[Paper Review] MLSolv-A: A Novel Machine Learning-Based Prediction of Solvation Free Energies from Pairwise Atomistic Interactions

Hyuntae Lim, YounJoon Jung|arXiv (Cornell University)|May 13, 2020

Machine Learning in Materials Science93 references50 citations

TL;DR

This paper introduces MLSolv-A, a novel machine learning model that predicts solvation free energies by computing pairwise atomic interactions through inner products of atomistic feature vectors encoded by neural networks. The model achieves high accuracy (MUE ≈ 0.2 kcal/mol) and strong transferability across solvents, while enabling interpretable group contribution analysis of solvation energy, revealing clear correlations between hydrophilic groups and hydration free energy contributions.

ABSTRACT

Recent advances in machine learning and their applications have lead to the development of diverse structure-property relationship models for crucial chemical properties, and the solvation free energy is one of them. Here, we introduce a novel ML-based solvation model, which calculates the solvation energy from pairwise atomistic interactions. The novelty of the proposed model consists of a simple architecture: two encoding functions extract atomic feature vectors from the given chemical structure, while the inner product between two atomistic features calculates their interactions. The results on 6,493 experimental measurements achieve outstanding performance and transferability for enlarging training data due to its solvent-non-specific nature. Analysis of the interaction map shows there is a great potential that our model reproduces group contributions on the solvation energy, which makes us believe that the model not only provides the predicted target property but also gives us more detailed physicochemical insights.

Motivation & Objective

• To develop a machine learning model for solvation free energy prediction that is both accurate and interpretable.
• To enable transferability across different solvents by avoiding solvent-specific parameterization.
• To decompose solvation free energy into contributions from individual atomic interactions for mechanistic insight.
• To evaluate model performance under extreme extrapolation using scaffold-based data splits.
• To compare the effectiveness of RNN-based (BiLM) and graph-based (GCN) encoders in capturing atomic connectivity and solvation behavior.

Proposed method

• Uses two encoder networks (BiLM or GCN) to extract atomistic feature vectors from solvent and solute molecular structures.
• Computes pairwise atomic interaction scores via the dot product of encoded feature vectors: Iαγ = −pα · qγ.
• Aggregates all pairwise interactions to compute the total solvation free energy: ΔG°_sol = Σαγ Iαγ.
• Employs a solvent-non-specific architecture to enhance transferability across different solvent types.
• Utilizes t-SNE for visualization of molecular feature vectors and their correlation with hydration energy.
• Applies K-fold cross-validation with both random and scaffold-based (clustered) data splits to assess generalization and extrapolation performance.

Experimental results

Research questions

• RQ1 Can a machine learning model based on pairwise atomic interactions accurately predict solvation free energies across diverse solvents?
• RQ2 How does the model's performance vary under extreme extrapolation conditions, such as novel molecular scaffolds?
• RQ3 To what extent can the model decompose solvation free energy into contributions from individual atoms or functional groups?
• RQ4 How do different encoder architectures (BiLM vs. GCN) compare in capturing atomic connectivity and solvation behavior?
• RQ5 Does the solvent-non-specific design enhance transferability and generalization when training data is expanded with new solvent systems?

Key findings

• The MLSolv-A model achieves a mean unsigned error (MUE) of 0.19 kcal/mol with BiLM and 0.23 kcal/mol with GCN on 6,493 experimental data points from FreeSolv and Solv@TUM.
• Incorporating data from the Solv@TUM database improved model accuracy by 20–30% in terms of MUE, demonstrating strong transferability across solvents.
• Under scaffold-based cross-validation, which simulates extreme extrapolation, the model maintained an MUE of 0.60 kcal/mol, within the acceptable range of chemical accuracy (1.0 kcal/mol).
• The GCN-based model outperformed the BiLM-based model in predicting symmetric systems like benzene, indicating that explicit bond information is essential for accurate group contribution analysis.
• Visualization of feature vectors revealed a strong geometric correlation between molecular representations and predicted solvation energies, confirming effective feature learning.
• The interaction map clearly showed that hydrophilic groups contribute more significantly to hydration free energy, validating the model’s interpretability and physical relevance.