On December 22, 2024, the Final of the 2nd AI Drug Discovery Algorithm Competition was successfully held at the School of Pharmaceutical Sciences, Tsinghua University.The main focus of this competition was predicting the quantum chemical properties of sesquiterpene molecules. It aimed to address the challenge of model generalization in the field of molecular property prediction. Participants were required to train models on a limited training dataset and evaluate their performance on a larger testing dataset.The competition encouraged contestants to design molecular representation learning algorithms with strong generalization abilities, capable of adapting to various types of molecules. Such breakthroughs are expected to become a significant advancement in the field of molecular representation learning, further promoting the application and development of AI in drug discovery.
The competition officially launched on July 11, 2024, and attracted 656 participants from 505 teams. After nearly four months of preliminary and semi-final rounds, a total of seven teams advanced to the grand finale, including teams from Tongji University, the Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Fudan University, and Nanjing Bingjian Information Technology Co., Ltd.
After an exciting afternoon of presentations, the TJ-AI4S team from Tongji University won the first prize of the competition. The team innovatively proposed a feature and label extension strategy, Stretching Features and Labels (SFL), for molecular representation, which diversifies the construction of out-of-distribution pseudo-data for both structure and property without disrupting the intrinsic semantic meaning of drug molecules. This method is simple, efficient, interpretable, and highly generalizable, seamlessly integrating into various pre-trained molecular models. The team won first place in both the semi-finals and finals.
The “Fight and Learn（打不过就学）” team from the Shanghai Institute of Materia Medica, Chinese Academy of Sciences, and the “Xiao Lao Zheng（小老正）” team from Fudan University won the second prize. The third prize was awarded to the teams “paipai” “I Only Slack Off, Not Work Hard（我只摸鱼不划水）” “Computers Are Also Chemistry（计算机也是化学）” and “Deep Metaphysics（深度玄学）”. Among them, the “Fight and Learn” team designed multiple pre-training tasks based on the VisNet model, allowing the model to learn basic physical knowledge. The team also used group information to construct invariant subgraphs and enhanced the model’s generalization ability through an encoding-then-separation strategy. The “Xiao Lao Zheng” team adopted an adaptive weighted average scheme during inference, dynamically adjusting the variance of prediction values based on the magnitude of model predictions, thereby improving the model’s ability to predict out-of-distribution data. The “I Only Slack Off, Not Work Hard” team used a pairwise learning strategy to predict molecular properties by predicting the property differences between sample pairs, significantly enhancing the model’s ability to predict YOOD molecules.
Other teams made innovations in molecular conformation optimization, molecular feature extraction, and model training strategies, achieving more accurate molecular prediction models compared to baseline models and other competitors.
Throughout the competition, the School of Pharmaceutical Sciences at Tsinghua University, Baidu, and Intel provided tremendous support. This competition not only fostered cooperation and development between AI and pharmaceutical fields but also introduced new ideas and methods for solving key challenges in drug discovery. It made a positive contribution to cultivating more interdisciplinary talents in AI and biomedicine.
For detail: https://mp.weixin.qq.com/s/GK18b-0SJZx_hYhYY672Kw
The judges took a group photo with all the participants and volunteers.