AI-Powered Gen-COMPAS Captures Rare Molecular Transitions in Record Time

Some of the most important events in biology happen in the blink of a molecular eye. A protein folds into its functional shape, a membrane transporter flips between two conformations, an ion channel snaps open to let charged particles flood across a cell membrane. These transitions are the machinery of life, yet they are precisely the events that molecular simulations have struggled to capture for decades. Now, a team of computational biophysicists reports a generative artificial intelligence framework, called Gen-COMPAS, that reconstructs these rare transition pathways from nothing more than the starting and ending structures of a molecule, at a computational cost that conventional approaches cannot approach.

The work, published in Nature by Chenyu Tang, Mayank Prakash Pandey and colleagues at the Université de Lorraine and CNRS in Nancy, France, with collaborators in Madrid, Chicago and Urbana-Champaign, addresses a stubborn bottleneck in computational biology. Standard molecular dynamics simulations propagate atoms step by step according to the laws of classical physics, and they do so faithfully. But the transitions that matter most, such as protein folding, allosteric signaling and membrane transport, are rare events. A simulation might churn through millions of time steps without ever witnessing the molecule cross from one stable state to another, because the probability of stumbling into the transition region by chance is vanishingly small.

Computational scientists have long responded with enhanced-sampling strategies, which apply artificial forces or cleverly designed biases to push molecules over energetic barriers. These methods have produced genuine insights, but they come with serious caveats. They are computationally demanding, and they typically require the researcher to choose, in advance, a set of collective variables, essentially a guess about which combination of atomic coordinates describes the reaction. If that guess is wrong or incomplete, the resulting free-energy landscape and apparent mechanism can be distorted. A recent study cited by the team put the problem bluntly in its title: convergence is not correctness, and enhanced-sampling methods can perform inconsistently as biological complexity increases.

Gen-COMPAS takes a different route. The framework couples a denoising diffusion probabilistic model, the same class of generative architecture that powers modern image synthesis, with a rigorous statistical concept from transition state theory known as the committor. The diffusion model is trained to produce structurally plausible intermediate conformations that sit between the two known end points of a transition. These generated structures are not accepted on faith. Instead, they are filtered using committor-based criteria, which identify the configurations from which a molecule is equally likely to fall forward to the product state or backward to the reactant state, the hallmark of a true transition state.

The committor is a powerful but demanding quantity. For any given molecular configuration, it answers a deceptively simple question: if the system were released from this point with random thermal velocities, what is the probability that it would commit to the product rather than the reactant? Configurations with a committor of one half define the transition state ensemble. Estimating this probability traditionally requires launching many short trajectories from each candidate configuration, which is exactly the kind of brute-force calculation that makes rare-event statistics so expensive. Gen-COMPAS turns this into an iterative loop: the generative model proposes candidate intermediates, short unbiased simulations from those intermediates test whether they genuinely belong to the transition region, and the results refine both the model and the ensemble.

The payoff is a dramatic compression of the required sampling. According to the authors, short unbiased simulations launched from the generated intermediates yield transition-region ensembles at aggregate sampling scales ranging from nanoseconds to submicroseconds, whereas conventional approaches would require orders of magnitude more simulation time to assemble equivalent statistics. In practical terms, the method replaces the hopeless task of waiting for a rare event to happen spontaneously with the tractable task of verifying that generated candidate structures are genuine members of the transition ensemble.

To demonstrate the framework, the team applied it to three systems of increasing size and biological complexity. The first was Trp-cage, a miniprotein of roughly twenty residues that has become a standard benchmark for folding studies because it folds quickly enough to be simulated directly. Decades of work, including landmark long-trajectory studies of fast-folding proteins, have mapped its folding landscape in detail, giving the team an independent yardstick against which to judge their results. Gen-COMPAS recovered committors, transition states and free-energy landscapes consistent with this established knowledge, starting only from the folded and unfolded end-point structures.

The second test system was the mitochondrial ADP/ATP carrier, a membrane transporter that shuttles the energy currency ATP across the inner mitochondrial membrane in exchange for ADP. This protein operates through a switching mechanism between conformational states that has been studied with structural biology and a variety of simulation techniques, but its full transition pathway remains challenging to sample because the carrier is embedded in a lipid membrane and undergoes large-scale domain movements. Gen-COMPAS reconstructed the transition pathway between the carrier’s known states without any predefined reaction coordinate, offering a mechanistic picture that previously would have demanded extensive biased simulations and careful, potentially contentious, choices of collective variables.

The third and most ambitious application was the nicotinic acetylcholine receptor, a pentameric ligand-gated ion channel that mediates synaptic transmission. Channel activation involves coordinated conformational changes across five protein subunits, and recent work has suggested that these subunit transitions occur asynchronously, priming the receptor for activation in a stepwise fashion. Simulating the full gating transition of such a large membrane protein is far beyond the reach of straightforward molecular dynamics. Gen-COMPAS nevertheless recovered the transition pathway and free-energy landscape from the known resting and activated structures alone, with no prior mechanistic knowledge encoded into the calculation.

The significance of the method lies not only in its speed but in what it does not require. No collective variables need to be chosen by hand, no reaction coordinates need to be guessed, and no prior hypothesis about the mechanism needs to be supplied. The generative model and the committor filter together let the data define the pathway. This removes one of the most persistent sources of bias in enhanced sampling, the researcher’s own assumptions about how the transition ought to proceed, and it opens the door to systems where no expert intuition exists to guide variable selection.

The method also arrives amid a broader convergence of machine learning and molecular simulation. Generative models have been used to emulate equilibrium protein ensembles, to sample molecular dynamics trajectories directly, and to guide transition path sampling in systems ranging from lipid flip-flop to membrane nanoporation. Gen-COMPAS distinguishes itself by combining generative proposal with a physically rigorous acceptance criterion rooted in transition rate theory, ensuring that the structures it produces are not merely plausible but statistically meaningful members of the true transition ensemble. The authors acknowledge prior machine-guided path-sampling work and committor-learning approaches as foundations for this synthesis.

Accessibility may prove as important as performance. Gen-COMPAS is released as an installable open-source package on GitHub, complete with a unified command-line interface, ready-to-use configuration files, molecular-simulation inputs for representative systems, and a graphical user interface that guides users through the available options. Input files for all systems analyzed in the paper are publicly available, and source data accompany the publication. The team also gratefully acknowledges reference molecular dynamics trajectories provided by D. E. Shaw Research, whose specialized hardware has long set the benchmark for long unbiased simulations that methods like this one can now be validated against.

For the field, the implications extend across drug discovery, basic biophysics and beyond. Ion channels and membrane transporters are among the most heavily targeted protein families in pharmacology, and understanding the exact pathway by which they switch states can reveal where small molecules might bind to stabilize one state or another. Allosteric transitions in signaling proteins, another class of rare events, could similarly be mapped without the guesswork that has historically limited such studies. If the framework generalizes as its demonstration systems suggest, the computational microscope that biophysicists have been building for half a century may finally be able to zoom in on the fleeting moments that matter most, the instants in which a molecule decides what it will become next.

Subject of Research: Generative committor-guided path sampling of biomolecular conformational transitions

Article Title: Breaking timescales with generative sampling of conformational transitions

Article References: Breaking timescales with generative sampling of conformational transitions. (n.d.). https://doi.org/10.1038/s41586-026-11025-1

Image Credits: AI Generated

DOI: 10.1038/s41586-026-11025-1

Keywords: Gen-COMPAS, molecular dynamics, transition path sampling, denoising diffusion model, committor, protein folding, enhanced sampling, ion channel, membrane transporter, free-energy landscape, machine learning, computational biophysics