Dr. Nagaraj Kulkarni

AI in Drug Discovery

What AI can, and cannot, solve in drug discovery

AI is already useful in drug discovery. It is not a shortcut around biology. A realistic look at both sides.

The conversation about AI in drug discovery tends to swing between two extremes: that it will replace the long, expensive process of finding medicines, or that it is hype with little to show. The more useful view sits between them. AI is a set of tools that are strong at some tasks and weak at others, and the difference comes down largely to data and to biology.

Where AI is genuinely useful

  • Finding patterns in large, messy datasets. Multi-omics, imaging and literature are too large for a person to read in full. Models can help prioritise targets and generate hypotheses worth testing.
  • Predicting molecular properties. Models trained on enough good data can estimate properties such as solubility, permeability or likely metabolic liabilities, helping chemists decide what to make next.
  • Exploring chemical space. Virtual screening and generative design can propose compounds that a team would not have thought to try, and narrow very large libraries to a testable set.
  • Understanding structure. Protein structure prediction has changed how quickly teams can form structural hypotheses about a target.
  • Reducing routine effort. Reading images, extracting data from papers and summarising results can free scientists' time for judgement.

Where it falls short

  • Data limits the model. Many drug discovery datasets are small, noisy, biased towards successful compounds, or generated under different conditions. A model can only be as reliable as the data it learned from.
  • Predictions outside the training data are shaky. A model that does well on compounds similar to those it has seen can fail on a genuinely new chemotype or target.
  • Biology is not fully captured. Disease mechanisms, compensation between pathways and differences between patients are only partly represented in the data. Predicting that a compound binds a target is a long way from predicting that it will help a patient.
  • It does not remove the need to test. A prediction is a hypothesis. It still has to be confirmed in an assay, then in a model, then in people. AI can change the order and the number of experiments, but not the need for them.
  • Clinical translation is the hardest step. The costliest failures happen in the clinic, where AI has so far had the least data and the least opportunity to prove itself.

How to judge a claim

When you hear that AI has made drug discovery faster or better, a few questions help separate substance from slogans:

  • Faster or better at which step, and compared with what baseline?
  • Was the prediction validated experimentally, and by whom?
  • How different was the test set from the training data?
  • Did the output lead to a molecule that progressed, and how far?
  • Is the claim about a measured result, or about potential?

The practical stance

The teams that get the most from AI tend to treat it as part of the scientific method, not a replacement for it. They use it to generate and rank hypotheses, test those hypotheses quickly, and feed the results back to improve the next prediction. Domain expertise stays central, because deciding which questions to ask, and which answers to trust, is still a scientist's job.

In short

AI can help us look at more possibilities and choose better experiments. It cannot tell us in advance what will work in a patient. The value lies in the loop between prediction and experiment, and in scientists who know the difference between the two.

Join the conversation

Pharmacology Insiders is a professional community for pharmacology, drug discovery and drug development. Members get case studies, expert discussion and career guidance.