Much of the last 10 years of AI drug discovery has focused on speed through lead optimisation: how quickly platforms can move from target to preclinical candidate.
That focus is understandable. Early discovery is slow, iterative, and expensive, and AI offers the potential to compress that process significantly.
But speed to candidate is not the same thing as delivering medicines. If the ambition for AI in drug discovery is to change outcomes for patients, the harder problem lies beyond candidate generation.
This focus on speed often leads to a broader claim: that the only metric that really matters is how quickly you can generate preclinical candidates, because once programmes reach IND-enabling studies and clinical trials, the process becomes constrained by regulatory requirements.
There is a thread of truth in this argument, but I think it is short sighted and missing the bigger picture. Clinical translation and probability of success through those regulatory trials is where the real impact lies. Clinical trials are the most expensive and the highest attrition, turning 10% success to 20% success would have a much bigger impact than halving the speed to a candidate.
If the goal is to produce large numbers of candidates and out-license them early, candidate throughput can be a reasonable metric.
This model is becoming increasingly common. Many organisations generate compounds against targets of interest and transfer the biological and clinical risk to partners who then take programmes forward.
This is a legitimate strategy that can create value and help test biological hypotheses quickly and efficiently. However, it does not necessarily demonstrate that AI drug discovery itself is solving the hardest problems in drug development.
But measuring success purely by the speed and volume of candidates risks missing the point.
A preclinical candidate is not the end of drug discovery; it is the beginning of the most difficult phase. The real test is whether those molecules translate into safe and effective medicines in patients.
If the ambition is to change outcomes for patients, the harder problem is clinical translation. This includes understanding why programmes succeed or fail, interpreting biological data correctly, and designing molecules that have patient impact - not just target engagement
AI drug discovery is sometimes treated as synonymous with candidate design. In reality, the field is much broader.
It includes understanding mechanisms, interpreting complex biological systems, identifying reasons for failure, and improving the probability that a programme succeeds once it reaches the stage that matters, patients.
Speed to candidate matters. But what ultimately matters more is what happens afterwards.
Patients are not treated with preclinical candidates. They are treated with medicines that survive the full journey from hypothesis to clinic
The Ignota Labs model learns from why drugs fail in the clinic to redesign them with an improved probability of success second time round - giving new hope to projects and patients.
Author: Sam Windsor
