Protein kinases are implicated in many serious diseases, prompting the development of targeted therapies. However, due to the high degree of similarity among kinases, it is crucial to systematically assess the polypharmacology of kinase small molecule inhibitors.
In a study published in Nucleic Acids Research, a research team led by LI Xutong, ZHENG Mingyue and ZHANG Sulin from Shanghai Institute of Materia Medica (SIMM) of the Chinese Academy of Sciences developed an AI platform, KinomeMETA, for kinase multi-target activity screening. This platform significantly expanded the range of predictable kinase and supported customization by leveraging users' proprietary data. ZHENG Mingyue’s group has been dedicated to the research on small molecule kinase inhibitor screening for years, and has developed KinomeX, a platform that could predict kinase inhibition profile using multi-task deep neural network, promoting the development in AI technology for kinase activity prediction research.
In this study, the developed KinomeMETA platform utilized meta-learning method Reptile and was built upon the molecular feature extraction graph neural network AttentiveFP. It learned patterns of active compounds in different kinases across more than 610,000 active data points by harnessing the rapid adaptation capabilities of meta-learning.
The platform was constructed through meta-training and fine-tuning processes, comprising two core modules, Customization and Prediction. The Customization module introduced a novel functionality within the field of kinase activity prediction. It allowed users to build new kinase models or enhance predictive capabilities of existing ones using proprietary data, thereby bridging computational and experimental research. The Prediction module could predict inhibition probabilities of compounds against 661 wild-type kinases and clinically relevant kinase mutants.
The platform provided three result analysis functions, i.e., kinase selectivity analysis, molecular property evaluation, and identification of similar kinase inhibitors. These features could assist kinase drug researchers in conducting extended studies.
Researchers explored the efficiency of KinomeMETA when fine-tuned with experimental data. Building on preliminary findings from KinomeMETA in screening small molecule probes for the "dark" kinase PKMYT1, they utilized the platform's Customization procedure to enhance the default model. The virtual screen by KinomeMETA and experimental validation was conducted for in-house compounds. The results demonstrated that KinomeMETA rapidly improved predictive capabilities with a small number of active compounds, consistently outperforming molecular docking methods.
The KinomeMETA platform offers a Customization module for efficiently utilizing proprietary data and a Prediction module for kinome activity assessment. Through an iterative cycle of dry and wet experiments, it effectively addresses low-data scenarios in drug discovery, significantly accelerating the discovery of novel kinase small molecule probes and druggable kinase targets.
