SolvAI

Read the manuscript and supplement · Explore the results · Run a prediction · Reproduce the paper

SolvAI turns existing physical and empirical solvent data into 15 molecular response descriptors predicted from structure. The descriptors can be combined with different endpoint models to predict hydration free energy. After training, a query starts with a SMILES string; the released predictor performs no new solvent simulation.

The research question is not whether one model can imitate a particular simulation. It is whether solvent-response information can be learned once, reused across molecules, and remain useful when the hydration predictor changes.

The results

The same 15 frozen response predictions were tested across tree ensembles, neural networks, message-passing graphs, pretrained molecular encoders, TabM and TabPFN-3.5. The lowest absolute MAE and the strongest benefit from adding responses are different questions. Seven of nine fixed endpoint configurations show a response-associated reduction on External-220 and its nested Strict-97 subset.

Endpoint with responses	ARROW-85	External-220	Strict-97	Size holdout
ExtraTrees	0.202	1.153	1.536	1.772
Residual MLP	0.340	1.428	2.016	1.813
Dual branch	0.288	1.234	1.616	1.560
Ridge + neural residual	0.336	1.172	1.429	1.391
Message passing	0.288	1.168	1.505	1.729
MoLFormer + neural head	0.273	1.111	1.450	1.609
TabM	0.291	1.132	1.454	1.389
TabPFN-3.5	0.199	1.134	1.493	1.318
MoLFormer + ExtraTrees	0.234	1.174	1.564	Not evaluated

All entries include responses and are MAE in kcal mol⁻¹. Bold marks the lowest displayed point estimate per column, not a statistically established winner. The size column is a separate exploratory refit on 1,160 training and 205 held-out molecules; fixed source surrogates and model-specific input recipes limit what can be attributed to endpoint architecture alone. Full no-response comparisons and intervals are in the paper, Supplementary Data 6 and the endpoint reproduction guide.

The experiments also reveal a source-composition trade-off: all 15 descriptors are better on ARROW-85, while a fixed ten-descriptor subset derived from computed sources has lower error externally. There is no universally preferred endpoint or source mixture in these data.

Matched response controls

In the ExtraTrees implementation, adding the 15 predicted responses changes only the response feature block; labels, folds, structural features, weights and model seeds remain fixed.

Evaluation	Structure only	+ Responses	Molecules
ARROW-85, out of fold	0.303	0.202	85
External-220, endpoint-disjoint	1.532	1.153	220
Strict-97, also response-source-disjoint	2.138	1.536	97

Values are MAE in kcal mol⁻¹. The strict 97 are a subset of the external 220, not a separate fit. The published ARROW/PIMD8 value is 0.205 on the same 85 molecules: an accuracy reference, not a SolvAI training target or evidence of physical equivalence. Shuffling molecule-to-response assignments removes the matched ARROW gain. These alignment and chemical-separation controls were run for ExtraTrees; the broader endpoint study uses its own matched response/no-response comparisons.

See the paper's full workflow figure

Quick start

Requires Python 3.11, uv and Git LFS. The released SMILES-only command uses the packaged ExtraTrees endpoint; the other endpoint families are research comparisons, not interchangeable public CLI checkpoints. On a fresh clone, pull the LFS model files before predicting.

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

The command prints the input SMILES, predicted hydration free energy and seed-ensemble spread, in kcal mol⁻¹. From Python:

from solv_ai import predict_smiles

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

The spread is not a calibrated per-molecule uncertainty estimate. On the release CPU host, the warm single-query median was 15.29 s; a batch of 32 took 15.82 s (0.494 s per molecule). These are measurements of the packaged ExtraTrees stack, not a same-hardware comparison with PIMD or the alternative endpoints. See the model card for the inference boundary.

Reproduce and inspect

The combined manuscript contains the main text, references and Supplementary Information in one continuously numbered PDF. Its figures are editable SVG/PDF assets; the Figure 1 illustrations are explicitly non-evidential. With a LaTeX installation, rebuild the reading copy using:

make journal-paper

The frozen-artifact route reproduces the original confirmatory metrics and displays without new physical calculations; the extended architecture study has separate scripts, results and licensed dependencies. Start with the quick reproduction guide, then the full protocol, endpoint-study guide and data provenance. The confirmatory freeze and external-cohort freeze record their respective evaluation boundaries.

Find	Location
Unified manuscript, source and vector figures	paper/
Supplementary Data 1–5	paper/supplementary_data/
Endpoint comparison and size diagnostics	results/endpoint_models_20261006/
Public inference code and model card	solv_ai/ · models/final/
Confirmatory predictions and audits	results/confirmatory/ · audits/confirmatory/
Reproduction instructions and source manifests	repro/

Scope and limits

The validated application is hydration free energy for neutral small molecules near room temperature. The architecture search reused already examined evaluation sets and is retrospective. The size holdout is exploratory and does not exclude its molecules from every upstream source dataset. The peptide series have no experimental reference values: a less saturated curve is not proof of physically accurate extrapolation or extensivity. Original pretraining membership of external molecular encoders was not audited, and native TabPFN predictive intervals under-cover the external observations. SolvAI does not yet provide conformer-resolved energies, forces, other solvents, temperatures, ions or biomacromolecular validation.

Citation and license

Gal Oren, Boris Fain and Michael Levitt. A reusable solvent-response framework for molecular hydration prediction. Manuscript and citation metadata are in paper/main.pdf and CITATION.cff. The code is MIT licensed; third-party data and model weights retain the terms listed in data provenance. TabPFN weights require the provider's license, and fitted contexts containing training labels are not redistributed.