Protein complex structure prediction model: an AlphaFold-Multimer reimplementation in MindSpore, retrained from scratch on Ascend chips.
No providers recorded yet. Browse all providers
When DeepMind released AlphaFold-Multimer, it released inference code and trained parameters, but anyone who wanted a second, independently trained set of complex-prediction weights, on different hardware or a different framework, had to reproduce the training run themselves. That is expensive: the network has to learn inter-chain pairing, chain-aware positional encoding and permutation-invariant losses from scratch, not just the monomer fold. MEGAFold-Multimer is one such reproduction. It reimplements the AlphaFold-Multimer architecture in Huawei's MindSpore framework and retrains it from random initialization on Ascend AI accelerators.
The model was trained by researchers at Huawei working with the Yi Qin Gao group at Changping Laboratory and Peking University. It is the multimer member of the MEGA family. The same group's earlier MEGA-Protein toolkit contains MEGA-Fold, a MindSpore monomer predictor trained on the group's PSP dataset, which topped the CAMEO-3D monthly ranking in April 2022. MEGAFold-Multimer shares that training corpus. It has no paper of its own and is described in the Methods of the GRASP paper, where it supplies one of GRASP's five starting checkpoints.
In spirit it resembles OpenFold, which retrained AlphaFold 2 in PyTorch. Having a from-scratch replica gives downstream fine-tuning a second starting point whose errors are not simply copies of the original checkpoint's.
MEGAFold-Multimer uses the AlphaFold-Multimer network: an Evoformer stack over paired and unpaired MSA features and templates, followed by a structure module with invariant point attention, trained with a crop size of 384 residues. Training data came from PSP, which contains 570k experimental structures released before 8 October 2021, with sequences, MSAs and templates, and a 760k-entry distillation set of predicted structures. Complex samples were built by pairing single-chain MSAs that share a PDB ID, using the AlphaFold-Multimer pairing protocol. Training ran in two stages over 17 days. The initial phase ran 33,000 steps over 12 days, warming the learning rate to 3e-4 over 1,000 steps and then decaying it by cosine. It sampled distillation, experimental and multimer data at 5:2:3. The second phase ran 13,200 steps over 5 days from a learning rate of 5e-5 with the ratio shifted to 3:2:5. The authors evaluated the model on 172 non-redundant complexes released between 20 July 2022 and 1 March 2023 and report accuracy comparable to AlphaFold-Multimer on that set. GRASP then fine-tuned this checkpoint for 22,000 steps on 64 Ascend 910A chips to produce its fifth inference model, MS-22k.
MEGAFold-Multimer does the same job as AlphaFold-Multimer: it predicts how several protein chains assemble, from sequences and their MSAs, for heterodimers, homo-oligomers and larger complexes. Its most concrete use so far is as a pretrained base for fine-tuning. GRASP adds restraint-aware inputs for crosslinking, covalent labeling, chemical shift perturbation and deep mutational scanning data on top of it. For groups working on Huawei Ascend hardware, it is an example of an AlphaFold-Multimer-class backbone trained natively in MindSpore.
MEGAFold-Multimer's main contribution is independence, not accuracy: a second AlphaFold-Multimer-class checkpoint trained on different hardware and software, which GRASP uses to diversify its ensemble. The model is documented only inside the GRASP paper (bioRxiv September 2024, Nature Methods 2025), and its evaluation is a single held-out set reported as parity with AlphaFold-Multimer rather than an improvement. The public GRASP weight deposit contains the fine-tuned MS-22k checkpoint, not the base MEGAFold-Multimer weights. The PSP training data are publicly distributed. Two Huawei authors on the GRASP paper, Min Wang and Yingtong Hu, trained and evaluated the model. Min Wang is also an author of the PSP and MEGA-Fold work, which ties the multimer model to that earlier monomer lineage.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.