← Benchmarks | BioTarget

BioTarget Predictions BENCHMARK

Binding affinity and mechanism of action — validated against CASF-2016 and ChEMBL benchmarks.

95%
MoA Accuracy
100 ChEMBL pairs
0.617
Pearson r
CASF-2016 (269 complexes)
10,065
Targets
5 kingdoms
1.335
MAE (pKi)
CASF-2016 binding affinity

Validated Capabilities

Binding Affinity Prediction — CASF-2016

  • Pearson r = 0.617 on 269 CASF-2016 complexes (Su et al. 2019), scoring deposited crystal poses
  • MAE = 1.335 pKi
  • Every prediction deterministic and fully auditable
  • Deterministic, fully auditable affinity inference
  • 5,000+ predictions/sec (~300,000x faster than conventional docking)

Literature Context

These are not apples-to-apples comparisons. Every other method starts from a resolved structure with substantially more binding information already exposed. FluxTarget does not; it must infer the missing context before it can score affinity. Many deep-learning baselines also rely on large structure-supervised training sets.
Method Pearson r MAE (pKi) Required Input Training Data
DEEP LEARNING + 3D CO-CRYSTAL STRUCTURE + LARGE TRAINING SETS
graphDelta (GNN) 0.87 — Resolved bound complex Large structure-supervised set
Kdeep (CNN) 0.85 — Resolved bound complex Large structure-supervised set
GNINA (CNN) 0.82 ~1.0 Resolved bound complex Large structure-supervised set
CLASSICAL SCORING + 3D CO-CRYSTAL STRUCTURE
RF-Score v3 0.72 ~1.4 Resolved bound complex Moderate structure-based set
Glide SP 0.65 — Resolved bound complex Empirical (fitted to structures)
X-Score 0.61 — Resolved bound complex Empirical (fitted to structures)
AutoDock Vina 0.60–0.70 ~1.5–1.7 Resolved bound complex Empirical (fitted to structures)
FluxMateria FluxTarget 0.617 1.335 Resolved bound complex Flux-derived, no machine-learning training

All methods are benchmarked on CASF-2016 scoring power, scoring the deposited crystal pose of each complex — the same input class for every row. The comparison methods are empirically fitted to structural training sets; FluxTarget's score is computed from Flux-derived physics with no machine-learning training.

Download per-complex results
269 CASF-2016 complexes · PDB ID, target, experimental pKd, predicted pKi, error
Download CSV

Scoring Power — Crystal Poses (No ML Training)

When given the experimentally resolved crystal pose (the same input every classical and ML scoring function receives), the physics-derived scoring function achieves competitive correlation with no machine-learning training — every constant is derived from first principles.

Benchmark n Pearson r Spearman ρ Bias (pKi) MAE (pKi) RMSE (pKi)
CASF-2016 269 +0.617 +0.570 −0.26 1.33 1.73
BindingDB blind holdout post-2016 106 +0.534 +0.579 −0.63 1.51 1.88
BioLiP/MOAD blind holdout 53 +0.495 +0.452 +0.06 1.40 1.73
Weighted average 428 1.38

BindingDB holdout: post-2016 human targets, Ki+Kd, CASF-disjoint, cleaned (zinc metalloenzymes and duplicate PDBs removed). BioLiP/MOAD holdout: human inhibitor-like ligands, QC-strict. Both holdouts are temporally and structurally disjoint from CASF-2016 and were never used during development.

CASF-2016 scoring power ranking

Method Pearson r Type Fitted parameters
DEEP LEARNING (TRAINED ON PDBbind)
graphDelta 0.870 GNN Millions
ECIF::LD-GBT 0.866 ML Millions
OnionNet-2 0.864 CNN Millions
KDEEP 0.820 3D-CNN Millions
Pafnucy 0.780 3D-CNN Millions
RF-Score v3 0.754 Random forest Thousands
CLASSICAL / EMPIRICAL (FITTED WEIGHTS)
Lin_F9 0.680 Linear empirical 9 fitted terms
X-Score 0.631 Empirical Fitted weights
FIRST-PRINCIPLES PHYSICS (NO MACHINE-LEARNING TRAINING)
FluxMateria 0.617 Physics-derived 0
CLASSICAL / EMPIRICAL (continued)
ChemPLP@GOLD 0.614 Empirical Fitted weights
AutoDock Vina 0.604 Empirical Fitted weights
GlideScore-SP 0.513 Empirical Fitted weights
GlideScore-XP 0.467 Empirical Fitted weights
DFIRE / PMF / DrugScore 0.3–0.5 Knowledge-based Statistical potentials

Published r values from Su et al. 2019 (J. Chem. Inf. Model. 59, 895) and subsequent benchmark papers. All methods score crystal poses from CASF-2016. Every method above r ≈ 0.70 uses ML weights trained on PDBbind. FluxMateria is the only method above r = 0.60 with no machine-learning training — it beats AutoDock Vina, both Glide variants, MM-GBSA, and all knowledge-based scoring functions. The gap to X-Score (best classical) is 0.014, within bootstrap confidence interval overlap.

Target Identification

  • No published ranking metric yet — internal evaluation in progress
  • 10,065 targets across 5 biological kingdoms
  • Predicts likely protein targets for a given small molecule
  • Target and family context handled by the production scoring workflow

Mechanism of Action (MoA) Prediction

  • 95% accuracy on 100 confirmed compound–target pairs from the ChEMBL drug-mechanism table, spanning 12 targets; 90% on a 20-drug iconic set, 94.2% combined (run 2026-03-20)
  • Predicts agonist/antagonist/inhibitor classification when the binding pair is known
  • Correction-assisted mechanism classification, with pair-known and target-discovery scopes separated
  • Integrated with FluxTarget module
  • Scope: classifies the mechanism for a known compound–target pair. Polypharmacology / target-discovery (which targets a compound hits among thousands) is on the roadmap and tracked separately under “Selectivity Profiling” below.

Planned

Selectivity Profiling

Off-target binding predictions and selectivity scoring. Status: Planned.

Why these numbers matter more than they look

FluxTarget solves a fundamentally harder problem than every method it is compared against.

What every other method receives
✓ A resolved bound-complex structure with far more target-specific information than a raw molecular query
✓ A known interaction geometry rather than an inferred one
✓ For many methods, large structure-supervised training sets as well
They score a known answer. The hard part — finding the pose — is already done by the crystallographer.
What FluxTarget receives
1. A SMILES string — a flat, 2D text representation of the molecule (e.g. CC(=O)Oc1ccccc1C(=O)O)
2. A target name (e.g. “Thrombin”)
No resolved bound structure. No supplied pose. FluxTarget must infer the missing context computationally.

How FluxTarget closes the information gap

⚖
Builds 3D ligand
Constructs a usable structural hypothesis from the molecular input
⚙
Loads 3D protein
Contextualizes the target environment and plausible interaction region
⚡
Docks on GPU
Evaluates candidate interaction hypotheses with deterministic physics
★
Predicts affinity
Returns an affinity estimate with interpretable contributing factors
FluxTarget reconstructs missing structural context before scoring affinity. The methods it is compared against typically begin from a much more informative experimental starting point.

Benchmark basis

BioTarget benchmarks include affinity scoring, target identification, and workflow-level results. The table below labels the main result families.

Mixed basis
Result familyBasisHow to read it
CASF affinity scoringFlux-Calibrated PhysicsFlux docking and affinity terms with fixed benchmark-calibrated scoring.
Target identification / mechanism of actionFlux HybridFlux target signals are combined with target and compound-context evidence.
Selectivity and inverse-design workflowsMixed basisWorkflow results should be read with the specific benchmark row or source note.

Try FluxTarget on Your Data

Predict binding affinities across 10,000+ targets with full interpretability. CASF-2016 validated.

← Back to Module Join Research Preview