UC student develops AI model for drug discovery and optimisation

Yanan Tian, a student in the dual doctoral programme jointly offered by the University of Coimbra’s Faculty of Sciences and Technology (FCTUC) and Macau Polytechnic University (MPU), has developed an AI model for drug discovery and optimisation.

The research, published in Nature Communications, was conducted under the supervision of Professors Joel P. Arrais, from the Department of Informatics Engineering at FCTUC, and Huanxiang Liu, from MPU. The work is part of the MPU‑UC Dual Doctoral Degree Programme, which promotes scientific cooperation and advanced training between Portugal and Macau.

Protein kinases are among the most important classes of therapeutic targets in biomedical research. Their potential lies in the central role they play in regulating key cellular processes, including proliferation, differentiation, and cell death. However, developing highly selective inhibitors remains a challenge because many kinases have very similar structures, and laboratory testing is expensive and time-consuming.

“This work introduces the MMCLKin model, a framework based on advanced AI methods, designed to predict with high accuracy and interpretability the activity and selectivity of kinase inhibitors, significantly speeding up the discovery and optimisation of new targeted drugs,” explains Yanan Tian.

MMCLKin combines geometric graph networks, language models for protein sequences, and multi-channel attention mechanisms to identify the key features of kinase–drug interactions. “The results show that the model outperforms existing methods in predicting inhibitor affinity and selectivity, even for unknown structures or mutated kinases,” says Joel P. Arrais.

According to the study authors, ADP‑Glo biological assays validated the model’s predictive power, demonstrating that five compounds suggested by MMCLKin effectively inhibit the LRRK2 G2019S mutation, associated with neurodegenerative diseases, with four of them active at nanomolar concentrations. These findings highlight MMCLKin’s potential as a tool to accelerate the development of targeted therapies, opening new possibilities for rational drug design with greater selectivity and clinical effectiveness.

“The approach represents a breakthrough in applying artificial intelligence to drug discovery, showing how next-generation computational models can reproduce complex biological processes in silico that would take years or even decades to study experimentally. MMCLKin demonstrates how AI-based models can simulate and understand molecular interactions in such detail that they can accurately predict inhibitor activity and selectivity,” the authors emphasise.

This capability allows researchers to quickly identify promising drug candidates, significantly reducing the time and cost of pharmaceutical research. Beyond its immediate impact, MMCLKin opens new research directions in kinase modelling, providing a unified framework to analyse structural, mutational, and functional patterns across the entire human kinase family.

“This type of approach could evolve into generalist models capable of anticipating the behaviour of new kinases — including those without known structures — and supporting the rational design of selective and personalised therapies in areas ranging from cancer to neurodegenerative diseases,” they conclude.

The scientific article “Enhancing Kinase‑Inhibitor Activity and Selectivity Prediction Through Contrastive Learning” is available here.