Indian researchers develop new AI framework to aid discovery of next-generation drugs

New Delhi: Researchers from IIT Madras’ Wadhwani School of Data Science and Artificial Intelligence (WSAI) and The Ohio State University, U.S., have developed a breakthrough Artificial Intelligence framework that can rapidly generate drug-like molecules that are easier to synthesise in real-world laboratory settings.

The system promises to significantly cut down the early-stage timelines of drug development — currently a billion-dollar, decade-long process — and could play a crucial role in addressing drug resistance in cancer and infectious diseases.

The new framework, called ‘PURE’ (Policy-guided Unbiased REpresentations for Structure-Constrained Molecular Generation (SCMG), stands apart from existing molecule-generation AI tools that rely on rigid scoring mechanisms or statistical optimisation.

PURE was evaluated on widely accepted molecule-generation benchmarks, including QED (drug-likeness), DRD2 (dopamine receptor activity), and solubility tests. It delivered higher diversity and novelty in generated molecules and generated possible synthetic routes without ever being trained on those specific scoring metrics. This makes PURE a general-purpose AI engine for molecular discovery, capable of working across multiple disease and property objectives using a single trained model.

The findings were published in the reputed, peer-reviewed Journal of Cheminformatics (https://doi.org/10.1186/s13321-025-01090-5), an open-access research on how computational methods, data science, and machine learning are used to analyse and design chemical systems

The Authors of this Research are Abhor Gupta, Barathi Lenin, Rohit Batra, Prof. B. Ravindran, and Prof. Karthik Raman from the Robert Bosch Centre for Data Science and AI, Wadhwani School of Data Science and AI (WSAI), IIT Madras, and Prof. Srinivasan Parthasarathy, and Sean Current, Department of Computer Science and Engineering, The Ohio State University, US

Elaborating on this Research, Prof. B. Ravindran, Head, Wadhwani School of Data Science and AI (WSAI), IIT Madras, said, “Artificial intelligence is increasingly reshaping how we think about discovery itself, and drug design offers a compelling example of that transformation. What’s unique about PURE is the way it uses reinforcement learning, not just to optimize specific metrics, but to learn how molecules transform. By treating chemical design as a sequence of actions guided by real reaction rules, PURE moves us closer to AI systems that can reason through synthesis steps much like a chemist would.”

Further, Prof. Karthik Raman, WSAI, IIT Madras, said, “PURE adopts a novel approach to mapping chemical space, without being biased towards a specific metric – a common failing of existing tools. Further, it grounds the search of the vast chemical space for novel molecules in synthesisability, by generating molecules that are likely to be synthesisable in the lab, through a novel reaction rule-based approach.”

Prof. Srinivasan Parthasarathy, Department of Computer Science and Engineering, The Ohio State University, US, added, “PURE offers game-changing early-stage discovery benefits for pharmaceutical research, with the capability to identify alternative (more effective) drug candidates in the face of resistance and hepatotoxicity. It blends cutting-edge self-supervised learning with policy-based reinforcement learning, using template-driven molecular simulations to navigate the discrete molecular search space while mitigating metric leakage.   In addition to drug discovery, the PURE framework provides a promising foundation for accelerating the discovery of new materials, an important future research direction.”

PURE draws inspiration from how drugs are actually synthesised in labs, simulating step-by-step molecular changes using templates derived from real chemical reactions. By blending self-supervised learning — which lets the model learn patterns from data without labels — with a policy-based reinforcement learning setup, PURE explores the chemical landscape more naturally. This avoids the need for external scores during training, reducing biases and allowing the model to build its own sense of molecular similarity.