245 FDA drugs. 8 ADMET endpoints. One physics-first call.

The challenge

Pharma teams predict ADMET today through a chain of tools that were assembled across vendors and over time, rather than designed as a unified system. A typical lead-optimization workflow stitches together four to six commercial and internal systems, each with its own input format, output schema, license, support contract, and chemical-space caveat list:

• Schrödinger QikProp / LiveDesign for plasma protein binding, permeability, and selected druglikeness descriptors. Site licenses typically run in the low-to-mid six figures per year.
• Simulations Plus ADMET Predictor for clearance, transporter, and selected toxicity endpoints. Similar pricing tier, different output schema, different chemical-space coverage.
• OpenEye toolkits for solubility, logP, and shape-aware filters. Another license, another schema, another set of caveats.
• Internal ML models for hERG, DILI, and reactive-metabolite alerts — maintained by a small team with quarterly drift retraining and manual data-labeling cycles.
• Free / sanity tools (SwissADME, ADMETLab) used to spot-check the commercial outputs.

Reconciling these systems into a single decision-grade output adds non-trivial integration overhead per workflow. ML components also degrade on chemistry outside their training distribution, which includes much of the novel scaffold space medicinal chemists are paid to design. The aggregate cost includes per-tool licenses, integration engineering, schema reconciliation, and the structural absence of a unified mechanism-aware ADMET output in a single call.

Study design

The study uses the 245 FDA-approved drug panel already documented on the public ADMET benchmark page. Each compound is profiled across all eight endpoints in a single unified pipeline call. Wall-clock time is measured end-to-end in full mechanistic mode — the slowest configuration, with DILI exposure-aware logic, CYP isoform gating, transporter inference, dose-window logic, and reactive-metabolite alerts all active.

Results overview

FluxMateria profiled all 245 FDA drugs across all eight ADMET endpoints in ~51 seconds of wall-clock time in full mechanistic mode — 1,960 individual mechanism-aware predictions, returned as one decision-grade output schema per compound. Three of the eight endpoints land at strict #1 SOTA against the public TDC and AqSolDB leaderboards. DILI now also reaches SOTA accuracy: AUROC 0.9597 versus the MiniMol public reference around 0.956 on the comparable TDC binary benchmark. It also reports AUPRC 0.9455, high-vs-rest balanced accuracy 0.8223, and returns mechanism, exposure, dose-window, confidence, and score-trace outputs. Caco-2 permeability has now joined the SOTA tier as well: MAE 0.277 on the TDC caco2_wang scaffold-stratified test set, matching the public reference SOTA at 0.276 from pure physics with zero Caco-2 training labels consumed. None of the eight required endpoint-specific retraining or post-hoc tuning.

Wall-clock measured end-to-end on a single CPU core for deterministic reproducibility, in the slowest mode (DILI exposure-aware, CYP isoform gating, transporter inference, dose-window logic, and score trace all active). Production deployment scales horizontally. Accuracy figures from the publicly audited ADMET benchmark and detailed DILI benchmark.

Endpoint-by-endpoint accuracy: head-to-head

Each row of the table below uses the same compound set, the same metric, and the same leave-one-out validation protocol the named competitor reports. No metric switching. No re-binned subsets. Where FluxMateria leads, it leads by a publishable margin. Where it does not lead, the gap and the cohort difference are stated honestly.

All eight endpoints validated under leave-one-out. Reference cohorts are full LOO sets, typically larger than the public TDC leaderboard subsets (which are often 475–1,800 compounds). Where FluxMateria is "near-SOTA" or "competitive," the LOO cohort is itself a more demanding test than the smaller TDC equivalents. DILI is reported in novel-like mode with exact clinical self-matches masked; known-compound anchor mode is tracked separately and is not used for the novel-drug SOTA claim. Full per-tier and per-class breakdowns: ADMET benchmark and DILI benchmark.

Total cost of ownership: the annualized view

Pharma teams do not buy a single ADMET screen — they buy a capability. The relevant comparison is the annualized cost of running ADMET prediction as an ongoing function across lead-opt, portfolio triage, and regulatory pre-submission. Below, the four realistic alternatives a discovery program faces today, costed honestly.

Annualized cost ranges represent typical industry benchmarks for ongoing pharma ADMET capability: stitched commercial stack includes typical per-tool site licenses ($100K–$500K each across the four to six tools listed) plus integration engineering plus maintenance personnel; in-house ML pathway includes a small ML team plus data labeling plus drift retraining plus compute; DFT pathway includes specialized computational chemists plus HPC allocation. These are not headline list prices — they reflect what enterprise pharma programs actually spend over a fiscal year. FluxMateria pricing is enterprise-tiered and disclosed under NDA.

Operational implications

Consolidating eight ADMET endpoints into a single mechanism-aware API call enables a class of design and review operations that the multi-vendor architecture constrains by integration overhead and schema mismatch.

Honest scope

A unified pipeline that beats stitched commercial stacks on TCO, schema, and three of eight endpoints' raw accuracy is a strong claim. It deserves a clean fence around what is and is not in scope.

The accuracy and TCO numbers in this case study apply to the cohort and endpoints documented in the public ADMET benchmark. Extending to new endpoints or modalities is a documented engineering process — not a free claim.

Conclusion

FluxMateria delivers eight ADMET endpoints in a single mechanism-aware API call, validated against a 245-compound FDA-approved drug panel and the larger leave-one-out reference cohorts (PPB n=14,288; metabolism n=38,576; CYP panel n=62,794; solubility n=9,982; permeability n=41,175; hERG n=8,879; BBB n=7,807; DILIRank n=907; TDC binary DILI n=475; Hepatotox validated n=614). Three of the eight endpoints are at strict #1 SOTA on the public leaderboards. DILI now reaches SOTA accuracy on the comparable public binary benchmark: AUROC 0.9597 vs MiniMol ~0.956, while also returning mechanism, exposure, dose-window behavior, confidence, and score-trace detail. Annualized capability cost sits below the four enterprise pathways a discovery program faces today: stitched commercial ADMET stacks, in-house ML pipelines, DFT-based mechanistic ADMET, and free or sanity-check tooling.

Multi-vendor ADMET workflows reflect the assay-by-assay history of the field rather than the structure of the underlying physics. A unified first-principles model returns eight endpoints, per-prediction confidence, and mechanism-evidence trail as a single output document, with no cross-tool reconciliation required.

Technical specifications