# Reverse Screen

> Reverse Screen answers the reverse screening question: given a molecule, which
> protein targets it might interact with. Every co-crystal ligand in the Protein
> Data Bank is indexed by the three-dimensional pharmacophore fingerprint it
> presents together with the UniProt accessions it was solved against, so a query
> retrieves candidate targets in 44 milliseconds. AutoDock Vina then docks into
> the sites retrieval returns. Built by Eidogen-Sertanty, Inc. Principal author
> Steven M. Muskal, ORCID 0000-0002-3487-270X
> (https://orcid.org/0000-0002-3487-270X).

## What it is

Forward screening asks which of many ligands bind one target. Reverse screening
asks which of many targets bind one ligand. That is the question that arises once
a molecule exists: what else does it hit, what explains a phenotype, what accounts
for a toxicity, what could it 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 molecule into every characterized site answers it directly and takes
40.5 hours. A pharmacophore retrieval step in front of the docking cuts the panel
to tens of proteins in 44 milliseconds.

## The measured numbers

- Indexed co-crystal ligands: rebuilt weekly; the live count is at /api/index-info
- Target sites they map to: rebuilt weekly; the live count is at /api/index-info
- Fingerprint the query with PharmCast: 4 ms
- Scan all indexed ligands by exact Tanimoto over 10,549 bits: 40 ms
- Pooling the 5 nearest ligands: 21.1 candidate proteins, target found 48.8 percent
- Pooling the 25 nearest ligands: 104.5 candidate proteins, target found 60.8 percent
- Pooling the 400 nearest ligands: 1,001.5 candidate proteins, target found 81.5 percent
- Five neighbors is the setting recommended for a library and 400 the setting for a
  single molecule, which is the default on the screening page
- Docking one site with AutoDock Vina at exhaustiveness 16: 5.1 seconds on twenty cores
- Redocking, 38 complexes: correct pose generated in 27, ranked first in 14
- Batched retrieval throughput on one core: 37.3 ms per molecule, about 96,000 per hour

Recall is measured across 3,000 held-out molecules, each removed from the index by
InChIKey connectivity layer first, so a deposited molecule cannot retrieve its own
target through its own deposition.

## Structure admission

Retrieval and site comparison admit structures under different rules on purpose.
Describing the atoms lining a pocket needs X-ray at 1.8 angstroms with a free R
factor at or below 0.23. Establishing which ligand sat in which protein does not:
a 3.1 angstrom cryo-EM map answers that perfectly well. Retrieval therefore admits
X-ray at 2.5 angstroms with R free at or below 0.28, and cryo-EM at 4.0 angstroms,
for components of 250 to 900 daltons. That takes coverage from 3,670 target sites
to 28,579, an increase of 7.8 fold, measured on the 8 September 2026 snapshot.

## Pages

- [Method](https://reversescreen.ai/): how it works, with the pipeline figure and an
  interactive three-dimensional map of the retrieval index
- [Screen a molecule](https://reversescreen.ai/try.html): draw a structure or paste a
  SMILES string, get candidate targets, drill into any of them, request a docking

## Related work by the same group

- [PharmCast](https://pharmcast.ai/): predicts the three-dimensional pharmacophore
  fingerprint, 10,549 bits wide, from a SMILES string. Reverse Screen uses it for the
  query and for every indexed ligand, so both sides of a comparison are the same
  calculation.
- [Family Foundation Model](https://familyfoundationmodel.com/): compound
  preference and target preference across 34 protein families, trained on ChEMBL 37
  alone. Given a molecule Reverse Screen has retrieved candidates for, the target
  preference model says which of two candidate targets from different families the
  molecule prefers. Video: https://www.youtube.com/watch?v=VWMWP5hi3uA
- [GPCR Foundation Model](https://gpcrfoundationmodel.com/): potency and
  selectivity rankings across human G protein-coupled receptors.
- [Kinase Foundation Model](https://kinasefoundationmodel.com/): sequence-conditioned
  ranking of kinase inhibitor potency and selectivity.
- [Eidogen-Sertanty](https://eidogen-sertanty.com/): curated kinase structure activity
  data and the models trained on it.
- [Steven M. Muskal](https://www.stevenmuskal.com/): principal author. Projects,
  publications and videos.

## Citation

Muskal, S. M. and Nicola, G. ReverseScreen.ai: Pharmacophore-Guided Reverse
Screening Across the Growing Co-Complex Proteome. bioRxiv, first posted
15 September 2026, version 1. doi: 10.64898/2026.09.09.750461
https://www.biorxiv.org/content/10.64898/2026.09.09.750461v1

Video overview: "Reverse Screen: Given One Molecule, Which Proteins Does It Bind? 40 Hours of Docking to 108 Seconds",
https://www.youtube.com/watch?v=mFRBfK6iv_w

Reverse Screen for a general reader, with the video and the paper:
https://www.stevenmuskal.com/projects.html#reverse-screen

The fingerprint method both sides of every comparison are made with:
Muskal, S. M. and McGregor, M. J. PharmCast: rapid generation of three-dimensional
pharmacophore fingerprints from two-dimensional structure without conformer
generation. bioRxiv, 2026. doi:10.64898/2026.09.02.748999
(https://www.biorxiv.org/content/10.64898/2026.09.02.748999v2)