Uni-Fold: An Open-Source Platform for Developing Protein Folding Models beyond AlphaFold

Key results

1. 01

   ShownUni-Fold Multimer, retrained from scratch, scores higher than AlphaFold-Multimer (v2 parameters; all 5 public models were evaluated) on RMSD and TM-Score for whole, recently released assemblies. AlphaFold-Multimer stays ahead on GDT-HA and Cα-LDDT, which the authors read as possibly better local accuracy.

   About 2% higher TM-Score on 162 assemblies (PDB, released 17 Jan to 14 Jul 2022, below 40% template identity); every method received the same features. On three example complexes: TM-score 0.984 vs 0.969 and Cα-LDDT 0.924 vs 0.968 (Uni-Fold vs AlphaFold-Multimer). The text does not reproduce the per-metric values from Table 2. · Abstract; Data; Baselines; Results

2. 02

   ShownRetraining the monomer model from scratch in PyTorch gave accuracy equal to or better than AlphaFold Model 2 and OpenFold Model 2 on recent single chains, which the authors take as evidence that the reimplementation is correct.

   Evaluated on 301 unique sequences (876 structures) with resolution below 3 Å, length 50 to 1024 and below 40% template identity; Table 2 is summarised as equivalent or better without numbers in the text. Accuracy dropped clearly when multimer models were run on monomer inputs. · 1 Introduction; Data; Baselines; Results

3. 03

   SuggestedEnd-to-end training is about 2.2 times faster than official AlphaFold under similar hardware and benchmark settings, and also ahead of the other recent baselines in Table 4. The multimer model reached better accuracy 1.9 times faster than AlphaFold-Multimer.

   Uni-Fold ran on 128 NVIDIA A100 GPUs, AlphaFold on 128 TPU cores and MEGA-Fold on 128 Ascend 910, all in bfloat16 (Table 4). OpenFold and HelixFold times were taken from their own reports, and the AlphaFold-Multimer finetuning steps are unknown. Under Uni-Fold's actual configuration, total training took about 4.11 days. · 3.3 Efficiency Benchmarks; Training

4. 04

   ShownPer Evoformer layer, Uni-Fold runs faster and uses less peak GPU memory than the baselines tested, including OpenFold, OpenFold with Low Memory Attention and FastFold, so longer sequences fit on the same GPU.

   One Evoformer layer on an A100 40GB with 512 MSAs and bfloat16: at chunk size 4, the maximum length is about 6720 residues for Uni-Fold vs about 5888 for OpenFold. In a forward-plus-backward test with 128 MSAs and 256 residues, Uni-Fold had the best time and memory (Table 5); OpenFoldLMA was slower and used more memory than plain OpenFold. · Inference; Evoformer Benchmark

5. 05

   SuggestedClipping gradients per sample, as AlphaFold does, converges better than the per-batch clipping used in most AlphaFold replicas such as OpenFold.

   Training Cα-LDDT of Uni-Fold Monomer with per-sample global-norm clipping at 0.1, compared with per-batch clipping at 0.1, 0.05 and 0.025: per-sample clipping was consistently better (Figure 3). No held-out test accuracy is reported for this comparison. · Per-Sample Gradient Clipping; 3.4 Effect of Per-Sample Gradient Clipping