Shiyun Wa

Active learning provides an efficient strategy for molecular property prediction by iteratively prioritizing compounds for experimental evaluation. However, the effectiveness of active learning pipelines depends strongly on the choice of molecular representation, and systematic understanding of how representation families affect the active learning process in terms of uncertainty and predictive performance remains limited. In this work, we introduce ActiveFusion, a framework for integrating heterogeneous molecular representations within active learning workflows for molecular property prediction. The framework enables systematic evaluation of physicochemical descriptors, molecular fingerprints, learned graph neural network (GNN) representations, and pretrained Transformer-based representations, as well as feature-level fusion strategies that combine complementary chemical information sources. ActiveFusion evaluates models across four molecular property prediction regression tasks. Across datasets, we demonstrate that feature fusion between learned graph representations and physicochemical descriptors consistently improves prediction performance and discovery (average final iteration R2 of 0.71, 0.67, and 0.54 for our overall best representations Chemprop+RDKit, Chemprop, and RDKit, respectively). We show that exploration-driven acquisition strategies enhance scaffold coverage and promote sampling of structurally novel regions of chemical space, and that model-agnostic acquisition of new compounds based on diversity has strong performance. Notably, both pretrained and finetuned Transformer-based embeddings do not consistently outperform physicochemical features or GNN-learned representations in our setting, highlighting the continued relevance of chemically interpretable features and learned features from supervised, task-specific models for active learning applications in molecular property prediction. Overall, ActiveFusion provides a systematic framework for studying representation-acquisition interactions in molecular discovery with representation fusion capabilities. Our study offers practical guidance for designing active learning pipelines that balance prediction accuracy with chemical space exploration.

Virtual screening (VS) on small molecules aims to identify promising drug candidates against protein targets from expansive chemical libraries by balancing the core requirements of accurate scoring and efficient search against the inherent trade‐off between accuracy and speed. This survey provides a comprehensive review of how Artificial Intelligence and Machine Learning (AI/ML) are redefining this landscape across three critical dimensions. First, we examine the evolution of AI‐driven scoring functions, which utilize AI/ML models to capture complex structure–activity relationships from massive biochemical datasets, significantly enhancing structure‐ and ligand‐based evaluations beyond traditional heuristics. Second, we summarize the emergence of efficient search algorithms that iteratively prioritize informative compounds to reduce search efforts by orders of magnitude. Third, we review the paradigm shift toward generative molecular design, making VS transition from screening fixed libraries to the de novo generation of molecules optimized for specific structural contexts and multi‐objective properties. This review outlines the transition toward end‐to‐end, adaptive discovery systems that ensure computational hits are biologically potent, structurally optimized, and synthetically accessible.

Active learning provides an efficient strategy for molecular property prediction by iteratively prioritizing compounds for experimental evaluation. However, the effectiveness of active learning pipelines depends strongly on the choice of molecular representation, and systematic understanding of how representation families affect the active learning process in uncertainty and predictive performance remain limited. In this work, we introduce ActiveFusion, a framework for integrating heterogeneous molecular representations within active learning workflows for molecular property prediction. The framework enables systematic evaluation of physicochemical descriptors, molecular fingerprints, learned graph neural network (GNN) representations, and pretrained Transformer-based representations, as well as feature-level fusion strategies that combine complementary chemical information sources. ActiveFusion evaluates models across four molecular property prediction regression tasks. Across datasets, we demonstrate that feature fusion between learned graph representations and physicochemical descriptors consistently improves prediction performance and discovery (average final iteration R ² of 0.71, 0.67 and 0.54 for our overall best representations Chemprop+RDKit, Chemprop, and RDKit respectively). We show that exploration-driven acquisition strategies enhance scaffold coverage and promote sampling of structurally novel regions of chemical space, and that model-agnostic acquisition of new compounds based on diversity has strong performance. Notably, both pretrained and fine-tuned Transformer-based embeddings do not consistently outperform physicochemical features or GNN-learned representations in our setting, highlighting the continued relevance of chemically interpretable features and learned features from supervised, task-specific models for active learning application in molecular property prediction. Overall, ActiveFusion provides a systematic framework for studying representation–acquisition interactions in molecular discovery with representation fusion capabilities. Our study offers practical guidance for designing active learning pipelines that balance prediction accuracy with chemical space exploration.

Active learning provides an efficient strategy for molecular property prediction by iteratively prioritizing compounds for experimental evaluation. However, the effectiveness of active learning pipelines depends strongly on the choice of molecular representation, and systematic understanding of how representation families affect the active learning process in uncertainty and predictive performance remain limited. In this work, we introduce ActiveFusion, a framework for integrating heterogeneous molecular representations within active learning workflows for molecular property prediction. The framework enables systematic evaluation of physicochemical descriptors, molecular fingerprints, learned graph neural network (GNN) representations, and pretrained Transformer-based representations, as well as feature-level fusion strategies that combine complementary chemical information sources. ActiveFusion evaluates models across four molecular property prediction regression tasks. Across datasets, we demonstrate that feature fusion between learned graph representations and physicochemical descriptors consistently improves prediction performance and discovery (average final iteration R ² of 0.62, 0.57 and 0.47 for our overall best representations Chemprop+RDKit, Chemprop, and RDKit respectively). We show that exploration-driven acquisition strategies enhance scaffold coverage and promote sampling of structurally novel regions of chemical space, and that model-agnostic acquisition of new compounds based on diversity has strong performance. Notably, pretrained Transformer-based embeddings do not consistently outperform physicochemical features or GNN-learned representations in our setting, highlighting the continued relevance of chemically interpretable features and learned features from supervised, task-specific models for active learning application in molecular property prediction. Overall, ActiveFusion provides a systematic framework for studying representation–acquisition interactions in molecular discovery with representation fusion capabilities. Our study offers practical guidance for designing active learning pipelines that balance prediction accuracy with chemical space exploration.

Machine learning (ML) techniques offer a promising path for accelerating antibiotic discovery by computationally predicting and optimizing desirable pharmaceutical properties of molecules. One difficult case for drug discovery is tuberculosis (TB), the disease caused byMycobacterium tuberculosis, a bacterium with intrinsic resistance to many antibiotics. The unique outer membrane of M. tuberculosis, called the mycomembrane, is thought to contribute to intrinsic antibiotic resistance by establishing a permeability barrier. While recent ML works have attempted to predict the permeability of compounds through the mycomembrane, the scarcity of labeled data has led to approaches that struggle to generalize. In this work, we propose a robust two-stage model, MycoPermeNet-v2, to predict compound permeability through the mycomembrane of M. tuberculosis. MycoPermeNet-v2 fuses molecular descriptors with learned graph-based embeddings and incorporates Noisy Student self-disTillation (NST) to improve performance even when labeled data is limited. Compared to prior work, we achieve significantly higher performance (RMSE from 0.755 ± 0.024 to 0.719 ± 0.021, adjusted p < 0.0001). We systematically evaluate the robustness of the overall model across different data split strategies, as well as the contribution of each component in the proposed feature fusion and NST framework. Furthermore, the interpretability analysis shows that the learned representations capture chemically meaningful features relevant to M. tuberculosis permeability. We show the generalizability of our approach over multiple models and physical chemistry property prediction tasks, demonstrating strong applicability in data-constrained molecular learning scenarios.

As Artificial Intelligence (AI) integrates deeper into diverse sectors, the quest for powerful models has intensified. While significant strides have been made in boosting model capabilities and their applicability across domains, a glaring challenge persists: many of these state-of-the-art models remain as black boxes. This opacity not only complicates the explanation of model decisions to end-users but also obstructs insights into intermediate processes for model designers. To address these challenges, we introduce InterpreTabNet, a model designed to enhance both classification accuracy and interpretability by leveraging the TabNet architecture with an improved attentive module. This design ensures robust gradient propagation and computational stability. Additionally, we present a novel evaluation metric, InterpreStability, which quantifies the stability of a model's interpretability. The proposed model and metric mark a significant stride forward in explainable models' research, setting a standard for transparency and interpretability in AI model design and application across diverse sectors. InterpreTabNet surpasses other leading solutions in tabular data analysis across varied application scenarios, paving the way for further research into creating deep-learning models that are both highly accurate and inherently explainable. The introduction of the InterpreStability metric ensures that the interpretability of future models can be measured and compared in a consistent and rigorous manner. Collectively, these contributions have the potential to promote the design principles and development of next-generation interpretable AI models, widening the adoption of interpretable AI solutions in critical decision-making environments.

This research primarily explores the application of Natural Language Processing (NLP) technology in precision financial fraud detection, with a particular focus on the implementation and optimization of the FinChain-BERT model. Firstly, the FinChain-BERT model has been successfully employed for financial fraud detection tasks, improving the capability of handling complex financial text information through deep learning techniques. Secondly, novel attempts have been made in the selection of loss functions, with a comparison conducted between negative log-likelihood function and Keywords Loss Function. The results indicated that the Keywords Loss Function outperforms the negative log-likelihood function when applied to the FinChain-BERT model. Experimental results validated the efficacy of the FinChain-BERT model and its optimization measures. Whether in the selection of loss functions or the application of lightweight technology, the FinChain-BERT model demonstrated superior performance. The utilization of Keywords Loss Function resulted in a model achieving 0.97 in terms of accuracy, recall, and precision. Simultaneously, the model size was successfully reduced to 43 MB through the application of integer distillation technology, which holds significant importance for environments with limited computational resources. In conclusion, this research makes a crucial contribution to the application of NLP in financial fraud detection and provides a useful reference for future studies.