Drug Discovery
Drug discovery is the process of finding a molecule that safely changes a disease-relevant biological target and then proving, through years of testing, that it helps patients; AI drug discovery applies machine learning to speed or improve parts of that process. The field's defining fact is that most of the cost and failure sits late, in clinical trials, while most machine-learning progress sits early, in choosing and designing molecules, so the honest question is how far better early choices translate into drugs that work in people.
The pipeline runs from target identification and validation through hit discovery, lead optimization, preclinical testing, three phases of human trials and regulatory review. Targets are proteins or pathways believed to drive a disease; hits are molecules that show some activity against them; lead optimization improves potency, selectivity and properties such as solubility, metabolism and toxicity; and preclinical studies in cells and animals precede Phase I (safety), Phase II (early efficacy) and Phase III (confirmation in large trials). The whole path commonly takes about a decade or more. Scannell and colleagues described in 2012 an apparent trend they called Eroom's law, Moore's law spelled backward: despite technological advances, the number of new drugs approved per billion dollars of inflation-adjusted research spending had fallen over decades.
Virtual screening uses computation to rank very large libraries of molecules so that only the best are synthesized and tested. Classical structure-based screening docks each molecule into a protein's binding pocket and scores the fit; ligand-based approaches predict activity from similarity to known actives; and learned models predict binding affinity or other properties from molecular graphs, fingerprints or protein-ligand structures. Structure predictions from AlphaFold widened the set of targets with a usable model, but docking scores remain imprecise, and a model trained on the public assay data can be biased by dataset artifacts so that apparent accuracy collapses on chemically new series. A well-known example of machine learning finding a new antibiotic is the 2020 Cell paper by Stokes and colleagues at MIT, in which a neural network trained on growth-inhibition data flagged the compound halicin, which was then shown to kill a range of bacteria in laboratory and mouse experiments; that is a validated discovery of a preclinical candidate, not an approved medicine.
Generative molecule design trains models to propose new structures that satisfy desired properties, using methods such as variational autoencoders, graph generators, reinforcement learning and diffusion models. A frequently cited demonstration is Zhavoronkov and colleagues in 2019 in Nature Biotechnology, where the company Insilico Medicine reported generating and experimentally testing candidate inhibitors of the kinase DDR1 within about three weeks of design, with one candidate showing activity in mice. The pitfalls are practical: generated molecules may be hard or impossible to synthesize, may exploit flaws in the scoring function rather than satisfy real biology, or may simply resemble known compounds, so chemists filter for synthetic accessibility and novelty before anything is made.
Clinical attrition is the sobering counterweight: roughly nine in ten drug candidates that enter clinical trials do not reach approval, and the leading causes are lack of efficacy and unmanageable toxicity. Better molecules do not help if the biological hypothesis about the target is wrong, which is the most common reason Phase II fails; a perfect binder to an irrelevant target is still a failed drug. For this reason, applying AI to target selection, patient stratification and trial design is arguably as important as molecule generation, though harder to validate because feedback takes years. Several companies have advanced AI-designed or AI-prioritized candidates into human trials, and some early candidates have been discontinued, but claims that AI has already produced a fully approved drug should be checked against regulatory records for the specific compound, and its role in the discovery stated precisely.
A frequent misconception is that AI will cut a decade-long process to months. Models can compress some early steps, such as screening and design iterations, but biology, safety testing in humans and regulatory requirements impose time that computation does not remove. The credible measure of success is not molecules generated or benchmark scores but the rate at which candidates survive clinical development compared with traditional approaches, and evidence on that rate is still accumulating.
Read drug discovery as a funnel in which AI mostly widens and sharpens the early, cheap stages, suggesting and ranking molecules faster than chemists alone, while the expensive verdict still comes from human trials where biology, not model accuracy, decides whether a candidate becomes a medicine.