Editor's note: AI drug research and development is one of the important directions for the future application of artificial intelligence. Since the first outbreak of the new coronavirus (SARS-CoV-2), the research and development of small molecule drugs for the new coronavirus has attracted much attention, and the first AI drug development algorithm competition held recently focused on this. In the competition, the team from the Scientific Intelligence Center of Microsoft Research achieved excellent results and won the championship with their innovative AI model systems AI2BMD and ViSNet.
Recently, the results of the first AI Drug Research and Development Algorithm Competition, co-organized by Tsinghua University School of Pharmacy, Baidu Paddle, Baidu Smart Cloud and Lingang Laboratory, were announced. A team from the Scientific Intelligence Center of Microsoft Research used the quantum precision power developed by The scientific simulation system AI2BMD and the universal molecular three-dimensional structure network ViSNet ranked first in the preliminary round, semi-finals, and finals, and won the overall championship of the competition, demonstrating the application potential of AI in promoting drug research and development.
The Microsoft Research Scientific Intelligence Center team won the first AI drug development algorithm competition
This competition is fully supported by industry authoritative organizations such as the Chinese Pharmaceutical Association, with a total of 878 teams from around the world participating. As a global technological innovation event, this competition focuses on the research and development of small molecule drugs for the new coronavirus (SARS-CoV-2). In fact, since the first outbreak of the new coronavirus, the development of small molecule drugs for the new coronavirus has attracted much attention. To combat the spread of COVID-19, it is crucial to have a deep understanding of the virus's replication and infection mechanisms. Among them, the new coronavirus main protease (Mpro) is a key enzyme, responsible for cutting the protein precursors produced by the virus during the infection process and promoting virus replication. Therefore, the main protease is a potential therapeutic target. Inhibiting its activity can effectively interfere with virus replication. process, providing a breakthrough for treatment methods. Therefore, in the preliminary stage of this competition, contestants need to use deep learning, molecular docking and other methods to conduct modeling to predict the probability of small molecules inhibiting the activity of the main protease. The semi-finals will focus on the probability of small molecules inhibiting the replication of the new coronavirus on Caco cells. .
In the preliminary round of drug prediction for the main protease of the new coronavirus, faced with the problem that commonly used molecular docking software cannot effectively distinguish between positive and negative samples and the binding free energy of target proteins, the Microsoft Research Scientific Intelligence Center team used the newly developed AI2BMD simulation system [ 1], significantly improving drug prediction accuracy. The AI2BMD simulation system achieves accurate calculations of various protein energies and forces over 10,000 atoms and has wide applicability. Compared with density functional theory (DFT), the calculation time of the AI2BMD simulation system is reduced by several orders of magnitude. With dynamics simulations of hundreds of nanoseconds, AI2BMD demonstrates its outstanding capabilities in exploring protein conformational space, predicting NMR experimental data, and simulating protein folding processes. Compared with traditional molecular docking and classical dynamics simulation methods, the AI2BMD system also has obvious advantages in calculating binding free energy.
AI2BMD simulation system paper link: https://www.biorxiv.org/content/10.1101/2023.07.12.548519v1
In the semi-finals, the team used the self-developed molecular modeling geometric deep learning model ViSNet [2] to conduct representation learning on compound molecules. ViSNet is a machine learning potential function in the AI2BMD simulation system. As an equivariant geometrically enhanced graph neural network, ViSNet can extract geometric features (distances, angles, dihedral angles, etc.) with the complexity of linear computation. ViSNet outperforms other state-of-the-art methods on multiple molecular dynamics benchmarks, including MD17, rMD17, and MD22, while also achieving excellent quantum chemical property predictions on QM9 and Molecule3D datasets.
In the semi-finals, the team also used the independently developed first self-developed protein macromolecule full conformation space data set AIMD-Chig [3] and the small molecule public data set OGB to pre-train the three-dimensional structure representation of proteins and small molecules respectively, and then used Multi-task learning fine-tunes the model. This method not only achieved the best prediction accuracy, but also led the second-place team in the competition by a large margin. In the final defense, the COVID-19 drug prediction algorithm solution of the Microsoft Research Scientific Intelligence Center team achieved an excellent score of 99.60 points in total, which was significantly superior to the final scores of 90.76 points for the runner-up and 85.31 points for the third runner-up in the competition.
The new coronavirus drug prediction algorithm proposed by the Microsoft Research Scientific Intelligence Center team
Through this drug research and development competition, the quantum precision dynamics simulation system AI2BMD developed by the Scientific Intelligence Center of Microsoft Research demonstrated excellent practical application potential. In the future, AI2BMD is expected to conduct more extensive exploration in the molecular mechanism explanation of life activities, drug design, enzyme catalysis, etc., and help accelerate the development of AI drug research and development.
[1] Wang T, He X, Li M, et al. AI2BMD: efficient characterization of protein dynamics with ab initio accuracy. bioRxiv, 2023: 2023.07. 12.548519.
[2] Wang Y, Li S, Wang T, et al. ViSNet: a scalable and accurate geometric deep learning potential for molecular dynamics simulation. arXiv preprint arXiv:2210.16518, 2022.
[3] Wang T, He X, Li M, et al. AIMD-Chig: Exploring the conformational space of a 166-atom protein Chignolin with ab initio molecular dynamics. Sci Data 10, 549 (2023).