Protein structure prediction model: AlphaFold 2 retrained from scratch in MindSpore on the million-sample PSP dataset, ranked first on CAMEO-3D.
No providers recorded yet. Browse all providers
When DeepMind released AlphaFold 2, it published inference code and trained parameters but not its training code or training set, and the public sequence-structure datasets available at the time were far too small to retrain it. The trRosetta set, the most widely used, held about 15,000 chains. A model trained on it plateaus early: in the authors' comparison, AlphaFold 2 trained on trRosetta levelled off at a CASP14 TM-score near 0.7, while the same network trained on a million-sample corpus reached 0.8 over the same number of samples. MEGA-Fold came out of closing that gap. It is AlphaFold 2 retrained from random initialization in Huawei's MindSpore framework on the purpose-built PSP dataset, with changes to the loss and the sampling scheme that make that training run converge and stay stable.
MEGA-Fold was developed by the Yi Qin Gao group at Changping Laboratory and Peking University with the MindSpore team at Huawei, and co-authors at Shenzhen Bay Laboratory, HKUST and Peng Cheng Laboratory. It is documented in the June 2022 PSP preprint, which calls it only "our model"; the group's MindSPONGE release names it MEGA-Fold (also written MEGAFold) and ships it as the structure predictor in MEGA-Protein, alongside the MSA generator MEGA-EvoGen and the quality scorer MEGA-Assessment.
It belongs to the same family of independent AlphaFold 2 reproductions as OpenFold, which retrained the network in PyTorch. Its multi-chain successor is MEGAFold-Multimer, trained on the same PSP corpus.
MEGA-Fold keeps the AlphaFold 2 network unchanged: an Evoformer over MSA and pair representations, template embedding, and a structure module with invariant point attention. The released checkpoint is 356 MB. Training ran on 128 Ascend 910 chips at batch size 1 with dropout removed. Initial training used a 256-residue crop, 128 MSA clusters, 1,024 extra sequences and four templates, and ran over 9.6 million samples in 11 days. Two fine-tuned variants followed over 3.0 million samples in 7 days each, with a 384-residue crop and structural violation loss switched on: model 1.1 with templates and model 1.2 without them. PSP's experimental set was built from PDB chains released up to 13 October 2021, with MSAs from MMseqs2 searches against UniClust30 and a ColabFold environmental database. The distillation set was selected from UniRef50 for diversity and dissimilarity to PDB70, then folded with AlphaFold 2.
On the 87 CASP14 targets the model reached an average TM-score of 0.86, the same as DeepMind's released checkpoint, and an all-atom lDDT of 0.80 on a 513-target validation set dated after the training cutoff, again matching the original. In the blind CAMEO-3D evaluation from 25 March to 16 April 2022, model 1.1 ranked first with an average lDDT of 87.3 and TM-score of 90.6, ahead of every compared server, including AlphaFold 2 baselines and RoseTTAFold.
MEGA-Fold predicts the three-dimensional structure of a single protein chain from its sequence and an MSA, with or without templates, and is the folding step in the MEGA-Protein pipeline. There, MEGA-EvoGen can generate or enrich the MSA for orphan and designed sequences that have few homologs, and MEGA-Assessment scores the predicted model per residue. The MindSPONGE documentation reports inference up to 2,048 residues on a 32 GB Ascend 910, or 3,072 in pipeline mode. For groups that want to retrain or fine-tune an AlphaFold 2-class model on their own data, particularly on Ascend hardware, the released training code gives a working recipe.
MEGA-Fold's main contribution is reproducibility. Alongside the PSP dataset, it gave the field a public, stage-by-stage training recipe for AlphaFold 2 at a time when DeepMind had released only inference code, and showed that an independent run on non-GPU hardware could match the original checkpoint. The group built MEGAFold-Multimer on the same corpus, and that multimer checkpoint in turn became a starting point for GRASP's restraint-aware models. The evidence for MEGA-Fold is a single preprint, not peer-reviewed, and its accuracy claims are parity with AlphaFold 2 rather than an improvement over it. Its CAMEO lead covers one month and 60 targets.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.