UFConf

Adenylate kinase is open when free and closed when bound, and both states sit in the PDB under one sequence; a predictor trained to emit one answer has to pick one. The usual workarounds perturb the input to a trained AlphaFold 2 — subsample the MSA, or cluster it and predict from each cluster — which does yield a spread of structures, but much of it lands far from any observed conformation, and pLDDT does not reliably separate the real states from the artifacts.

UFConf converts the network itself into a generative model instead. It defines an explicit diffusion process over backbone geometry — Euclidean on each residue's position, SO(3) on its local frame rotation, with sidechain torsions carried alongside — and fine-tunes AlphaFold2 to denoise it. The architecture change is deliberate and small: the structural-template channel is removed, a channel carrying a time embedding for the noise scale is added, and the recycling channel that normally receives AlphaFold2's own intermediate structure takes the diffused one. Because the pretrained structure module already refines rough structures into real ones, most weights transfer straight to denoising, and every sample is an independent draw rather than a reweighted prediction.

UFConf was developed by Jiahao Fan, Ziyao Li, Eric Alcaide, Guolin Ke, Huaqing Huang and Weinan E at Peking University, DP Technology and the University of Barcelona, posted to bioRxiv in May 2024 and published in Journal of Chemical Information and Modeling that year. It sits beside AlphaFlow, which fine-tunes the same backbone under flow matching, and Distributional Graphormer.

#Key Features

• Structural diffusion on SE(3): Position and rotation of every residue frame diffuse independently, on the Euclidean and SO(3) manifolds, giving a well-defined score to learn instead of an ad hoc input perturbation.
• Hierarchical reweighting of training samples: PDB sequences are clustered into families, pairwise TM-scores build a graph within each, and Louvain community detection splits it into fold sub-clusters, with weights drawn from the logarithm of cluster sizes at both levels. Removing it degrades precision and recall.
• Continuous noise schedule: Time is continuous during training, so the reverse-diffusion step count is a run-time choice — one step already gives usable structures, and recall gains flatten past 30.
• Fixed-motif partial sampling: Random span-masking during training lets a user hold a domain or loop fixed and resample the rest, recovering accuracy where an unconstrained flexible linker dominates the diversity.
• Langevin and interpolation modes: The same checkpoint runs correlated MD-like sampling via Langevin dynamics at fixed diffusion time, and interpolates between two conformations by diffusing both and integrating back.

#Technical Details

Parameters were initialized from the AlphaFold-Multimer v3 weights and fine-tuned on all PDB chains released before 2022-04-30, with no self-distillation samples or structural templates; the implementation builds on Uni-Fold, DP Technology's open-source AlphaFold2 training platform, and inference MSAs come from the MMseqs2 server used by ColabFold. Fine-tuning took 16 NVIDIA A100 80GB GPUs about 84 hours at batch size 64 for 20,000 steps. Evaluation uses RAC-47, curated from PDB chains released between 2022-04-30 and 2024-03-29: 100%-identity sequence clusters of 2–10 chains, length 128–768, maximum intra-cluster RMSD above 2 Å, yielding 47 cases. Sampling 100 conformations per case, UFConf reached a median-of-best TM-score of 0.915/0.896 against the two most distant experimental structures, versus 0.916/0.876 for AlphaFlow and 0.866/0.808 for AlphaFold 2, and sampled successfully — best TM-score above 0.90 — in 19 of 47 cases against 18 for AlphaFlow, 11 for AlphaFold 3 and 2 for AF-cluster, with lower violation, C–N bond and clash losses than AlphaFlow's.

#Applications

The clearest use is drug discovery against targets whose druggable state is not the crystal one. In a blind test on the lysosomal transporter SLC15A4, the authors sampled 100 conformations, found pockets with FPocket, and docked the ligand feeblin with Uni-Mol Docking and Uni-Dock; the best pose by smoothed confidence landed within 2.0 Å ligand RMSD of the experimental complex, which conventional docking tools missed even when handed that structure, and a 100 ns MD run from the other conformer never reached the bound state. It also supports cryptic-pocket screening, allosteric studies and seeding targeted simulations.

#Impact

UFConf shows that AlphaFold2's pretrained weights absorb a diffusion objective with a handful of architectural edits, and that the harder problem is the training distribution, not the generative framework — hierarchical reweighting is what the ablation credits most. The tradeoff shows in the numbers: broader sampling improves recall but leaves mean MAT-P precision behind AlphaFold 2, AlphaFold 3, AlphaFlow and MSA-subsampling at depth 256. It inherits AlphaFold2's dependence on MSA search at inference and ships as one checkpoint, not a family of scales. Code lives on a branch of the Uni-Fold platform with the checkpoint on Zenodo, and follow-on work already uses UFConf to study fold-switching pathways in metamorphic proteins.