🧬 FLUXMATERIA — LIFE SCIENCE

Binding affinity, off-target, microbiome,
from a SMILES and a target name

BioTarget scores pathogen efficacy, human off-target risk, and microbiome impact across 10,065 curated targets in 5 biological kingdoms. CASF-2016 validated. No crystal structure required. No training set. The whole multi-panel screen runs at 5,000+ predictions per second.

10,065 targets CASF-2016 validated Multi-panel selectivity ADMET fusion No ML
Open BioTarget benchmark Read the iptacopan case study
10,065
Curated targets across 5 biological kingdoms
r = 0.537
Pearson correlation on 270 CASF-2016 complexes
AUC 0.980
Target identification across the panel
5,000+ / s
Predictions per second — ~300,000× vs docking
0
Trained parameters · no crystal structure required
The breakthrough

CASF-2016-validated affinity, from a SMILES and a target name.

Every other method on the CASF-2016 leaderboard starts from a resolved bound complex and a large structure-supervised training set. BioTarget takes the flat SMILES and the target name, builds the 3D ligand, loads the protein environment, docks on GPU, and scores affinity — all with deterministic physics. Same benchmark, a fundamentally harder input — and 5,000+ predictions per second across pathogen, human-safety, and microbiome panels.

What BioTarget does

Not just a docking score — a program-aware decision bundle.

🎯

Targeted efficacy scoring

Rank compounds by predicted binding strength across 10,065 targets spanning 5 biological kingdoms from minimal molecular input.

🧷

Human off-target triage

Counter-screen across safety-critical human targets to catch liabilities before the wet lab sees them.

🦠

Microbiome impact lens

Estimate collateral binding across gut-relevant targets so microbiome disruption risk is a first-class output, not an afterthought.

📏

Selectivity & TI metrics

Therapeutic-index style margins (pathogen vs human) with every contribution traceable. 5,000+ predictions per second, roughly 300,000× vs conventional docking.

🧠

Mechanism of action

Agonist / antagonist / inhibitor classification at 91% accuracy on confirmed compound–target pairs from the ChEMBL drug-mechanism table — no trained parameters.

🔎

Target identification

AUC 0.980 on target ranking across the 10,065-target panel — given a molecule, surface the likely primary target and its kingdom.

💊

ADMET fusion

One-run developability gating (hERG / CYP / solubility / permeability / DILI) via the FluxMateria ADMET panel — “potent but undruggable” falls out automatically.

🧪

Auto test plan

Suggested counter-screens and confirmatory assays ranked by predicted liability — the decision artifact the wet lab actually acts on.

How a program screen is built

From SMILES + target name to a de-risked shortlist in a single pipeline.

1

Define program

Select pathogen target(s), choose counter-screen panels (human off-target, microbiome), set program priorities and ADMET guardrails.

2

Build 3D

Construct a usable 3D ligand hypothesis from SMILES; load the protein target environment and candidate interaction region.

3

Multi-target screen

GPU search and score plausible interaction hypotheses across the chosen panels — 5,000+ predictions per second.

4

ADMET fusion

Automatically gate by developability. Remove “potent but undruggable” candidates from the shortlist early.

5

Decision bundle

Ranked shortlist + risk map + CSV exports + narrative report. Counter-screen suggestions included for the next wet-lab round.

Why you can trust it

Benchmarked on published CASF-2016 and ChEMBL datasets, not on internal splits.

10,065
Curated targets across 5 biological kingdoms (bacterial, fungal, viral, parasitic, human).
r = 0.537
Pearson correlation on 270 CASF-2016 complexes (Su et al. 2019). MAE 1.90 pKi, bias −0.03 pKi, 62% within 2.0 pKi.
AUC 0.980
Target identification across the panel — given a molecule, the right target surfaces near the top.
91%
Mechanism-of-action accuracy (agonist / antagonist / inhibitor) on confirmed compound–target pairs from the ChEMBL drug-mechanism table.
5,000+ / s
Predictions per second on a single GPU — roughly 300,000× vs conventional docking.
0
Trained parameters · zero required crystal structures · deterministic, re-runnable.

How FluxMateria compares

Head-to-head on CASF-2016, with the required-input column that usually gets hidden.

MethodFluxMateriagraphDelta (GNN)GNINA (CNN)RF-Score v3AutoDock Vina
CASF-2016 Pearson r0.5370.870.820.720.60–0.70
Required inputSMILES + target nameResolved bound complexResolved bound complexResolved bound complexResolved bound complex
Training dataNoneLarge structure-supervised setLarge structure-supervised setModerate structure-based setEmpirical (fitted)
Target IDAUC 0.980 (10,065)Not providedNot providedNot providedNot provided
MoA classification91% on confirmed ChEMBL pairsNot providedNot providedNot providedNot provided
Multi-panel (pathogen / human / microbiome)Built-inOne run per targetOne run per targetOne run per targetOne run per target
ADMET fusionBuilt-inSeparate toolSeparate toolSeparate toolSeparate tool
Throughput (GPU)5,000+ / sSeconds eachSeconds eachMilliseconds (rescoring)Sub-minute pose set

The key insight: Every other method on CASF-2016 gets handed a resolved bound complex — BioTarget starts from a flat SMILES and has to build the whole structural context itself. On a fundamentally harder input, it still hits r = 0.537 and adds target-ID, MoA, multi-panel screening, and ADMET fusion that the ML methods don’t provide at all. See the full benchmark →

Where BioTarget wins

Drug-discovery workflows where program-level selectivity — not just binding — is the prize.

Use case 1

Antibiotic discovery

Pathogen efficacy + human off-target + microbiome-aware selection in one report. Ship only the molecules that clear all three filters.

Use case 2

Antiviral & antiparasitic

Rapid counter-screening to reduce host liabilities while maintaining target engagement across the 5 biological kingdoms in the panel.

Use case 3

Host-targeted therapy

Optimise selectivity within human target families and de-risk early ADMET failure modes in the same pipeline pass.

Use case 4

Lead optimisation

Modify the scaffold, re-run the panels, see exactly which risk moved and why. The feedback loop is seconds per iteration, not hours.

Use case 5

Pathogen-to-pipeline

Pair with the FluxTarget Discovery Wizard: pathogen in, BioTarget screens the suggested targets, ADMET filters the output, and the pipeline hands off to Workspace.

Use case 6

Audit & reproducibility

Every prediction is deterministic and grounded in physics. Re-running the same SMILES + target returns the same affinity bit-for-bit — IP reviewers and regulators see the same numbers.

Use cases 1 & 2 above sit inside the broader Flux Microbial Pharmacology adjacent domain — antibacterial, antiviral, antifungal, and antiparasitic pharmacology built on the same cascade architecture. The binding-affinity layer reaches most pathogen targets today via this engine; envelope crossing, MIC determination, and host-vs-pathogen selectivity composition are on the roadmap. See the full microbial cascade roadmap →

BioTarget in the product

Real captures from the live application. Click any image to zoom.

BioTarget program configuration with pathogen target, counter-screen panels, and ADMET guardrails
Program configurationPathogen target(s), counter-screen panels (human off-target, microbiome), and ADMET guardrails on one screen.
Ranked candidate shortlist with efficacy, selectivity, ADMET composite score
Ranked shortlistCandidates sorted by efficacy + selectivity + ADMET-weighted composite, with per-panel contributions broken out.
Risk map showing top off-targets, microbiome risk drivers, ADMET flags
Risk mapTop off-targets + gut-risk drivers + ADMET flags, annotated with the pose and contact that drove each.
Auto test-plan with recommended counter-screens and confirmatory assays
Auto test planRecommended counter-screens and confirmatory assays, ranked by predicted liability — the handoff the wet lab acts on.

Scope & Limitations

Strengths

  • CASF-2016 validated binding affinity from SMILES + target name alone — no crystal structure, no training set.
  • 10,065 curated targets across bacterial, fungal, viral, parasitic, and human panels.
  • Multi-panel screening (pathogen + human off-target + microbiome) in a single run — not three separate workflows.
  • ADMET fusion is built-in, so “potent but undruggable” candidates drop out of the shortlist automatically.
  • Deterministic and auditable — same SMILES + target returns the same affinity, MoA, and risk map every time.

Known limitations

  • Beta status. The target panel and workflow surfaces are actively expanding; published validations are added as they clear our criteria.
  • Affinity Pearson r (0.537) trails ML + crystal-structure methods (0.7–0.87) — by design, given the much harder input. Treat as a ranker, not as absolute pKi.
  • Induced-fit / ensemble handling is planned. Single-conformer target today.
  • Kinetics and mechanism-discovery integration for program-level hypotheses is on the near-term roadmap, not yet shipping.

Bring BioTarget to your program

Bring a target + a compound set (or a library). Pilot access returns a structured shortlist, risk map, and test plan. FluxTarget, ADMET, and Workspace are included.

Request Pilot Access →