A unified geometric engine powering hardware-accelerated drug discovery at LatticeZero.
Three foundational advances that enable physics-accurate molecular scoring at speeds previously impossible.
Most docking pipelines use fixed Cartesian sampling and spend substantial compute in low-value pocket volume. RFT uses a unified geometric framework with proprietary isotropic spatial packing mathematics to evaluate continuous protein pockets with higher volumetric efficiency. This reduces dead-space computation and enables sub-angstrom conformational searches.
Rather than patching together disconnected empirical approximations, the RFT engine evaluates 14 distinct physical interactions, including precise coordination geometries and second-shell hydrophobic effects, from a unified mathematical foundation.
The RFT engine was designed for matrix-native GPU parallelization, allowing complex energy tensor calculations to run client-side via WebGPU compute shaders at HTVS-class throughput. No server queues are required. Internal RTX 3090 benchmarks exceed 10,000 ligands per second, and internal mobile WebGPU tests on recent iPhone-class hardware reach multi-thousand ligands per second, including runs around 5,000 ligands per second. Workflows that once required cloud farms can now be achieved on a device in your pocket.
Validated against DUD-E and DEKOIS2 benchmarks with property-matched decoys. Every result includes holdout cross-validation and Y-scramble null controls.
| Target | Family | Validated AUC |
|---|---|---|
| HMGR | Reductase | 0.980 |
| PDE5 | Phosphodiesterase | 0.883 |
| Neuraminidase | Glycosidase | 0.875 |
| MCL1 | Apoptosis Regulator | 0.868 |
| ESR1 | Nuclear Receptor | 0.858 |
| ADRB2 | GPCR | 0.834 |
| CATL | Cysteine Protease | 0.829 |
Validated via 5-fold cross-validation with Y-scramble null controls. All AUC values are holdout-validated on DEKOIS 2.0 with property-matched decoys.
Beyond ranking poses, the same engine computes quantum-mechanical interaction energies directly from geometry, with zero fitted parameters and no training data. On gold-standard quantum-chemistry benchmarks it reaches the accuracy tier of machine-learning potentials trained on millions of reference calculations. Because the physics is computed rather than learned, accuracy holds across chemistry, not only the molecules it was built on.
| Method | Correlation (r) | MAE (kcal/mol) | Trained Parameters |
|---|---|---|---|
| eSEN | 0.996 | 14.4 | millions |
| GFN2-xTB | 0.992 | 10.6 | semi-empirical |
| UMA-S | 0.991 | 16.3 | millions |
| AIMNet2 | 0.985 | 35.8 | millions |
| RFT Engine | 0.92 | 21.7 | none |
| ANI-2x | 0.71 | 73.0 | millions |
| MACE | 0.78 | 112.2 | millions |
PLA15 protein-ligand interaction-energy benchmark (Kříž & Řezáž, J. Chem. Inf. Model. 2020), reference CCSD(T)/CBS. Comparator predictions from Rowan Scientific's published pla15-benchmarks (MIT-licensed); correlation and mean absolute error computed on the identical 15 structures. Every method at or above the RFT engine was trained on millions of quantum-chemistry calculations. The RFT engine uses none.
Cross-checked on standard physics benchmarks: S66 dimers ρ = 0.92 (vs SAPT2+), A24 ρ = 0.84 (vs near-full-CI), S66×8 / 528 points ρ = 0.90.
The RFT engine is available exclusively through the LatticeZero platform.
Explore LatticeZero