Liwei Liu

AI-assisted reading interfaces often present a few passages from a known scientific document. We evaluated how much annotated evidence remains visible under fixed selection budgets using SciFact and QASPER. Without fine-tuning, we compared leading passages, lexical overlap, document-local BM25, MiniLM, and BGE; SPECTER provided an additional exploratory scientific-text comparison. On 338 SciFact development evidence sets, BGE top-4 sentence recall was 0.788 versus 0.663 for BM25 (paired difference 0.125; 95% claim-cluster bootstrap interval [0.078, 0.171]), and complete coverage was 77.2% versus 64.2%. Only 12 of 25 multi-sentence sets were completely covered (48.0%). In an internal document-disjoint subset, BGE recall gains under 96-, 128-, and 160-word caps ranged from 0.086 to 0.097. QASPER results include an original 89-question yes/no comparison and sensitivities across 883 answerable questions and 714 uniquely text-mapped questions. SPECTER showed mixed results across the two datasets. These retrospective analyses concern evidence visibility within known documents. The illustrative prototype returns selected text and source positions; neither the results nor the prototype establishes improved reader understanding or decisions.

Fundus images from the two eyes of one patient can share disease and acquisition cues, making matching performance difficult to interpret. We developed a controlled evaluation protocol for same-patient cross-eye matching that combines disease matching, image quality controls, and patient-level bootstrap resampling. Frozen features from an ImageNet-pretrained 50-layer residual network (ResNet-50) were evaluated in 300 patients from DeepDRiD, a retinal dataset with two photographs per eye and patient-disjoint splits. In a post hoc 34-patient test subset matched on ordered bilateral diabetic retinopathy grades and right-eye quality tier, the area under the receiver operating characteristic curve (AUROC) was 0.861 (interval envelope, 0.681–0.995). A separate 33-patient four-view analysis showed higher cross-eye than matched-negative similarity; the interval for the same-eye–cross-eye difference included zero. External validation in 244 patients from the independent PAPILA glaucoma dataset yielded AUROC 0.859 under bilateral diagnosis matching (interval envelope, 0.804–0.907). A learned hierarchical head achieved a higher validation AUROC point estimate than frozen features in DeepDRiD (0.882 versus 0.843) but a lower estimate in a previously inspected 137-patient fundus–optical coherence tomography cohort (0.849 versus 0.902). These results support cross-eye correspondence under the tested controls and show that source-cohort optimization does not ensure better transfer. The protocol provides a reproducible assessment of bilateral fundus representations across cohorts.

The human ether-a-go-go-related gene (hERG) cardiac toxicity of a compound refers to its inhibitory effect on the hERG potassium channel. The hERG channel is crucial for cardiac depolarization, and its blockage can lead to prolongation of the QT interval, triggering arrhythmias and posing life-threatening risks. Therefore, assessing hERG cardiac toxicity is a vital consideration in drug development. Traditional assessment methods are complex and have low throughput, making the development of deep learning models to predict this toxicity essential for enhancing drug development efficiency, reducing risks, and promoting personalized treatment. In this paper, we propose a novel multi-type feature fusion framework, MTF-hERG, for accurately predicting the cardiac toxicity of hERG compounds. This framework integrates various molecular features such as molecular fingerprints, 2D molecular images, and 3D molecular graphs to comprehensively capture the intrinsic structures and properties of compounds. By utilizing fully connected neural networks, DenseNet, and Equivariant Graph Neural Networks for feature extraction, we ensure that the model can precisely identify molecular characteristics associated with hERG blocking activity. Through deep fusion of extracted features and the construction of fully connected layers with different activation functions, we achieve classification predictions of whether a compound is an hERG blocker and regression predictions of its hERG inhibitory capacity. When comparing MTF-hERG with other state-of-the-art methods using benchmark datasets, we found that the average ACC, AUC, AUPR, RMSE, and R² values of MTF-hERG were 0.926, 0.943, 0.913, 0.453, and 0.681, respectively. The results demonstrate that MTF-hERG exhibits excellent predictive performance in various scenarios, significantly outperforming the existing baseline models. Furthermore, the visualization results of MTF-hERG not only reveal the key features and decision mechanisms of the model but also provide valuable support for further optimization of molecular structures. Therefore, the MTF-hERG framework is poised to become a powerful tool for predicting the hERG cardiac toxicity of compounds, offering robust support for drug development and exerting a profound impact on human health.

Molecular carcinogenicity is a crucial factor in the development of cancer, and accurate prediction of it is vital for cancer prevention, treatment, and drug development. In recent years, deep learning has been applied to predict molecular carcinogenicity, but due to limitations in data quality and feature richness, these methods still need improvement in terms of accuracy, robustness, and interpretability. In this article, we propose a deep transfer learning model based on multimodal information fusion, called MMF-MCP, for molecular feature extraction and carcinogenicity prediction. We extract molecular graph features and fingerprint features using graph attention networks and convolutional neural networks, respectively, and process molecular images through a deep residual network, SE-ResNet18, equipped with a squeeze-and-excitation module. To more effectively utilize limited carcinogenicity data and enhance the model's predictive performance and generalization ability, we further apply a transfer learning strategy by pretraining the model on a molecular mutagenicity data set and then fine-tuning it on the carcinogenicity data set, enabling knowledge transfer and significant improvement in model performance. MMF-MCP achieves average ACC, AUC, SE, and SP scores of 0.8452, 0.8513, 0.8571, and 0.8333 on benchmark data sets for molecular carcinogenicity, significantly outperforming state-of-the-art molecular carcinogenicity prediction methods. Additionally, the visualization results of MMF-MCP on molecular images demonstrate its strong interpretability, providing significant assistance in visually observing and understanding the critical structures and features of molecular carcinogenicity. The source code for MMF-MCP is available at https://github.com/liuliwei1980/MCP.

CircRNAs are a type of single-stranded non-coding RNA molecules, and their unique feature is their closed circular structure. The interaction between circRNAs and RNA-binding proteins (RBPs) plays a key role in biological functions and is crucial for studying post-transcriptional regulatory mechanisms. The genome-wide circRNA binding event data obtained by cross-linking immunoprecipitation sequencing technology provides a foundation for constructing efficient computational model prediction methods. However, in existing studies, although machine learning techniques have been applied to predict circRNA-RBP interaction sites, these methods still have room for improvement in accuracy and lack interpretability. We propose CR-deal, which is an interpretable joint deep learning network that predicts the binding sites of circRNA and RBP through genome-wide circRNA data. CR-deal utilizes a graph attention network to unify sequence and structural features into the same view, more effectively utilizing structural features to improve accuracy. It can infer marker genes in the binding site through integrated gradient feature interpretation, thereby inferring functional structural regions in the binding site. We conducted benchmark tests on CR-deal on 37 circRNA datasets and 7 lncRNA datasets, respectively, and obtained the interpretability of CR-deal and discovered functional structural regions through 5 circRNA datasets. We believe that CR-deal can help researchers gain a deeper understanding of the functions and mechanisms of circRNA in living organisms and its critical role in the occurrence and development of diseases. The source code of CR-deal is provided free of charge on https://github.com/liuliwei1980/CR.

Accurate prediction of drug-drug interaction (DDI) is essential to improve clinical efficacy, avoid adverse effects of drug combination therapy, and enhance drug safety. Recently researchers have developed several computer-aided methods for DDI prediction. However, these methods lack the substructural features that are critical to drug interactions and are not effective in generalizing across domains and different distribution data. In this work, we present SAGAN, a domain adaptive interpretable substructure-aware graph attention network for DDI prediction. Based on attention mechanism and unsupervised clustering algorithm, we propose a new substructure segmentation method, which segments the drug molecule into multiple substructures, learns the mechanism of drug interaction from the perspective of interaction, and identifies important interaction regions between drugs. To enhance the generalization ability of the model, we improve and apply a conditional domain adversarial network to achieve cross-domain generalization by alternately optimizing the cross-entropy loss on the source domain and the adversarial loss of the domain discriminator. We evaluate and compare SAGAN with the state-of-the-art DDI prediction model on four real-world datasets for both in-domain and cross-domain scenarios, and show that SAGAN achieves the best overall performance. Moreover, the visualization results of the model show that SAGAN has achieved pharmacologically significant substructure extraction, which can help drug developers screen for some undiscovered local interaction sites, and provide important information for further drug structure optimization. The codes and datasets are available online at https://github.com/wyx2012/SAGAN.

Noncoding RNA refers to RNA that does not encode proteins. The lncRNA and miRNA it contains play crucial regulatory roles in organisms, and their aberrant expression is closely related to various diseases. Traditional experimental methods for validating the interactions of these RNAs have limitations, and existing prediction models exhibit relatively limited functionality, relying on isolated feature extraction and performing poorly in handling various types of small sample tasks. This paper proposes an improved de Bruijn graph that can inject RNA structural information into the graph while preserving sequence information. Furthermore, the improved de Bruijn graph enables graph neural networks to learn broader dependencies and correlations among data by introducing richer edge relationships. Meanwhile, the multitask learning model, DVMnet, proposed in this paper can handle multiple related tasks, and we optimize model parameters by integrating the total loss of three tasks. This enables multitask prediction of RNA interactions, disease associations, and subcellular localization. Compared with the best existing models in this field, DVMnet has achieved the best performance with a 3% improvement in the area under the curve value and demonstrates robust results in predicting diseases and subcellular localization. The improved de Bruijn graph is also applicable to various scenarios and can unify the sequence and structural information of various nucleic acids into a single graph.

The prediction of drug-target affinity (DTA) plays a crucial role in drug development and the identification of potential drug targets. In recent years, computer-assisted DTA prediction has emerged as a significant approach in this field. In this study, we propose a multi-modal deep learning framework called MMD-DTA for predicting drug-target binding affinity and binding regions. The model can predict DTA while simultaneously learning the binding regions of drug-target interactions through unsupervised learning. To achieve this, MMD-DTA first uses graph neural networks and target structural feature extraction network to extract multi-modal information from the sequences and structures of drugs and targets. It then utilizes the feature interaction and fusion modules to generate interaction descriptors for predicting DTA and interaction strength for binding region prediction. Our experimental results demonstrate that MMD-DTA outperforms existing models based on key evaluation metrics. Furthermore, external validation results indicate that MMD-DTA enhances the generalization capability of the model by integrating sequence and structural information of drugs and targets. The model trained on the benchmark dataset can effectively generalize to independent virtual screening tasks. The visualization of drug-target binding region prediction showcases the interpretability of MMD-DTA, providing valuable insights into the functional regions of drug molecules that interact with proteins