Designing a drug-like molecule for a given protein target has historically been a craft. Skilled chemists draw on decades of experience to propose new compounds, balancing potency, selectivity, and developability against an ever-changing data landscape. The bottleneck is human judgement applied one molecule at a time.
Our latest research describes a generative model that learns to propose ligands directly from a target's structure, fine-tuned on internal assay data and informed by physics-based scoring. The result is a system that suggests candidates orders of magnitude faster than manual design, while keeping a chemist firmly in the loop.
How the model works
The architecture combines a structure-aware protein encoder with an autoregressive ligand decoder, supervised across a curated set of public and proprietary binding data. We then condition the decoder with property constraints — solubility, permeability, off-target liabilities — so that the molecules we evaluate in the lab are already pre-filtered for what matters.
What's next
Generative chemistry is a tool, not a destination. The real measure is how quickly we can move from target to candidate and from candidate to clinic. We're using this model today across several internal programs and plan to share more results — and the data behind them — in the months to come.