AI2BMD

AI²BMD: AI-Powered Ab Initio Biomolecular Dynamics Simulation

Overview

AI²BMD is a cutting-edge program designed to simulate the dynamics of proteins with high accuracy using ab initio methods. This project offers various resources including the simulation software, datasets, and related materials to support the research community in molecular dynamics simulations.

AI²BMD Setup Guide

The AI²BMD software is available through a dedicated repository, and its deployment is simplified with Docker and Python. This makes it accessible even to those not deeply familiar with complex software setups. Users can download a single script to begin simulations without needing to clone the entire repository. Key simulation parameters can be adjusted as needed, such as temperature, the number of simulation steps, and GPU usage.

Running Simulation

To simulate molecular dynamics with AI²BMD, users can follow straightforward commands to initialize the software and input data. For example, in the case of the Chignolin protein, users merely need to download the protein structure and execute a script. The results will be stored in a designated directory, containing detailed trajectory information for analysis.

Datasets

AI²BMD provides access to extensive datasets critical for protein dynamics research. Two core datasets are available:

1. Protein Unit Dataset: Offers a comprehensive range of dipeptide conformations.
2. AIMD-Chig Dataset: This dataset features 2 million conformations of the Chignolin protein, complete with potential energy and atomic force calculations at a quantum mechanical level.

System Requirements

For optimal performance, AI²BMD requires specific hardware and software configurations:

• Hardware: A multicore CPU, significant memory (32+ GB), and a CUDA-enabled GPU.
• Tested GPUs: A100, V100, RTX A6000, Titan RTX.
• Operating Systems: Compatible with Ubuntu 20.04 and ArchLinux, using Docker.

Related Research

AI²BMD is at the forefront of several exciting areas in molecular dynamics and machine learning:

• ViSNet: A graph neural network that captures molecular geometries efficiently, balancing computational cost and data utility.
• Geoformer: A Transformer model that predicts molecular properties by taking into account atomic environments.
• MLFF: A method to improve machine learning force fields through detailed force metrics, enhancing robustness and accuracy in simulations.
• MSMs: Use of stochastic parameters to improve the modeling of protein dynamics under Markov state models, increasing prediction robustness.

Citation and License

Researchers contributing to AI²BMD include leaders in the field, and the project is shared under the MIT license, encouraging further exploration and application.

Disclaimer and Contacts

AI²BMD is a research initiative by Microsoft, not a commercial product. For inquiries or feedback, the AI2BMD team can be reached via email, with the primary contacts being specialists such as Tong Wang and Yatao Li.