🧬 FLUXMATERIA — LIFE SCIENCE

Docking, without force-field parameterisation

Pose and affinity for any ligand in any published pocket. Covalent warheads, metal-site geometry, bridging waters, kinetics, and flexible side-chain refinement — in one studio, no force field to fit.

Pose & affinity Covalent warheads Metal sites Bridging waters Flexible residues

See FluxTarget (MoA / target) Request Pilot Access

47.4%

Top-1 pose accuracy on PoseBusters-428, blind

Parameters fitted — no training data, no reference values

Interaction types profiled per pose

Covalent warhead chemistries supported

200 FU

Metered per docking run — predictable pricing

The breakthrough

The ligand settles. Nothing was fitted to make it.

FLUX Settling does not search a scoring function someone tuned — the ligand relaxes into the pocket's physical field until it stops moving, the way it would in solution. The field is derived from first-principles physics and textbook atomic data. Nothing in it was regressed against binding data, pose data, or any benchmark, and no reference structure is consulted at prediction time. Measured blind on PoseBusters-428 it places the correct pose first in 47.4% of cases — and, unusually, it also tells you when to trust that answer.

Blind benchmark: PoseBusters-428

The standard blind pose-prediction set. Official scoring, all 428 complexes, no per-target tuning.

Method	Top-1	Top-5	Top-10	Fitted to data?
FluxMateria — FLUX Settling	47.4%	72.1%	76.1%	No
AutoDock Vina	60%	—	—	Force field fitted
GOLD	58%	—	—	Force field fitted
DiffDock	38%	—	—	Trained on the PDB

Read this honestly: Vina and GOLD are more accurate at top-1, and we do not claim otherwise. They are also fitted to binding data, and DiffDock is trained on the PDB itself. FluxMateria is fitted to nothing — the same field that predicts a pose was derived without ever seeing one. That is the trade we are making, and it is the reason the method behaves predictably on chemistry no training set covered.

67.9%

A correct pose is in the top 3 — the curve is steep to k=3 and flat after, so three poses capture what the method knows.

76.1%

A correct pose is somewhere in the returned set. Ranking it first is the open problem, not finding it.

~12%

Of correct poses lost to physical-validity checks — against 68% for the deep-learning entrant. Physical geometry by construction, not by filtering.

8–22 s

Per ligand, single core. Free-energy methods take hours for the same question.

It tells you when to trust it

A per-case confidence, validated on held-out complexes the threshold was never tuned on.

If you take the most confident…	Top-1 is correct
20% of cases	69.4%
40% of cases	66.7%
50% of cases	60.5%
every case (no filter)	46.3%

Why this matters more than the headline: A method that is right half the time is hard to act on. A method that is right half the time and knows which half is a triage tool. On the most confident 40% of targets the pose is correct two times in three; everywhere else the module returns a short ranked set instead of a single answer, so you know when to look at three poses rather than one. The threshold is set on one half of the benchmark and the accuracy measured on the other — it is not tuned on the number it reports.

Note on the baseline: this table is computed over the held-out split, so its unfiltered rate (46.3%) sits marginally below the 47.4% full-benchmark headline. Both are reported as measured; neither is rounded toward the other.

What Docking does

One studio for pose, affinity, kinetics, and the chemistry that lives inside the pocket.

🎯

Pose prediction

FLUX Settling — the ligand relaxes into the pocket's field, with a multi-conformer ensemble and flexible ligand torsions. Benchmarked blind at 47.4% top-1 on PoseBusters-428.

📏

Affinity & breakdown

ΔG = ΔH − TΔS + ΔG_solv. Every term is reported — van der Waals, electrostatic, H-bond, desolvation, conformational entropy, translational / rotational penalty.

🤝

Interaction profile

Hydrogen bonds, salt bridges, hydrophobic contacts, π-stacking, cation-π, halogen bonds — with geometric criteria and per-residue contributions.

🔗

Covalent warheads

Michael addition, acylation, disulfide, vinyl sulfone, epoxide chemistries. Returns bond energy plus the effective K_i.

⚙️

Metal-site geometry

Detects Zn, Fe, Mg, Cu, and other metal centres, identifies coordinating residues, and reports the coordination geometry.

💧

Bridging waters

Finds water-mediated hydrogen-bond bridges between ligand and pocket — the interactions structure-based tools often miss.

🧪

Flexible side-chain refinement

Rotamer-library sweep on residues within 5 Å of the ligand, re-scored per iteration. Rigid pocket by default, opt-in flexible.

⏱️

Kinetics

k_on, k_off, and residence time from the activation-barrier estimate — useful for slow-off-rate drug design programs.

How a docking job is built

From a SMILES and a PDB to a ranked pose set in one workflow.

Prepare receptor

PDB in, grids out. Pre-compute the van der Waals, electrostatic, H-bond donor / acceptor, and desolvation grids around the binding site.

Generate 3D ligand

Single conformation from the SMILES, or a multi-conformer ensemble for flexible ligands ranked by score.

Settle into the field

Each starting placement relaxes into the pocket's physical field — position, orientation and ligand torsions moving together — until it stops. No scoring function is searched; the pose is where the physics comes to rest.

Rank & score confidence

Distinct settled poses are ranked and de-duplicated, and the case is assigned a confidence so you know whether to act on the top pose or review the short list.

Profile interactions

Hydrogen bonds, salt bridges, hydrophobic, π-stacking, cation-π, halogen bonds — per pose, with per-residue attribution.

Affinity & kinetics

Full ΔG breakdown plus k_on / k_off / residence time. Export CSV / JSON / PDB poses or hand off to Workspace.

Why you can trust it

Deterministic physics, auditable formulas, and honest scope.

Deterministic

Same ligand + pocket + seed returns the same pose set and the same ΔG, run after run.

Auditable

Every contribution to ΔG is reported — VDW, electrostatic, H-bond, desolvation, entropy terms. No black-box score.

Geometry-validated

H-bond and hydrophobic criteria match literature geometric ranges (2.8–3.0 Å / 155–165° / 3.4–4.0 Å).

6 chemistries

Covalent warhead coverage includes Michael, acylation, disulfide, vinyl sulfone, and epoxide.

5 Å flex

Flexible side-chain refinement on residues within 5 Å of the ligand, with up to three iterations.

Force-field parameters fit. No training data; no per-target re-parameterisation.

How FluxMateria compares

Head-to-head against the common sources of docking poses and affinities.

Metric	FluxMateria	AutoDock Vina	FEP / MM-PBSA	ML scoring
Training / fitting required	None	Force-field parameters	Force-field parameters	Thousands of poses
Blind PoseBusters top-1	47.4%	60%	Not a pose method	Depends on input poses
Pose engine	Field settling	Iterated local search	Usually reuses pose	Rescoring only
Per-case confidence	Yes, validated	Not provided	Not provided	Score only
Interaction profile built-in	6 types + metal + water	Limited	Limited	Not provided
Covalent warheads	5 chemistries	Custom	Manual	Not provided
Kinetics (k_on / k_off)	Built-in	Not provided	Separate workflow	Not provided
Bridging waters	Detected automatically	No	Explicit modelling	No
Runtime per ligand	Sub-minute pose set	Sub-minute	Hours	Milliseconds (rescore)
Out-of-distribution behaviour	Degrades gracefully	Force-field-limited	Force-field-limited	Confidently wrong

The key insight: Classic docking is fast but depends on hand-tuned force fields. Free-energy methods are accurate but take hours per ligand. ML rescoring is fast but only as good as the poses someone else generated — and as good as its training set, which is why it fails quietly on novel chemistry. FluxMateria lets the ligand settle into first-principles physics, profiles the full interaction pattern, and returns affinity + kinetics in one pass — with nothing fitted anywhere in the pipeline, and a confidence score telling you which answers to act on. See FluxTarget for the MoA context →

Where Docking wins

Workflows where the pocket chemistry is as important as the pose.

Use case 1

Covalent warhead triage

Michael acceptors, acyl fluorides, vinyl sulfones — pick the warhead, let the engine score pose and bond energy, rank by effective K_i.

Use case 2

Metalloenzyme inhibitor design

Detects the metal site and reports coordination geometry, so ligands are evaluated in the chemistry that actually sits in the pocket.

Use case 3

Water-bridge-sensitive series

Pockets where the crystal water is the key interaction. Bridging-water detection surfaces the ligands that keep it versus the ones that displace it.

Use case 4

Off-rate programs

k_on / k_off / residence time per pose, useful for the kinetics-driven discovery programs where slow-off-rate is the goal.

Use case 5

Flexible-pocket challenges

GPCRs and kinases with mobile side chains. 5 Å flexible refinement pass re-scores the top poses with the correct rotamers.

Use case 6

FluxTarget hand-off

When FluxTarget flags an off-target of interest, open it in Docking to confirm the pose and affinity before chasing it into the lab.

Docking in the product

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

Target selection panel with search and filter-by-category — **Target selection**Search and filter the curated pocket library, preview PDB metadata and binding-site residues.

New docking dialog with ligand entry, exhaustiveness, pose count, and flexible residue toggle — **New job**SMILES / SDF / CSV ligand entry, exhaustiveness slider, pose count, and flexible-residue toggle.

Active jobs table with status, progress bar, cancel button — **Jobs**Active and completed docking jobs with status, progress, best-score badge, and cancel / view / download actions.

Results dialog with ligand rows, score, RMSD, interaction summary icons — **Results**Ligand rows with score, RMSD, pose count, and interaction-summary badges (H / hydrophobic / π / salt / halogen).

Scope & Limitations

Strengths

Validated blind on a public benchmark — 47.4% top-1 on PoseBusters-428 — with zero parameters fitted to any binding or pose data.
Per-case confidence that holds up on held-out complexes: 66.7% correct on the most confident 40% of targets.
Physically valid geometry by construction, not by filtering — ~12% of correct poses lost to validity checks, against 68% for the deep-learning entrant.
Covalent warheads, metal sites, and bridging waters are first-class features, not plug-ins.
Full ΔG breakdown + kinetics from one pass, deterministic and re-runnable.

Known limitations

Top-1 accuracy trails fitted methods by roughly 13 points (47.4% vs ~60% for Vina). We publish this rather than omit it; closing it is active work and no route to parity is claimed today.
A correct pose is present in the returned set 76.3% of the time but ranked first far less often — ranking, not sampling, is the open problem.
Affinity, kinetics and interaction profiling are reported from the pose; they have their own accuracy characteristics and are not covered by the pose benchmark above.
Rigid pocket is the default; flexible side-chain refinement is opt-in. Large induced-fit motions and full-protein dynamics belong in the MD module.
Peptide / macrocycle docking works best with an explicit multi-conformer seed rather than a single SMILES.

Dock your next series

Pilot access includes Docking, FluxTarget, the ADMET suite, and a Workspace seat to keep every run auditable.

Request Pilot Access →