Reverse Screen
Given a molecule, the targets it might interact with
A screening campaign asks which of many ligands binds one target. The reverse question asks which of many targets binds one ligand, and it is the question that arises once a molecule exists. What else does it hit, what explains a phenotype, what accounts for a toxicity finding, and what could the compound be repurposed against. It is also the only question available for a molecule a generative method proposes, because a designed structure has no measured target at all.
Docking one compound into every characterized binding site answers it directly and takes 40.5 hours. We put a retrieval step in front of the docking, which cuts the panel to tens of proteins in 44 milliseconds, and then dock only what retrieval returns.
Four stages, and only the last is expensive
Retrieval is purely ligand based. PharmCast predicts the query's 10,549-bit pharmacophore fingerprint from its two-dimensional structure, and that is compared by Tanimoto coefficient against every indexed ligand. The scan is exhaustive, with no approximate index and no clustering, which keeps it exact and leaves no parameter that could be tuned to flatter a benchmark. The sites carried by the nearest ligands become the candidate pool, and docking is the drill-down step on that pool rather than the first thing that happens.
Every co-crystal ligand in the archive, embedded by the fingerprint it presents and colored by target class. Drag to rotate.
What it costs and what it finds
Three thousand molecules were held out and screened. Each query was removed from the index by InChIKey connectivity layer first, so a molecule that had itself been deposited could not retrieve its own target through its own deposition. The question asked of each pool is whether the molecule's known target is inside it.
| Nearest ligands pooled | Proteins in the pool | Fraction finding the target | Docking time |
|---|---|---|---|
| 1 | 3.1 | 0.363 | 16 s |
| 5 | 21.1 | 0.488 | 108 s |
| 25 | 104.5 | 0.608 | 8.9 min |
| 400 | 1,001.5 | 0.815 | 1.42 h |
Recall rises with the pool and the pool grows faster than the recall does, so the operating point is set by the docking budget rather than by one best answer. Five neighbors is the setting we recommend for a library and 400 for a single molecule run on its own.
Run in batches on a single core the retrieval stage averages 37.3 milliseconds per molecule, about 96,000 an hour. A generative run can therefore be triaged in full, with docking spent only on what survives.
Why the admission rule is split
Comparing one binding site to another needs coordinates good enough to place every atom that lines the pocket, and the rule for that work is X-ray at 1.8 angstroms with a free R factor at or below 0.23. Establishing which ligand sat in which protein needs none of it. A 3.1 angstrom cryo-EM map answers that perfectly well.
Applying the stricter rule to both questions loses coverage silently, and it loses it unevenly: the loss falls on the targets and large complexes whose structures are cryo-EM. Separating the two rules took coverage from 3,670 target sites to 28,579. The glucagon-like peptide 1 receptor is the clearest case. It carries 31 entries holding a drug-like ligand and not one satisfies the 1.8 angstrom rule, so under that rule the target is absent from the panel and no amount of searching can return it.
Related work
PharmCast is the fingerprint every comparison is made with, on both sides. Our foundation models answer the adjacent question once you have a target in mind: which of two compounds is more potent against it, and which of two proteins prefers one compound.