Leveraging patients' longitudinal data to improve the Hospital One-year Mortality Risk

Purpose

Predicting medium-term survival after admission is necessary for identifying end-of-life patients who may benefit from goals of care (GOC) discussions. Considering that several patients have multiple hospital admissions, this study leverages patients' longitudinal data and information collected routinely at admission to predict the Hospital One-year Mortality Risk.

Methods

We propose the Ensemble Longitudinal Network (ELN) to predict one-year mortality using patients' longitudinal records. The model was evaluated: (i) with only predictors reported upon admission (AdmDemo); and (ii) also with diagnoses available later during patients' stay (AdmDemoDx). Using records of 123,646 patients with 250,812 hospitalizations from 2011 to 2021, our dataset was split into a learning set (2011-2017) to compare models with and without longitudinal information using nested cross-validation, and a holdout set (2017-2021) to assess clinical utility towards GOC discussions.

Results

The ELN achieved a significant increase in predictive performance using longitudinal information (p-value < 0.05) for both the AdmDemo and AdmDemoDx predictors. For randomly selected hospitalizations in the holdout set, the ELN showed: (i) AUROCs of 0.83 (AdmDemo) and 0.87 (AdmDemoDx); and (ii) superior decision-making properties, notably with an increase in precision from 0.25 for the standard process to 0.28 (AdmDemo) and 0.36 (AdmDemoDx). Feature importance analysis confirmed that the utility of the longitudinal information increases with the number of patient hospitalizations.

Conclusion

Integrating patients' longitudinal data provides better insights into the severity of illness and the overall patient condition, in particular when limited information is available during their stay.

Utilizing semantically enhanced self-supervised graph convolution and multi-head attention fusion for herb recommendation

Traditional Chinese herbal medicine has long been recognized as an effective natural therapy. Recently, the development of recommendation systems for herbs has garnered widespread academic attention, as these systems significantly impact the application of traditional Chinese medicine. However, existing herb recommendation systems are limited by data sparsity, insufficient correlation between prescriptions, and inadequate representation of symptoms and herb characteristics. To address these issues, this paper introduces an approach to herb recommendation based on semantically enhanced self-supervised graph convolution and multi-head attention fusion (BSGAM). This method involves efficient embedding of entities following fine-tuning of BERT; leveraging the attributes of herbs to optimize feature representation through a residual graph convolution network and self-supervised learning; and ultimately employing a multi-head attention mechanism for feature integration and recommendation. Experiments conducted on a publicly available traditional Chinese medicine prescription dataset demonstrate that our method achieves improvements of 6.80%, 7.46%, and 6.60% in F1-Score@5, F1-Score@10, and F1-Score@20, respectively, compared to baseline methods. These results confirm the effectiveness of our approach in enhancing the accuracy of herb recommendations.

Coherence-Based Graph Convolution Network to Assess Brain Reorganization in Spinal Cord Injury Patients

Motor imagery (MI) engages a broad network of brain regions to imagine a specific action. Investigating the mechanism of brain network reorganization during MI after spinal cord injury (SCI) is crucial because it reflects overall brain activity. Using electroencephalogram (EEG) data from SCI patients, we conducted EEG-based coherence analysis to examine different brain network reorganizations across different frequency bands, from resting to MI. Furthermore, we introduced a consistency calculation-based residual graph convolution (C-ResGCN) classification algorithm. The results show that the [Formula: see text]- and [Formula: see text]-band connectivity weakens, and brain activity decreases during the MI task compared to the resting state. In contrast, the [Formula: see text]-band connectivity increases in motor regions while the default mode network activity declines during MI. Our C-ResGCN algorithm showed excellent performance, achieving a maximum classification accuracy of 96.25%, highlighting its reliability and stability. These findings suggest that brain reorganization in SCI patients reallocates relevant brain resources from the resting state to MI, and effective network reorganization correlates with improved MI performance. This study offers new insights into the mechanisms of MI and potential biomarkers for evaluating rehabilitation outcomes in patients with SCI.

KNU-DTI: KNowledge United Drug-Target Interaction prediction

MOTIVATION

Accurately predicting drug-target protein interactions (DTI) is a cornerstone of drug discovery, enabling the identification of potential therapeutic compounds. Sequence-based prediction models, despite their simplicity, hold great promise in extracting essential information directly from raw sequences. However, the focus in recent DTI studies has increasingly shifted toward enhancing algorithmic complexity, often at the expense of fully leveraging robust sequence representation learning methods. This shift has led to the underestimation and gradual neglect of methodologies aimed at effectively capturing discriminative features from sequences. Our work seeks to address this oversight by emphasizing the value of well-constructed sequence representation algorithms, demonstrating that even with simple interaction mapping algorithm techniques, accurate DTI models can be achieved. By prioritizing meaningful information extraction over excessive model complexity, we aim to advance the development of practical and generalizable DTI prediction frameworks.

RESULTS

We developed the KNowledge Uniting DTI model (KNU-DTI), which retrieves structural information and unites them. Protein structural properties were obtained using structural property sequence (SPS). Extended-connectivity fingerprint (ECFP) was used to estimate the structure-activity relationship in molecules. Including these two features, a total of five latent vectors were derived from protein and molecule via various neural networks and integrated by elemental-wise addition to predict binding interactions or affinity. Using four test concepts to evaluate the model, we show that the model outperforms recently published competitors. Finally, a case study indicated that our model has a competitive edge over existing docking simulations in some cases.

A Novel session-based recommendation system using capsule graph neural network

Session-based recommendation systems (SBRS) are essential for enhancing the customer experience, improving sales and loyalty, and providing the possibility to discover products in dynamic and real-world scenarios without needing user history. Despite their importance, traditional or even current SBRS algorithms face limitations, notably the inability to capture complex item transitions within each session and the disregard for general patterns that can be derived from multiple sessions. This paper proposes a novel SBRS model, called Capsule GraphSAGE for Session-Based Recommendation (CapsGSR), that marries GraphSAGE's scalability and inductive learning capabilities with the Capsules network's abstraction levels by generating multiple integrations for each node from different perspectives. Consequently, CapsGSR addresses challenges that may hinder the optimal item representations and captures transitions' complex nature, mitigating the loss of crucial information. Our system significantly outperforms baseline models on benchmark datasets, with improvements of 8.44% in HR@20 and 4.66% in MRR@20 , indicating its effectiveness in delivering precise and relevant recommendations.

A universal strategy for smoothing deceleration in deep graph neural networks

Graph neural networks (GNNs) have shown great promise in modeling graph-structured data, but the over-smoothing problem restricts their effectiveness in deep layers. Two key weaknesses of existing research on deep GNN models are: (1) ignoring the beneficial aspects of intra-class smoothing while focusing solely on reducing inter-class smoothing, and (2) inefficient computation of residual weights that neglect the influence of neighboring nodes' distributions. To address these weaknesses, we propose a novel Smoothing Deceleration (SD) strategy to reduce the smoothing speed rate of nodes as information propagates between layers, thereby mitigating over-smoothing. Firstly, we analyze the smoothing speed rate of node representations between layers by differential operations. Subsequently, based on this analysis, we introduce two innovative modules: Class-Related Smoothing Deceleration (CR-SD) loss and Smooth Deceleration Residual (NAR). CR-SD loss first takes into account the duality of smoothing, reducing inter-class smoothing while preserving the benefits of intra-class smoothing, thus reducing over-smoothing while maintaining model performance. NAR is specifically designed for graph-structured data, integrating the distribution of neighboring nodes, and is a novel method for computing residual weights. Finally, the comparative experimental results demonstrate that our SD strategy can extend existing shallow GNNs to deeper and delivers superior performance compared to both vanilla models and existing deep GNNs. And, a series of analytical experiments be conducted to prove that our proposed SD strategy effectively mitigates over-smoothing in deep GNNs. The source code for this work is available at https://github.com/cheng-qi/sd.

A simple yet effective approach for predicting disease spread using mathematically-inspired diffusion-informed neural networks

The COVID-19 outbreak has highlighted the importance of mathematical epidemic models like the Susceptible-Infected-Recovered (SIR) model, for understanding disease spread dynamics. However, enhancing their predictive accuracy complicates parameter estimation. To address this, we proposed a novel model that integrates traditional mathematical modeling with deep learning which has shown improved predicted power across diverse fields. The proposed model includes a simple artificial neural network (ANN) for regional disease incidences, and a graph convolutional neural network (GCN) to capture spread to adjacent regions. GCNs are a recent deep learning algorithm designed to learn spatial relationship from graph-structured data. We applied the model to COVID-19 incidences in Spain to evaluate its performance. It achieved a 0.9679 correlation with the test data, outperforming previous models with fewer parameters. By leveraging the efficient training methods of deep learning, the model simplifies parameter estimation while maintaining alignment with the mathematical framework to ensure interpretability. The proposed model may allow the more robust and insightful analyses by leveraging the generalization power of deep learning and theoretical foundations of the mathematical models.

RT-Pred: A web server for accurate, customized liquid chromatography retention time prediction of chemicals

High-performance liquid chromatography (HPLC) together with mass spectrometry (MS) is routinely used to separate, identify and quantify chemicals. HPLC data also provides retention time (RT) which can be aligned with structural data. Recent developments in machine learning (ML) have improved our ability to predict RTs from known or postulated chemical structures, allowing RT data to be used more effectively in LC-MS-based compound identification. However, RT data is highly specific to each chromatographic method (CM) and hundreds of different CMs with interdependent parameters are used in separations. This has limited the application of ML-based RT predictions in compound identification. Here we introduce an easy-to-use RT prediction webserver (called RT-Pred) that predicts RTs for molecules across most chromatographic setups. RT-Pred not only supports its own in-house CM-specific RT predictors, it allows users to easily train a custom RT-Pred model using their own RT data on their own CM and to predict RTs with that custom model. RT-Pred also supports RT and compound searches against its own database of millions of predicted RTs spanning >40 different CMs. RT-Pred is also uniquely capable of accurately identifying compounds that will elute in the void volume or be retained on the column. Including this void/retained/eluted classifier significantly improves RT-Pred's performance. Tests indicate that RT-Pred had an average coefficient of determination (R²) of 0.95 over 20 different CMs. Comparisons of RT-Pred against other RT predictors showed that RT-Pred achieved lower mean absolute errors and higher R² scores than any other published RT predictor. RT-Pred is freely available at https://rtpred.ca.

SGTB: A graph representation learning model combining transformer and BERT for optimizing gene expression analysis in spatial transcriptomics data

In recent years, spatial transcriptomics (ST) has emerged as an innovative technology that enables the simultaneous acquisition of gene expression information and its spatial distribution at the single-cell or regional level, providing deeper insights into cellular interactions and tissue organization, this technology provides a more holistic view of tissue organization and intercellular dynamics. However, existing methods still face certain limitations in data representation capabilities, making it challenging to fully capture complex spatial dependencies and global features. To address this, this paper proposes an innovative spatial multi-scale graph convolutional network (SGTB) based on large language models, integrating graph convolutional networks (GCN), Transformer, and BERT language models to optimize the representation of spatial transcriptomics data. The Graph Convolutional Network (GCN) employs a multi-layer architecture to extract features from gene expression matrices. Through iterative aggregation of neighborhood information, it captures spatial dependencies among cells and gene co-expression patterns, thereby constructing hierarchical cell embeddings. Subsequently, the model integrates an attention mechanism to assign weights to critical features and leverages Transformer layers to model global relationships, refining the ability of learned representations to reflect variations in spatial patterns. Finally, the model incorporates the BERT language model, mapping cell embeddings into textual inputs to exploit its deep semantic representation capabilities for high-dimensional feature extraction. These features are then fused with the embeddings generated by the Transformer, further optimizing feature learning for spatial transcriptomics data. This approach holds significant application value in improving the accuracy of tasks such as cell type classification and gene regulatory network construction, providing a novel computational framework for deep mining of spatial multi-scale biological data.

Shapley Fields Reveal Chemotopic Organization in the Mouse Olfactory Bulb Across Diverse Chemical Feature Sets

Representations of chemical features in the neural activity of the olfactory bulb (OB) are not well-understood, unlike the neural code for stimuli of the other sensory modalities. This is because the space of olfactory stimuli lacks a natural coordinate system, and this significantly complicates characterizing neural receptive fields (tuning curves), analogous to those in the other sensory modalities. The degree to which olfactory tuning is spatially organized across the OB, often referred to as chemotopy, is also not well-understood. To advance our understanding of these aspects of olfactory coding, we introduce an interpretable method of Shapley fields, as an olfactory analog of retinotopic receptive fields. Shapley fields are spatial distributions of chemical feature importance for the tuning of OB glomeruli. We used this tool to investigate chemotopy in the OB with diverse sets of chemical features using widefield epifluorescence recordings of the mouse dorsal OB in response to stimuli across a wide range of the chemical space. We found that Shapley fields reveal a weak chemotopic organization of the chemical feature sensitivity of dorsal OB glomeruli. This organization was consistent across animals and mostly agreed across very different chemical feature sets: (i) the expert-curated PubChem database features and (ii) features derived from a Graph Neural Network trained on human olfactory perceptual tasks. Moreover, we found that the principal components of the Shapley fields often corresponded to single commonly accepted chemical classification groups, that therefore could be "recovered" from the neural activity in the mouse OB. Our findings suggest that Shapley fields may serve as a chemical feature-agnostic method for investigating olfactory perception.

Deep learning in neurosurgery: a systematic literature review with a structured analysis of applications across subspecialties

Objective

This study systematically reviewed deep learning (DL) applications in neurosurgical practice to provide a comprehensive understanding of DL in neurosurgery. The review process included a systematic overview of recent developments in DL technologies, an examination of the existing literature on their applications in neurosurgery, and insights into the future of neurosurgery. The study also summarized the most widely used DL algorithms, their specific applications in neurosurgical practice, their limitations, and future directions.

Materials and methods

An advanced search using medical subject heading terms was conducted in Medline (via PubMed), Scopus, and Embase databases restricted to articles published in English. Two independent neurosurgically experienced reviewers screened selected articles.

Results

A total of 456 articles were initially retrieved. After screening, 162 were found eligible and included in the study. Reference lists of all 162 articles were checked, and 19 additional articles were found eligible and included in the study. The 181 included articles were divided into 6 categories according to the subspecialties: general neurosurgery (n = 64), neuro-oncology (n = 49), functional neurosurgery (n = 32), vascular neurosurgery (n = 17), neurotrauma (n = 9), and spine and peripheral nerve (n = 10). The leading procedures in which DL algorithms were most commonly used were deep brain stimulation and subthalamic and thalamic nuclei localization (n = 24) in the functional neurosurgery group; segmentation, identification, classification, and diagnosis of brain tumors (n = 29) in the neuro-oncology group; and neuronavigation and image-guided neurosurgery (n = 13) in the general neurosurgery group. Apart from various video and image datasets, computed tomography, magnetic resonance imaging, and ultrasonography were the most frequently used datasets to train DL algorithms in all groups overall (n = 79). Although there were few studies involving DL applications in neurosurgery in 2016, research interest began to increase in 2019 and has continued to grow in the 2020s.

Conclusion

DL algorithms can enhance neurosurgical practice by improving surgical workflows, real-time monitoring, diagnostic accuracy, outcome prediction, volumetric assessment, and neurosurgical education. However, their integration into neurosurgical practice involves challenges and limitations. Future studies should focus on refining DL models with a wide variety of datasets, developing effective implementation techniques, and assessing their affect on time and cost efficiency.

Hybrid intelligence for environmental pollution: biodegradability assessment of organic compounds through multimodal integration of graph attention networks and QSAR models

Computational methods are crucial for assessing chemical biodegradability, given their significant impact on both environmental and human health. Organic compounds that are not biodegradable can persist in the environment, contributing to pollution. Our novel approach leverages graph attention networks (GATs) and incorporates node and edge attributes for biodegradability prediction. Quantitative Structure-Activity Relationship (QSAR) models using two-dimensional descriptors alongside weighted average and stacking approaches were employed to generate ensemble models. The GAT models demonstrated a stable function and generally higher specificity on the validation set compared to a graph convolutional network, although definitive superiority is challenging to establish owing to overlapping standard deviations. However, the sensitivities tended to decrease with potential performance overlap owing to the interval intersection. Ensemble learning enhanced several performance metrics compared with individual models and base models, with the combination of extreme Gradient Boosting and GAT achieving the highest precision and specificity. Combining GAT with random forest and Gradient Boosting may be preferable for accurately predicting biodegradable molecules, whereas the stacking approach may be suitable for prioritizing the correct classification of nonbiodegradable substances. Important descriptors, such as SpMax1_Bh(m) and SAscore, were identified in at least two QSAR models. Despite inherent complexities, the ease of implementation depends on factors such as data availability, and domain knowledge. Assessing the biodegradability of organic compounds is essential for reducing their environmental impact, assessing risks, ensuring regulatory compliance, promoting sustainable development, and supporting effective pollution remediation. It assists in making informed decisions about chemical use, waste management, and environmental protection.

Geometric deep learning framework for de novo genome assembly

The critical stage of every de novo genome assembler is identifying paths in assembly graphs that correspond to the reconstructed genomic sequences. The existing algorithmic methods struggle with this, primarily due to repetitive regions causing complex graph tangles, leading to fragmented assemblies. Here, we introduce GNNome, a framework for path identification based on geometric deep learning that enables training models on assembly graphs without relying on existing assembly strategies. By leveraging only the symmetries inherent to the problem, GNNome reconstructs assemblies from PacBio HiFi reads with contiguity and quality comparable to those of the state-of-the-art tools across several species. With every new genome assembled telomere-to-telomere, the amount of reliable training data at our disposal increases. Combining the straightforward generation of abundant simulated data for diverse genomic structures with the AI approach makes the proposed framework a plausible cornerstone for future work on reconstructing complex genomes with different degrees of ploidy and aneuploidy. To facilitate such developments, we make the framework and the best-performing model publicly available, provided as a tool that can directly be used to assemble new haploid genomes.

Reinforcement learning for healthcare operations management: methodological framework, recent developments, and future research directions

With the advancement in computing power and data science techniques, reinforcement learning (RL) has emerged as a powerful tool for decision-making problems in complex systems. In recent years, the research on RL for healthcare operations has grown rapidly. Especially during the COVID-19 pandemic, RL has played a critical role in optimizing decisions with greater degrees of uncertainty. RL for healthcare applications has been an exciting topic across multiple disciplines, including operations research, operations management, healthcare systems engineering, and data science. This review paper first provides a tutorial on the overall framework of RL, including its key components, training models, and approximators. Then, we present the recent advances of RL in the domain of healthcare operations management (HOM) and analyze the current trends. Our paper concludes by presenting existing challenges and future directions for RL in HOM.

Accurate and Rapid Prediction of Protein pKa: Protein Language Models Reveal the Sequence-pKa Relationship

Protein pKa prediction is a key challenge in computational biology. In this study, we present pKALM, a novel deep learning-based method for high-throughput protein pKa prediction. pKALM uses a protein language model (PLM) to capture the complex sequence-structure relationships of proteins. While traditionally considered a structure-based problem, our results show that a PLM pretrained on large-scale protein sequence databases can effectively learn this relationship and achieve state-of-the-art performance. pKALM accurately predicts the pKa values of six residues (Asp, Glu, His, Lys, Cys, and Tyr) and two termini with high precision and efficiency. It performs well at predicting both exposed and buried residues, which often deviate from standard pKa values measured in the solvent. We demonstrate a novel finding that predicted protein isoelectric points (pI) can be used to improve the accuracy of pKa prediction. High-throughput pKa prediction of the human proteome using pKALM achieves a speed of 4,965 pKa predictions per second, which is several orders of magnitude faster than existing state-of-the-art methods. The case studies illustrate the efficacy of pKALM in estimating pKa values and the constraints of the method. pKALM will thus be a valuable tool for researchers in the fields of biochemistry, biophysics, and drug design.

Enhanced inhibitor-kinase affinity prediction via integrated multimodal analysis of drug molecule and protein sequence features

The accurate prediction of inhibitor-kinase binding affinity is pivotal for advancing drug development and precision medicine. In this study, we developed predictive models for human kinases, including cyclin-dependent kinases (CDKs), mitogen-activated protein kinases (MAP kinases), glycogen synthase kinases (GSKs), CDK-like kinases (CMGC kinase group) and receptor tyrosine kinases (RTKs)-key regulators of cellular signaling and disease progression. These kinases serve as primary drug targets in cancer and other critical diseases. To enhance affinity prediction precision, we introduce an innovative multimodal fusion model, KinNet. The model integrates the GraphKAN network, which effectively captures both local and global structural features of drug molecules. Furthermore, it leverages kernel functions and learnable activation functions to dynamically optimize node and edge feature representations. Additionally, the model incorporates the Conv-Enhanced Mamba module, combining Conv1D's ability to capture local features with Mamba's strength in processing long sequences, facilitating comprehensive feature extraction from protein sequences and molecular fingerprints. Experimental results confirm that the KinNet model achieves superior prediction accuracy compared to existing approaches, underscoring its potential to elucidate inhibitor-kinase binding mechanisms. This model serves as a robust computational framework to support drug discovery and the development of kinase-targeted therapies.

Quantum Chemical Density Matrix Renormalization Group Method Boosted by Machine Learning

The use of machine learning (ML) to refine low-level theoretical calculations to achieve higher accuracy is a promising and actively evolving approach known as Δ-ML. The density matrix renormalization group (DMRG) is a powerful variational approach widely used for studying strongly correlated quantum systems. High computational efficiency can be achieved without compromising accuracy. Here, we demonstrate the potential of a simple ML model to significantly enhance the performance of the quantum chemical DMRG method.

Structure-Informed Insights into Catalytic Mechanism and Multidomain Collaboration in α-Agarase CmAga

α-Agarases are glycoside hydrolases that cleave α-1,3-glycosidic bonds in agarose to produce bioactive agarooligosaccharides. Despite their great industrial potential, the structures and functional mechanisms of α-agarases remain unclear due to their complex and flexible architecture. Here, we investigated the structure-based catalytic mechanism of α-agarase CmAga from Catenovulum maritimum STB14 by integrated Cryo-EM and AlphaFold2. D994 and E1129 were identified as catalytic residues, with E1129 selectively recognizing α-1,3-glycosidic bonds. Y858, W1201, Y1164, and W1166 facilitate preferential substrate binding at the -3 ∼ +3 subsites. Molecular dynamics simulations and neural relational inference modeling revealed a cooperative mechanism involving the catalytic domain (CD) and four carbohydrate-binding modules (CBMs), with CBM6-1 and CBM6-2 capturing substrates, CBM_like transferring them to the CD, and CBM6-3 stabilizing the active site. D149 and L608 served as pivotal nodes within the interdomain communication pathways. These insights provide a foundation for mechanistic investigations and rational engineering of carbohydrate-active enzymes (CAZymes) with multiple CBMs.

ToxDL 2.0: Protein toxicity prediction using a pretrained language model and graph neural networks

Motivation

Assessing the potential toxicity of proteins is crucial for both therapeutic and agricultural applications. Traditional experimental methods for protein toxicity evaluation are time-consuming, expensive, and labor-intensive, highlighting the requirement for efficient computational approaches. Recent advancements in language models and deep learning have significantly improved protein toxicity prediction, yet current models often lack the ability to integrate evolutionary and structural information, which is crucial for accurate toxicity assessment of proteins.

Results

In this study, we present ToxDL 2.0, a novel multimodal deep learning model for protein toxicity prediction that integrates both evolutionary and structural information derived from a pretrained language model and AlphaFold2. ToxDL 2.0 consists of three key modules: (1) a Graph Convolutional Network (GCN) module for generating protein graph embeddings based on AlphaFold2-predicted structures, (2) a domain embedding module for capturing protein domain representations, and (3) a dense module that combines these embeddings to predict the toxicity. After constructing a comprehensive toxicity benchmark dataset, we obtained experimental results on both an original non-redundant test set (comprising pre-2022 protein sequences) and an independent non-redundant test set (a holdout set of post-2022 protein sequences), demonstrating that ToxDL 2.0 outperforms existing state-of-the-art methods. Additionally, we utilized Integrated Gradients to discover known toxic motifs associated with protein toxicity. A web server for ToxDL 2.0 is publicly available at www.csbio.sjtu.edu.cn/bioinf/ToxDL2/.

Nellie: automated organelle segmentation, tracking and hierarchical feature extraction in 2D/3D live-cell microscopy

Cellular organelles undergo constant morphological changes and dynamic interactions that are fundamental to cell homeostasis, stress responses and disease progression. Despite their importance, quantifying organelle morphology and motility remains challenging due to their complex architectures, rapid movements and the technical limitations of existing analysis tools. Here we introduce Nellie, an automated and unbiased pipeline for segmentation, tracking and feature extraction of diverse intracellular structures. Nellie adapts to image metadata and employs hierarchical segmentation to resolve sub-organellar regions, while its radius-adaptive pattern matching enables precise motion tracking. Through a user-friendly Napari-based interface, Nellie enables comprehensive organelle analysis without coding expertise. We demonstrate Nellie's versatility by unmixing multiple organelles from single-channel data, quantifying mitochondrial responses to ionomycin via graph autoencoders and characterizing endoplasmic reticulum networks across cell types and time points. This tool addresses a critical need in cell biology by providing accessible, automated analysis of organelle dynamics.

Modeling document causal structure with a hypergraph for event causality identification

Document-level event causality identification (ECI) aims to detect causal relations in between event mentions in a document. Some recent approaches model diverse connections in between events, such as syntactic dependency and etc., with a graph neural network for event node representation learning. However, not all such connections contribute to augment node representation for causality identification. We argue that the events' causal relations in a document are often interdependent, i.e., multiple causes with one effect, and identifying one cause for an effect may facilitate the identification of another cause of the same effect. In this paper, we use a hypergraph to model such events' causal relations as the document causal structure, and propose a neural causal hypergraph model (NCHM) for event causality identification. In NCHM, we design a pairwise event semantics learning module (PES) based on prompt learning to learn the pairwise event representation as well as the pairwise causal connections between two events. A document causal hypergraph is then constructed based on pairwise causal connections. We also design a document causal structure learning module (DCS) with a hypergraph convolutional neural network to learn document-wise events' representations. Finally, two kinds of representations are concatenated for document-level event causality identification. Experiments on both EventStoryLine and English-MECI corpus show that our NCHM significantly outperforms the state-of-the-art algorithms.

Dual-tower model with semantic perception and timespan-coupled hypergraph for next-basket recommendation

Next basket recommendation (NBR) is an essential task within the realm of recommendation systems and is dedicated to the anticipation of user preferences in the next moment based on the analysis of users' historical sequences of engaged baskets. Current NBR models utilise unique identity (ID) information to represent distinct users and items and focus on capturing the dynamic preferences of users through sequential encoding techniques such as recurrent neural networks and hierarchical time decay modelling, which have dominated the NBR field more than a decade. However, these models exhibit two significant limitations, resulting in suboptimal representations for both users and items. First, the dependence on unique ID information for the derivation of user and item representations ignores the rich semantic relations that interweave the items. Second, the majority of NBR models remain bound to model an individual user's historical basket sequence, thereby neglecting the broader vista of global collaborative relations among users and items. To address these limitations, we introduce a dual-tower model with semantic perception and timespan-coupled hypergraph for the NBR. It is carefully designed to integrate semantic and collaborative relations into both user and item representations. Specifically, to capture rich semantic relations effectively, we propose a hierarchical semantic attention mechanism with a large language model to integrate multi-aspect textual semantic features of items for basket representation learning. Simultaneously, to capture global collaborative relations explicitly, we design a timespan-coupled hypergraph convolutional network to efficiently model high-order structural connectivity on a hypergraph among users and items. Finally, a multi-objective joint optimisation loss is used to optimise the learning and integration of semantic and collaborative relations for recommendation. Comprehensive experiments on two public datasets demonstrate that our proposed model significantly outperforms the mainstream NBR models on two classical evaluation metrics, Recall and Normalised Discounted Cumulative Gain (NDCG).

Graph neural networks in histopathology: Emerging trends and future directions

Histopathological analysis of whole slide images (WSIs) has seen a surge in the utilization of deep learning methods, particularly Convolutional Neural Networks (CNNs). However, CNNs often fail to capture the intricate spatial dependencies inherent in WSIs. Graph Neural Networks (GNNs) present a promising alternative, adept at directly modeling pairwise interactions and effectively discerning the topological tissue and cellular structures within WSIs. Recognizing the pressing need for deep learning techniques that harness the topological structure of WSIs, the application of GNNs in histopathology has experienced rapid growth. In this comprehensive review, we survey GNNs in histopathology, discuss their applications, and explore emerging trends that pave the way for future advancements in the field. We begin by elucidating the fundamentals of GNNs and their potential applications in histopathology. Leveraging quantitative literature analysis, we explore four emerging trends: Hierarchical GNNs, Adaptive Graph Structure Learning, Multimodal GNNs, and Higher-order GNNs. Through an in-depth exploration of these trends, we offer insights into the evolving landscape of GNNs in histopathological analysis. Based on our findings, we propose future directions to propel the field forward. Our analysis serves to guide researchers and practitioners towards innovative approaches and methodologies, fostering advancements in histopathological analysis through the lens of graph neural networks.

Another Perspective of Over-Smoothing: Alleviating Semantic Over-Smoothing in Deep GNNs

Graph neural networks (GNNs) are widely used for analyzing graph-structural data and solving graph-related tasks due to their powerful expressiveness. However, existing off-the-shelf GNN-based models usually consist of no more than three layers. Deeper GNNs usually suffer from severe performance degradation due to several issues including the infamous "over-smoothing" issue, which restricts the further development of GNNs. In this article, we investigate the over-smoothing issue in deep GNNs. We discover that over-smoothing not only results in indistinguishable embeddings of graph nodes, but also alters and even corrupts their semantic structures, dubbed semantic over-smoothing. Existing techniques, e.g., graph normalization, aim at handling the former concern, but neglect the importance of preserving the semantic structures in the spatial domain, which hinders the further improvement of model performance. To alleviate the concern, we propose a cluster-keeping sparse aggregation strategy to preserve the semantic structure of embeddings in deep GNNs (especially for spatial GNNs). Particularly, our strategy heuristically redistributes the extent of aggregations for all the nodes from layers, instead of aggregating them equally, so that it enables aggregate concise yet meaningful information for deep layers. Without any bells and whistles, it can be easily implemented as a plug-and-play structure of GNNs via weighted residual connections. Last, we analyze the over-smoothing issue on the GNNs with weighted residual structures and conduct experiments to demonstrate the performance comparable to the state-of-the-arts.

IBPL: Information Bottleneck-based Prompt Learning for graph out-of-distribution detection

When training and test graph samples follow different data distributions, graph out-of-distribution (OOD) detection becomes an indispensable component of constructing the reliable and safe graph learning systems. Motivated by the significant progress on prompt learning, graph prompt-based methods, which enable a well-trained graph neural network to detect OOD graphs without modifying any model parameters, have been a standard benchmark with promising computational efficiency and model effectiveness. However, these methods ignore the influence of overlapping features existed in both in-distribution (ID) and OOD graphs, which weakens the difference between them and leads to sub-optimal detection results. In this paper, we present the Information Bottleneck-based Prompt Learning (IBPL) to overcome this challenging problem. Specifically, IBPL includes a new graph prompt that jointly performs the mask operation on node features and the graph structure. Building upon this, we develop an information bottleneck (IB)-based objective to optimize the proposed graph prompt. Since the overlapping features are inaccessible, IBPL introduces the noise data augmentation which generates a series of perturbed graphs to fully covering the overlapping features. Through minimizing the mutual information between the prompt graph and the perturbed graphs, our objective can eliminate the overlapping features effectively. In order to avoid the negative impact of perturbed graphs, IBPL simultaneously maximizes the mutual information between the prompt graph and the category label for better extracting the ID features. We conduct experiments on multiple real-world datasets in both supervised and unsupervised scenarios. The empirical results and extensive model analyses demonstrate the superior performance of IBPL over several competitive baselines.

Deep reinforcement learning can promote sustainable human behaviour in a common-pool resource problem

A canonical social dilemma arises when resources are allocated to people, who can either reciprocate with interest or keep the proceeds. The right resource allocation mechanisms can encourage levels of reciprocation that sustain the commons. Here, in an iterated multiplayer trust game, we use deep reinforcement learning (RL) to design a social planner that promotes sustainable contributions from human participants. We first trained neural networks to behave like human players, creating a stimulated economy that allows us to study the dynamics of receipt and reciprocation. We use RL to train a mechanism to maximise aggregate return to players. The RL mechanism discovers a redistributive policy that leads to a large but also more equal surplus. The mechanism outperforms baseline mechanisms by conditioning its generosity on available resources and temporarily sanctioning defectors. Examining the RL policy allows us to develop a similar but explainable mechanism that is more popular among players.

DeeWaNA: An Unsupervised Network Representation Learning Framework Integrating Deepwalk and Neighborhood Aggregation for Node Classification

This paper introduces DeeWaNA, an unsupervised network representation learning framework that unifies random walk strategies and neighborhood aggregation mechanisms to improve node classification performance. Unlike existing methods that treat these two paradigms separately, our approach integrates them into a cohesive model, addressing limitations in structural feature extraction and neighborhood relationship modeling. DeeWaNA first leverages DeepWalk to capture global structural information and then employs an attention-based weighting mechanism to refine neighborhood relationships through a novel distance metric. Finally, a weighted aggregation operator fuses these representations into a unified low-dimensional space. By bridging the gap between random-walk-based and neural-network-based techniques, our framework enhances representation quality and improves classification accuracy. Extensive evaluations on real-world networks demonstrate that DeeWaNA outperforms four widely used unsupervised network representation learning methods, underscoring its effectiveness and broader applicability.

Graph-based deep fusion for architectural text representation

Amidst the swift global urbanization and rapid evolution of the architecture industry, there is a growing demand for the automated processing of architectural textual information. This demand arises from the abundance of specialized vocabulary in architectural texts, posing a challenge for accurate representation using traditional models. To address this, we propose a novel fusion method that integrates Transformer-based models with graph neural networks (GNNs) for architectural text representation. While independently utilizing Bidirectional Encoder Representations from Transformers (BERT) and the robustly optimized BERT approach (RoBERTa) to generate initial document representations, we also employ term frequency-inverse document frequency (TF-IDF) to extract keywords from each document and construct a corresponding keyword set. Subsequently, a graph is created based on the keyword vocabulary and document embeddings, which is then fed into the graph attention network (GAT). The final document embedding is generated by GAT, and the text embedding is crafted by the attention module and neural network structure of the GAT. Experimental results from comparison studies show that the proposed model outperforms all baselines. Additionally, ablation studies demonstrate the effectiveness of each module, further reinforcing the robustness and superiority of our approach.

Multilevel Context Learning with Large Language Models for Text-Attributed Graphs on Social Networks

There are complex graph structures and rich textual information on social networks. Text provides important information for various tasks, while graph structures offer multilevel context for the semantics of the text. Contemporary researchers tend to represent these kinds of data by text-attributed graphs (TAGs). Most TAG-based representation learning methods focus on designing frameworks that convey graph structures to large language models (LLMs) to generate semantic embeddings for downstream graph neural networks (GNNs). However, these methods only provide text attributes for nodes, which fails to capture the multilevel context and leads to the loss of valuable information. To tackle this issue, we introduce the Multilevel Context Learner (MCL) model, which leverages multilevel context on social networks to enhance LLMs' semantic embedding capabilities. We model the social network as a multilevel context textual-edge graph (MC-TEG), effectively capturing both graph structure and semantic relationships. Our MCL model leverages the reasoning capabilities of LLMs to generate semantic embeddings by integrating these multilevel contexts. The tailored bidirectional dynamic graph attention layers are introduced to further distinguish the weight information. Experimental evaluations on six real social network datasets show that the MCL model consistently outperforms all baseline models. Specifically, the MCL model achieves prediction accuracies of 77.98%, 77.63%, 74.61%, 76.40%, 72.89%, and 73.40%, with absolute improvements of 9.04%, 9.19%, 11.05%, 7.24%, 6.11%, and 9.87% over the next best models. These results demonstrate the effectiveness of the proposed MCL model.

Prediction of herbal compatibility for colorectal adenoma treatment based on graph neural networks

Colorectal adenoma is a common precancerous lesion with a high risk of malignant transformation. Traditional Chinese medicine and its complex prescriptions have shown promising efficacy in the treatment of adenomas; however, there remains a lack of systematic understanding regarding the compatibility patterns within these prescriptions, as well as an effective model for predicting therapeutic outcomes. In this study, we collected numerous TCM prescriptions and their components, recommended by experts for the treatment of colorectal adenoma, and developed a heterogeneous graph neural network model to predict the compatibility strength and probability among the herbs within these prescriptions. This model delineates the complex relationships among herbs, active compounds, and molecular targets, allowing for a quantification of the interactions and compatibility potential among the herbs. Using this model, we identified high-potential therapeutic prescriptions from clinical prescription records and identified their active components through network pharmacology. Through this approach, we aim to provide a theoretical foundation for the clinical TCM treatment of colorectal adenoma, foster the discovery of new prescriptions to optimize the therapeutic efficacy of TCM, and ultimately advance the field of cancer prevention and treatment based on traditional Chinese medicine.

Graph neural networks for single-cell omics data: a review of approaches and applications

Rapid advancement of sequencing technologies now allows for the utilization of precise signals at single-cell resolution in various omics studies. However, the massive volume, ultra-high dimensionality, and high sparsity nature of single-cell data have introduced substantial difficulties to traditional computational methods. The intricate non-Euclidean networks of intracellular and intercellular signaling molecules within single-cell datasets, coupled with the complex, multimodal structures arising from multi-omics joint analysis, pose significant challenges to conventional deep learning operations reliant on Euclidean geometries. Graph neural networks (GNNs) have extended deep learning to non-Euclidean data, allowing cells and their features in single-cell datasets to be modeled as nodes within a graph structure. GNNs have been successfully applied across a broad range of tasks in single-cell data analysis. In this survey, we systematically review 107 successful applications of GNNs and their six variants in various single-cell omics tasks. We begin by outlining the fundamental principles of GNNs and their six variants, followed by a systematic review of GNN-based models applied in single-cell epigenomics, transcriptomics, spatial transcriptomics, proteomics, and multi-omics. In each section dedicated to a specific omics type, we have summarized the publicly available single-cell datasets commonly utilized in the articles reviewed in that section, totaling 77 datasets. Finally, we summarize the potential shortcomings of current research and explore directions for future studies. We anticipate that this review will serve as a guiding resource for researchers to deepen the application of GNNs in single-cell omics.

Deep learning in single-cell and spatial transcriptomics data analysis: advances and challenges from a data science perspective

The development of single-cell and spatial transcriptomics has revolutionized our capacity to investigate cellular properties, functions, and interactions in both cellular and spatial contexts. Despite this progress, the analysis of single-cell and spatial omics data remains challenging. First, single-cell sequencing data are high-dimensional and sparse, and are often contaminated by noise and uncertainty, obscuring the underlying biological signal. Second, these data often encompass multiple modalities, including gene expression, epigenetic modifications, metabolite levels, and spatial locations. Integrating these diverse data modalities is crucial for enhancing prediction accuracy and biological interpretability. Third, while the scale of single-cell sequencing has expanded to millions of cells, high-quality annotated datasets are still limited. Fourth, the complex correlations of biological tissues make it difficult to accurately reconstruct cellular states and spatial contexts. Traditional feature engineering approaches struggle with the complexity of biological networks, while deep learning, with its ability to handle high-dimensional data and automatically identify meaningful patterns, has shown great promise in overcoming these challenges. Besides systematically reviewing the strengths and weaknesses of advanced deep learning methods, we have curated 21 datasets from nine benchmarks to evaluate the performance of 58 computational methods. Our analysis reveals that model performance can vary significantly across different benchmark datasets and evaluation metrics, providing a useful perspective for selecting the most appropriate approach based on a specific application scenario. We highlight three key areas for future development, offering valuable insights into how deep learning can be effectively applied to transcriptomic data analysis in biological, medical, and clinical settings.

Leveraging deep-learning and unconventional data for real-time surveillance, forecasting, and early warning of respiratory pathogens outbreak

BACKGROUND

Controlling re-emerging outbreaks such as COVID-19 is a critical concern to global health. Disease forecasting solutions are extremely beneficial to public health emergency management. This work aims to design and deploy a framework for real-time surveillance, prediction, forecasting, and early warning of respiratory disease. To this end, we selected southern African countries and Canadian provinces, along with COVID-19 and influenza as our case studies.

METHODOLOGY

Six different datasets were collected for different provinces of Canada: number of influenza cases, number of COVID-19 cases, Google Trends, Reddit posts, satellite air quality data, and weather data. Moreover, five different data sources were collected for southern African countries whose COVID-19 number of cases were significantly correlated with each other: number of COVID-19 infections, Google Trends, Wiki Trends, Google News, and satellite air quality data. For each infectious disease, i.e. COVID-19 and Influenza for Canada and COVID-19 for southern African countries, data was processed, scaled, and fed into the deep learning model which included four layers, namely, a Convolutional Neural Network (CNN), a Graph Neural Network (GNN), a Gated Recurrent Unit (GRU), and a linear Neural Network (NN). Hyperparameters were optimized to provide an accurate 56-day-ahead prediction of the number of cases.

RESULT

The accuracy of our models in real-time surveillance, prediction, forecasting, and early warning of respiratory diseases are evaluated against state-of-the-art models, through Root Mean Square Error (RMSE), coefficient of determination (R2-score), and correlation coefficient. Our model improves R2-score, RMSE, and correlation by up to 55.98 %, 39.71 %, and 44.47 % for 56 days-ahead COVID-19 prediction in Ontario, 34.87 %, 25.52 %, 50.91 % for 8 weeks-ahead influenza prediction in Quebec, and 51.04 %, 32.04 %, and 28.74 % for 56 days-ahead COVID-19 prediction in South Africa, respectively.

CONCLUSION

This work presents a framework that automatically collects data from unconventional sources, and builds an early warning system for COVID-19 and influenza outbreaks. The result is extremely helpful to policy-makers and health officials for preparedness and rapid response against future outbreaks.

CoNglyPred: Accurate Prediction of N-Linked Glycosylation Sites Using ESM-2 and Structural Features With Graph Network and Co-Attention

N-Linked glycosylation is crucial for various biological processes such as protein folding, immune response, and cellular transport. Traditional experimental methods for determining N-linked glycosylation sites entail substantial time and labor investment, which has led to the development of computational approaches as a more efficient alternative. However, due to the limited availability of 3D structural data, existing prediction methods often struggle to fully utilize structural information and fall short in integrating sequence and structural information effectively. Motivated by the progress of protein pretrained language models (pLMs) and the breakthrough in protein structure prediction, we introduced a high-accuracy model called CoNglyPred. Having compared various pLMs, we opt for the large-scale pLM ESM-2 to extract sequence embeddings, thus mitigating certain limitations associated with manual feature extraction. Meanwhile, our approach employs a graph transformer network to process the 3D protein structures predicted by AlphaFold2. The final graph output and ESM-2 embedding are intricately integrated through a co-attention mechanism. Among a series of comprehensive experiments on the independent test dataset, CoNglyPred outperforms state-of-the-art models and demonstrates exceptional performance in case study. In addition, we are the first to report the uncertainty of N-linked glycosylation predictors using expected calibration error and expected uncertainty calibration error.

AEGNN-M:A 3D Graph-Spatial Co-Representation Model for Molecular Property Prediction

Improving the drug development process can expedite the introduction of more novel drugs that cater to the demands of precision medicine. Accurately predicting molecular properties remains a fundamental challenge in drug discovery and development. Currently, a plethora of computer-aided drug discovery (CADD) methods have been widely employed in the field of molecular prediction. However, most of these methods primarily analyze molecules using low-dimensional representations such as SMILES notations, molecular fingerprints, and molecular graph-based descriptors. Only a few approaches have focused on incorporating and utilizing high-dimensional spatial structural representations of molecules. In light of the advancements in artificial intelligence, we introduce a 3D graph-spatial co-representation model called AEGNN-M, which combines two graph neural networks, GAT and EGNN. AEGNN-M enables learning of information from both molecular graphs representations and 3D spatial structural representations to predict molecular properties accurately. We conducted experiments on seven public datasets, three regression datasets and 14 breast cancer cell line phenotype screening datasets, comparing the performance of AEGNN-M with state-of-the-art deep learning methods. Extensive experimental results demonstrate the satisfactory performance of the AEGNN-M model. Furthermore, we analyzed the performance impact of different modules within AEGNN-M and the influence of spatial structural representations on the model's performance. The interpretability analysis also revealed the significance of specific atoms in determining particular molecular properties.

Machine Learning in Polymer Research

Machine learning is increasingly being applied in polymer chemistry to link chemical structures to macroscopic properties of polymers and to identify chemical patterns in the polymer structures that help improve specific properties. To facilitate this, a chemical dataset needs to be translated into machine readable descriptors. However, limited and inadequately curated datasets, broad molecular weight distributions, and irregular polymer configurations pose significant challenges. Most off the shelf mathematical models often need refinement for specific applications. Addressing these challenges demand a close collaboration between chemists and mathematicians as chemists must formulate research questions in mathematical terms while mathematicians are required to refine models for specific applications. This review unites both disciplines to address dataset curation hurdles and highlight advances in polymer synthesis and modeling that enhance data availability. It then surveys ML approaches used to predict solid-state properties, solution behavior, composite performance, and emerging applications such as drug delivery and the polymer-biology interface. A perspective of the field is concluded and the importance of FAIR (findability, accessibility, interoperability, and reusability) data and the integration of polymer theory and data are discussed, and the thoughts on the machine-human interface are shared.

Intradialytic Hypotension Frequency Prediction Using Generalizable Neighborhood Reasoning on Temporal Patient Knowledge Graph

Intradialytic hypotension (IDH) is a common complication among hemodialysis patients, adversely affecting quality of life and elevating mortality risk. IDH prediction enables physicians to take proactive measures, effectively reducing its occurrence. However, most prediction works rely on machine learning models, with a focus on real-time or session-level IDH. Hemodialysis patient data is multi-type and temporal, necessitating research on patient condition representation and temporal information utilization. Knowledge graphs (KGs) offer flexible data modeling and encompass rich structured information. This study represents patients using KGs and reason on graph structures to predict IDH. To study monthly IDH and utilize temporal information, a temporal patient KG is constructed. Patient KGs are first built at the monthly granularity based on data of 532 patients between January 2017 and August 2022. Six sequential monthly KGs are then combined into an observation window, resulting in a temporal KG dataset of 15,807 independent windows from 458 patients. The aim of this study is to utilize information from multiple months within a window to predict frequent IDH in the last month. However, the characteristics of IDH scenario and generalizability requirement pose challenges for the application of general KG reasoning models. Therefore, we adopt neighborhood-based KG reasoning and devise a visible feature guided patient-centric graph convolution to obtain patients' generalizable representations. Finally, patient representations in a window are fused using a sequential model, and processed by a prediction MLP to obtain the prediction results. Compared to 7 classic machine learning models, our model demonstrates superior performance in comprehensive metrics such as accuracy and F1 score.

Multimodal generative AI for medical image interpretation

Accurately interpreting medical images and generating insightful narrative reports is indispensable for patient care but places heavy burdens on clinical experts. Advances in artificial intelligence (AI), especially in an area that we refer to as multimodal generative medical image interpretation (GenMI), create opportunities to automate parts of this complex process. In this Perspective, we synthesize progress and challenges in developing AI systems for generation of medical reports from images. We focus extensively on radiology as a domain with enormous reporting needs and research efforts. In addition to analysing the strengths and applications of new models for medical report generation, we advocate for a novel paradigm to deploy GenMI in a manner that empowers clinicians and their patients. Initial research suggests that GenMI could one day match human expert performance in generating reports across disciplines, such as radiology, pathology and dermatology. However, formidable obstacles remain in validating model accuracy, ensuring transparency and eliciting nuanced impressions. If carefully implemented, GenMI could meaningfully assist clinicians in improving quality of care, enhancing medical education, reducing workloads, expanding specialty access and providing real-time expertise. Overall, we highlight opportunities alongside key challenges for developing multimodal generative AI that complements human experts for reliable medical report writing.

DropNaE: Alleviating irregularity for large-scale graph representation learning

Large-scale graphs are prevalent in various real-world scenarios and can be effectively processed using Graph Neural Networks (GNNs) on GPUs to derive meaningful representations. However, the inherent irregularity found in real-world graphs poses challenges for leveraging the single-instruction multiple-data execution mode of GPUs, leading to inefficiencies in GNN training. In this paper, we try to alleviate this irregularity at its origin-the irregular graph data itself. To this end, we propose DropNaE to alleviate the irregularity in large-scale graphs by conditionally dropping nodes and edges before GNN training. Specifically, we first present a metric to quantify the neighbor heterophily of all nodes in a graph. Then, we propose DropNaE containing two variants to transform the irregular degree distribution of the large-scale graph to a uniform one, based on the proposed metric. Experiments show that DropNaE is highly compatible and can be integrated into popular GNNs to promote both training efficiency and accuracy of used GNNs. DropNaE is offline performed and requires no online computing resources, benefiting the state-of-the-art GNNs in the present and future to a significant extent.

A graph neural network-based model with out-of-distribution robustness for enhancing antiretroviral therapy outcome prediction for HIV-1

Predicting the outcome of antiretroviral therapies (ART) for HIV-1 is a pressing clinical challenge, especially when the ART includes drugs with limited effectiveness data. This scarcity of data can arise either due to the introduction of a new drug to the market or due to limited use in clinical settings, resulting in clinical dataset with highly unbalanced therapy representation. To tackle this issue, we introduce a novel joint fusion model, which combines features from a Fully Connected (FC) Neural Network and a Graph Neural Network (GNN) in a multi-modality fashion. Our model uses both tabular data about genetic sequences and a knowledge base derived from Stanford drug-resistance mutation tables, which serve as benchmark references for deducing in-vivo treatment efficacy based on the viral genetic sequence. By leveraging this knowledge base structured as a graph, the GNN component enables our model to adapt to imbalanced data distributions and account for Out-of-Distribution (OoD) drugs. We evaluated these models' robustness against OoD drugs in the test set. Our comprehensive analysis demonstrates that the proposed model consistently outperforms the FC model. These results underscore the advantage of integrating Stanford scores in the model, thereby enhancing its generalizability and robustness, but also extending its utility in contributing in more informed clinical decisions with limited data availability. The source code is available at https://github.com/federicosiciliano/graph-ood-hiv.

Chemical Space Networks Enhance Toxicity Recognition via Graph Embedding

Chemical space networks (CSNs) are a new effective strategy for detecting latent chemical patterns irrespective of defined coordinate systems based on molecular descriptors and fingerprints. CSNs can be a new powerful option as a new approach method and increase the capacity of assessing potential adverse impacts of chemicals on human health. Here, CSNs are shown to effectively characterize the toxicity of chemicals toward several human health end points, namely chromosomal aberrations, mutagenicity, carcinogenicity, developmental toxicity, skin irritation, estrogenicity, androgenicity, and hepatoxicity. In this work, we report how the content from CSNs structure can be embedded through graph neural networks into a metric space, which, for eight different toxicological human health end points, allows better discrimination of toxic and nontoxic chemicals. In fact, using embeddings returns, on average, an increase in predictive performances. In fact, embedding employment enhances the learning, leading to an increment of the classification performance of +12% in terms of the area under the ROC curve. Moreover, through a dedicated eXplainable Artificial Intelligence framework, a straight interpretation of results is provided through the detection of putative structural alerts related to a given toxicity. Hence, the proposed approach represents a step forward in the area of alternative methods and could lead to breakthrough innovations in the design of safer chemicals and drugs.

TransPeakNet for solvent-aware 2D NMR prediction via multi-task pre-training and unsupervised learning

Nuclear Magnetic Resonance (NMR) spectroscopy is essential for revealing molecular structure, electronic environment, and dynamics. Accurate NMR shift prediction allows researchers to validate structures by comparing predicted and observed shifts. While Machine Learning (ML) has improved one-dimensional (1D) NMR shift prediction, predicting 2D NMR remains challenging due to limited annotated data. To address this, we introduce an unsupervised training framework for predicting cross-peaks in 2D NMR, specifically Heteronuclear Single Quantum Coherence (HSQC). Our approach pretrains an ML model on an annotated 1D dataset of 1H and 13C shifts, then finetunes it in an unsupervised manner using unlabeled HSQC data, which simultaneously generates cross-peak annotations. Our model also adjusts for solvent effects. Evaluation on 479 expert-annotated HSQC spectra demonstrates our model's superiority over traditional methods (ChemDraw and Mestrenova), achieving Mean Absolute Errors (MAEs) of 2.05 ppm and 0.165 ppm for 13C shifts and 1H shifts respectively. Our algorithmic annotations show a 95.21% concordance with experts' assignments, underscoring the approach's potential for structural elucidation in fields like organic chemistry, pharmaceuticals, and natural products.

Artificial Intelligence in Retrosynthesis Prediction and its Applications in Medicinal Chemistry

Retrosynthesis is a strategy to analyze the synthetic routes for target molecules in medicinal chemistry. However, traditional retrosynthesis predictions performed by chemists and rule-based expert systems struggle to adapt to the vast chemical space of real-world scenarios. Artificial intelligence (AI) has revolutionized retrosynthesis prediction in recent decades, significantly increasing the accuracy and diversity of predictions for target compounds. Single-step AI-driven retrosynthesis models can be generalized into three types based on their dependence on predefined reaction templates (template-based, semitemplate-based methods, template-free models), with respective advantages and limitations, and common challenges that limit their medicinal chemistry applications. Moreover, there are relatively inadequate multi-step retrosynthesis methods, which lack strong links with single-step methods. Herein, we review the recent advancements in AI applications for retrosynthesis prediction by summarizing related techniques and the landscape of current representative retrosynthesis models and propose feasible solutions to tackle existing problems and outline future directions in this field.

Two-stream spatio-temporal GCN-transformer networks for skeleton-based action recognition

For the purpose of achieving accurate skeleton-based action recognition, the majority of prior approaches have adopted a serial strategy that combines Graph Convolutional Networks (GCNs) with attention-based methods. However, this approach frequently treats the human skeleton as an isolated and complete structure, neglecting the significance of highly correlated yet indirectly connected skeletal parts, finally hindering recognition accuracy. This study proposes a novel architecture addressing this limitation by implementing a parallel configuration of GCNs and the Transformer model (SA-TDGFormer). This parallel structure integrates the advantages of both the GCN model and the Transformer model, facilitating the extraction of both local and global spatio-temporal features, leading to more accurate motion information encoding and improved recognition performance. The proposed model distinguishes itself through its dual-stream structure: a spatiotemporal GCN stream and a spatiotemporal Transformer stream. The former focuses on capturing the topological structure and motion representations of human skeletons. In contrast, the latter seeks to capture motion representations that consist of global inter-joint relationships. Recognizing the unique feature representations generated by these streams and their limited mutual understanding, the model also incorporates a late fusion strategy to merge the results from the two streams. This fusion allows the spatiotemporal GCN and Transformer streams to complement each other, enriching action features and maximizing information exchange between the two representation types. Empirical validation on three established benchmark datasets, NTU RGB + D 60, NTU RGB + D 120, and Kinetics-Skeleton, substantiates the model's effectiveness. The experimental results indicate that, compared to existing classification frameworks, the method proposed in this paper improves the accuracy of human action recognition by 1-5% (NTU RGB + D 60 dataset). This improvement demonstrates the superior performance of the model in action recognition.

Conotoxins: Classification, Prediction, and Future Directions in Bioinformatics

Conotoxins, a diverse family of disulfide-rich peptides derived from the venom of Conus species, have gained prominence in biomedical research due to their highly specific interactions with ion channels, receptors, and neurotransmitter systems. Their pharmacological properties make them valuable molecular tools and promising candidates for therapeutic development. However, traditional conotoxin classification and functional characterization remain labor-intensive, necessitating the increasing adoption of computational approaches. In particular, machine learning (ML) techniques have facilitated advancements in sequence-based classification, functional prediction, and de novo peptide design. This review explores recent progress in applying ML and deep learning (DL) to conotoxin research, comparing key databases, feature extraction techniques, and classification models. Additionally, we discuss future research directions, emphasizing the integration of multimodal data and the refinement of predictive frameworks to enhance therapeutic discovery.

Machine learning approaches for predicting protein-ligand binding sites from sequence data

Proteins, composed of amino acids, are crucial for a wide range of biological functions. Proteins have various interaction sites, one of which is the protein-ligand binding site, essential for molecular interactions and biochemical reactions. These sites enable proteins to bind with other molecules, facilitating key biological functions. Accurate prediction of these binding sites is pivotal in computational drug discovery, helping to identify therapeutic targets and facilitate treatment development. Machine learning has made significant contributions to this field by improving the prediction of protein-ligand interactions. This paper reviews studies that use machine learning to predict protein-ligand binding sites from sequence data, focusing on recent advancements. The review examines various embedding methods and machine learning architectures, addressing current challenges and the ongoing debates in the field. Additionally, research gaps in the existing literature are highlighted, and potential future directions for advancing the field are discussed. This study provides a thorough overview of sequence-based approaches for predicting protein-ligand binding sites, offering insights into the current state of research and future possibilities.

BalancerGNN: Balancer Graph Neural Networks for imbalanced datasets: A case study on fraud detection

Fraud detection for imbalanced datasets is challenging due to machine learning models inclination to learn the majority class. Imbalance in fraud detection datasets affects how graphs are built, an important step in many Graph Neural Networks (GNNs). In this paper, we introduce our BalancerGNN framework to tackle with imbalanced datasets and show its effectiveness on fraud detection. Our framework has three major components: (i) node construction with feature representations, (ii) graph construction using balanced neighbor sampling, and (iii) GNN training using balanced training batches leveraging a custom loss function with multiple components. For node construction, we have introduced (i) Graph-based Variable Clustering (GVC) to optimize feature selection and remove redundancies by analyzing multi-collinearity and (ii) Encoder-Decoder based Dimensionality Reduction (EDDR) using transformer-based techniques to reduce feature dimensions while keeping important information intact about textual embeddings. Our experiments on Medicare, Equifax, IEEE, and auto insurance fraud datasets highlight the importance of node construction with features representations. BalancerGNN trained with balanced batches consistently outperforms other methods, showing strong abilities in identifying fraud cases, with sensitivity rates ranging from 72.87% to 81.23% across datasets while balancing specificity. Additionally, BalancerGNN achieves impressive accuracy rates, ranging from 73.99% to 94.28%. These outcomes underscore the crucial role of graph representation and neighbor sampling techniques in optimizing BalancerGNN for fraud detection models in real-world applications.

VC dimension of Graph Neural Networks with Pfaffian activation functions

Graph Neural Networks (GNNs) have emerged in recent years as a powerful tool to learn tasks across a wide range of graph domains in a data-driven fashion. Based on a message passing mechanism, GNNs have gained increasing popularity due to their intuitive formulation, closely linked to the Weisfeiler-Lehman (WL) test for graph isomorphism, to which they were demonstrated to be equivalent (Morris et al., 2019 and Xu et al., 2019). From a theoretical point of view, GNNs have been shown to be universal approximators, and their generalization capability - related to the Vapnik Chervonekis (VC) dimension (Scarselli et al., 2018) - has recently been investigated for GNNs with piecewise polynomial activation functions (Morris et al., 2023). The aim of our work is to extend this analysis on the VC dimension of GNNs to other commonly used activation functions, such as the sigmoid and hyperbolic tangent, using the framework of Pfaffian function theory. Bounds are provided with respect to the architecture parameters (depth, number of neurons, input size) as well as with respect to the number of colors resulting from the 1-WL test applied on the graph domain. The theoretical analysis is supported by a preliminary experimental study.

Weighted Echo State Graph Neural Networks Based on Robust and Epitaxial Film Memristors

Hardware system customized toward the demands of graph neural network learning would promote efficiency and strong temporal processing for graph-structured data. However, most amorphous/polycrystalline oxides-based memristors commonly have unstable conductance regulation due to random growth of conductive filaments. And graph neural networks based on robust and epitaxial film memristors can especially improve energy efficiency due to their high endurance and ultra-low power consumption. Here, robust and epitaxial Gd: HfO2-based film memristors are reported and construct a weighted echo state graph neural network (WESGNN). Benefiting from the optimized epitaxial films, the high switching speed (20 ns), low energy consumption (2.07 fJ), multi-value storage (4 bits), and high endurance (109) outperform most memristors. Notably, thanks to the appropriately dispersed conductance distribution (standard deviation = 7.68 nS), the WESGNN finely regulates the relative weights of input nodes and recursive matrix to realize state-of-the-art performance using the MUTAG and COLLAB datasets for graph classification tasks. Overall, robust and epitaxial film memristors offer nanoscale scalability, high reliability, and low energy consumption, making them energy-efficient hardware solutions for graph learning applications.