Protein conformational ensemble generator that samples heavy-atom structures in a latent space, with a variant conditioned on temperature.
No providers recorded yet. Browse all providers
Structure predictors return one conformation, but proteins exist as populations whose relative weights determine function. Molecular dynamics (MD) simulation samples those populations from physics, at a cost that makes exhaustive sampling impractical for most systems. A growing class of deep generative models is trained on MD trajectories to reproduce ensembles in seconds instead of GPU-months, with AlphaFlow and BioEmu-1 as prominent examples.
aSAM — atomistic Structural Autoencoder Model — is the entry in that class from Giacomo Janson, Alexander Jussupow, and Michael Feig at Michigan State University. It differs from its peers in two ways. First, it is a latent diffusion model: an autoencoder compresses heavy-atom coordinates into an SE(3)-invariant per-residue latent representation, and the diffusion process runs entirely in that latent space, with an AlphaFold 2 Structure Module serving as the decoder. Because every heavy atom is represented in the latent code, the model learns joint distributions of backbone and side chain torsions directly, rather than generating a backbone and appending side chains afterwards.
Second, one variant takes temperature as an explicit input. That variant, aSAMt, is trained on the mdCATH dataset of simulations at five temperatures and generates ensembles conditioned on a continuous temperature value, including values it never saw in training. A companion constant-temperature model, aSAMc, is trained on ATLAS at 300 K and serves as the head-to-head comparison against AlphaFlow. (The reference implementation lives in a repository named sam2, unrelated to the Segment Anything family of image models.)
The encoder is a four-block transformer fed radial basis expansions of Cα–Cα, N–O, and side chain centroid distance matrices plus φ, ψ, ω, and χ1–χ4 torsion features, producing a latent of dimension 32 per residue; the decoder is a five-block AlphaFold 2 Structure Module. A denoising diffusion probabilistic model with a Diffusion Transformer backbone, a maximum of 1,000 timesteps, and a sigmoid noise schedule learns the latent distribution, conditioned on sequence, an initial 3D structure, and — for aSAMt — temperature. aSAMc was trained on 1,174 ATLAS chains (0.35 ms cumulative simulation time) and aSAMt on 5,268 mdCATH domains at 320, 348, 379, 413, and 450 K (55.0 ms cumulative), with 90 held-out test domains sharing at most 20% sequence identity with training. On the ATLAS test set, AlphaFlow holds a small edge in Cα RMSF correlation (0.904 versus 0.886), while aSAMc is better on backbone and side chain torsion distributions and runs 17.4× to 28.0× faster at the default 100 diffusion steps. Against MD run under a matched protocol, aSAMt was 1,000× to 2,317× faster. On 62 monomeric proteins with measured melting temperatures, fitted values correlated with experiment at r = 0.510 while sitting systematically about 130 K high.
aSAM suits any workflow that needs a conformational ensemble rather than a single structure and can tolerate MD-quality statistics: estimating residue flexibility, generating starting structures for targeted simulations, examining side chain rotamer populations at a binding site, and screening designed proteins for conformational stability. Temperature conditioning opens a use case its peers do not cover — comparing the thermal behavior of homologs or variants in silico, which the authors demonstrate on mesophile/hyperthermophile protein pairs and on de novo designs from RFdiffusion characterized experimentally as highly thermostable.
aSAM was published in Communications Chemistry in November 2025, following a March 2025 bioRxiv preprint; the version of record carries CC BY-NC-ND 4.0 terms while the preprint is CC BY. Its main contribution is showing that environmental conditions can be folded into a transferable ensemble generator at all, and that latent-space diffusion is a practical route to atomistic detail without an all-atom denoiser. Real limitations remain and are documented by the authors: relative free energies between metastable states are not quantitatively reproduced, since mdCATH trajectories are too short to observe many transitions; accuracy degrades at 450 K, where the folded-fraction error rises to 0.141; and the model handles only single-chain proteins of standard amino acids with no ligands or cofactors. Code and both checkpoints are released under Apache 2.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.