BioUnfold #3 — The AI Use Cases for Drug Discovery
Making a difference with AI in drug discovery is not about finding more places to apply it — but about choosing deliberately where it adds real value.
Drug discovery feeds on multimodal data. Public datasets keep growing and improving in accuracy: pathways are mapped mechanistically, genes are classified by function and disease relevance, and every experiment contributes to a shared pool of knowledge. Where there is data, there is an opportunity for computation.
For target selection, omics and pathway databases are a good starting point. For screening, imaging often provides the most efficient readout. For design, generative chemistry can help explore molecular space.
When I say “AI,” I do not mean a particular algorithm — not LLMs, XGBoost, KMeans, or GNNs in isolation. I mean the deliberate use of the right computational method to extract actionable information for the next experimental step. Any time there is knowledge or observation, algorithms can support discovery — but only if the data and context are strong enough.
Why selectivity matters
Algorithms do not understand context. Humans do. There is always a gap between what we think we give to a model and what the model actually receives. It does not have the accumulated experience that makes decisions obvious to a scientist.
Building effective systems costs time — collecting the right data, shaping the right model, and iterating. Real leverage does not come from applying AI everywhere. It comes from choosing one place where it matters most.
Most experiments need analysis, but few deserve full automation. A team may run many analyses, yet build only one or two true AI platforms — the systems that learn across projects.
Target-based vs phenotypic platforms
Target-based discovery has a structural advantage: it starts from a defined mechanism and models how a molecule can alter it. That clarity makes it easier to build digital workflows early.
Phenotypic discovery is different. After observing a biological effect, it is harder to build a digital platform without introducing assumptions. The art is to restrict those assumptions without collapsing the biological space.
For example, assuming that most disease models can be screened through imaging lets you specialize in a single, fast readout — enabling multi-parameter selection and counter-screens in parallel. Transcriptomics can still be layered in later, but at higher cost and longer cycle time.
Of course, phenotypic campaigns remain unpredictable. A hit molecule might point to an easy-to-optimize pathway — or to a mechanism that is much harder to model. Some modern campaigns now mix modes: for example, Recursion uses label-free imaging to prioritize phenotypes, then applies transcriptomic profiling or cell sorting on a smaller set of candidates.
The real AI advantage
The power of AI in drug discovery is not in applying it everywhere, but in knowing where to build depth.
One strong, integrated computational capability is worth more than ten scattered tools.