No items found.

AI in drug discovery: what has changed, and what hasn't

November 25, 2026
Pharmaceutical pills next to a vial of medicine

The latest advances in AI-driven drug discovery, and why scientific data quality and experimental validation still matter

‍

By Dr. Neil Taylor

‍

Few areas of science have attracted as much attention, or as much investment, as AI in biotech. Hardly a week passes without a new model, partnership or funding round. For research leaders, the harder task is separating what is genuinely changing how medicines are discovered from what is still a promise.

‍

This article looks at the innovations that matter most in AI-driven drug discovery, and at the practical lessons emerging from them.

‍

1. Protein structure prediction became infrastructure

‍

Five years ago, accurate protein structure prediction was a research breakthrough. Today, it is routine. AlphaFold's predicted models are available for more than 200 million proteins, and the work behind it received the 2024 Nobel Prize in Chemistry.

‍

The frontier has moved from single proteins to complexes. AlphaFold 3 and newer models such as Boltz and Chai can predict how proteins interact with other proteins, nucleic acids and small molecules. Open alternatives, including OpenFold, are also expanding access to AI-based protein structure prediction. Several of these models are freely available, meaning research groups can generate plausible models of drug–target interactions.

‍

2. The race to predict binding affinity

‍

Predicting that a molecule binds is useful. Predicting how strongly it binds is what medicinal chemists really need, and it remains one of the hardest problems in computational chemistry.

‍

In 2025, Boltz-2 combined protein structure prediction with binding affinity estimation in a single open model, reporting accuracy approaching physics-based free energy methods at a fraction of the computing cost.

‍

Other groups are pursuing the same goal. Independent evaluations have also highlighted limitations in how reliably these predictions translate across different datasets and molecular systems.

‍

Binding affinity prediction is now one of the most closely watched areas in AI drug discovery because reliable predictions could reduce the number of compounds that need to be made and tested.

‍

3. AI-designed molecules reached patients

‍

The question "has AI produced a drug?" now has a more concrete answer. Several molecules discovered or optimised with AI have entered clinical trials.

‍

In 2025, Insilico Medicine reported positive Phase 2a results for rentosertib, an AI-discovered drug candidate for idiopathic pulmonary fibrosis.

‍

However, AI-discovered medicines have yet to demonstrate the full journey from discovery to regulatory approval. Clinical success rates remain the measure that will ultimately judge the field.

‍

4. Investment and consolidation in AI drug discovery

‍

Capital has followed the science. Isomorphic Labs, the AI drug discovery company spun out of Google DeepMind, raised US$600 million in its first external funding round in 2025.

‍

Recursion merged with Exscientia in 2024, combining two of the best-known AI drug discovery companies. Large pharmaceutical companies are signing partnerships with AI biotech companies while also building substantial in-house AI teams.

‍

For pharma industry observers, the pattern is clear. AI capability is moving from a differentiator to an expectation.

‍

Scientist inspecting a server

What hasn't changed

‍

Amid the momentum, some fundamentals of pharmaceutical research remain firmly in place:

  • Data quality still decides outcomes. Models trained on public data perform best on targets that resemble public data. A company's own experimental structures and assay results, curated and consistent, are often its most valuable AI asset.
  • Experiments remain the final judge. Predicted structures and binding affinities guide decisions; X-ray crystallography, cryo-electron microscopy (cryo-EM) and experimental assays confirm them.
  • Integration is the bottleneck. A model running in isolation helps one scientist. Value comes when predictions sit alongside experimental results, where whole teams can compare, question and build on them.
  • Biology is still hard. Most drug candidates fail in clinical trials for reasons of efficacy and safety that early-stage AI does not yet address.

The practical takeaway for AI-driven drug discovery

‍

For biotech and pharmaceutical research teams, the question is no longer whether to use AI. It is whether their data, workflows and people are ready to get value from it.

‍

The organisations seeing results tend to share three traits: well-managed proprietary data, clear ways to compare AI predictions with experimental results, and scientists who understand both the strengths and limits of the models.

‍

About DesertSci

‍

DesertSci has worked at the intersection of structural biology, computational chemistry and software for more than 25 years. Its Proasis platform helps leading pharmaceutical companies manage the experimental and predicted structural data that AI-driven drug discovery depends on.

‍

Learn more at desertsci.com and follow Dr Neil Taylor on LinkedIn for more insights.

‍