GLG Frontier
>
Research

Where AI Is Actually Changing Drug Discovery – and Where It Isn't

Five practitioners – from AI-native biotech, platform business development, computational biology, small-biotech leadership, and clinical research – walk the pipeline from target to clinic and rate where AI is doing real work, where it’s still emerging, and where the claims run ahead of the evidence.
Drug Discovery
Frontier Briefing
GLG Frontier
November 2026
CONTRIBUTORS
GLG Frontier Research Team

By now AI is ubiquitous in preclinical discovery. Where biotech teams vary is on depth of use: Some run a model on one narrow step, while others have rebuilt their workflow around it.

So, where is AI changing clinical work? We spoke to five practitioners – from AI-native biotech, platform business development, computational biology, small-biotech leadership, and clinical research – to understand where AI is doing real work from target to clinic, where it’s still emerging, and where the claims run ahead of the evidence.

All five agree that AI has changed the front end – finding targets and generating candidates faster – and that its value fades as work moves into cells, animals, and patients. AI gets a molecule from target to clinic faster and cheaper, but, as the AI discovery executive said, “the jury is very much out on whether it actually increases probability of success.”

Below, we walk the pipeline stage by stage and rate each use of AI based on their perspectives. In real use means teams rely on it for decisions today. Emerging means it works in places but isn’t routine. Mostly hype means the claim remains ahead of the evidence.

The Scorecard

StageUse of AIRating
Target identification and validationIntegrating genetic, perturbation, and literature evidence to rank targetsIn real use
Calling a target causal from model output aloneMostly hype
Molecule designStructure and binding predictionIn real use
Generative design of antibodies and other biologicsIn real use
Novel small molecules generated from scratchEmerging
Lab and preclinical validationClosed lab-in-the-loop cycles that feed results back to the modelEmerging
Predicting in vivo or human biology from models aloneMostly hype
Into the clinicA faster, cheaper path to first-in-human trialsIn real use
Patient stratification by disease mechanismEmerging
Higher clinical success ratesMostly hype (unproven so far)

Target Identification: Real Use, but the Causal Logic Stays Human

Programs live or die in target selection. “If you have picked the wrong target, there’s no amount of good chemistry that is going to save it,” the former genomics lead said.

Her workflow starts with human genetics, and then perturbation biology. In both, AI integrates the layers and ranks targets, organizing the evidence, while the causal argument still depends on the team’s interpretation and analysis.

Indeed, the experts agree AI is most live in high-scale data analysis. The former computational sciences VP points to proprietary perturbation data generated at scale in disease-relevant cells. The biotech CEO uses general-purpose AI agents for literature search. The former CRO executive describes AI pulling literature and historical data together and taking over spreadsheet-heavy screening analysis.

They offer differing perspectives on trust, though. The former CRO executive argues that a well-specified screening task removes the fatigue errors a technician might make, but warns that AI mistakes are harder to catch and that a wrong early direction wastes everything built on it. The biotech CEO has a committee of senior scientists set guardrails and judge every output. “What’s most important is the molecule, the disease, and the patient,” he said.

Molecule Design: Real for Structures and Biologics, Emerging for New Chemistry

Structure prediction has made real progress, and the best models are now largely open and nearly free – so having one is no longer an edge.

Weak spots remain, though, including antibody-antigen interfaces, disordered regions, and induced fit, where a protein changes shape as something binds. Membrane targets such as G protein-coupled receptors (GPCRs) and ion channels are hardest, and for those the former genomics lead wants human genetic evidence before spending structural effort. The AI discovery executive adds that better structure prediction isn’t the rate-limiting step in drug discovery.

The current utility of generative design varies by modality. Designing antibodies and other biologics is the most mature work and delivers value today. Novel small molecule generation is less reliable. The near-term payoff, in her view, is faster hit-to-lead work, as opposed to finished drugs invented from scratch. With no standardized benchmarks, the AI discovery executive said, buyers end up judging platforms by partnerships and pipelines rather than technical merit.

Lab and Preclinical Validation: Where Predictions Meet Biology

Moving from in silico to in vivo is, in the former genomics lead’s view, the central unsolved problem. Models learn from cell lines, animals, and observational data, not human disease. Batch effects are the subtlest: data from different days, instruments, or reagent lots carries non-biological variation that a pattern-hungry model learns as signal.

There are unglamorous methodological fixes, like randomizing across batches, benchmarking correction methods, and testing on biology the model hasn’t seen. Virtual cell models interpolate well but are weaker on novel perturbations, and the benchmarks for judging them remain immature. That’s why standalone predictions of human biology still rate as mostly hype for now.

What does work is the closed loop, where predictions identify experiments and clean results retrain the model. A model pilot survives when its output changes a real go/no-go decision and is measured on program outcomes, not a score on a held-out set. The former computational sciences VP sees these lab-in-the-loop workflows becoming important but not yet standard, and the biotech CEO treats every prediction as provisional until in vitro and in vivo work confirms it.

Into the Clinic: Faster to First Dose, Unproven After That

The experts are aligned here. AI has shortened the path to first-in-human trials, and many AI-derived molecules have cleared Phase 1. Phase 2, which tests efficacy, has produced only a few single-point examples. And in patients, a molecule’s origin stops mattering. “There are no shortcuts, really, once you get into the clinic,” one expert said.

Patient stratification may be the most promising clinic-adjacent use for AI. Splitting a disease label into mechanistic subtypes can raise trial success, shrink trials, and define a biomarker population.

The Open Question: Are AI-Designed Drugs Holding Up in the Clinic?

AI demonstrably speeds up early discovery, the former genomics lead said, but it’s still unproven when it comes to raising clinical success rates. AI has changed the front end, she said, but “the hard part of this business was never really the front end.”

When we spoke to them, the experts were still waiting on the first approval of an AI-discovered drug. At the time, the former computational sciences VP expected late-stage readouts to set the direction of investment: clear differentiation would pull in capital, and results no better than conventional pipelines would slow it. The AI discovery executive warned against reading any one trial as a verdict, since most Phase 2 programs fail however they were discovered and few companies run enough trials to be judged alone. He also questions the label itself, because nearly every program now uses AI somewhere.

Our Take: What This Means for Teams Building AI for Drug Discovery

What holds models back from clinical use is rarely their architecture. All five experts agree it’s evidence: causal, human-relevant data, prospective validation, and transfer to biology the model hasn’t seen.

AI developers should test a few critical questions before a discovery model informs a real decision:

  • Does the eval hold out new perturbations, cell contexts, and targets, or does it reward interpolation?
  • Were predictions locked before the lab run, and is the confirmation rate reported with failures included?
  • Could the model be learning batch, instrument, or reagent-lot artifacts?
  • Does each ranked target show what kind of evidence supports it – genetic, interventional, or observational?
  • For generative chemistry, is synthesizability a hard constraint rather than an afterthought?
  • Is the model wired to a real go/no-go decision, with lab results flowing back as labeled data?

These questions become useful evals only with domain input. What counts as out of distribution, which assessment captures the right functional state, and what makes a causal argument credible are judgments that come from people who have run discovery programs. Practitioner judgment turns a generic benchmark into one a discovery team would actually trust.

Go Deeper

This Briefing draws on five GLG Library interviews: AI in Drug Discovery: Biotech Perspective (August 27, 2026), AI Driven Drug Discovery (August 27, 2026), and three Drug Discovery Lab Channel Check interviews (October 6, 2026; October 12, 2026; and October 12, 2026). GLG clients can read the full transcripts in myGLG. To pressure-test your own approach, GLG can connect you with practitioners in target discovery, computational biology, and preclinical development through a one-on-one expert call. Teams designing evaluations for discovery AI can work with GLG Frontier to build benchmarks and test scenarios grounded in real practitioner judgment.

About this piece: This Frontier Briefing draws on five GLG Library interviews recorded in August and October 2026. The experts are a former principal scientist and therapeutic genomics lead at an AI-native biotech; a chief business officer at an AI-driven drug discovery company; a computational biology consultant and former vice president of computational sciences at a biotech; the CEO of a small therapeutics biotech; and a former senior vice president at a clinical research organization (CRO) with a background in pharma biostatistics. The last three answered the same questions in a GLG channel check, so we compare them directly. The frameworks and recommendations for drug developers are theirs. The ratings and builder takeaways are GLG Frontier’s.

Get connected to the world's leading platform for trusted human expertise

Enter your information below to get in touch with our team.