BioNeMo Framework：用于药物发现中 AI 模型开发的模块化高性能库
机器学习
2025-09-10 v5 生物大分子
摘要
编码生物学和化学的人工智能模型正在为高通量、高质量的计算机模拟药物开发开辟新途径。然而，它们的训练越来越依赖计算规模，最近的蛋白质语言模型（pLM）在数百个图形处理单元（GPU）上训练。我们引入 BioNeMo Framework 以促进跨数百 GPU 的计算生物学和化学 AI 模型训练。其模块化设计允许将数据加载器等单个组件集成到现有工作流中，并开放给社区贡献。我们通过蛋白质语言模型预训练和微调等用例详述了 BioNeMo Framework 的技术特性。在 256 块 NVIDIA A100 上，BioNeMo Framework 在 4.2 天内使用超过一万亿个 token 训练了一个 30 亿参数的基于 BERT 的 pLM。BioNeMo Framework 是开源的，供所有人免费使用。
引用
@article{arxiv.2411.10548,
title = {BioNeMo Framework: a modular, high-performance library for AI model development in drug discovery},
author = {Peter St. John and Dejun Lin and Polina Binder and Malcolm Greaves and Vega Shah and John St. John and Adrian Lange and Patrick Hsu and Rajesh Illango and Arvind Ramanathan and Anima Anandkumar and David H Brookes and Akosua Busia and Abhishaike Mahajan and Stephen Malina and Neha Prasad and Sam Sinai and Lindsay Edwards and Thomas Gaudelet and Cristian Regep and Martin Steinegger and Burkhard Rost and Alexander Brace and Kyle Hippe and Luca Naef and Keisuke Kamata and George Armstrong and Kevin Boyd and Zhonglin Cao and Han-Yi Chou and Simon Chu and Allan dos Santos Costa and Sajad Darabi and Eric Dawson and Kieran Didi and Cong Fu and Mario Geiger and Michelle Gill and Darren J Hsu and Gagan Kaushik and Maria Korshunova and Steven Kothen-Hill and Youhan Lee and Meng Liu and Micha Livne and Zachary McClure and Jonathan Mitchell and Alireza Moradzadeh and Ohad Mosafi and Youssef Nashed and Saee Paliwal and Yuxing Peng and Sara Rabhi and Farhad Ramezanghorbani and Danny Reidenbach and Camir Ricketts and Brian C Roland and Kushal Shah and Tyler Shimko and Hassan Sirelkhatim and Savitha Srinivasan and Abraham C Stern and Dorota Toczydlowska and Srimukh Prasad Veccham and Niccolò Alberto Elia Venanzi and Anton Vorontsov and Jared Wilber and Isabel Wilkinson and Wei Jing Wong and Eva Xue and Cory Ye and Xin Yu and Yang Zhang and Guoqing Zhou and Becca Zandstein and Alejandro Chacon and Prashant Sohani and Maximilian Stadler and Christian Hundt and Feiwen Zhu and Christian Dallago and Bruno Trentini and Emine Kucukbenli and Saee Paliwal and Timur Rvachov and Eddie Calleja and Johnny Israeli and Harry Clifford and Risto Haukioja and Nicholas Haemel and Kyle Tretina and Neha Tadimeti and Anthony B Costa},
journal= {arXiv preprint arXiv:2411.10548},
year = {2025}
}