Molecular representation plays a central role in computational drug discovery. Pharmacophores, functional groups responsible for molecular bioactivity, have been widely studied in cheminformatics. However, their incorporation into molecular representation learning, particularly in a context reasoning or generalization, remains relatively limited. To address this gap, we propose PharmaQA, a pharmacophore oriented question answering framework that formulates tailored prompts to extract context-aware molecular semantics. Rather than encoding pharmacophore features, PharmaQA learns to answer pharmacophore related queries. This design enables flexible reasoning across diverse tasks, including molecular property prediction, compound-target interaction prediction, and binding affinity estimation. Experimental results on benchmark datasets demonstrate that PharmaQA achieves competitive performance. In a ligand discovery case study using FDA-approved compounds, the framework identified potential inhibitors for three therapeutic targets, with strong docking performance. As a generalizable and modular solution, PharmaQA incorporates pharmacophoric knowledge into molecular embeddings, enhancing both predictive accuracy and interpretability in drug discovery applications.

Vision foundation models (e.g., SAM2, CLIP) show strong generalization in natural image analysis but degrade significantly in specialized domains like medical imaging. This is critical for tasks such as brain tumor segmentation, where errors directly affect surgical planning and patient outcomes. In such contexts, segmentation must be highly reliable and structurally precise, underscoring the need for adaptable methods with low error tolerance. While fine-tuning is the dominant strategy, it is computationally expensive and prone to forgetting. To address this, we propose CausalBridgeNet, a causality-guided correction framework for medical image segmentation. Inspired by predictive coding theories of the Bayesian brain, our method introduces a Predictive Causal Reasoning Unit (PCRU) that estimates structured error maps and delivers targeted feedback to iteratively refine predictions. This forms a closed-loop, error-aware correction mechanism without modifying the foundation model. By keeping the backbone frozen, CausalBridgeNet preserves general visual priors while enhancing task-specific accuracy. On the BraTS 2025 benchmark, it achieves an average Dice score of 84.48 and HD95 of 5.48 across tumor subregions, demonstrating its effectiveness for high-precision medical segmentation.

Three-dimensional medical image segmentation plays a significant role in clinical diagnosis, treatment planning, and disease research, as it provides doctors with precise anatomical and lesion information and improves the accuracy and efficiency of medical decision-making. However, most existing 3D segmentation approaches rely heavily on densely volumetric data and often fail to perform segmentation properly for incomplete 3D volume acquisition, i.e., missing slices. In this work, we present InterFrameNet, a framework designed to predict intermediate lesion structures by modeling spatial relationships across frames, enabling robust segmentation performance under sparse acquisition conditions, without requiring full-volume information. Our method explicitly models cross-frame spatial continuity and leverages structural relationships between available frames to accurately infer missing lesion regions. This design significantly reduces the dependence on consecutive frames while fully exploiting contextual anatomical information. Extensive experiments on brain lesion datasets demonstrate that our approach achieves robust segmentation performance under sparse acquisition settings, offering a practical solution to maximize usability of incomplete clinical imaging data.

Facial Expression Recognition (FER) has vital applications in computer vision, multimedia, and human behavior analysis. The real-world environment presents a significant challenge for FER due to variations such as face pose changes and occlusions. Many existing methods rely on features extracted from the end of the backbone model, which often capture high-level semantic information, to make final predictions. In this paper, we proposed a novel Cascaded Feature Fusion Network (CFFN) for FER in the wild, where different levels of semantic features are learned to enrich the information flow to the downstream modules. To effectively use the facial cues, a multi-branch cascaded network structure was proposed, where a novel semantic feature fusion block (SFFB) was developed to integrate semantic features from neighboring branches and a novel multi-branch fusion block (MBFB) was proposed to obtain the final semantic features by integrating all branches. Experimental results on two benchmark datasets show that the proposed method utilizes image features more effectively, achieves state-of-the-art FER performance, and generalizes well across different demographic groups.

Accurate segmentation of lesions is crucial for disease diagnosis and treatment planning. However, blurring and low contrast in the imaging process can affect segmentation results. We have observed that noninvasive medical imaging shares considerable similarities with natural images under low light conditions and that nocturnal animals possess extremely strong night vision capabilities. Inspired by the dark vision of these nocturnal animals, we proposed a novel plug-and-play dark vision network (DVNet) to enhance the model's perception for low-contrast medical images. Specifically, by employing the wavelet transform, we decompose medical images into subbands of varying frequencies, mimicking the sensitivity of photoreceptor cells to different light intensities. To simulate the antagonistic receptive fields of horizontal cells and bipolar cells, we design a Mamba-Enhanced Fusion Module to achieve global information correlation and enhance contrast between lesions and surrounding healthy tissues. Extensive experiments demonstrate that the DVNet achieves SOTA performance in various medical image segmentation tasks.

MOTIVATION: Metabolic stability is crucial in the early stage of drug discovery and development. Drug candidate screening and optimization can be streamlined through the accurate prediction of stability. Functional groups within drug molecules are known as pharmacophores, which bind directly to receptors or biological macromolecules to produce biological effects, thereby affecting metabolic stability. Therefore, determining metabolic stability via the pharmacophore groups remains a significant challenge. RESULTS: To address these issues, we propose a Pharmacophore-guided Hypergraph representation framework for predicting metabolic Stability (HyperPhS). In this study, we introduce a hypergraph-based method to extract features from metabolic pharmacophores with multi-view representation and contrastive learning. In particular, we introduce a pharmacophore-based contrastive learning encoder that captures the consistency between functional and nonfunctional structures. Our method applies ChatGPT simultaneously to metabolites and heterogeneous encoders and integrates multimodal representations by using attention-driven fusion modules coupled with fully connected neural networks. On the HLM dataset, HyperPhS achieves outstanding performance with 87.6% in AUC and 62.6% in MCC, alongside an external test AUC of 88.3%. In addition, pharmacophore groups studied by HyperPhS are validated for their interpretability through case studies. Overall, HyperPhS is an effective and interpretable tool for determining metabolic stability, identifying critical functional groups, and optimizing compounds. AVAILABILITY AND IMPLEMENTATION: The code and data are available at https://github.com/xiaoyiliu-usc/HyperPhS.

MOTIVATION: Retrosynthesis identifies available precursor molecules for various and novel compounds. With the advancements and practicality of language models, Transformer-based models have increasingly been used to automate this process. However, many existing methods struggle to efficiently capture reaction transformation information, limiting the accuracy and applicability of their predictions. RESULTS: We introduce RetroCaptioner, an advanced end-to-end, Transformer-based framework featuring a Contrastive Reaction Center Captioner. This captioner guides the training of dual-view attention models using a contrastive learning approach. It leverages learned molecular graph representations to capture chemically plausible constraints within a single-step learning process. We integrate the single-encoder, dual-encoder, and encoder-decoder paradigms to effectively fuse information from the sequence and graph representations of molecules. This involves modifying the Transformer encoder into a uni-view sequence encoder and a dual-view module. Furthermore, we enhance the captioning of atomic correspondence between SMILES and graphs. Our proposed method, RetroCaptioner, achieved outstanding performance with 67.2% in top-1 and 93.4% in top-10 exact matched accuracy on the USPTO-50k dataset, alongside an exceptional SMILES validity score of 99.4%. In addition, RetroCaptioner has demonstrated its reliability in generating synthetic routes for the drug protokylol. AVAILABILITY AND IMPLEMENTATION: The code and data are available at https://github.com/guofei-tju/RetroCaptioner.

De novo drug design is crucial in advancing drug discovery, which aims to generate new drugs with specific pharmacological properties. Recently, deep generative models have achieved inspiring progress in generating drug-like compounds. However, the models prioritize a single target drug generation for pharmacological intervention, neglecting the complicated inherent mechanisms of diseases, and influenced by multiple factors. Consequently, developing novel multi-target drugs that simultaneously target specific targets can enhance anti-tumor efficacy and address issues related to resistance mechanisms. To address this issue and inspired by Generative Pre-trained Transformers (GPT) models, we propose an upgraded GPT model with generative adversarial imitation learning for multi-target molecular generation called MTMol-GPT. The multi-target molecular generator employs a dual discriminator model using the Inverse Reinforcement Learning (IRL) method for a concurrently multi-target molecular generation. Extensive results show that MTMol-GPT generates various valid, novel, and effective multi-target molecules for various complex diseases, demonstrating robustness and generalization capability. In addition, molecular docking and pharmacophore mapping experiments demonstrate the drug-likeness properties and effectiveness of generated molecules potentially improve neuropsychiatric interventions. Furthermore, our model's generalizability is exemplified by a case study focusing on the multi-targeted drug design for breast cancer. As a broadly applicable solution for multiple targets, MTMol-GPT provides new insight into future directions to enhance potential complex disease therapeutics by generating high-quality multi-target molecules in drug discovery.

Deep learning has made significant advancements in supervised learning. However, models trained in this setting often face challenges due to domain shift between training and test sets, resulting in a significant drop in performance during testing. To address this issue, several domain generalization methods have been developed to learn robust and domain-invariant features from multiple training domains that can generalize well to unseen test domains. Data augmentation plays a crucial role in achieving this goal by enhancing the diversity of the training data. In this paper, inspired by the observation that normalizing an image with different statistics generated by different batches with various domains can perturb its feature, we propose a simple yet effective method called NormAUG (Normalization-guided Augmentation). Our method includes two paths: the main path and the auxiliary (augmented) path. During training, the auxiliary path includes multiple sub-paths, each corresponding to batch normalization for a single domain or a random combination of multiple domains. This introduces diverse information at the feature level and improves the generalization of the main path. Moreover, our NormAUG method effectively reduces the existing upper boundary for generalization based on theoretical perspectives. During the test stage, we leverage an ensemble strategy to combine the predictions from the auxiliary path of our model, further boosting performance. Extensive experiments are conducted on multiple benchmark datasets to validate the effectiveness of our proposed method.

Domain generalization (DG) aims at learning a model on source domains to well generalize on the unseen target domain. Although it has achieved great success, most of the existing methods require the label information for all training samples in source domains, which is time-consuming and expensive in the real-world application. In this article, we resort to solving the semi-supervised domain generalization (SSDG) task, where there are a few label information in each source domain. To address the task, we first analyze the theory of multi-domain learning, which highlights that (1) mitigating the impact of domain gap and (2) exploiting all samples to train the model can effectively reduce the generalization error in each source domain so as to improve the quality of pseudo-labels. According to the analysis, we propose MultiMatch, i.e., extending FixMatch to the multi-task learning framework, producing the high-quality pseudo-label for SSDG. To be specific, we consider each training domain as a single task (i.e., local task) and combine all training domains together (i.e., global task) to train an extra task for the unseen test domain. In the multi-task framework, we utilize the independent batch normalization and classifier for each task, which can effectively alleviate the interference from different domains during pseudo-labeling. Also, most of the parameters in the framework are shared, which can be trained by all training samples sufficiently. Moreover, to further boost the pseudo-label accuracy and the model’s generalization, we fuse the predictions from the global task and local task during training and testing, respectively. A series of experiments validate the effectiveness of the proposed method, and it outperforms the existing semi-supervised methods and the SSDG method on several benchmark DG datasets.

Development of robust and effective strategies for synthesizing new compounds, drug targeting and constructing GEnome-scale Metabolic models (GEMs) requires a deep understanding of the underlying biological processes. A critical step in achieving this goal is accurately identifying the categories of pathways in which a compound participated. However, current machine learning-based methods often overlook the multifaceted nature of compounds, resulting in inaccurate pathway predictions. Therefore, we present a novel framework on Multi-View Multi-Label Learning for Metabolic Pathway Inference, hereby named MVML-MPI. First, MVML-MPI learns the distinct compound representations in parallel with corresponding compound encoders to fully extract features. Subsequently, we propose an attention-based mechanism that offers a fusion module to complement these multi-view representations. As a result, MVML-MPI accurately represents and effectively captures the complex relationship between compounds and metabolic pathways and distinguishes itself from current machine learning-based methods. In experiments conducted on the Kyoto Encyclopedia of Genes and Genomes pathways dataset, MVML-MPI outperformed state-of-the-art methods, demonstrating the superiority of MVML-MPI and its potential to utilize the field of metabolic pathway design, which can aid in optimizing drug-like compounds and facilitating the development of GEMs. The code and data underlying this article are freely available at https://github.com/guofei-tju/MVML-MPI. Contact: [email protected], [email protected] or [email protected].

Detecting potential associations between drugs and diseases plays an indispensable role in drug development, which has also become a research hotspot in recent years. Compared with traditional methods, some computational approaches have the advantages of fast speed and low cost, which greatly accelerate the progress of predicting the drug-disease association. In this study, we propose a novel similarity-based method of low-rank matrix decomposition based on multi-graph regularization. On the basis of low-rank matrix factorization with$L_{2}$regularization, the multi-graph regularization constraint is constructed by combining a variety of similarity matrices from drugs and diseases respectively. In the experiments, we analyze the difference in the combination of different similarities, resulting that combining all the similarity information on drug space is unnecessary, and only a part of the similarity information can achieve the desired performance. Then our method is compared with other existing models on three data sets (Fdataset, Cdataset and LRSSLdataset) and have a good advantage in the evaluation measurement of AUPR. Besides, a case study experiment is conducted and showing that the superior ability for predicting the potential disease-related drugs of our model. Finally, we compare our model with some methods on six real world datasets, and our model has a good performance in detecting real world data.

MOTIVATION: Metabolomics has developed rapidly in recent years, and metabolism-related databases are also gradually constructed. Nowadays, more and more studies are being carried out on diverse microbes, metabolites and diseases. However, the logics of various associations among microbes, metabolites and diseases are limited understanding in the biomedicine of gut microbial system. The collection and analysis of relevant microbial bioinformation play an important role in the revelation of microbe-metabolite-disease associations. Therefore, the dataset that integrates multiple relationships and the method based on complex heterogeneous graphs need to be developed. RESULTS: In this study, we integrated some databases and extracted a variety of associations data among microbes, metabolites and diseases. After obtaining the three interconnected bilateral association data (microbe-metabolite, metabolite-disease and disease-microbe), we considered building a heterogeneous graph to describe the association data. In our model, microbes were used as a bridge between diseases and metabolites. In order to fuse the information of disease-microbe-metabolite graph, we used the bipartite graph attention network on the disease-microbe and metabolite-microbe bipartite graph. The experimental results show that our model has good performance in the prediction of various disease-metabolite associations. Through the case study of type 2 diabetes mellitus, Parkinson's disease, inflammatory bowel disease and liver cirrhosis, it is noted that our proposed methodology are valuable for the mining of other associations and the prediction of biomarkers for different human diseases.Availability and implementation: https://github.com/Selenefreeze/DiMiMe.git.

RNA binding protein (RBP) is extensively involved in various cellular regulatory processes through the interaction with RNAs. Capturing the RBP binding preferences is fundamental for revealing the pathogenesis of complex diseases. Many experimental detection techniques are still time-consuming and labor-intensive, therefore, it is indispensable to develop a computational method with convincing accuracy. In this study, we proposed a CNN-BLSTM hybrid deep learning framework, named DeepDW, for predicting the RBP binding sites on RNAs with high-order encoding features of RNA sequence and secondary structure. The high-order encoding strategy was used to characterize the dependencies among adjacency nucleotides. For CNN-BLSTM hybrid model, DeepDW first employed two 1-D convolutional neural networks (CNNs) for learning the local features from high-order encoded matrices of RNA sequence and structure separately, and then applied two bidirectional long short-term memory networks (BLSTMs) to capture the global information in a higher level. Moreover, a series of experiments were carried out on 31 public datasets to evaluate our proposed framework, and DeepDW achieved superior performance than the state-of-the-art methods. The results indicated that the combination of high-order encoding method and CNN-BLSTM hybrid model had advantages in identifying RBP-RNA binding sites.