AQPotency

Early-stage virtual screening has mostly been a structure-based exercise: dock a library into a pocket, then spend free-energy perturbation cycles on the survivors. Both steps need a reliable three-dimensional structure of the target, and many programs never get one — membrane proteins, mutant panels, and proteins nobody has crystallized are exactly the cases where the first triage decision is hardest. AQPotency, released by SandboxAQ, takes the other route. It is a proteochemometric model: it encodes the protein and the ligand separately, learns their interaction jointly, and predicts potency on the pIC₅₀ scale from a UniProt accession and a SMILES string.

What makes that practical is how cheap the protein side is. The accession is resolved against a table of precomputed ESM-2 embeddings, so no protein language model runs at inference; the ligand becomes a small-molecule fingerprint on the fly; and a fused potency head reads the pair and emits a predicted pIC₅₀. The whole thing runs on ordinary CPUs — roughly 2,000 protein-ligand pairs in about 15 seconds — with no per-target fitting step, so one checkpoint serves targets it has never seen. Reverse screening a compound across a 20,431-protein panel becomes a routine call, not a compute project.

AQPotency reached general availability on August 19, 2026, delivered as four tools on SandboxAQ's hosted Model Context Protocol server — a distribution channel rather than part of the model, since no language model sits in the inference path. It is distinct from AQAffinity, SandboxAQ's other protein-ligand potency model, which runs an OpenFold3 trunk and folds the complex internally before an affinity head reads it out. AQPotency never builds a structure at any stage.

#Key Features

• Structure-free ranking: Potency is predicted from sequence identity and ligand chemistry alone, opening targets that docking and physics-based methods cannot reach without a structure.
• Frozen protein embeddings: Resolving each UniProt accession to a precomputed embedding removes the protein encoder from the inference path, which is what allows proteome-scale scans on CPU hardware.
• Always-on confidence signals: Every prediction returns potency_sigma (a per-pair standard deviation from a Gaussian-mixture head), ligand_similarity (maximum Tanimoto similarity to the training set) and protein_in_training, so out-of-domain extrapolation is visible rather than silent.
• Four screening modes: single-pair prediction, a library screen of up to 10,000 SMILES, a panel scan against the Bowes safety panel (46 proteins), human kinome (604) or full proteome (20,431), and off-target fold-selectivity.
• No per-target training: One served checkpoint transfers to new targets and new chemistry rather than being refit for each program.

#Technical Details

The architecture pairs a ligand fingerprint with a frozen ESM-2 protein embedding in an interaction head that predicts pIC₅₀ as the mean of a Gaussian mixture, giving both a point estimate and a per-pair spread. Inputs are restricted to reviewed UniProt (Swiss-Prot) accessions; unreviewed TrEMBL entries and raw protein sequences are not accepted. SandboxAQ evaluates the model on LIT-PCBA early enrichment across 15 targets, reporting an enrichment factor at 5% of 2.2 against 1.5 for Vinardo docking on the full set and 2.1 against 1.3 after filtering training overlap and PDB selection, with Wilcoxon signed-rank p = 0.21–0.54 — statistically indistinguishable on ranking, at orders of magnitude less cost. Mean maximum training-ligand similarity in the leakage-controlled arm was 0.31. The model is calibrated for rank correlation, not absolute affinity: SandboxAQ states that R² is often poor and can be negative out of domain, and warns against quoting a predicted potency_mean as a binding constant. No preprint or peer-reviewed paper describes the model, and the training corpus, weights and code are not disclosed.

#Applications

The intended use is the cheapest first pass of a small-molecule campaign: ranking a focused or purchasable library to decide what enters docking, FEP, or an assay; profiling selectivity across a target family; scanning a compound against safety and kinome panels for liabilities; and running the problem backwards, taking a phenotypic hit and scoring it proteome-wide for candidate targets. Because the inputs are an accession and a SMILES string, medicinal chemists and screening groups can use it on programs that begin before any structural biology does.

#Impact

AQPotency is a commercial, closed-weight release whose significance lies in the delivery model as much as the science: it packages a proteochemometric ranker behind a hosted, metered service priced per protein-ligand pair. SandboxAQ reports use in eight customer programs with experimentally validated impact, but names none of them and publishes no prospective hit rates; the LIT-PCBA comparison is the only public benchmark, so independent evaluation is not possible — the standing limitation of paperless commercial releases. The design choice worth carrying forward is the confidence layer: shipping applicability-domain and uncertainty signals with every score, rather than a bare number, is what lets a ranking model serve as triage without being mistaken for a measurement.