sci-form

High-performance 3D molecular conformer generation using ETKDG distance geometry, written in Rust.

Generates chemically valid 3D coordinates from SMILES strings, matching RDKit's ETKDGv2 quality while offering native bindings for Rust, Python, TypeScript/JavaScript (WASM), and a cross-platform CLI.

Features

ETKDG Distance Geometry — Cambridge Structural Database torsion preferences (837 SMARTS patterns)
High Accuracy — 0.00% heavy-atom RMSD > 0.5 Å vs RDKit on GDB-20 (2000 molecules, ensemble comparison)
Fast — 60+ molecules/second in Rust, parallel batch processing via rayon
Multi-platform — Rust lib, Python (PyO3), TypeScript/JS (WASM), CLI (Linux/macOS/Windows)
Zero dependencies at runtime — pure Rust, no C++ toolchain needed
SMILES + SMARTS — full SMILES parser and SMARTS pattern matching engine

Quick Start

Rust

[dependencies]
sci-form = "0.1"

let result = sci_form::embed("CCO", 42);
println!("Atoms: {}, Coords: {:?}", result.num_atoms, result.coords);

Python

pip install sci-form

import sci_form

result = sci_form.embed("CCO")
print(f"Atoms: {result.num_atoms}, Time: {result.time_ms:.1f}ms")
positions = result.get_positions()  # [(x, y, z), ...]

TypeScript / JavaScript

npm install sci-form

import { embed } from 'sci-form';

const result = JSON.parse(embed("CCO", 42));
console.log(`Atoms: ${result.num_atoms}`);

CLI

# Single molecule
sci-form embed "CCO" --format xyz

# Batch processing
sci-form batch -i molecules.smi -o output.sdf --format sdf --threads 8

# Parse only (no 3D)
sci-form parse "c1ccccc1"

Benchmark Results

Diverse Molecules (131 molecules, all chemical functional groups)

Metric	Value
Parse success	100%
Embed success	97.7%
Geometry quality	97.7%
Throughput	60 mol/s

RDKit Comparison (heavy-atom pairwise-distance RMSD)

Metric	Value
Average RMSD	0.064 Å
Median RMSD	0.011 Å
< 0.5 Å	98.4%
< 0.3 Å	94.4%

GDB-20 Ensemble (2000 molecules × 10 seeds vs 21 RDKit seeds)

Metric	All-atom	Heavy-atom
Avg RMSD	0.035 Å	0.018 Å
> 0.5 Å	0.95%	0.00%

Algorithm

sci-form implements the ETKDGv2 (Experimental Torsion Knowledge Distance Geometry) algorithm:

SMILES Parsing → Molecular graph with atoms, bonds, hybridization
Bounds Matrix → 1-2, 1-3, 1-4, and VdW distance bounds from topology
Triangle Smoothing → Floyd-Warshall triangle inequality enforcement
Distance Picking → Random distances from smoothed bounds (MinstdRand)
Metric Matrix Embedding → Eigendecomposition → 4D coordinates
Bounds Force Field → BFGS minimization in 4D to satisfy distance constraints
Projection to 3D → Drop lowest-variance dimension
ETKDG 3D Refinement — Force field with CSD torsion preferences (837 patterns)
Validation — Tetrahedral centers, planarity, double-bond geometry

See documentation for detailed algorithm descriptions with mathematical derivations.

Building from Source

# Library + CLI
cargo build --release

# Python bindings
cd crates/python && maturin develop --release

# WASM bindings
cd crates/wasm && wasm-pack build --target bundler --release

Testing

# Unit tests
cargo test --lib

# Diverse molecule benchmark
cargo test --release --test test_diverse_molecules -- --nocapture

# Geometry quality (requires GDB20.50000.smi)
cargo test --release --test test_geometry_quality -- --nocapture

# Gradient correctness
cargo test --release --test test_gradient_check -- --nocapture

License

MIT

sci-form 0.1.7