TAIPEI — The inaugural Global AI Drug R&D Algorithm Competition finals, a prominent event that attracted interest from both academic and business circles, were recently held at Tsinghua University. The joint team of IceKredit and Nanjing University emerged as a standout force, winning the coveted third place, after a hard competition amongst 878 teams from universities, research institutes, and enterprises around the world.
This innovative competition received strong support from the Chinese Pharmaceutical Association and other important organizations. It was co-sponsored by Baidu PaddlePaddle, the Tsinghua University School of Pharmacy, and Lingang Laboratory. The competition’s outstanding jury was composed by a group of eminent professors and experts in the biopharmaceutical field who contributed their knowledge.
The competition attracted 1105 participants from 878 teams across the globe, collecting 6080 algorithmic submissions in total. The partnership between IceKredit and Nanjing University produced a strong competitor. The final confrontation included in-person defenses and in-depth conversations about the tactics for answering competition questions, fundamental ideas, data manipulation and analytics, and algorithmic solutions. The partnership between IceKredit and Nanjing University emerged victorious from this test.
The focus of the competition was on urgent issues, such as the creation of small molecule drugs to combat the new coronavirus. The main objective was to inspire people to use artificial intelligence to its fullest capacity in revealing new virus treatments. To predict and assess interactions between tiny compounds and significant proteases, this involved using molecular docking and deep learning approaches. Participants also examined these compounds’ capacity to prevent viral multiplication in cells, essentially looking for good therapeutic possibilities. This daring project aimed to establish a solid foundation for future disease prevention and treatment while elevating innovation in the field of medication research and development.
The preliminary rounds marked the beginning of the glorious collaboration between IceKredit and the Christopher J. Butch research team at Nanjing University. The researchers used both traditional machine learning methods including Bayesian docking, SVM, LightGBM, and GBDT in this case, as well as cutting-edge deep neural network models like Transformer-CNN, GCN, and D-MPNN. To predict enzyme activity, the team experimented with a variety of molecular representations, including graph characteristics, 3D molecular conformation data, and molecular characterization techniques like Morgan Fingerprinting. Notably, the SVM model developed with Morgan fingerprinting stood out among the competition and successfully predicted enzyme activity.
Participants in the semi-final round struggled to estimate molecular activity in Caco2 cell tests, which was a challenging job. The IceKredit-Nanjing University cooperation rose to the occasion and developed a method of feature fusion, skillfully upgrading the GEM baseline model. The outcome was a more comprehensive molecular data representation, which improved the model’s capability for categorization and prediction.
Since March 2022, IceKredit and Nanjing University have been working together to explore the potential uses of AI in the medical industry. Their efforts center on computer-aided drug molecular design approaches, fusing conventional chemistry and biology labs with artificial intelligence, molecular dynamics simulation, and computational biology. The result is an expedited search for possible medication compounds.
This alliance has produced notable results in just one year. This new approach to drug development, which combines reinforcement learning and deep generative models, has shown tremendous promise. The technology creates novel compounds that outperform docking scores of comparable sized molecules by 10–20% while accelerating the process 130 times faster than traditional docking methods. It uses approximations of docking scores predicted by a Bayesian regression model. The unique method holds out the possibility of quickly identifying novel chemical molecular structures with potential medicinal uses.
