DISSOLVR: An Interpretable and Fast Framework for Aqueous and Organic Solubility Prediction

As a researcher, I have meticulously analyzed both provided texts regarding DISSOLVR. The information is highly consistent but presented from two different perspectives: Paper A focuses on the framework's architecture, mathematical constraints, and core contributions, while Paper B details the implementation of its LLM-assisted explanation pipeline.

My synthesis will integrate these elements to create a comprehensive, high-fidelity summary suitable for a detailed technical review.

Comprehensive Research Summary: DISSOLVR Framework

DISSOLVR is introduced as a novel, transparent framework designed for molecular solubility prediction that achieves state-of-the-art performance without incurring the computational overhead or complexity associated with deep learning architectures. The core philosophy of DISSOLVR is to replace opaque models with explicitly engineered, interpretable molecular descriptors derived directly from structural and physicochemical information.

Core Architectural Philosophy and Objectives

The framework is fundamentally guided by three primary objectives:

1. Accuracy: Achieving predictive accuracy approaching the aleatoric limit (0.6–0.8 10(S) units).

2. Efficiency: Utilizing computationally inexpensive features and maintaining low training complexity, scaling favorably as O(T NKL) for training and O(T L) for inference, enabling deployment on commodity CPUs.

3. Symbolic Accessibility: Being expressed in a manner that is symbolically accessible, specifically as a weighted sum of logical predicates.

DISSOLVR achieves this transparency by employing a Gradient Boosted Decision Tree (GBDT) ensemble as its primary regressor, which is explicitly contrasted with non-transparent deep learning baselines.

Molecular Descriptor Engineering: The Three Pillars

The framework constructs a comprehensive state vector for each molecule by encoding molecular properties across three distinct pillars:

1. Moietal Identity (Subgraph Isomorphism Census): This pillar quantifies functional moieties by counting subgraph isomorphisms, providing a structural census of the molecule's functional groups.

2. Topological Structure (Homomorphism Profiles): This pillar captures steric and connectivity patterns using Motif Structural Encoding (MoSE).

3. Thermodynamic Reconstruction (Energetics): This is the most complex component, utilizing Joback group-contribution methods to generate lattice energy proxies, alongside physicochemical proxies to reconstruct Abraham solvation parameters. This is concatenated with electronic structure data, surface area distributions, and druglikeness indices to form the final input vector.

Predictive Regimes and Interaction Modeling

DISSOLVR distinguishes its methodology based on the solvent environment:

• Regime (i) - Fixed Solvent Prediction: For prediction in a single, fixed solvent (e.g., aqueous solubility), the framework utilizes a descriptor-only regressor trained on the solute-centric descriptor vector.

• Regime (ii) - Multiple Solvent Prediction: To model complex, solvent-dependent interactions governed by the like dissolves like principle, DISSOLVR introduces an Interaction Layer. This layer treats the solute and solvent as distinct entities to compute a 24 times 24 matrix quantifying feature-level coupling. The final prediction fuses this interaction-aware representation with thermodynamic state variables (temperature, inverse temperature, predicted melting point, reduced temperature) within a physics-constrained framework assuming an endothermic dissolution regime.

Physics-Informed Constraints and Model Capacity Control

A critical architectural feature is the application of physics-aware monotonic constraints on temperature features. These constraints enforce physical consistency with the standard endothermic dissolution regime by ensuring that the predicted solubility obeys coordinate-wise thermodynamic constraints (Equation 7). This constraint serves a dual purpose: it restricts the model's capacity, preventing it from fitting arbitrary noise beyond the aleatoric limit, and guarantees physical plausibility.

LLM-Assisted Post-Hoc Explanation Pipeline

To bridge the gap between complex model evidence and human understanding, DISSOLVR incorporates an advanced LLM-assisted post-hoc explainer. This pipeline employs a six-stage progressive refinement strategy designed to ensure explanations are chemically grounded and sensitive to model evidence:

1. Stage 1 (Molecular Characterization): Generates unbiased structural descriptions from the input SMILES string.

2. Stage 2 (Evidence Synthesis): Analyzes SHAP values and group contributions to identify salient patterns and detect anomalies within the model's decision process.

3. Stage 3 (Decision Logic Analysis): Interprets decision splits and coupling effects derived from the Interaction Layer matrix, focusing on how solute-solvent interactions influence predictions.

4. Stage 4 (Mechanistic Integration): Synthesizes all synthesized evidence into a coherent explanation grounded in established chemical concepts, deliberately avoiding raw feature names to prioritize chemical intuition.

5. Stage 5 (Validation): A crucial self-correction step where the integrated explanation is cross-referenced against the model evidence supplied in earlier stages, flagging and rewriting any claims that lack direct support from the model's internal logic, thereby preventing hallucinations.

6. Stage 6 (Condensation): Condenses the detailed narrative into a concise, information-dense summary focusing on prediction summary, solute-solvent dynamics, mechanistic interpretation, and confidence/uncertainty.

A key design principle of this pipeline is staged disclosure: the LLM is intentionally blinded to the final numerical prediction quality until Stage 3. This forces the explainer to first summarize available model evidence before rationalizing the final output, significantly reducing confirmation bias. Furthermore, analysis (H.1) reveals a hierarchy of importance: Tier 1 drivers are global thermodynamic factors, while Tier 2 fine-tuning is achieved via interaction terms.

Experimental Validation and Conclusion

Experiments validate DISSOLVR across both regimes using datasets such as ESOL and AqSolDB for Regime I, and BigSolDB 1.0/2.0 for Regime II (multi-solvent). The framework demonstrates superior performance over various baselines in both environments, including strong generalization to out-of-distribution (OOD) solvent conditions. Empirical analysis confirms that the LLM explanations are highly faithful to the model's evidence; specifically, lower agreement between trained and untrained models is observed when checking against ablation studies, indicating that the explainer relies on specific structural evidence rather than generic parametric priors. Human expert evaluations confirm that the generated narratives are chemically coherent and persuasive, leading to a statistically significant increase in expert agreement with model predictions.

In conclusion, DISSOLVR establishes itself as a reproducible benchmark prioritizing interpretability and theory-grounded efficiency. It proves that current performance ceilings in solubility prediction are increasingly constrained by experimental noise rather than the depth of the underlying model architecture. While its limitations exist regarding stereochemistry, 3D conformation, and inorganic salts, DISSOLVR offers a powerful paradigm for building reliable, transparent predictive models.

1. The DISSOLVR framework can achieve state-of-the-art performance while remaining computationally efficient by using a Gradient Boosted Decision Tree framework that replaces non-transparent architectures with explicitly engineered molecular descriptors. This allows for faster training and inference than deep learning baselines and maintains a complexity of O(T · NKL) for training, making it deployable on commodity CPUs.

2. The model can enforce thermodynamic consistency under the standard endothermic-dissolution regime by using hard monotonic constraints on temperature features, which ensures the predicted solubility obeys the coordinate-wise thermodynamic constraints in Equation 7 and prevents thermodynamically inverted predictions that plague unconstrained ensembles.

3. The system can provide chemically grounded narratives for model decisions through an LLM-assisted post-hoc explanation pipeline, which uses a six-stage guardrail pipeline (Sec. 4) to map symbolic model artifacts, such as decision paths and interaction layer coefficients, into chemically grounded narratives.

4. The AI system can identify the specific drivers of solubility prediction by performing mechanistic feature importance analysis via SHAP on test set predictions, allowing researchers to map statistical feature importance directly to the four physicochemical pillars defined in our hypothesis.

5. The framework can generalize robustly to out-of-distribution (OOD) solvent environments through its structural invariance, as DISSOLVR's performance is not merely a stylistic choice: by collapsing the D×D cross-product into a 24×24 structured matrix, it both sharpens generalisation.

High-fidelity solubility prediction is fundamental to pharmaceutical development and environmental partitioning, where accurate modeling must couple molecular structure with thermodynamic behavior across diverse chemical environments. However, recent advancements have been dominated by deep learning architectures that often sacrifice physical interpretability for predictive power. We challenge this trend by showing that state-of-the-art performance does not require such non-transparent architectures. To address this, we introduce DISSOLVR, a transparent framework for molecular solubility prediction. In addition, we perform a comprehensive literature review and a benchmarking study against various methods. We show that DISSOLVR approaches the aleatoric limit of experimental uncertainty and achieves OOD generalization through structural invariance, derived by mapping molecules to physically-grounded descriptors. Then, we present an LLM-assisted post-hoc explanation pipeline that bridges the gap between symbolic model artifacts and chemically grounded narratives. Finally, a comparative benchmark of a survey involving 22 expert chemists reveals that expert evaluators provide deep insights.