Scientists and algorithms: Finding the right balance in modern drug development

Every few years, drug development goes through a moment where a new technology promises to change everything. Artificial intelligence is having that moment now. Unlike some previous waves of hype, this one is grounded in something real, but only if we are honest about what it can and cannot do.

I have spent my career in pharmaceutical development, much of it focused on formulation science. Over the past two years, I have also spent a considerable amount of time applying machine learning to that same discipline. What I have learned is simple to state, but easy to get wrong in practice. The future of drug development is not scientists versus algorithms. It is scientists and algorithms working together, each contributing what they do best.

Why drug development is fertile ground for AI

Formulation development, and pharmaceutical development more broadly, is a field defined by complexity. We are routinely asked to make decisions with incomplete information, under commercial pressure, and against timelines that leave little room for trial and error.

Consider the number of variables involved in formulating a single molecule. Particle size, excipient selection, processing method, dose strength, release profile, stability under different conditions. Each variable interacts with the others in ways that are rarely linear or intuitive. The experimental space is vast, and fully understanding the Critical Quality Attributes and their correlations to get a drug to market is an immense effort.

This is precisely the kind of problem where AI adds genuine value. Not because it is clever in some abstract sense, but because it can process patterns across large, multidimensional datasets far faster than a human can, surfacing options and relationships that would otherwise take months of trial and error to uncover.

Why AI should not replace scientists

That value only holds, however, if we are clear about the limits of the technology. A model only knows what it has been trained on. It cannot (yet) come up with hypotheses to test and how to test them. It has no understanding of regulatory context, no appreciation of clinical risk, and no ability to weigh trade-offs the way an experienced formulator does when a result does not quite fit expectations.

Drug development remains, at its core, a scientific discipline built on risk analysis. Every development programme involves uncertainty and difficult trade-offs, and every regulatory submission demands explainability and accountability that a model alone cannot provide. A regulator does not want to hear that an algorithm made a decision. They want to understand the scientific rationale behind it.

This is why I have come to see AI as a decision-support technology, rather than a decision-making technology. It should inform judgement, not replace it.

The clearest illustration of this comes from our own work applying machine learning to formulation development, developed in collaboration with Intrepid Labs. Rather than relying on generic, shared datasets, the model is trained from scratch on each client's own experimental data. As real experimental data is generated, the model learns the behaviour of that specific molecule against pre-determined CQAs and CPPs on formulations manufactured on the same equipment and analysed with the same methods.

Throughout this process, scientists remain firmly in control. They determine the objectives. They design the experiments that generate the data the model learns from. They interpret the outputs, and they decide what happens next. The role of the AI is to reveal patterns and relationships that would be genuinely difficult to identify through manual experimentation alone, allowing the team to explore the formulation design space more efficiently and learn faster with each cycle.

The case for human in the loop

The greatest value we have seen has not come from an algorithm alone, and it has not come from scientific expertise alone. It has come from combining scientific expertise, experimental data, and AI-driven insight. AI can explore a broader design space than a person can manually, and it can surface viable, non-obvious formulations along with the underlying relationships between composition, properties, and performance. But none of that removes the need for scientific judgement. If anything, it increases the value of that judgement.

Interpretability and scientific credibility matter. When a development programme moves toward the clinic, someone needs to be accountable for the decisions that got it there, and that accountability sits with the scientist, not the software. Human expertise provides context that no algorithm can replicate, understanding not just what the data shows, but what it means for a specific molecule, a specific patient population, and a specific regulatory pathway.

The scientist remains responsible for the decision. AI simply makes that decision better informed.

The scientists who embrace these tools well will become more effective, not obsolete. The organisations that succeed will be the ones that bring data science and pharmaceutical science closer together, rather than treating them as separate disciplines competing for influence.

In modern drug development, the question is no longer whether scientists or algorithms are better. That framing misses the point entirely. The real opportunity lies in ensuring they work together, each contributing what they do best, so that patients ultimately benefit from medicines developed faster and with greater confidence.

About the author

Dr Andrew (Andy) Lewis is chief scientific officer at Quotient Sciences, leading the company's scientific teams and drug development consultants to drive innovation and operational excellence. With more than 25 years' experience in pharmaceutical development and drug delivery, Lewis previously held senior roles at Ipsen. He holds a PhD in Tissue Engineering and a Bachelor of Pharmacy from the University of Nottingham.