Skip to content

Repository files navigation

SolvAI

SolvAI learns reusable solvent-response coordinates from calculated, empirical and corrected solvation data and predicts hydration free energy directly from molecular structure—without running simulation at inference.

SolvAI concept

The released system maps one SMILES string to one hydration free energy. Its response surrogates were trained on benchmark-disjoint quantum-continuum, alchemical, empirical and conformational data; the expensive source calculations are not rerun for a query. PIMD-derived features were tested but are not present in the final model.

Confirmatory result

On the 85-solute neutral-hydration reference set introduced with ARROW:

Method Simulation at inference? MAE (kcal/mol)
Classical ARROW yes 0.785
ARROW/PIMD8 yes 0.205
Matched structure-only endpoint no 0.303
SolvAI, fixed five-fold OOF no 0.202
SolvAI, five complete partitions no 0.207 ± 0.004
SolvAI, no ARROW labels in training no 0.257

The matched endpoint uses exactly the same experimental labels, descriptors, ExtraTrees architecture, weights, folds and seeds; only the 15 response priors are removed. The paired OOF improvement is −0.101 kcal/mol (95% bootstrap interval, −0.215 to −0.020). Shuffled priors do not improve the endpoint, and the advantage survives global family, scaffold, molecular-cluster and nearest-neighbour exclusions.

In a prospectively frozen external molecule-disjoint cohort, the same matched comparison lowers MAE from 1.532 to 1.153 kcal/mol (N=220). The advantage also survives in 97 molecules absent from all six response-teacher source tables (2.138 to 1.536 kcal/mol). These broader molecules establish transfer of the response-layer advantage, not PIMD8-level absolute accuracy outside ARROW-85.

The supported conclusion is PIMD8-level accuracy on this reference chemistry, not a general sub-0.20 claim. Global family and scaffold separation remain harder at 0.468 and 0.376 kcal/mol, respectively.

Install and predict

Python 3.11, uv and Git LFS are required.

git lfs install
git clone https://github.com/Scientific-Computing-Lab/SolvAI.git
cd SolvAI
make setup
uv run solvai predict 'CCO'

The command returns the ensemble-mean hydration free energy and ensemble spread in kcal/mol:

CCO    -5.012566    0.004714

The API is equally small:

from solv_ai import predict_smiles

prediction, spread = predict_smiles(["CCO", "c1ccccc1"])

On the release CPU host, the packaged artifact has a 15.29 s warm single-molecule median and processes a batch of 32 in 15.82 s (0.494 s per molecule). Startup of the two response D-MPNNs dominates single-query latency; no simulation is performed.

Reproduce the paper

make test && make verify && make figures && make paper

This quick path uses frozen, hash-verified artifacts to recompute predictions, metrics, tables, figures and PDFs. It does not rerun physical calculations or model training. The preregistered confirmation protocol is in release/CONFIRMATORY_FREEZE.md, with results in reports/CONFIRMATORY_ANALYSIS.md. The prospectively frozen external protocol and report are release/TIER_A_EXTERNAL_VALIDATION_FREEZE.md and reports/TIER_A_EXTERNAL_VALIDATION.md.

The compiled manuscript, Supplementary Information, including five supporting figures, and machine-readable Supplementary Data are included. See repro/QUICK_REPRODUCTION.md, repro/FULL_REPRODUCTION.md and repro/DATA_PROVENANCE.md.

Scientific safeguards

  • Exact and standardized benchmark equivalents are absent from all supervised external training sources used by the confirmatory model.
  • Every reported accuracy value is held out; the all-data deployment refit is never used as evidence.
  • Shuffled-prior, global chemical-separation and zero-ARROW-label controls are included molecule by molecule.
  • Tier-A eligibility was frozen before evaluation; all 220 rows are endpoint-disjoint and the strict 97-molecule subset is also response-source-disjoint.
  • Inference requires no experimental target, family/scaffold label, MD, PIMD, ARROW trajectory, probe or routing policy.
  • The released artifact contains no retained PIMD-trained feature.
  • The returned ensemble spread is not a calibrated applicability or reliability score; the validated scope is neutral small-molecule hydration chemistry.

Repository map

  • solv_ai/ — SMILES-only inference and metric code
  • models/final/ — standardized-exclusion response surrogates and endpoint ensemble
  • results/confirmatory/ — preregistered predictions, comparisons and statistics
  • audits/confirmatory/ — identity, similarity and refit audits
  • paper/ — Nature Communications manuscript and Supplementary files
  • repro/ — quick/full reproduction and data provenance

Citation metadata are provided in CITATION.cff. Code is MIT licensed; external datasets retain the terms listed in the provenance record.

About

Physics-distilled, structure-only hydration free-energy prediction without simulation at inference

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages