STAR

A small molecule does not only touch the protein it was designed for. It brushes against ion channels in the heart, enzymes in the liver, transporters in the kidney — and whether that matters depends on where those proteins are expressed. Screening answers this one target at a time, affordable for a handful of anti-targets and impossible for the twenty thousand proteins a compound meets in a patient.

STAR — for structure–activity–tissue relationship — is Omic Inc.'s answer to that problem: a pretrained drug–target affinity model that takes a compound structure and returns predicted binding scores across the human proteome in a single pass. The name traces the chain it models, from a molecule's structure through its activity at each protein to the tissues in which that protein is present. Its outputs are a fixed derived product — computed once per compound, frozen, and read by downstream analyses as static artifacts rather than regenerated for each new question.

STAR has no publication of its own. Its entire public record is the Methods and availability statements of Omic's 2026 MERIT preprint, which describes a different model: a clinical-trial outcome classifier that consumes STAR's binding profiles as its molecular feature layer. That report characterizes STAR as trained against its own objective, with no trial outcome or clinical annotation entering training, and applied to new compounds as a fixed scorer rather than refit per cohort. That sets it apart from the per-pair scorers this catalog carries — CASTER-DTA, HydrAffinity and the affinity head of Boltz-2 — run on demand once a target of interest is known.

#Key Features

• Proteome-scale profiles: One pass yields a compound's predicted binding across the human proteome, so off-target hypotheses come from the same object as on-target ones. The frozen profile archive from the MERIT study totals roughly 3 TB.
• Outcome-blind objective: No clinical outcome, approval status or trial annotation enters STAR's training. A downstream model reading its profiles cannot inherit clinical labels through the binding layer, which is what lets a trial-outcome classifier hold out a compound honestly.
• Tissue-resolved readout: Predicted binding is interpreted against tissue-specific expression from the Human Protein Atlas across thirteen organ systems, from heart and brain to liver, kidney, immune and skin.
• Safety-target coverage: Profiles span known liability targets — cardiac ion channels such as hERG, hepatic enzymes, renal transporters, DNA-damage machinery — alongside essential housekeeping genes, supporting per-organ burden summaries.
• Human proteome only: STAR predicts interactions with human proteins, not pathogen proteins, so anti-infective compounds fall outside its coverage and were excluded from the downstream efficacy evaluation.

#Technical Details

Omic has not published STAR's architecture, training corpus, parameter count or held-out affinity benchmarks. What is on record is its class and its mode of use: a pretrained drug–target affinity model producing continuous per-compound, per-protein scores, applied unchanged to new compounds. Downstream, those scores are summarized as proteome-wide distribution statistics — percentiles, concentration, entropy — plus direct-target engagement, tissue-weighted organ profiles and toxicity-target panels, together supplying 172 of the 285 candidate features in the MERIT classifier.

The published evidence on STAR's accuracy is indirect, measured through those derived features. Cardiac binding features predicted experimental hERG blockade across 105 compounds (65 blockers) above a baseline of general promiscuity and lipophilicity, with organ specificity confirmed by residualizing on those confounds; hepatic binding features were tested against DILIrank liver-injury concern across 213 compounds. The binding-derived safety score is uncorrelated with black-box-warning or withdrawal status (AUC 0.50 and 0.48), evidence that it reflects predicted molecular interactions rather than existing regulatory flags.

#Applications

STAR suits proteome-wide off-target and safety triage: ranking which human proteins a candidate is likely to engage, quantifying selectivity as a distribution rather than a single ratio, and flagging organ-specific liability before a compound reaches an assay. Its demonstrated use is as the molecular feature layer beneath a clinical-trial outcome model, where one profile per compound serves every downstream indication. Because the profiles are frozen artifacts, groups can reuse them without running the model; Omic supplies the archive on request, and the pipeline for non-commercial academic use on request.

#Impact

STAR is an unusual case of a closed model published as a reproducibility contract. Rather than release the pipeline, Omic distributes its frozen outputs so that results built on them can be reconstructed without the proprietary core; only profiling a new compound requires the model. That makes the derived work checkable while leaving STAR itself unaudited. With no released weights, code, model card or standalone evaluation, its predictions cannot be independently reproduced, benchmarked against established affinity baselines, or examined for the bias a proteome-wide scorer might carry toward well-studied targets — and its authors are explicit that the analysis leaning most heavily on the binding layer is the weakest they report.