Developing a new drug remains one of the industry's longest and most expensive bets: roughly 2.6 billion dollars and nearly ten years, for one candidate in ten that clears clinical trials. Artificial intelligence promises to shorten the first half of that journey, the discovery part. Between AlphaFold, awarded the 2024 Nobel Prize in Chemistry, and the first drug whose target and molecule were found by AI now in phase 2, the topic has left the lab. For Belgium, Europe's leading pharmaceutical hub per capita, the stakes are industrial.
A drug starts from tens of thousands of molecules screened in the lab. Only a few clear preclinical testing, fewer still enter clinical trials, and a handful obtain marketing approval. The average cost to carry a candidate to market, accounting for every failure, is estimated at about 2.6 billion dollars by the Tufts Center for the Study of Drug Development, over a ten to fifteen year cycle.
The bottleneck is failure. According to data from BIO (the biotech trade body) compiled with Informa, close to 9 of every 10 candidates that enter clinical trials never reach approval. Each late failure, in phase 2 or 3, costs years and hundreds of millions. It is precisely this first half of the journey, finding a good target and a good molecule, that AI seeks to make faster and less uncertain.
AI does not invent a drug on its own. It steps in at specific stages of research, where a vast space of possibilities must be explored and leads ranked. Four uses dominate today.
Models sift through the literature, omics data and structures to propose plausible biological targets for a disease.
novel targetsGenerative chemistry proposes molecules with the desired properties, synthesizing dozens of candidates rather than thousands.
fewer synthesesModels of absorption, toxicity and efficacy (ADMET) rule out candidates bound to fail early, before the lab.
earlier failuresPatient selection, site choice and protocol design from real-world data, for better-targeted trials.
targeted trialsFor fifty years, predicting the three-dimensional shape of a protein from its sequence remained an open problem. Yet that shape governs function, and therefore how a molecule can bind to it. In 2021, AlphaFold 2, developed by Google DeepMind, predicted the structure of nearly all 200 million known proteins and published those structures in open access, through a database hosted by EMBL-EBI. In May 2024, AlphaFold 3 extended prediction to complexes between proteins and DNA, RNA or small molecules.
Recognition came fast: the 2024 Nobel Prize in Chemistry honored Demis Hassabis and John Jumper for structure prediction, and David Baker for protein design. Knowing the structure of a target before any experiment shortens the design of molecules that bind to it, even though a predicted structure remains a hypothesis to validate in the lab.
Nearly 200 million protein structures predicted and made freely available. A resource used by millions of researchers worldwide.
Prediction of interactions between proteins and DNA, RNA, ligands and ions, at the heart of drug design.
The most advanced case comes from Insilico Medicine. Its molecule rentosertib (ISM001-055), an inhibitor of the TNIK protein for idiopathic pulmonary fibrosis, is the first drug whose target and molecule were both identified by generative AI. Across its 22 nominated candidates between 2021 and 2024, the company says it moved from project start to preclinical candidate in 12 to 18 months, against 2.5 to 4 years usually, synthesizing only 60 to 200 molecules per project.
Above all, rentosertib delivered clinical results. The phase 2a trial, published in Nature Medicine on 3 June 2025, covered 71 patients across 22 sites in China. The group treated with 60 mg per day gained 98.4 mL of forced vital capacity over twelve weeks, against a decline of 20.3 mL on placebo, a benchmark metric for this disease. Insilico is now discussing a phase 3 with regulators. The company itself notes that the sample is limited and will need confirmation in larger cohorts.
The market for AI applied to drug discovery is still young and estimates vary widely depending on scope, from under 3 to over 20 billion dollars for 2026. Precedence Research places this market at around 6.9 billion dollars in 2025 and 7.6 billion in 2026, with annual growth of roughly 10% through 2035. Beyond the figure, the clearest signal comes from partnerships: large pharma companies are striking deals with AI-focused biotechs, which validates the approach without yet proving its final return.
The measured advantage sits at the start of the journey. According to the Boston Consulting Group analysis, AI-discovered molecules pass phase 1 in 80 to 90% of cases, against 40 to 65% historically. That phase mainly tests safety: a higher rate suggests AI designs better-tolerated and better-targeted molecules, which reduces early failures and their associated costs.
Clinical proof remains partial. In phase 2, which tests efficacy, AI-discovered molecules succeed at about 40%, in line with the historical average (BCG): AI accelerates discovery, it does not yet guarantee efficacy. Failures exist: several high-profile AI-discovered candidates failed in phase 2, and some pioneering biotechs have scaled back their ambition. Data quality is decisive: a model is only as good as the biological data feeding it. Regulation does not change: the EMA and FDA require the same evidence, and health data remains governed by the GDPR and the EU AI Act.
Belgium is a pharmaceutical heavyweight: it ranks at the top of the European Union for pharma R&D investment relative to population, and exports about 216 million euros of biopharmaceutical products every day (pharma.be). But 2024 sent warning signals, with exports down 6.8% and patent applications down 15.3%. In this context, R&D efficiency becomes a matter of competitiveness, not a laboratory topic.
UCB, headquartered in Brussels, and Janssen (J&J group), with much of its R&D in Beerse, are among the major players. The 2024 downturn is a reminder that leadership is not a given.
Shortening discovery, ruling out failures earlier and better targeting trials can defend this position, provided there is investment in clean data and skills.
A hard target, a screening bottleneck, a trial cohort to optimize. AI serves a measurable goal, not the other way around.
Models are only as good as the available biological and chemical data: clean, structured and usable.
Specialized platforms, academic collaborations or in-house tools, depending on maturity and the skills in place.
Health data under the GDPR, model traceability, EMA and FDA regulatory dossier, documented governance.
Molderez Consult helps Belgian health and pharma players frame an AI use case, audit their data and deploy in line with the GDPR and regulatory requirements, on infrastructure they control.
Discuss my use case