The Geometric Bypass
Physics-derived molecular scoring. 0.87 AUC on BACE1. 50-80 ligands/secโ5ร faster than Smina.
The "Fast & Smart" Region
LatticeZero achieves FEP-level accuracy at docking-level speed
Hero target average: EGFR (0.80), AA2AR (0.85), BACE1 (0.87) AUC
Empirical overfitting
Physics-derived scoring (vdW โ HBQ) grounded in calibrated sterics and electrostatics. No training data needed.
Steric clashes
Docking-calibrated sR softening reduces clash artifacts while keeping pose discrimination.
One-size-fits-all
Pocket-aware scoring with optional family ML heads for kinases, proteases, and GPCRs.
LatticeZero v1.8 Benchmark Results
9-target validation ยท December 2025 ยท +31% mean AUC over Smina
| Target | Class | AUC (vdW) | AUC (vdW+ML) | Status |
|---|---|---|---|---|
| BACE1 | Protease | 0.41 | 0.87 | Hero |
| AA2AR | GPCR | 0.64 | 0.85 | Hero |
| EGFR | Kinase | 0.58 | 0.80 | Hero |
Physics (vdW) vs Physics+ML (v1.2/v1.3) on hero targets
Runtime Performance
5-25ร faster than Smina on a 2-core VM
LatticeZero rescores 50-80 ligands/sec on a 2-core VMโ5ร faster than Smina score-only mode
Scoring Modes
Choose your physics-ML tradeoff
vdW
Pure physics: lattice-derived dispersion + CCSD(T)-calibrated sterics.
--score-mode vdw
AUC: 0.41-0.64
vdW + ML
Physics features + family-specific gradient boosting heads.
--score-mode vdw_ml
AUC: 0.78-0.86
HBQ
H-bond geometry + Coulomb. Under validation.
--score-mode hbq
Status: Quarantined
How It Works
Four layers from first principles to production scoring
The Lattice
Physical constants from geometric first principles. No fitting. No training. Just math.
RFT FoundationvdW Engine
Lattice-derived Cโ coefficients + CCSD(T)-calibrated sterics with sR softening.
5ร vs SminaHBQ Mode
Directional H-bonds + Coulomb electrostatics. 2,300x accelerated geometry detection.
Polar PocketsML Heads
Family-specific GBMs refine ranking. Physics stays interpretable.
Optional$ lz-score \
--receptor data/BACE1/receptor.pdb \
--poses-sdf data/BACE1/poses.sdf \
--score-mode vdw_ml \
--target-class protease \
--output results/bace1_scored.csv
# Output: 0.863 AUC on BACE1 benchmarkRequest Research Access
LatticeZero is in active development. Join the research preview for early access, benchmark drops, and priority support.
Open Source
Coming Q1 2026
- vdW scoring engine
- CLI interface
- Community benchmarks
Research Preview
Available Now
- Full HBQ mode
- Family ML heads
- Priority updates
- Direct support channel
Resources
Documentation, papers, and source code