A typical AI Drug Discovery pipeline involves: (1) target identification โ models mine omics data and literature to pinpoint disease-associated proteins/genes; (2) target-structure prediction โ tools like AlphaFold model the protein's 3D structure; (3) compound generation and screening โ generative models (e.g. variational autoencoders, diffusion models, graph networks) design molecules de novo while virtual docking/QSAR scores their fit to the target; (4) lead optimisation โ models predict potency, selectivity and ADMET profile to guide iterative refinement; (5) experimental validation โ top candidates are synthesised and tested, with results fed back as training data (active-learning loop).
Traditional drug discovery is slow, costly and inefficient: searching an enormous chemical space (on the order of 10^60 possible molecules), identifying the right biological target and optimising a compound by trial and error takes years, and most candidates fail in preclinical and clinical testing. AI shrinks that search space, predicts binding affinity, selectivity and ADMET properties, and prioritises the most promising compounds before they reach the lab bench.
Deep-learning models predicting a protein's 3D structure from its amino-acid sequence, providing the structural target for drug design.
Models (VAEs, GANs, diffusion, graph-based) that design novel chemical structures with desired properties.
Computational scoring of how well compounds fit the target's binding pocket, narrowing millions of compounds to the most promising ones.
Models predicting absorption, distribution, metabolism, excretion and toxicity, reducing late-stage failures.
High-quality, consistent bioactivity data is scarce and noisy; models easily learn artefacts instead of real biology.
Models fail on novel chemical classes and at 'activity cliffs', where a small structural change drastically alters activity.
A deep-neural-network team wins the Kaggle QSAR challenge, showing deep learning's edge over classical methods for predicting compound activity.
DeepMind presents AlphaFold 1, winning the CASP13 protein-structure-prediction competition.
AlphaFold2 wins CASP14 scoring above 90 on the global distance test, effectively solving the protein-structure-prediction problem.
DSP-1181 (Exscientia and Sumitomo Dainippon Pharma) enters Phase I trials as one of the first molecules designed using AI.
Insilico Medicine announces INS018_055 for idiopathic pulmonary fibrosis โ a candidate with both an AI-discovered target and an AI-designed molecule.
DeepMind announces AlphaFold 3, predicting complexes of proteins with ligands, DNA and RNA; Hassabis and Jumper receive the 2024 Nobel Prize in Chemistry for protein-structure prediction.
