Demonstrated on two-state (A↔B) transitions. The authors explicitly frame extension "far beyond the simple two-state Markov-jump paradigm" — to multiple metastable states — as future work arxiv.org.
- It does not replace brute-force or enhanced sampling. In their words, Gen-COMPAS "does not entirely replace these approaches" but shows the essence of rare-event dynamics can be revealed without extreme cost arxiv.org.
Primary sources
- The paper: Tang, Pandey, Chen, Megías, Dehez & Chipot, "Breaking timescales with generative sampling of conformational transitions," Nature (2026), doi:10.1038/s41586-026-11025-1, PMID 42717085 [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl]; open preprint (full text): arXiv:2510.24979 arxiv.org. Code/data:
github.com/Tangcyu/Gen-COMPAS.
- Committor / transition-state foundations: Kang, Trizio & Parrinello, "Computing the committor with the committor," Nat. Comput. Sci. 4, 451–460 (2024), DOI 10.1101/2022.01.08.475519 context aside — direct record DOI 10.1038/s43588-024-00645-0; Contreras Arredondo et al., "Learning the committor without collective variables," Nat. Comput. Sci. 6, 350–357 (2026); Megías et al., Nat. Comput. Sci. 5, 592–602 (2025) [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl].
- Building blocks: Ho, Jain & Abbeel, denoising diffusion probabilistic models (NeurIPS 2020); Schlitter, Engels & Krüger, targeted MD (1994); Lindorff-Larsen, Piana, Dror & Shaw, "How fast-folding proteins fold," Science 334, 517–520 (2011) — the DESRES reference for the Trp-cage comparison [source:web-www.nature.com-aHR0cHM6Ly93d3cubmF0dXJl] arxiv.org.
A note on process: my independent second-pass review tool was unavailable this turn, so I re-checked the load-bearing claims directly against the full-text preprint and the Nature abstract/figures myself. The one claim I could not verify against full journal text — the nAChR-specific results — is flagged as resting on the abstract and figure captions rather than a full-text read. If you'd like, I can pull the RBP or AAC mechanism into a labeled pathway diagram, or attempt the paper's Supplementary Information (NANMA/trialanine benchmarks) if you can supply access.