molecular featurization

Molecular property annotation scarcity severely limits AI applications in drug and materials discovery, making few-shot molecular property prediction (FSMPP) a critical challenge. This paper presents the first systematic survey of FSMPP, identifying two fundamental bottlenecks: cross-property generalization—hindered by distributional shifts and weak biochemical correlations among properties—and cross-molecule generalization—undermined by high structural heterogeneity. To address these, we propose a multi-level taxonomy encompassing data, models, and learning paradigms; integrate graph neural networks, meta-learning, knowledge transfer, and chemical priors to enhance prediction robustness under extreme label scarcity; and unify mainstream benchmarks, evaluation protocols, and method performance. Our work establishes the first scalable research framework for FSMPP and delivers a clear, actionable technical roadmap—bridging foundational gaps and accelerating progress in low-data molecular AI.

Current evaluations of pretrained molecular embedding models lack statistical rigor and fair benchmarking, often assuming neural models inherently outperform traditional fingerprints like ECFP without empirical validation. Method: We systematically evaluate 25 pretrained molecular embedding models across 25 benchmark datasets and introduce a hierarchical Bayesian statistical testing framework to control for multiple hypothesis testing bias in a unified manner. Contribution/Results: Contrary to prevailing assumptions, most pretrained models fail to achieve statistically significant improvements over the ECFP baseline; only CLAMP demonstrates robust superiority. This challenges the implicit “neural models are necessarily better” assumption pervasive in molecular representation learning and exposes systemic issues—including methodologically unsound evaluation protocols, weak baselines, and reporting bias. Based on these findings, we propose a standardized evaluation protocol, recommend strong baselines (e.g., ECFP with optimized hyperparameters), and advocate for small-sample validation to ensure reliability. Our work establishes a methodological foundation and practical guidelines for trustworthy molecular AI research.

To address substantial experimental noise in molecular property data and systematic biases inherent in quantum simulations, this paper introduces CheMeleon, a novel molecular foundation model. CheMeleon pioneers a descriptor-driven, noise-robust pretraining paradigm that jointly learns robust and generalizable molecular representations using Mordred descriptors and a directed message-passing neural network (D-MPNN). Under few-shot settings, it achieves significant performance gains: 79% win rate on the Polaris benchmark—43 percentage points higher than Chemprop—and 97% on MoleculeACE. t-SNE visualizations confirm clear separation of chemically coherent structural series. This work establishes a new paradigm for high-accuracy, low-data-dependency molecular modeling. However, accurately capturing subtle structure–activity relationships—such as activity cliffs—remains challenging.

Molecular property prediction is crucial in drug discovery yet often hindered by limited data availability and the high barrier to entry posed by existing foundation models that require task-specific fine-tuning. This work introduces tabular foundation models (TFMs) combined with in-context learning to this domain for the first time, enabling highly effective predictions without any fine-tuning. The proposed approach integrates frozen molecular embeddings—such as those from CheMeleon—with classical descriptors (RDKit2D, Mordred) and fingerprints. Evaluated across 30 tasks from MoleculeACE, the method achieves a 100% win rate using CheMeleon embeddings, substantially outperforming conventional approaches while maintaining high accuracy and significantly reducing both computational cost and accessibility barriers.

This study addresses three critical challenges in chemical process design: low accuracy in molecular property prediction, inefficient discovery of novel molecules, and the absence of molecular–process co-design. To tackle these, we propose a hybrid architecture integrating physics- and chemistry-informed graph neural networks (GNNs) with Transformer modules, enabling high-accuracy, interpretable prediction of thermodynamic and transport properties for both pure components and mixtures. Building upon this, we establish a closed-loop framework unifying molecular generation, property prediction, and process optimization—facilitating targeted exploration of chemical space and cross-scale co-design. Furthermore, we introduce the first unified benchmark spanning molecular modeling, process simulation, and industrial validation, substantially enhancing model generalizability and engineering applicability. The resulting paradigm provides a scalable, experimentally verifiable methodology for the co-development of novel functional molecules and low-carbon chemical processes.

This work addresses the high computational cost and complex data engineering requirements of conventional molecular property prediction methods, which typically rely on molecular graphs, 3D conformations, or large language models. For the first time, it systematically investigates a purely vision-based paradigm by evaluating ten visual architectures and seven pretraining strategies across ten downstream tasks using a dataset of two million molecular scaffold images. The study introduces a chemistry-informed curriculum learning strategy that dynamically orders training samples according to molecular structural complexity. Experimental results demonstrate that accurate predictions can be achieved using only a single molecular image, with the proposed approach ranking first on five out of ten benchmarks and placing within the top two on all tasks, while reducing computational costs by up to 80× compared to state-of-the-art multimodal methods.

Existing molecular generation methods rely heavily on surrogate metrics and pretraining data from drug discovery, limiting their generalizability to structurally distinct domains such as nanotechnology. This work introduces the Nanomaterial Molecular Optimization (NMO) benchmark, which, for the first time, employs quantum simulations as a ground-truth objective function and incorporates structural constraints alongside domain-agnostic pretraining strategies to formulate a generative task with a rugged fitness landscape. By establishing a new paradigm that balances scientific utility with rigorous model evaluation, the proposed framework not only surpasses state-of-the-art physical performance metrics through its baseline methods but also uncovers novel structural motifs, exposes limitations of advanced optimization algorithms in this emerging domain, and demonstrates the potential of machine learning to drive authentic scientific discovery.

This work proposes Pretrained Embedding Distance (PED), a general and tuning-free molecular similarity metric that leverages distances in the embedding space of pretrained molecular models. Traditional similarity measures often rely on handcrafted features or incur high computational costs, while existing deep learning approaches typically require task-specific supervision or large amounts of labeled data, limiting their generalizability. In contrast, PED eliminates the need for both manual feature engineering and task-specific fine-tuning. It effectively ranks active compounds in virtual screening and successfully guides goal-directed molecular generation. Experimental results demonstrate that PED exhibits strong correlation with conventional similarity metrics across multiple tasks, while offering superior scalability and practical utility.

This work proposes a unified molecular machine learning framework that overcomes the limitations of existing models, which are often confined to specific codebases and struggle to generalize across the full periodic table or diverse molecular properties. The framework supports elements 1–100, encompassing organic, inorganic, coordination, and biomolecular systems, and enables predictions at atomic, bond, molecular, and functional group levels. It natively incorporates conditional modeling of charge and spin states and uniquely integrates E(3)-equivariant networks, Transformers, and 2D graph neural networks within a single architecture, while combining both aleatoric and epistemic uncertainty quantification. Evaluated across multiple chemical benchmarks, the model matches or exceeds state-of-the-art performance, scales to datasets containing millions of molecules, and significantly lowers the barrier to entry for researchers without computational expertise.

Existing benchmarks for chemical property prediction suffer from limited task diversity, fragmented datasets, and inconsistent evaluation protocols, hindering comprehensive assessment of model generalization. To address these limitations, this work introduces Chem World, a large-scale, unified benchmark comprising 17 diverse tasks and over 800,000 molecules. Furthermore, the study proposes Mixture-PINN, a physics-informed neural network that integrates chemical prior knowledge through domain-specific inductive biases and multitask learning. Evaluated under a standardized protocol, Mixture-PINN demonstrates significantly improved accuracy, robustness, and reliability across tasks, thereby advancing the development of trustworthy AI for computational chemistry.