AQPotency ranks how strongly candidate compounds bind their protein targets, fast and cheaply, without a crystal structure, and it matches docking on early-enrichment ranking. It is a proteochemometric (PCM) model: it takes a target UniProt ID and ligand SMILES strings, runs on standard CPUs, and scores pairs in seconds. On leakage-controlled virtual screening, its ranking is statistically equivalent to physics-based docking, at a fraction of the time and cost. Because it needs no 3D structure, it also reaches targets docking cannot, including membrane proteins, mutant panels, and proteins no one has crystallized. Every prediction carries an uncertainty estimate and an applicability score, so a scientist knows which rankings to trust and which to validate first.
Potency ranking sits upstream of the most capital-intensive steps in small-molecule discovery. Docking and free energy perturbation (FEP) estimate affinity rigorously, but they run slowly, cost heavily, and require a reliable 3D protein structure many programs never obtain. Machine-learning potency models add their own constraints: most need custom training per target, depend on GPUs that resist scaling, and hide how far to trust any single prediction. Teams working with limited structural information have to choose between waiting for structural data and advancing compounds on thin evidence.
AQPotency turns the first pass into fast, inexpensive triage. It scores roughly 2,000 pairs in about 15 seconds and ranks 1,000 for about a dollar in Claude, all on CPUs, so a team screens broadly, shortens the shortlist, and saves its FEP and wet-lab capacity for the compounds that clear the filter.
AQPotency is a proteochemometric model, an approach documented in the peer-reviewed literature. It represents both the ligand and the protein and learns their interaction jointly, so one model generalizes across many targets, including those with little measured data, and no one maintains a separate model per target.
The architecture pairs two encoders. A molecule encoder converts each compound, supplied as a SMILES string, into a learned representation; a protein encoder does the same for the target from its UniProt accession. An interaction model combines the two and predicts binding strength as a pIC50 value. Grounding the model in a documented method gives reviewers a familiar basis for judging where its assumptions hold.
Two boundaries follow from a structure-free, sequence-plus-ligand approach, and they are why the confidence signals below matter:
Every AQPotency prediction comes with two signals that tell you how far to trust it: an applicability score between 0 and 1, and an uncertainty estimate. The applicability score flags whether a molecule falls inside the chemical space the model handles well (below 0.4, treat the prediction with caution). The uncertainty estimate rates the model's confidence in that specific value.
Read together, they turn a raw ranking into a risk-managed one: act on high-confidence predictions inside the applicability domain, and set the rest aside for review. This is necessary because accuracy varies with familiarity. The more a candidate resembles what the model was trained on, the better it ranks, so performance is strongest inside the applicability domain and drops off beyond it. The two signals surface that pattern rather than hiding it.
Most scoring tools return a number and nothing else. AQPotency exposes its confidence machinery as a tool: teams can benchmark it against their own data and check for data leakage, whether a model was tested on chemistry it already saw in training. That programmatic audit is unusual, and paired with an honest operating contract, ranking rather than absolute accuracy, reviewed UniProt targets across species, and performance that tracks chemical similarity, it lets teams prioritize broadly and early while reserving the expensive tools for where they matter most
