MEGA-Protein: Full‑Process Protein Structure Prediction Tool Powered by MindSpore AI Framework

Protein structure prediction is a crucial step for understanding protein function, and for decades it has been regarded as one of the most important challenges in biophysics.

Historically, the massive number of possible conformations and the complexity of calculations meant that AI‑based prediction made little progress, and experimental methods such as cryo‑EM and X‑ray crystallography remained the primary approaches, often costing months and millions of yuan per protein.

The emergence of AlphaFold 2 brought a breakthrough, achieving near‑experimental accuracy and winning CASP14, an achievement described by Nature as an unprecedented advance.

In July 2021, DeepMind released the inference source code of AlphaFold 2. Huawei, together with Beijing Changping Laboratory, the BIOPIC center at Peking University, the School of Chemistry and Molecular Engineering, and the Shenzhen Bay Laboratory’s Gaoyi Qin group quickly reproduced and optimized the code, releasing a MindSpore‑based inference tool in November that improved efficiency by 2–3×.

Recently, Huawei and its partners launched the full‑process protein structure prediction tool MEGA‑Protein on the all‑scenario AI framework MindSpore.

AlphaFold 2's shortcomings

1. Orphan or synthetic sequences often lack multiple sequence alignments (MSA), causing a sharp drop in prediction accuracy for models like AlphaFold 2.

2. Standard MSA retrieval involves massive databases and long search times, hindering research progress.

AI MSA Engine

1. For proteins with low‑quality or few MSAs, integrating the AI MSA engine with AlphaFold 2 noticeably improves prediction quality.

2. End‑to‑end inference performance is greatly enhanced, and the trained AI MSA engine requires no additional database configuration.

3. The AI MSA engine serves as a universal pre‑training scheme that can be directly plugged into downstream structure prediction models.

MindSpore AI Framework

1. Deep integration with Ascend AI hardware and CANN provides high‑performance operators that fully exploit hardware compute power.

2. A multi‑stage parallel pipeline dramatically increases data processing throughput.

3. Support for large‑cluster efficient training and a three‑layer distributed programming paradigm significantly boost parallel program development efficiency.

Full‑process Protein Structure Prediction Tool – MEGA‑Protein

MEGA‑Protein incorporates the AI MSA engine, protein folding training and inference pipelines, structure scoring, and the PSP dataset, enabling high‑precision, high‑performance prediction of protein structures and functions. The AI MSA engine can maintain or even improve inference accuracy in low‑sample or zero‑sample MSA scenarios, effectively overcoming AlphaFold 2’s limitations, and the overall workflow achieves a 2–3× efficiency gain.

Event Announcement

The MSG Enterprise Tour and Hangzhou Developer Day will be held on April 27, co‑organized by the MindSpore community and the DataFun developer community. The afternoon features a MindSpore SPONGE session with Huawei experts sharing AI+life‑science integration methods and hands‑on experience. The morning session includes talks on smart medical, industrial operation, smart finance, and speech recognition, addressing enterprise AI transformation needs. Registration is available via the “Read Original” link.