Abstract
Molecular property prediction is essential for drug discovery. In recent years, deep learning methods have been introduced to this area and achieved state-of-the-art performances. However, most of existing methods ignore the intrinsic relations between molecular properties which can be utilized to improve the performances of corresponding prediction tasks. In this paper, we propose a new approach, namely Physics properties Enhanced Molecular Property prediction (PEMP), to utilize relations between molecular properties revealed by previous physics theory and physical chemistry studies. Specifically, we enhance the training of the chemical and physiological property predictors with related physics property prediction tasks. We design two different methods for PEMP, respectively based on multi-task learning and transfer learning. Both methods include a model-agnostic molecule representation module and a property prediction module. In our implementation, we adopt both the state-of-the-art molecule embedding models under the supervised learning paradigm and the pretraining paradigm as the molecule representation module of PEMP, respectively. Experimental results on public benchmark MoleculeNet show that the proposed methods have the ability to outperform corresponding state-of-the-art models.
Introduction
In the world of drug discovery, predicting how a molecule behaves—its solubility, toxicity, or ability to cross the blood-brain barrier—is a high-stakes challenge. While deep learning models like Graph Neural Networks (GNNs) have shown promise, they face a massive bottleneck: data scarcity.
Measuring physiological properties requires expensive, time-consuming "wet-lab" experiments or animal studies, resulting in tiny datasets. Conversely, quantum physics properties (like a molecule’s dipole moment or internal energy) are intrinsic to the molecule's structure and can be calculated accurately using theoretical chemistry. Most existing AI models ignore the fundamental link between these calculable physics properties and the complex biological outcomes we actually care about. PEMP (Physics properties Enhanced Molecular Property prediction) is a framework designed to bridge this gap by leveraging "easy-to-calculate" quantum physics properties as auxiliary tasks.
Methodology
The PEMP framework is model-agnostic, designed to inject physical chemistry domain knowledge into the learning process. The authors developed two primary strategies to utilize these relationships:
- PEMP-MTL (Multi-Task Learning): The model jointly learns to predict the target property (e.g., toxicity) and several auxiliary physics properties simultaneously. This forces the shared molecule representation module to capture features that are physically meaningful.
- PEMP-TransL (Transfer Learning): The model is first pre-trained on a large dataset of physics properties to develop a "contextualized" understanding of molecular structures. It is then fine-tuned on the specific, smaller target dataset.
The researchers tested these strategies using two state-of-the-art backbones: AttentiveFP (a supervised graph attention network) and GROVER (a massive, pre-trained transformer-style model). They specifically selected five auxiliary physics properties based on chemical theory: dipole moment, isotropic polarizability, HOMO-LUMO gap, internal energy, and heat capacity.
Experimental Setup
Datasets
- Auxiliary Dataset: QM9, containing approximately 130,000 molecules with 13 quantum properties calculated via density functional theory.
- Target Datasets (from MoleculeNet):
- Physical Chemistry (Regression): FreeSolv (hydration energy), ESOL (solubility), and Lipophilicity.
- Physiology (Classification): BBBP (blood-brain barrier penetration), ClinTox (clinical trial toxicity), and Tox21 (environmental toxicity).
Key Insights
Physics as a Prior
The core innovation is that physics properties act as a bridge. For instance, a molecule's polarity (dipole moment) directly influences its intermolecular forces, which in turn determines its solubility. By teaching a model physics, researchers effectively provide the "rules of the game" before asking the model to predict complex biological outcomes.
Solving the Small-Data Problem
On datasets like FreeSolv (only 642 molecules), standard deep learning often overfits. PEMP uses the 130,000-sample QM9 dataset to regularize the model, ensuring the learned representations are grounded in physical reality. Ablation studies revealed that using a combination of different physics properties (e.g., both polarity-related and scaffold-related) performed better than using any single property alone, suggesting that the model benefits from a holistic physical profile of the molecule.
Results
PEMP consistently outperformed state-of-the-art baselines across nearly all tested datasets:
- Performance Boost: PEMP achieved an average relative improvement of 6.9% over the original AttentiveFP and 9.1% over the original GROVER.
- Regression Success: On the FreeSolv dataset, GROVER-TransL reduced the error (RMSE) from 1.898 to 1.659.
- Classification Success: On the BBBP dataset, the model improved the prediction accuracy (AUC-ROC) from 0.893 to 0.928.
- Stability: PEMP-TransL proved to be more stable than the MTL version, likely because the two-step process prevents the large physics dataset from "overwhelming" the smaller target dataset during training.
Conclusion and Future Work
PEMP demonstrates that integrating domain knowledge from physical chemistry can significantly enhance the predictive power of molecular AI. While the authors used standard Multi-Task and Transfer Learning, they note that more sophisticated "task-aware" learning methods could potentially yield even better results. Future research may extend this "physics-enhanced" approach to molecule generation or molecule-protein interaction tasks, further refining the model's understanding of molecular behavior through additional physical descriptors.
Summary updated