Chem World: A Large-Scale Benchmark and Physics-Aware Framework for Trustworthy Molecular Property Prediction

Tianyou Bai, Huan Wang, Mingchen Gao, Tongrui Yang, Fangyue Lin, Pinze Ren, Zhenlin Zhao, Siming DongarXiv 2026 · arXiv

TL;DR

Chem World is a massive new benchmark for molecular property prediction that integrates 17 datasets and 800,000+ instances to evaluate how AI models handle both single molecules and complex mixtures. Alongside it, the authors introduce PACM (Physics-Aware Compositional Modeling), a framework that combines chemical language models with domain-specific physical constraints to ensure predictions are not only accurate but also physically plausible.

What Problem Does This Paper Address?

Predicting the properties of molecules is the bread and butter of drug discovery and materials science. However, the field faces two major hurdles:

  1. Fragmentation: Existing benchmarks are often "siloed," focusing on narrow tasks or single-molecule systems. In reality, most chemical applications involve mixtures (like electrolytes or drug formulations) where properties emerge from complex intermolecular interactions.
  2. Physical Unreliability: Standard "black-box" deep learning models often ignore the laws of physics. They might predict a negative viscosity or suggest that a liquid becomes more viscous as it gets hotter—results that are mathematically possible for a neural network but physically impossible in a lab.

Methods

The authors propose a two-part solution: a unified benchmark and a novel modeling framework.

The PACM Framework

PACM (Physics-Aware Compositional Modeling) moves away from simple statistical correlation toward a more principled approach to chemical systems. Its pipeline consists of three main stages:

  1. Molecular Encoding: It uses pretrained chemical language models (like MolFormer or MolT5) to extract rich semantic representations from SMILES strings (text-based molecular representations).
  2. Composition-Aware Interaction Modeling: Instead of just averaging the embeddings of components in a mixture, PACM explicitly models pairwise interactions. It learns how component A and component B influence each other, which is critical for capturing non-additive mixture behaviors.
  3. Physics-Aware Learning: This is the "trustworthy" core. The model is regularized by differentiable physical priors. These constraints discourage the model from violating basic chemical principles, such as:
  4. Composition Consistency: Ensuring learned weights align with actual mole fractions.
  5. Temperature Response: Enforcing known trends (e.g., conductivity generally increases with temperature, while viscosity decreases).
  6. Physical Boundaries: Preventing impossible values like negative solubility.
The PACM pipeline integrates encoding, interaction modeling, and physical constraints.

Datasets

The authors introduced Chem World, a large-scale benchmark that unifies 17 public datasets. It covers over 800,000 instances across 10 property tracks, including:

  • Transport/Electrochemical: Conductivity, Viscosity.
  • Thermodynamic: Density, Enthalpy of Mixing, Fusion Enthalpy, Heat of Vaporization.
  • Physicochemical: Melting Temperature, Solubility, Motor Octane Number.
  • Perceptual: Olfactory Similarity.

This variety allows for a systematic evaluation of how well models generalize from simple molecules to multi-component systems.

Figure 2: Dataset statistics and structural composition of Chem World.

Key Insights and Ideas

  • LLMs are Chemical Experts: Pretrained molecular language models provide surprisingly strong and transferable representations, outperforming traditional graph neural networks and descriptor-based models (like RDKit + XGBoost).
  • Interactions Matter: Simply aggregating molecular embeddings (e.g., mean pooling) fails for mixtures. Explicitly modeling pairwise interactions is essential for capturing the "emergent" properties of chemical systems.
  • Physics as a Regularizer: Incorporating physical constraints doesn't just make the model "safer"—it actually improves predictive accuracy by providing a better inductive bias, especially in data-sparse regions.
  • DL vs. Classical Thermodynamics: Deep learning models like PACM can now match the accuracy of heavily engineered, parameter-intensive classical models (like COSMO-RS or NRTL) without needing specialized physical parameters for every new mixture.

Key Results

  • State-of-the-Art Performance: The MolFormer + PACM combination achieved the lowest error across the board, with a Macro NRMSE of 0.367.
  • Significant Gains: PACM reduced the error by approximately 15.2% compared to standard DeepSets aggregation.
  • Improved Plausibility: PACM drastically reduced physical constraint violations. For example, boundary violations (like negative predictions for non-negative properties) dropped from 9.7% in baselines to just 1.8%.
  • Broad Coverage: Unlike classical thermodynamic models which only supported a few tracks (like density and solubility), PACM provided competitive predictions across all 10 tracks in the benchmark.

Limitations and Future Work

While Chem World is a major step forward, the authors acknowledge several limitations:

  • Data Heterogeneity: The integrated datasets come from different sources with varying measurement protocols and data quality, which can introduce noise.
  • Simplified Physics: The current physical constraints rely on general assumptions (like monotonicity) and may not capture highly complex, non-linear thermodynamic phase behaviors.
  • Generalization: While group-aware splitting reduces data leakage, truly generalizing to entirely unseen chemical families remains a significant challenge for future research.

Future work will focus on expanding dataset coverage, incorporating more sophisticated physical models, and strengthening out-of-distribution evaluation.

Summary updated

Abstract

Molecular property prediction is fundamental to computational chemistry and drug formulation research, yet existing benchmarks remain fragmented across datasets, properties, and evaluation protocols, limiting systematic evaluation of model generalization. We introduce Chem World, a large-scale benchmark integrating 17 public datasets with over 800,000 molecular and chemical-system instances across 10 property tracks, covering single-molecule and multi-component systems, including drug solubility. Chem World provides a unified testbed for evaluating molecular representations and compositional modeling across diverse chemical properties. Building on the benchmark, we propose PACM (Physics-Aware Compositional Modeling), which combines molecular representations with composition-aware modeling and domain-specific physical constraints. Extensive experiments show that molecular language models provide strong and transferable representations, while PACM consistently improves prediction across diverse chemical systems. Chem World and our PACM-powered drug-property prediction Agent support generalizable chemical modeling and offer a foundation for AI-assisted biomedical research and drug formulation development.
BibTeX
@misc{bai2026chem,
  author = {Tianyou Bai and Huan Wang and Mingchen Gao and Tongrui Yang and Fangyue Lin and Pinze Ren and Zhenlin Zhao and Siming Dong},
  title = {{Chem World: A Large-Scale Benchmark and Physics-Aware Framework for Trustworthy Molecular Property Prediction}},
  year = {2026},
  eprint = {2611.13945},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13945},
}
Figure