Hundreds to thousands of ligands per second on a single GPU — with full physics accuracy and zero trained parameters. No fitted parameters. No training data. No black box. Every term traces back to a physical interaction.
LatticeZero computes physics-first binding interactions at hundreds to thousands of ligands per second on a single GPU. No fitted parameters. No training data. No black box. Every term traces back to a physical interaction.
A complete molecular docking platform built on physics first principles. Works on any target without calibration. No external dependencies.
Universal discrimination scorer. Adaptive pocket filtering matches each target's binding regime from population statistics alone. Zero per-target calibration. 0.916 mean AUC across 178 validated targets.
GPU-accelerated docking with proprietary AMR placement. Hundreds to thousands of ligands per second per GPU. No force field required.
Absolute binding affinity from 18 physics energy terms (Coulomb, dispersion, repulsion, H-bonds, burial, strain). R = 0.843 on CASF-2016.
Relative binding free energy prediction with pocket-geometry adaptive empirical routing. Predicts which physics formula works best for each congeneric series by analyzing pocket geometry on the fly. Mean Spearman ρ = 0.814 across 87 public FEP benchmarks (Schrödinger, OpenFF, Merck, JACS). Outperforms industry simulation methods at screening speed.
Automated receptor preparation. PDB cleanup, protonation, binding site detection, grid generation. Ready in minutes.
Bring your own labeled data to build custom scoring profiles for your target. Maximizes per-target AUC with cross-validated physics weights. Optional when you have active/decoy labels.
Adaptive geometric filtering matches each pocket's binding regime from population statistics alone. No per-target calibration.
Absolute and relative binding affinity from the same physics engine, at screening speed.
| Method | CASF-2016 R | Speed |
|---|---|---|
| LatticeZero | 0.843 | 100s–1,000s lig/sec |
| RF-Score v3 | 0.75 | ~1,000 lig/sec |
| FEP+ (AB-FEP) | 0.69 | hours per molecule |
| AutoDock Vina | 0.60 | ~50 lig/sec |
| Glide SP | 0.57 | ~100 lig/sec |
Predicts experimental relative binding free energies (ΔΔG) for congeneric ligand series using pocket-geometry adaptive empirical routing.
Beats or ties industry FEP+ references across Schrödinger, OpenFF, Merck, and JACS benchmarks. Geometry-aware routing automatically selects the optimal physics pathway for each target’s binding pocket.
From first-principles physics to GPU-accelerated scoring in four steps.
Scoring grounded in calibrated sterics and electrostatics. No parameters trained on binding data. Every weight traces to a physical interaction term.
Coulomb, dispersion, repulsion, H-bonds, pi-stacking, burial penalties, strain, and more. All computed from our geometric physics engine.
GPU-accelerated scoring via WebGPU compute shaders. Score hundreds of molecules per second on RTX-series GPUs, directly in the browser.
Pocket-geometry adaptive empirical routing — the docked ensemble reveals the pocket's binding regime and automatically routes each prediction through the most accurate physics formula. Zero calibration required.
Coulomb, dispersion, repulsion, strain, and hydrogen bond geometry computed per atom pair at hundreds to thousands of ligands per second on a single GPU. Not parameterized force fields. Not learned potentials. Computed from atomic coordinates and partial charges.
Population statistics of the docked ensemble reveal the pocket's binding regime. The optimal physics formula is selected automatically. Rigid pockets, flexible grooves, metal sites, allosteric cavities: each gets the physics it needs.
The same physics engine that discriminates actives from decoys also predicts absolute binding free energy (R = 0.843, CASF-2016) and relative affinity (ρ = 0.814, 87 FEP benchmarks). Screening, ranking, and affinity in one platform.
No training set required. No active/decoy labels needed. Upload a receptor PDB and a ligand library. Score in minutes.
Physics-derived scoring grounded in calibrated sterics and electrostatics. No parameters trained on binding data. No decoy bias.
Docking-calibrated softening derived from potential matching, not AUC optimization. Reduces clash artifacts while preserving pose discrimination.
Most ML scoring functions need active/decoy labels for every new target. IsoScreen scores any target with zero calibration - docking strain and physics terms determine the scoring pathway automatically.
Every score traces to interpretable physics terms. Coulomb, dispersion, repulsion, H-bonds, burial - all inspectable. No neural network hidden layers.
Full platform access during open beta. No credit card required.
Dock and score molecules in your browser. No signup required. Try IsoScreen on validated targets.
Launch demo →178 targets across DUD-E, DEKOIS2, and LIT-PCBA. Full methodology and comparison tables.
View results →Full platform access. Upload receptors, dock ligand libraries, export results. No credit card required.
Get started →
LatticeZero