Biomolecules are not single structures. They are ensembles, and the function — catalysis, binding, allostery — lives in the distribution over conformations, not in any one snapshot. This is why generative models matter: they naturally parameterize distributions, and the distribution is exactly what biology needs. Diffusion and, more recently, flow matching are the current best tools for learning those distributions from data. This article connects the two to my work on RNA conformational ensembles in the GeneFold AI project.
Why biology needs a distribution
A protein or RNA in solution samples a Boltzmann-weighted ensemble. AlphaFold gives a single high-confidence structure, which is a triumph for folded proteins but a poor description of intrinsically disordered regions, RNA, and any molecule whose function involves a conformational transition. Predicting the ensemble — the set of populated states and their relative weights — is the harder and more biologically relevant problem. It is also a generative modeling problem: learn p(conformation | sequence, environment).
Diffusion models in one page
Denoising diffusion probabilistic models (Ho et al. 2020) define a forward process that gradually adds Gaussian noise to data over many steps, and train a network to reverse it. Sampling starts from noise and iteratively denoises to a sample. The score-based formulation (Song et al. 2021) generalizes this to continuous-time stochastic differential equations and shows the equivalence with score matching. For biomolecules, the crucial feature is equivariance: a molecule has no absolute frame, so the network must be SE(3)- or E(3)-equivariant, which architectures like EGNN and IPA enforce. RFdiffusion and Chroma demonstrated that diffusion can generate protein backbones with realistic geometry, and RFdiffusion2 extended it to enzyme active-site scaffolding.
Flow matching: a straighter path
Flow matching (Lipman et al. 2023) reframes generation as learning a time-dependent vector field that transports a simple prior (Gaussian) to the data distribution. Instead of a stochastic denoising chain, you regress the velocity along a probability path — most simply the linear interpolation between noise and data. Training is simulation-free and stable, and sampling integrates an ODE, which is typically much faster than the hundreds of steps a diffusion sampler needs. Rectified flow makes the path even straighter, enabling one- or few-step generation. For ensembles, the ODE viewpoint is a gift: fast sampling means you can generate thousands of conformations in the time a diffusion model produces dozens.
Flow matching for RNA conformational ensembles
In GeneFold AI we use flow matching to model RNA conformational ensembles conditioned on sequence. RNA is the hard case: few experimental structures, high flexibility, metal and Mg²⁺ dependence, and non-canonical base pairs that structure-prediction pipelines handle poorly. The training signal comes from molecular dynamics trajectories and the sparse experimental structures that exist. The model learns a field that transports noise to physically plausible conformations, and by sampling many trajectories we recover an ensemble rather than a point estimate. Downstream, we cluster the samples to identify populated states and estimate relative populations — the quantities that matter for riboswitch and ribozyme function.
The relationship to molecular dynamics
Flow matching does not replace molecular dynamics; it learns from it and accelerates it. MD gives physically grounded trajectories but is limited by timestep and by the millisecond timescale barrier: a system that folds on 10 ms is effectively unreachable. Generative models compress that process into a learned field that samples equilibrium-like conformations directly, bypassing the kinetic bottleneck. The trade-offs are real — generated samples respect learned statistics, not the true force field — so the honest framing is that these models are surrogate ensemble generators whose physical validity must be checked against MD and experiment, not assumed.
Validation: how do you know an ensemble is right?
Validation is the weakest link in generative structural biology. Reasonable-looking samples are not evidence. Useful criteria include: reproducing experimental observables (SAXS profiles, cryo-EM density, NMR chemical shifts, FRET efficiencies); matching MD-derived free-energy landscapes where MD is affordable; and internal consistency of populations across independent samples. Reporting only a structural plausibility score or an RMSD to one reference is not enough. I would rather see a model that reproduces three orthogonal experimental constraints than one that generates prettier pictures.
Current limitations
- Data scarcity. RNA and flexible-protein ensembles have orders of magnitude less training data than images.
- Physical fidelity. Learned fields are not force fields; energies and transition rates are not directly available.
- Conditioning. Response to mutations, ligands, ions, and temperature is hard to control and rarely validated.
- Evaluation. No community-standard benchmark for ensemble quality; metrics are fragmented.
- Compute. Equivariant sampling at scale remains expensive.
Diffusion and flow matching have given biology a genuinely new capability: fast, controllable sampling of molecular distributions. The next five years will be less about new architectures and more about physical grounding and rigorous evaluation. That is the part that turns a demo into a tool.
References
- Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models. Adv Neural Inf Process Syst. 2020;33:6840–6851.
- Song Y, Sohl-Dickstein J, Kingma DP, Kumar A, Ermon S, Poole B. Score-based generative modeling through stochastic differential equations. Int Conf Learn Represent. 2021.
- Lipman Y, Chen RTQ, Ben-Hamu H, Nickel M, Le M. Flow matching for generative modeling. Int Conf Learn Represent. 2023.
- Watson JL, Juergens D, Bennett NR, et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023;620(7976):1089–1100.
- Satorras VG, Hoogeboom E, Welling M. E(n) equivariant graph neural networks. Proc 38th Int Conf Mach Learn. 2021;139:9323–9332.
- Liu X, Gong C, Liu Q. Flow straight and fast: learning to generate and transfer data with rectified flow. Int Conf Learn Represent. 2023.
- Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583–589.
- Lindorff-Larsen K, Piana S, Dror RO, Shaw DE. How fast-folding proteins fold. Science. 2011;334(6055):517–520.
- Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493–500.