CataCon: a contrastive graph representation learning framework for catalyst prediction.

Accurate prediction of molecular toxicity is vital for drug development. Most mainstream methods rely on fingerprints or graph-based feature extraction, the emergence of large language models (LLMs) offers new prospects for molecular representation learning in toxicity prediction. Although several studies attempt to leverage LLMs to integrate molecular sequence data for pretraining molecular representations, certain limitations remain. Current LLM-based approaches usually utilize solely on class embedding features, overlooking the rich information in sequence embedding. Moreover, integrating pre-trained molecular representations with multi-modal molecular data may further enhance performance in toxicity prediction. To address these challenges, we propose BFGTP, a BERT-guided two-stage molecular representation learning framework for toxicity prediction. Firstly, we design independent encoders for molecular descriptions of three modalities, where the fingerprint encoder with dual level attention mechanisms effectively integrates multi-category fingerprints. Then, the two-stage guide strategy is introduced to fully utilize the prior knowledge of LLMs, employing contrastive learning to align and fuse the tri-modal representations and knowledge distillation to align predicted value distributions. BFGTP ultimately combines fingerprint and graph representations to predict molecular toxicity. Experiments on seven toxicity datasets show that BFGTP outperforms baselines, achieving the highest AUC on five datasets and the best average performance across five evaluation metrics. Ablation studies, t-SNE visualization and case study confirm the effectiveness of BFGTP's components and its ability to capture meaningful molecular representations.

G-K BertDTA: A graph representation learning and semantic embedding-based framework for drug-target affinity prediction

The oxidative addition and migratory 1,1-insertion in the Monsanto and Cativa processes. A density functional study of the catalytic carbonylation of methanol

Graph representation learning is crucial for many real-world applications (e.g. social relation analysis). A fundamental problem for graph representation learning is how to effectively learn representations without human labeling, which is usually costly and time-consuming. Graph contrastive learning (GCL) addresses this problem by pulling the positive node pairs (or similar nodes) closer while pushing the negative node pairs (or dissimilar nodes) apart in the representation space. Despite the success of the existing GCL methods, they primarily sample node pairs based on the node-level proximity yet the community structures have rarely been taken into consideration. As a result, two nodes from the same community might be sampled as a negative pair. We argue that the community information should be considered to identify node pairs in the same communities, where the nodes insides are semantically similar. To address this issue, we propose a novel Graph Communal Contrastive Learning (gCooL) framework to jointly learn the community partition and learn node representations in an end-to-end fashion. Specifically, the proposed gCooL consists of two components: a Dense Community Aggregation (DeCA) algorithm for community detection and a Reweighted Self-supervised Cross-contrastive (ReSC) training scheme to utilize the community information. Additionally, the real-world graphs are complex and often consist of multiple views. In this paper, we demonstrate that the proposed gCooL can also be naturally adapted to multiplex graphs. Finally, we comprehensively evaluate the proposed gCooL on a variety of real-world graphs. The experimental results show that the gCooL outperforms the state-of-the-art methods.

With the continuous development of Electronic Health Record (EHRs), medical heterogeneous data have become increasingly abundant, containing diverse types of entities and complex semantic relationships that provide essential support for disease diagnosis. However, conventional heterogeneous graph neural networks struggle to distinguish semantic differences among multi-type nodes and k-hop neighbors, often leading to semantic confusion and vulnerability to noise, which limits their classification performance and generalization capability. To tackle these challenges, we present a novel framework named Heterogeneous Graph Transformer and Diffusion mechanism for Disease Diagnosis (TD4DD), which captures multi-scale semantic dependencies across k-hop neighborhoods, while the diffusion module performs latent-space denoising to alleviate noise interference. Specifically, a k-hop hierarchical Transformer is introduced to capture multi-scale dependencies across different hop layers, enabling the differentiation of fine-grained semantic variations among neighbors at various distances. Additionally, a diffusion module is designed to handle noise in the data by performing denoising in the latent space through auxiliary subgraphs constructed using different meta-paths, thereby generating more discriminative node representations. Finally, the model fuses structural information with denoised embeddings to accomplish disease classification. Experiments conducted on two real-world clinical datasets, MIMIC-III(7,000 patients, 5 disease categories) and MIMIC-IV(8,331 patients, 6 disease categories), demonstrate that TD4DD consistently outperforms existing baseline methods in terms of both Micro-F1 and Macro-F1 scores, showing strong generalization ability. On MIMIC-III, TD4DD achieves a Micro-F1 of 88.29 and a Macro-F1 of 86.11, while on MIMIC-IV, it reaches 83.60 and 83.94, respectively. Furthermore, ablation studies and t-SNE visualizations further validate the effectiveness of each module and the distinguishing capability of the learned embeddings.

In real-world scenarios, sparse or unlabeled occlusion data significantly limits the optimization of pedestrian localization and identification models. This limitation is particularly evident in intelligent urban systems and security applications, where manual labeling is not only costly but also hindered by challenges such as multi-camera coverage and varying degrees of occlusion. The absence of features in occluded regions further exacerbates performance degradation, especially in multi-camera surveillance settings. This study proposes a novel architecture, MaskFormer, based on an Adaptive Masking Mechanism (AMM), which dynamically generates occlusion masks and integrates a local-global feature interaction module to effectively address occlusion recovery and pedestrian feature extraction. Experimental results on the Occluded-Duke dataset demonstrate that MaskFormer significantly outperforms traditional methods in both occlusion recovery performance (PSNR and SSIM) and pedestrian identification metrics (mAP and Rank-1 accuracy), achieving an mAP of 52.8% and a Rank-1 accuracy of 65.4%. Additionally, t-SNE visualization and ablation studies further validate the contributions of each module to the overall performance. The findings not only highlight MaskFormer’s robustness in complex occlusion scenarios but also provide an efficient solution for pedestrian identification tasks in large-scale unlabeled data environments. Future work will focus on enhancing the model's real-time performance, cross-domain generalization capabilities, and integration with multimodal data.

Graph representation learning has become a ubiquitous component in many scenarios, ranging from social network analysis to energy forecasting in smart grids. In several applications, ensuring the fairness of the node (or graph) representations with respect to some protected attributes is crucial for their correct deployment. Yet, fairness in graph deep learning remains underexplored, with few solutions available. In particular, the tendency of similar nodes to cluster on several real-world graphs (i.e., homophily) can dramatically worsen the fairness of these procedures. In this article, we propose a novel biased edge dropout algorithm (FairDrop) to counter-act homophily and improve fairness in graph representation learning. FairDrop can be plugged in easily on many existing algorithms, is efficient, adaptable, and can be combined with other fairness-inducing solutions. After describing the general algorithm, we demonstrate its application on two benchmark tasks, specifically, as a random walk model for producing node embeddings, and to a graph convolutional network for link prediction. We prove that the proposed algorithm can successfully improve the fairness of all models with only a small or negligible drop in accuracy, and compares favourably with existing state-of-the-art solutions. In an ablation study, we demonstrate that our algorithm can flexibly interpolate between biasing towards fairness and an unbiased edge dropout. Furthermore, to better evaluate the gains, we propose a new dyadic group definition to measure the bias of a link prediction task when paired with group-based fairness metrics. In particular, we extend the metric used to measure the bias in node embeddings by taking account of the graph structure.

Ammonia decomposition has gained significant attention as an eco-friendly method for hydrogen production because it creates no carbon dioxide emissions. While Ru catalysts are known for their high activity in ammonia decomposition, their high cost makes them uneconomical for commercial use. Therefore, it is essential to explore novel alloy catalysts composed of inexpensive elements with high catalytic performance. Nitrogen adsorption energies serve as key descriptors indicating the catalytic performance for ammonia decomposition, and first-principle calculations can compute these energies. However, the screening of numerous alloy catalyst candidates through extensive first-principle calculations and experimental validations remains time-consuming due to the vast number of potential candidates. To address this, artificial intelligence and machine learning models are being developed to quickly predict catalyst performance, efficiently searching for promising catalyst candidates. In this study, we developed a machine-learning-based method to rapidly predict nitrogen adsorption energies using a graph-based artificial neural network, thereby efficiently searching for novel catalysts for ammonia decomposition. Our training dataset included the nitrogen adsorption energies of 30 pure transition metal catalyst candidates, as well as binary alloy catalyst candidates, including core-shell and intermetallic compounds. As a result, we successfully identified 12 catalyst candidates composed of inexpensive elements that are likely to exhibit catalytic performance comparable to Ru catalysts.

No-reference image quality assessment (NR-IQA) aims to predict perceptual quality in alignment with the human visual system (HVS), yet existing methods face challenges in capturing long-range dependencies across distortion types and levels while preserving content fidelity during preprocessing. This paper presents a perceptually-driven NR-IQA framework that integrates meta-learning, graph representation learning, and multi-scale feature fusion to address these limitations. First, a meta-learning paradigm is employed to pre-train a self-calibrated convolutional backbone, which adaptively models spatial and channel-wise dependencies across scales, thereby enhancing the extraction of distortion-aware features while mitigating information loss caused by fixed-input preprocessing. Second, a graph representation learning module is introduced to explicitly encode the hierarchical relationships among distortion types, distortion levels, and image content. Nodes in the graph correspond to distorted images, while edges capture inter-sample similarities; these are jointly optimized via a graph convolutional network under dual supervision from a triplet-based distortion-type discriminator and a probabilistic distortion-level regressor that accounts for content-induced uncertainty. Extensive experiments on four benchmark datasets demonstrate that our method achieves better performance, with average SROCC and PLCC improvements of 3.6–36.6% over hand-crafted feature-based methods and consistent gains over deep learning-based approaches. Ablation studies and visualizations confirm that the proposed components collectively yield a more discriminative and generalizable distortion representation, closely mirroring human perceptual judgments.

Real-valued inter-residue distance maps provide essential spatial information for understanding protein folding mechanisms and guiding downstream applications such as function annotation, drug discovery, and structural modeling. However, existing prediction methods often struggle to capture long-range dependencies and to maintain topological consistency across different structural scales. To address these challenges, we propose a novel prediction framework that integrates a Mamba architecture, based on a selective state space model, to effectively model global interactions, and incorporates the Kolmogorov-Arnold Network (KAN) to enhance nonlinear structural representation. Extensive experiments on standard benchmark datasets, including CASP13, CASP14, and CASP15, demonstrate prediction accuracies of 86.53%, 85.44%, and 82.77%, respectively, outperforming state-of-the-art approaches. These results indicate that the proposed framework substantially improves the fidelity of real-valued distance prediction and offers a promising tool for downstream structural and functional studies.

Given a known protein sequence, predicting its secondary structure can help understand its three-dimensional (tertiary) structure, i.e., the folding. In this paper, we present an approach for predicting protein secondary structures. Different from the existing prediction methods, our approach proposes an encoding schema that weaves physio-chemical information in encoded vectors and a prediction framework that combines the context information with secondary structure segments. We employed Support Vector Machine (SVM) for training the CB513 and RS126 data sets, which are collections of protein secondary structure sequences, through sevenfold cross validation to uncover the structural differences of protein secondary structures. Hereafter, we apply the sliding window technique to test a set of protein sequences based on the group classification learned from the training set. Our approach achieves 77.8% segment overlap accuracy (SOV) and 75.2% three-state overall per-residue accuracy (Q3), which outperform other prediction methods.

KFPredict: An ensemble learning prediction framework for diabetes based on fusion of key features

Here, I describe a theoretical approach to the structure and chemical composition of minerals based on their bond topology. This approach allows consideration of many aspects of minerals and mineral behaviour that cannot be addressed by current theoretical methods. It consists of combining the bond topology of the structure with aspects of graph theory and bond-valence theory (both long range and short range), and using the moments approach to the electronic energy density-of-states to interpret topological aspects of crystal structures. The structure hierarchy hypothesis states that higher bond-valence polyhedra polymerize to form the (usually anionic) structural unit, the excess charge of which is balanced by the interstitial complex (usually consisting of large low-valence cations and (H2O) groups). This hypothesis may be justified within the framework of bond topology and bond-valence theory, and may be used to hierarchically classify oxysalt minerals. It is the weak interaction between the structural unit and the interstitial complex that controls the stability of the structural arrangement. The principle of correspondence of Lewis acidity–basicity states that stable structures will form when the Lewis-acid strength of the interstitial complex closely matches the Lewis-base strength of the structural unit, and allows us to examine the factors that control the chemical composition and aspects of the structural arrangements of minerals. It also provides a connection between a structure, the speciation of its constituents in aqueous solution and its mechanism of crystallization. The moments approach to the electronic energy density-of-states provides a link between the bond topology of a structure and its thermodynamic properties, as indicated by correlations between average anion coordination number and reduced enthalpy of formation from the oxides for [6]Mg [4] Si n O(m+2n) and MgSO4(H2O) n .

This study explores the integration of the Variational Information Bottleneck (VIB) and the Perceiver model in a Hybrid Bottleneck (HB) framework for disease prediction, specifically focusing on the challenges of convergence speed and generalization in small, structured datasets. Leveraging VIB's information compression and the Perceiver's cross- and self-attention mechanisms, we hypothesize that this hybrid approach enhances model performance. We conducted ablation studies using three open datasets to assess the model's effectiveness across different disease prediction tasks. Our ablation study examines the impact of removing components such as the Transformer's auxiliary layers, residual connections, and VIB, evaluating their roles in model convergence and generalization. Results indicate that integrating VIB enhances predictive accuracy and stabilizes learning by filtering noise and retaining essential features, while the Transformer's attention mechanisms quicken convergence. The HB framework, particularly with the inclusion of Add & Norm components, demonstrates improved stability and robustness, particularly in noisy datasets. These findings underscore the synergistic potential of combining VIB and Perceiver components in biomedical predictive modeling, offering significant improvements in rapid convergence and training stability.

Crime reduction and prevention strategies are essential to increase public safety and reduce the crime costs to society. Law enforcement agencies have long realized the importance of analyzing co-offending networks---networks of offenders who have committed crimes together---for this purpose. Although network structure can contribute significantly to co-offence prediction, research in this area is very limited. Here we address this important problem by proposing a framework for co-offence prediction using supervised learning. Considering the available information about offenders, we introduce social, geographic, geo-social and similarity feature sets which are used for classifying potential negative and positive pairs of offenders. Similar to other social networks, co-offending networks also suffer from a highly skewed distribution of positive and negative pairs. To address the class imbalance problem, we identify three types of criminal cooperation opportunities which help to reduce the class imbalance ratio significantly, while keeping half of the co-offences. The proposed framework is evaluated on a large crime dataset for the Province of British Columbia, Canada. Our experimental evaluation of four different feature sets show that the novel geo-social features are the best predictors. Overall, we experimentally show the high effectiveness of the proposed co-offence prediction framework. We believe that our framework will not only allow law enforcement agencies to improve their crime reduction and prevention strategies, but also offers new criminological insights into criminal link formation between offenders.