EnSol: an environment-aware graph neural network for molecular solubility prediction
Thao Nguyen, Thao Nguyen, Saman Shafaei, Zhengyi Zhang, Huimin Zhao, Heng Ji
cs.LG · cs.AI
Quick verdict
EnSol adds cross-attention, FiLM temperature and MDN uncertainty to solubility GNNs; results are credible, but Leeds gains are marginal and stats thin.
Summary
The paper introduces EnSol, a probabilistic graph neural network for solubility prediction that jointly models solute, solvent, and temperature. Solute and solvent molecules are encoded by AttentiveFP GNNs, coupled through 4-head cross-attention, conditioned on temperature via RBF features and FiLM modulation of the solvent embedding, and parameterize a 3-component mixture density network trained by negative log-likelihood. Trained on BigSolDB (100,570 measurements; 1,375 solutes; 70 solvents; 243-426 K) with solute-level splits, EnSol is evaluated on deduplicated external benchmarks: on SolProp it reaches Spearman 0.876±0.005 and RMSE 0.824±0.032, versus 0.509/1.303 (FASTSOLV) and 0.569/1.700 (Vermeire); on Leeds it reaches 0.602±0.013/0.944±0.017, marginally ahead of FASTSOLV in ranking (0.593) but behind in RMSE (0.922). Ablations attribute gains to the encoder, cross-attention, FiLM conditioning, and the MDN head. Fine-tuning to aqueous data (AqSolDB, ESOL) yields modest, statistically insignificant gains over random initialization. Experimental validation on 78 measurable solute-solvent pairs (of 100, visual dissolution assay, 1 mg/mL detection limit) gives Spearman 0.715±0.035 and RMSE 0.699±0.062, beating both baselines. Main limitations: no significance testing for headline comparisons, inconsistent seed counts (three vs five) in the transfer section, exclusion of 22 censored pairs from metrics, no uncertainty-calibration evaluation, and a coarse experimental assay.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- EnSol achieves Spearman 0.876±0.005 on SolProp and 0.602±0.013 on Leeds on deduplicated external benchmarks, outperforming FASTSOLV and the Vermeire model, with the Leeds ranking gain marginal over FASTSOLV.
- Each architectural component (AttentiveFP encoder, cross-attention, FiLM temperature conditioning, MDN head) contributes measurably, per one-factor ablation deltas of -0.042 to -0.088 Spearman on SolProp.
- In prospective experiments on 78 measurable solute-solvent pairs at 298.15 K, EnSol reached Spearman 0.715±0.035 and RMSE 0.699±0.062, beating both baselines and supporting reliable solvent ranking.
- BigSolDB pretraining provides a modest, statistically insignificant transfer benefit for aqueous solubility prediction, most visible when target data are limited.
Strengths
Leakage-aware evaluation design
Solute-level 90:10 splits (Sec. 4.3), removal of all benchmark solutes from BigSolDB before training (Sec. 4.1), and InChIKey-based deduplication for transfer datasets (Sec. 4.1) directly address the most common failure mode in solubility ML.
Component-isolating ablations
Sec. 2.4 (Table S3) swaps exactly one component at a time over three seeds: MPNN encoder costs -0.088 Spearman/+0.187 RMSE, removing cross-attention -0.088/+0.104, temperature concatenation -0.042/+0.092, MSE head -0.043/+0.100 - a clean one-factor-at-a-time design.
Prospective wet-lab validation with honest failure analysis
Honest transfer-learning reporting
Well-engineered conditioning design
Weaknesses
No statistical inference for headline comparisons
Abstract overclaims on Leeds
Uncertainty - a headline feature - is never evaluated
Censored, replicate-free experimental assay
Internal inconsistencies and citation errors
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
EnSol (arXiv:2609.21151, cs.LG) is a method paper proposing an environment-aware probabilistic GNN for solubility prediction across solute-solvent-temperature triplets, paired with a prospective wet-lab solvent-screening validation. Its central claims: (1) joint graph encoding with cross-attention, FiLM temperature conditioning, and a mixture density head outperforms FASTSOLV and the Vermeire model on deduplicated SolProp and Leeds benchmarks; (2) each architectural component contributes measurably (Sec. 2.4); (3) predictions transfer to prospective experiments (Spearman 0.715 on 78 measured pairs, Sec. 2.5). The claims are mostly supported at the level of point estimates, but statistical support is thin, and the abstract's phrase outperforming state-of-the-art models across both benchmarks outruns the Leeds evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EnSol and what does it predict?
EnSol is a deep learning model that predicts the solubility of a molecule in a specified organic solvent at a given temperature. It takes the molecular graphs of the solute and solvent plus the temperature as input and outputs a full probability distribution over solubility (as log values), giving both a point estimate and an uncertainty.
How accurate is EnSol compared to other solubility models?
On the deduplicated external SolProp benchmark it reached Spearman 0.876 and RMSE 0.824, versus 0.509/1.303 for FASTSOLV and 0.569/1.700 for the Vermeire model. On the Leeds benchmark it reached 0.602/0.944, roughly on par with FASTSOLV (0.593/0.922 - FASTSOLV slightly better on RMSE) and far ahead of Vermeire (0.245/2.026). In 78 experimentally measured solute-solvent pairs it reached Spearman 0.715, above both baselines.
What data was EnSol trained on?
BigSolDB: about 100,570 experimental solubility measurements covering 1,375 solutes, 70 organic solvents, and temperatures from 243.15 to 425.77 K. Solutes appearing in the SolProp and Leeds test benchmarks were removed from the training set to prevent data leakage.
Is the EnSol code and data available?
Yes - the paper states that source code and training data are available at github.com/thaonguyen217/EnSol. The underlying datasets (BigSolDB, SolProp, Leeds, AqSolDB, ESOL) are public, so the computational results are in principle independently checkable.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-5 and supplementary tables S1-S5 are referenced but not extractable from text; all numeric claims rely on in-text values only.
- The wet-lab protocol's replicate structure and raw measurement data cannot be verified from the extracted text.
- Reference numbering contains a collision ([26] used for both AttentiveFP and AqSolDB), so citation-based cross-checks are partly unreliable.
- The paper is a preprint (arXiv:2609.21151) and has apparently not yet completed peer review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.