Thinking Upstream
This work began from an apparently simple observation: in biology, genes encode proteins, and proteins encode function. When nature evolves or optimises a function, it is the encoding gene that has been mutated.
In small-molecule discovery, we often work differently. Reactants (and reagents) give rise to products, and products exhibit function. Yet, if that function needs improvement, we rarely change the reactants directly – the “code” that generated the molecules. Instead, we redesign the products themselves.
If we wanted to more closely mirror nature’s strategy for discovering and evolving function, shouldn’t we also be thinking upstream? Perhaps optimisation should lie not only in the products, but also in the reactant combinations from which they emerged.
Moving Beyond Targets
With this upstream perspective in mind, we then considered how function itself is usually assessed.
Drug discovery tends to begin with a predefined biological target – usually a protein implicated in disease – and molecules are evaluated according to how effectively they modulate its activity. This approach is powerful when structural or biophysical information is available.
Alternatively, we can look directly at observable changes in cells and organisms. Cells may alter their shape, size, internal organisation, or overall morphology in response to treatment, and these changes can be measured systematically. Using phenotypic effects as a readout of biological impact allows compounds to be evaluated without assumptions about molecular targets, sometimes enabling unexpected mechanisms to emerge.
Bringing these perspectives together, we arrived at a guiding question:
Could we design a system in which neither molecular structures nor biological targets are specified up front, and in which functional outcomes emerge as reactions are explored? As one might imagine, early conversations often ended with raised eyebrows. The idea felt a little like setting off without a map or GPS, trusting that direction would only become clear once the journey was underway.
Automating Discovery
Framed in this way, one practical constraint became immediately obvious: scale. The number of possible reactant combinations quickly exceeds what can realistically be explored through intuition, manual experimentation or conventional synthesis planning.
Automation therefore became not only a way to accelerate workflows, but also a viable path forward. By automating synthesis, purification, and phenotypic evaluation, we might systematically sample reaction space, while maintaining consistency and throughput.
Automation also allowed us to streamline some time-consuming steps. Reaction mixtures were analysed rapidly (around three minutes per reaction) to identify new products and estimate their abundance. This information was then passed directly to the purification system, which automatically separated and collected the products. The purified products could therefore move straight into biological testing, while full structural elucidation was reserved for products showing interesting biological effects.
As with any attempt to automate discovery, not everything behaved as intended on the first pass. Or the second…
On one occasion, a sudden loss of mass spectrometry signal prompted what we believed to be a display of technical competence. The instrument was vented, dismantled, cleaned, and treated with the level of care normally reserved for historic artefacts. The original error suddenly disappeared – only to be replaced by an entirely new one. The eventual diagnosis was far from rocket science (or automation science!): a single checkbox in the software had been accidentally left unticked. Progress, it turned out, was less about eliminating errors and more about transforming them into recognisable ones.
Letting Algorithms Learn
Once chemistry and biology data were being generated, the next challenge was deciding how to present that information so an algorithm could act on it effectively to complete the loop.
After extensive brainstorming, and occasional detours into complexity, we ultimately chose to define our problem simply, but through explicit rules. We specified: a reaction space in which the algorithm could operate; how reactant combinations should be represented; the maximum permitted total reactant cost; and how similarity and diversity should be assessed. In effect, we defined the playground and its rules, and allowed the algorithm to navigate it.
Within this framework, morphological scores were propagated back into reaction space, enabling outcomes from one discovery round to influence the next. The chosen chemical fingerprinting provided a consistent measure of similarity, allowing exploitation of regions related to previously-active combinations, while still exploring unsampled areas.
Once this structure was in place, functional organisation began to emerge. In the first round, several reactions produced compounds whose morphological effects closely resembled those of known microtubule disruptors involved in cell division, despite no biological target having been specified in advance. These signals then guided the next iteration: reactant combinations were scored according to the strength and similarity of phenotypic effects to the emerging tubulin-like cluster. In the second round, this led to four-fold higher reaction productivity and enrichment of tubulin-associated hits.
Looking Forward
Although we focused here on phenotype-directed discovery, the framework itself is not limited to this mode of operation. Any outcome that can be measured, scored, and meaningfully propagated through reaction space – including target-based readouts – could in principle be integrated into the same closed-loop workflow. In that sense, the approach is less about choosing a single discovery strategy, and more about building flexible systems capable of making good use of data wherever it comes from.
And despite the automation, the journey itself was deeply human. It was shaped by the collective effort of many people: those who built and connected the systems, brainstormed, wrote the code, tested the chemistry, ran and validated the biology, ran NMR late into the evenings, and offered patience and advice throughout.
In the end, the loop may have been closed by an algorithm, but it was built, shaped, and understood by people who shared ideas and kindness within and beyond the lab. Perhaps the clearest sign of this was the gift of a T-shirt printed with the discovery cycle itself – a lovely gesture, but also a gentle indication that we might have talked about it just enough!