Artificial intelligence is fundamentally changing how drug development is done. It’s moving from slide decks into standard operating procedures. Alphabet’s Isomorphic Labs began 2024 by inking discovery deals with Eli Lilly and Novartis worth “nearly $3 billion,” then followed up with AlphaFold 3, a model that predicts not just protein shapes but how proteins, DNA, RNA and small-molecule ligands interact—exactly the kind of physics that governs whether a drug actually binds. In parallel, regulators like the FDA have started to publish rules for when and how AI-generated evidence can be used in filings. The result is a pipeline that’s faster at the front end, more instrumented in the lab, and increasingly algorithmic in clinical development.

On the design bench, the leap isn’t only structural prediction; it’s generative. EvolutionaryScale’s ESM3 model can design proteins to a prompt—a capability its creators dramatized by producing a novel green fluorescent protein that natural evolution would have taken hundreds of millions of years to find. Open-source efforts are hardening too: OpenFold2 has been packaged by Nvidia into a production-ready service, making high-accuracy structure prediction a callable API rather than a research project. The upshot: medicinal chemists increasingly begin with an AI candidate that is “drug-like” by construction, then iterate with models that learn from each new assay.

The wet lab is becoming software-addressable. A growing slice of preclinical work is handled by “self-driving labs”—robotic facilities orchestrated by scheduling software and active-learning loops. Emerald Cloud Lab and Strateos let scientists specify experiments remotely and receive machine-readable data back, with reviewers noting that such autonomy can compress timelines and tame variability. The punchline for drug makers is not just speed; it’s reproducibility, which makes downstream regulatory audits easier.

Early clinical proofs are arriving. Insilico Medicine reported positive Phase 2a data for ISM001-055, a first-in-class fibrosis program it says was generated by a generative-AI design loop—evidence that AI-originated assets can clear the translational hurdle into patients. Meanwhile, platform companies are scaling industrial “phenomics” to map cell responses across millions of perturbations; Recursion, which completed its merger with Exscientia in late 2024, has since logged partnership milestones with Sanofi and others. These are still early innings, but they mark a shift from isolated demos to pipelines with multiple clinical shots on goal.

Clinical development is where AI may save the most money. Sponsors are using real-world data to pre-qualify sites and patients, trimming screen-fail rates and smoothing enrollment. Vendors such as Medidata and TriNetX now sell AI layers for feasibility, site selection and patient-finding, while startups like Unlearn build “digital twins” that can reduce the size of control arms by modeling what would have happened to enrolled patients without therapy. Nature Biotechnology argued this year that synthetic controls will likely earn broader near-term adoption than full patient digital twins—partly because they dovetail with the data regulators already accept.

Regulators are no longer on the sidelines. In January 2025 the U.S. FDA issued draft guidance laying out a risk-based credibility framework for AI models used to support safety, efficacy and quality decisions—a practical playbook that nudges sponsors to document data provenance, model validation and monitoring. Europe’s EMA finalized a reflection paper in September 2024 spanning discovery through manufacturing, and the UK MHRA in June 2025 joined a new international network focused on safe AI use in healthcare. The common thread: AI is welcome, but only if sponsors can show their math.

Hardware matters, and the stack is professionalizing. Nvidia has stitched together model hubs and sovereign supercomputing agreements with pharmas such as Novo Nordisk, packaging cheminformatics and structural biology models into services that R&D teams can actually deploy. That infrastructure reduces the distance from an academic preprint to a validated internal tool—and it brings procurement and IT into the AI conversation earlier.

Big pharma’s operating model is changing along with the tooling. Sanofi has been explicit about becoming an “AI-first” drug company—announcing a three-way alliance with Formation Bio and OpenAI in 2024 and, in 2025, a second accelerator to embed AI across development and manufacturing. Isomorphic’s Lilly and Novartis deals show that top-tier pharmas will rent platforms they can’t—or don’t want to—build in-house, while the Recursion–Exscientia combination signaled consolidation among platform players aiming for end-to-end coverage from target ID to clinical trial simulation. The through-line: more build-partner-buy optionality at every step.

There are limits. Models trained on biased or low-quality data can underperform in under-represented populations; generative systems can propose molecules that are synthetically impractical or metabolically messy. And while AI can sharpen trial design, regulators will test whether sponsor-specific models generalize beyond the data they were trained on. That’s why the FDA’s credibility framework and the EMA’s lifecycle guidance matter: they formalize expectations for documentation, validation, and ongoing monitoring—turning AI from a black box into an auditable instrument.

The bottom line: AI is compressing the distance between a biological idea and a registrational study. It’s doing so by proposing better starting molecules, routing them through robotic labs that learn as they go, and then carrying those gains into trial design and conduct. If the last wave of drug productivity came from platform biology, the next could come from platform intelligence—provided companies pair the models with clean data, automated labs, and regulatory-grade validation.