LURA screens a compound against genetic context, moves it through tissue in three dimensions, and returns why it wins or fails.
Drug–cell pairs screened. 2.5 million drug–cell interactions evaluated in 1.5 seconds on a single NVIDIA A100.
Screening is not the bottleneck it used to be. The bottleneck is that the answer has to survive contact with data the model has never seen — and most of what is fast is fast because it has stopped being careful.
So the number that matters is not the first one. It is the two that follow it. We can compute a massive space, we can compute it quickly, and the signal holds outside the set it was trained on. Those three facts in that order are the whole thesis.
Into the tissuePotency in a well says nothing about the tissue a compound never reaches. LURA solves transport, gradients and pressure in three dimensions, so the regions a compound fails to enter are part of the answer instead of being absent from it.
A compound enters tissue by diffusion, against the interstitial pressure pushing back out. LURA solves that field in three dimensions, so how far a molecule travels before it is cleared or consumed is a computed quantity rather than an assumption.
None of these is uniform across a tumour, and each one changes what a drug does when it arrives. The engine carries them as coupled fields, so the local chemistry of a region is part of that region’s answer instead of an average taken across the whole mass.
Potency only counts where the drug arrives. The engine reports the share of tumour volume that reaches a therapeutic concentration, and it names the regions that never do — which is the part a well plate cannot tell you at all.
Every claim on a LURA surface is routed to the engine that can actually make it. An engine that cannot answer a question does not get asked it, and the surface says so rather than guessing.
Which compounds are selective for this genotype. The drug-differentiation engine, and the only one allowed to rank one compound against another.
Decision-gradePenetration, hypoxia and kill pressure solved in three dimensions. Physics, not a proxy for it.
Decision-gradeWhich pathways are active in a cell context. Context only — it is not used to separate one compound from another.
ContextCompound to target to pathway, from a verified knowledge graph. A prior with a citation, never a substitute for a measurement.
Verified priorA compound outside the catalog is carried as a read-across hypothesis with its nearest analog and an estimate range, labelled as such. It is never quietly promoted to a measurement.
ScopedThe output is not a dashboard. It is a governed artifact: the decision, the evidence under it, the engine that produced each line, and the provenance chain that lets a reviewer check any of it. If a claim is read-across rather than direct, the artifact says so on its face.
Start a conversationThe fastest way to understand what LURA does is to point it at a real question: a compound series that is not separating, a genotype you cannot explain, or a candidate that works in a well and fails in tissue.
We read every message. If a study is already in flight and the timing is tight, say so in the last field and we will answer the same day.