Generative model of multiple sequence alignments that denoises or invents homologs so AlphaFold2 can fold orphan and poor-MSA proteins.
No providers recorded yet. Browse all providers
AlphaFold 2 reads a raw multiple sequence alignment (MSA) rather than precomputed coevolution statistics. In principle, then, it does not need a deep alignment. In practice its accuracy falls steeply as the number of homologs drops. When the EvoGen authors fed AlphaFold2 shallow random subsamples of a target's MSA, the predicted structure depended on which sequences were drawn; with 64 or more, predictions converged. A few dozen homologs give AlphaFold2 a rugged, noisy folding landscape, and orphan proteins, viral proteins, toxins and de novo designs rarely have more than that.
EvoGen fixes the MSA instead. Given a query sequence and whatever homologs are available as context, this deep generative model samples new aligned sequences. It can calibrate an existing alignment, rewriting each sequence to strengthen the signals that are relevant to folding and to suppress noise. It can also augment an alignment with virtual homologs, or create one from the query sequence alone. Because its input and output are both MSAs, it slots in front of AlphaFold2 as a "prompt", and in calibration mode the structure predictor is used exactly as released.
EvoGen was developed by Jun Zhang, Yi Qin Gao and colleagues at Changping Laboratory and Peking University, with the MindSpore team at Huawei and co-authors at Shenzhen Bay Laboratory. The arXiv preprint appeared in August 2022, and the work was published in the Journal of Chemical Theory and Computation in 2023. Within the group's MindSPONGE toolkit it ships as MEGA-EvoGen, the MSA generator of the MEGA-Protein suite, alongside MEGA-Fold and MEGAFold-Multimer.
EvoGen is a hierarchical variational model with a U-shaped encoder–decoder. Each half stacks 12 Hyperformer blocks, which are Evoformer-style blocks with 256-dimensional sequence and 128-dimensional pair representations. They add rotary "hyper-attention" and a query-conditioning module. Three latent modules (64, 128 and 256 dimensions) connect the encoder and decoder. All residues of a sequence are generated at once rather than autoregressively. Training maximizes a variational lower bound on a multi-sequence conditional likelihood, and uses NVAE-style KL warm-up. The unsupervised stage ran for 150K steps on alignments for 447K PDB chains plus 648K UniRef50 sequences from the PSP dataset, each trimmed to a depth of 256 and kept only if at least 128 homologs remained. A further 50K steps on the PDB chains backpropagated AlphaFold2's FAPE, torsion, violation and confidence losses into EvoGen through a Gumbel-softmax straight-through estimator, with the EvoGen likelihood kept as a regularizer. AlphaFold2's Evoformer and confidence head were also unfrozen in this stage, because virtual MSAs otherwise led it to overestimate its own confidence. Training used 128 Ascend 910 NPUs.
The authors evaluated EvoGen with AlphaFold2 model 3, running without templates, on 84 CASP14 domains, 292 CAMEO targets and a curated set of 1,074 poor-MSA chains with fewer than 30 homologs. With only tens of input sequences, augmentation kept AlphaFold2's CASP14 and CAMEO accuracy near its full-MSA level. In the zero-shot setting, it turned more than half of the single-sequence misfolds into correct folds (TM-score above 0.5). On a de novo design set, every target reached the correct fold. Gains were largest for X-ray-determined viral and toxin proteins and smallest for NMR structures.
EvoGen is aimed at proteins that database search serves poorly: orphan and metagenomic sequences, viral proteins, toxins and designed proteins. For designed proteins, single-sequence prediction is often the only option. It also offers a way to explore conformational diversity: the authors' protocol calibrates subsampled alignments under several random seeds and clusters the confident AlphaFold2 outputs, which surfaced a candidate STAT1-bound state of Ebola VP24. The released pipeline runs in MindSpore on Ascend or GPUs.
EvoGen was an early demonstration that a pretrained generator can supply the evolutionary context a structure predictor lacks, rather than retraining the predictor for single-sequence input. The MindSPONGE documentation reports that MEGA-EvoGen placed first in the CAMEO-3D monthly ranking in July 2022. The benchmarks are computational only, the MindSPONGE release covers inference code and a checkpoint but not EvoGen's training code, and the alternative conformations remain unconfirmed experimentally.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.