In the treatment of complex diseases, single drug therapy is limited by resistance and tolerance, making the exploration of efficient drug combinations crucial in cancer research. When evaluating drug efficacy, traditional deep learning methods analyze single-drug sensitivity and drug combination synergy in isolation, which cannot capture the complex internal relationship between them, resulting in prediction deviation. To overcome these limitations, this article proposes a novel drug combination synergy prediction model called MTDSN (Multi-Task Deep Synergy Network), which integrates multi-task learning and deep neural networks to simultaneously predict single-drug sensitivity and drug combination synergy. During the model construction, the Autoencoder integrated with Convolutional Block Attention Module (CBAM) is used to reduce the dimension of input features, and then the drug features and cell line features are connected and input into the shared module embedded with cross-stitch mechanism to exchange information. Finally, the features of each task are input into different task-specific branches to obtain the synergy score of the drug combination, the sensitivity score of the single drug and their corresponding classification results. Evaluated on the O'Neil dataset, MTDSN achieves the lowest mean squared error (MSE) and highest Pearson correlation coefficient (PCC) in drug synergy prediction, with an ROC-AUC of 0.92 and accuracy of 0.95 in the classification task, demonstrating substantial improvements in predictive efficacy.
We proposed a practical non-recursive parallel list ranking algorithm, NR-Ranking. NR-Ranking adopts an independent set to avoid potential operation contention between neighbor nodes. In each communication round, every node in an independent set bridges its left neighbor and right neighbor by adding edges with new distance; then all nodes in this independent set are excluded from previous linked lists. The probability of one node being selected into the independent set is about 1/3. According to the stop criterion of selecting an independent set, the number of communication rounds of NR-Ranking is different. It is bounded by log(p) if the selection step stops when the number of nodes in the reminder lists is less than n/p, where n is the number of nodes in the linked lists and p is the number of processors, or O(logw) if all nodes in the remaining lists are end nodes, where w is the length of the longest linked list. The complexity of computation and communication on both stop criteria is bounded by O(n). Experimental results confirm the above complexity analysis, and the implementation of NR-Ranking in GPS has achieved a speed increase of 4X when the number of workers increases from 8 to 48.
The rapid development of joint profiling methods makes it possible to simultaneously measure multi-modal data from the same cell. It enables the more comprehensive insight of cellular heterogeneity. However, due to the heterogeneity of features across different modalities and the complexity of each modal data, effectively integrating multi-modal data to achieve more accurate cell heterogeneity analysis remains challenging. Here, we propose a novel contrastive learning-based approach for single-cell multi-omics data clustering, named scCLC. Taking contrastive learning as backbone, scCLC integrates scRNA-seq data and scATAC-seq data for cell presentation learning with dedicated data augmentation strategy and self-supervised learning model. scCLC leverages the topological structure derived from single-cell multi-omics data to determine positive pairs, which not only increases the training sample size but also provides basic label priors for the model to learn more cluster-friendly cell representations. The experimental results on single-cell multi-omics datasets show the superior clustering performance of scCLC. When applied to visualization, scCLC can effectively separate the cell subtypes.
Depression, a widespread mental health condition, presents significant challenges due to stigma, concealed symptoms, and high treatment costs, underscoring the need for effective Automatic Depression Detection (ADD) systems. However, existing methods often face difficulties in modeling long-term dependencies and directional information. To address these limitations, we propose DepMambaformer, a novel multimodal depression detection framework. Our model integrates CNNs for local feature extraction, the designed DepBiMamba2 module to capture global dependencies and directional information, and multi-head attention for multimodal fusion of audio and video features. The combination of state space duality module with attention mechanisms demonstrates superior performance. Extensive experiments validate the effectiveness and scalability of DepMambaformer, outperforming state-of-the-art methods. Furthermore, ablation studies and visualizations provide deeper insights into the model's functionality. This work highlights the potential of advanced global feature modeling in enhancing ADD systems for early and efficient depression screening.
Neoantigens, defined as tumor-specific peptides triggering immune responses, have emerged as a promising therapeutic target. Recently, proteogenomic analysis employing liquid chromatography-tandem mass spectrometry (LC-MS/MS) has become a standard approach for neoantigen identification. However, the accuracy of this method is limited by inconsistent false discovery rate (FDR) estimation and inherent biases in the target-decoy method. Retention time (RT), a key physicochemical property of peptides, provides an effective metric for evaluating identified neoantigens' quality. Therefore, we propose PepHiFuse, a deep learning framework incorporating global biochemical semantics and local physical interactions to predict peptide RT. Moreover, it employs an adaptive weighting mechanism to balance these complementary feature domains dynamically. Evaluated on three large public datasets, PepHiFuse outperforms three recently published methods (AutoRT, AlphaPeptDeep, PepMNet). The significant improvement in RT prediction accuracy demonstrates the efficacy of PepHiFuse as a prediction tool. PepHiFuse also enhances neoantigen identification by selectively filtering low-confidence candidates while retaining high-quality targets across multiple FDR estimation methods when applied to clinical proteogenomic datasets. This ability emphasizes the practical utility of the PepHiFuse method, offering a critical step toward reliable immunotherapeutic discovery by incorporating accurate RT prediction into neoantigen validation. The code is available at https://github.com/lyotvincent/PepHiFuse.
Drug repositioning aims to discover new therapeutic pathways for approved drugs, thereby reducing drug development costs. However, in computational drug repositioning, validated drug-disease associations are sparse, and negative samples can only be randomly selected from unvalidated associations that inherently conflate truly non-interacting pairs with undetected therapeutic relationships. This fundamental challenge particularly affects deep learning models, which require high-quality training samples to achieve optimal performance. To transcend this basic limitation, we propose a novel sample selection strategy tailored explicitly for deep learning models to filter reliable negative samples and potential positive samples. First, grounded in the low-rank structure of the drug-disease association matrix, we employ matrix completion algorithms to infer unvalidated association information. Subsequently, we develop an unsupervised clustering algorithm that comprehensively considers the completed probability scores and ranking positions of candidate diseases for each drug, categorizing all samples into distinct categories based on their confidence levels. Finally, we select reliable negative samples and potential positive samples from the clustering results to train deep neural networks for drug repositioning. Substantial experimental results verify that our proposed sample selection strategy enhances the performance of deep learning-based computational drug repositioning methods.
In this paper, we initiate the study on some problems related to multiple protein scaffold filling, with or without references. The objective is to maximize the sum of the Blosum62 scores of the filled sequences when no reference is given, or to maximize the Blosum62 score between the filled sequence and a reference. We present the following results: (1) given n scaffolds generated from the top-down tandem mass spectra, finding k scaffolds whose corresponding contents can be used to fill into a target sequence (or, a sequence whose Blosum62 score with a reference is maximized) takes (n^k-ε) time, unless the SETH (Strong Exponential Time Hypothesis) fails. (2) given two or more protein scaffolds and the corresponding multisets of amino acids to be filled accordingly, the corresponding optimization problem can be solved in polynomial time with dynamic programming. (3) Due to the high (and impractical) running times of the algorithms in (2), we implement several heuristic algorithms for the special cases when three scaffolds are given, and the corresponding empirical results are quite promising—although we also find that for the greedy algorithm more biological information needs to be incorporated to generate biologically meaningful results.
Drug synergy is crucial for developing effective combination therapies, but traditional screening methods suffer from inefficiency and high costs. While deep learning shows promise for predicting drug synergy, current approaches using Transformers and graph neural networks focus on combining drug and cell line features without modelling how genes causally influence drug responses. To address this limitation, we propose CADS (Causal Adjustment for Drug Synergy), a deep learning framework that integrates causal relationships between genes and drug responses. Leveraging multi-omics data, CADS uses a learnable mask mechanism to identify key causal genes while filtering out irrelevant genetic factors through backdoor adjustment-a causal inference technique. Our model achieves two key objectives simultaneously: accurate prediction of drug synergy and interpretable causal gene discovery. Experiments on multiple datasets show that CADS consistently outperforms state-of-the-art methods across multiple metrics. Case studies demonstrate that CADS can reduce unnecessary complexity while providing more biological insights through its gene importance scores, which help identify clinically validated cancer-related genes that mediate drug interactions. Taken together, CADS advances combination therapy prediction by explicitly modelling drug synergy causal genes, offering enhanced interpretability for AI-based drug development.
Spatial transcriptomics has transformed tissue analysis by preserving spatial context in gene expression data, enabling deeper insights into tissue microenvironments. However, current spatial domain identification methods largely focus on adjacent cellular similarities, limiting their ability to capture long-range spatial dependencies and identify identical cell types distributed across distant and non-contiguous areas. To address these challenges, we introduce MambaST, a hybrid deep-learning framework that integrates selective state space modeling (Mamba) and self-supervised learning for Spatial Transcriptomics data analysis. Specifically, MambaST incorporates a Six-Directional Selective Scan (SS6D) algorithm to convert graph-structured spatial data into topology-preserving pseudo-sequences, effectively bridging sequential modeling with spatial topology. Additionally, we propose HexMambaBlock (HMB), which applies Mamba to simultaneously denoise gene expression data and capture global spatial dependencies. Furthermore, contrastive learning enhanced with a Dynamic Context-aware Readout (DCR) module improves the biological specificity of local representations. Comprehensive evaluations across three spatial transcriptomic datasets demonstrate MambaST's superior performance in spatial domain identification, achieving a 0.58 average Adjusted Rand Index (ARI) on the DLPFC dataset, which surpasses state-of-the-art methods by 2.7%.
Microsatellite instability (MSI) results from deficient mismatch repair (dMMR) and plays a crucial role in tumorigenesis and treatment, particularly in colon cancer. The gold standard for MSI classification relies on multiplex fluorescent PCR with capillary electrophoresis (CE); however, manual interpretation is time-consuming and subjective due to panel variability. No dedicated tools currently exist for MSI classification from CE profiles. We developed Automated Classification of Microsatellite Instability (ACMSI) to streamline fragment analysis and automate MSI classification. ACMSI includes size calling, manual calibration, and automated classification (Fig. 1). Evaluated on 322 electrophoresis profiles (129 tumor-normal pairs, 774 markers) using the 2B3D NCI Panel (BAT25, BAT26, D2S123, D5S346, D17S250), ACMSI achieved a 99.07 https://github.com/OpenGene/ACMSI .
The PM1 criterion, which states that a variant is located in a mutational hot spot and/or critical and well-established functional domain without benign variation, is considered moderate evidence for assessing its pathogenicity. The application of PM1 criterion is limited due to the lack of a comprehensive and reliable database of variant hotspot regions. Compared to hotspots, coldspots have been neglected by the ACMG guidelines. In order to improve variant classification, we suggest including coldspots in the ACMG guidelines that support the classification of benignity. Consequently, we have developed the HCSeer tool to provide data support for PM1 and the ‘Benign’ criteria. HCSeer employs the Kernel Density Estimation (KDE) algorithm and the Expectation-Maximization (EM) algorithm to identify potential hotspot and coldspot regions. Through our HCSeer, we successfully identified 423 hotspots and 3,942 coldspots regions in 1,523 genes. We then provided a database for general geneticists and clinicians to easily query whether a variant is located in a hotspot or coldspot region ( http://www.genemed.tech/hcseer/ ), so as to determine if it can apply the PM1 of ACMG or the Benign criteria. Availability: Code and Supplementary data are available at https://github.com/xq-xia/HCSeer .
Although the medical image segmentation model SCTransNet based on convolutional neural network and Transformer performs well, SCTransNet has the problems of large number of parameters and computation. When SCTransNet processes the 3D shadow data, the 3D convolution operation in the CNN encoder and the decoder, and the Transformer model needs a large number of parameters to learn when operating on the 3D feature image, leading to the increase in the number of model parameters and the rising computational cost. Considering the problem of large number of parameters and calculation, this paper proposed the lightweight model LiteSCTransNet to effectively reduce the parameters and computation of the model while ensuring the segmentation accuracy. The encoder module in the Transformer model is refined using the channel separation idea, replacing the conventional convolution operation using deeply separable convolution. Experiments show that the calculated lightweight method reduced the number of parameters and calculations by 74.43
Covariance or its inverse (called precision) matrix estimation is very useful in data analyses. In principle, the empirical covariance matrix calculated with the unlimited samples is an unbiased estimation of the covariance matrix of the distribution from which samples are drawn, but there are only a very limited samples available in many real-life applications. Therefore, ℓ _1 -regularized estimation methods are developed in recovering a sparse precision matrix. However, the ℓ _1 -regularization has its own drawbacks. To address these drawbacks and the issue of limited samples with multi-classes in biomedical applications, we propose a joint sparse precision matrix estimation method, in which the SCAD-regularization is used for sparsity and the Frobenius norm of the between-class precision matrix difference is adopted to reinforce the similarity among classes. An alternating direction method of multipliers and an iterative weighted penalized method are developed to optimize the objective function. Fisher’s linear discriminant analysis with the estimated precision matrices is applied to two gene expression datasets for lung cancer diagnosis. The diagnosis results indicate the excellent performance of our proposed method for estimating precision matrices.
The miniaturization of EEG devices is essential for the development of consumer-grade brain-computer interface technology. However, low-channel EEG signals exacerbate the inherent disadvantages of low signal-to-noise ratio and low spatial resolution, making the decoding of neural activity even more challenging. To overcome these limitations, we propose an advanced multi-model fusion network that combines temporal-spatial and spectral-spatial features, referred to as the temporal-spectral cross-fusion network (TSCF-Net). This novel architecture consists of two parallel models, i.e., the spectral-spatial model and the temporal-spatial model. In the spectral-spatial model, the one-dimensional EEG time series is first transformed into a two-dimensional time-frequency representation to reveal its intrinsic time-varying characteristics. The time-frequency representation is then extended into the depth dimension to capture the spatial characteristics. On the other hand, the temporal-spatial model directly applies the one-dimensional EEG time series as input, and similarly extends it into the depth dimension to extract spatiotemporal features. To constrain the distribution of these features, a maximum mean discrepancy loss is introduced for feature fusion during the training. Finally, a weighted fusion method is employed to integrate these features. Experimental results based on the BCI Competition IV 2a and IV 2b datasets demonstrate that the TSCF-Net outperforms other baseline methods in low-channel EEG decoding tasks, achieving the highest average accuracy and kappa across all datasets. Additionally, a series of ablation experiments further confirm the effectiveness of the multimodal fusion structure.
Accurate identification of DNA/RNA modification sites is imperative for the study of their biological functions. Machine learning algorithms are unable to utilize sequence data directly to construct models for predicting these sites. Consequently, the development of DNA/RNA sequence feature representation algorithms is paramount for the effective encoding of sequence data into a usable numerical format, thereby facilitating the construction of high-performance machine learning predictive models. Existing DNA/RNA sequence feature representation algorithms suffer from the problems of simple extracting information, failure to take into account the sequence position and order information, and irrelevant or redundant features brought about by multi-methods together. This leads to the inability of machine learning models to break through the bottleneck of prediction performance. To address the aforementioned issues, this paper introduces the BiPSDP (Bidirectional Position-Specific Dinucleotide Propensities) algorithm. This algorithm extracts dinucleotide position-specific propensities from both forward and backward sequence directions. The incorporation of a parameter for dinucleotide spacing is pivotal in capturing global order information, with DNA/RNA sequences being encoded as numerical features that are rich in class-distinguishing information. The validity of the BiPSDP algorithm was tested by constructing a DNA/RNA modification site prediction model using the SVM learning machines, and a comparison was made with seven existing representation algorithms across seven modification types. The experimental results demonstrate that the prediction model with BiPSDP consistently outperforms the comparison models, thus validating its value as a tool for constructing prediction models for various DNA/RNA modification sites. The code of BiPSDP is available at https://github.com/Mingzhao2017/BiPSDP.
Single-cell RNA sequencing and spatial transcriptomics(ST) have provided new perspectives for exploring the mechanisms of cancer development, diagnosis, and treatment. However, due to the highly sparse nature of single-cell and spatial transcriptome data, it remains a challenge to extract the cell type composition information of each point in the spatial transcriptome data from single-cell data. To address this issue, this study proposes an integrated analysis method called LDADW, which is based on the Latent Dirichlet Allocation (LDA) topic model and damped weighted least squares. By using LDA to mine the cell function topics in single-cell data, a topic-gene feature matrix is constructed. Combined with the damped weighted least squares deconvolution technique, the cell composition of spatial sites is analyzed. This method significantly outperforms the current RCTD, Seurat, Spotlight, and Tangram models. Verified on the simulated dataset of the mouse pancreas, the recall rate of LDADW reaches 94
Automatic detection of wound healing progress is essential for assisting clinicians in evaluating wound conditions, guiding clinical treatment, and preventing infections and complications. However, current methods for detecting wound healing progress are often susceptible to interference from irrelevant information, leading to low detection accuracy, and the existing techniques generally lack interpretability. In this study, we construct a new wound healing dataset, the first to provide continuous temporal observations of skin wound healing status. Based on this dataset, we employ a task-related knowledge learning framework for extracting wound-related features of the skin wound images. Utilizing these features, we design several simple and highly interpretable machine learning models for classification study of the wound healing progress. In comparison to existing deep learning models, which lack interpretability, our simple models achieve superior accuracies in predicting wound healing progress. Furthermore, we analyze the learned wound-related features and find that their spatial distribution aligns with established medical principles, further confirming the features we learn are both interpretable and reliable.
Identifying prokaryotic and eukaryotic viruses is key to understanding host-environment interactions and ecological differences. Viral genomes, marked by rapid mutation and diversity, challenge accurate identification. We propose ViDSG, which hypothesizes that local nucleotide preferences and non-linear dependencies in viral sequences determine virus types. ViDSG quantifies nucleotide pair frequency and weighted distance to assess local combination preferences, while a masked language model captures global semantic features from non-linear dependencies. A dual-channel deep learning model integrates these statistical and semantic features for classification. We tested ViDSG using viral genome datasets from monkeys and pigs, comparing it with methods like IPEV and HTP for accuracy, robustness, and efficiency. ViDSG outperforms, especially in the 1200–1800 bp range, improving accuracy by 0.3
Parkinson's Disease (PD) is a typical neurodegenerative disease, and early diagnosis is crucial to delaying the progression of the disease. However, existing PD diagnosis methods using sMRI and fMRI fail to design modality-specific feature extraction networks tailored to the unique characteristics of each modality. In addition, most of these methods do not deeply explore the feature correlation of different modalities in the feature fusion module. To tackle these challenges, we propose a Spatial-temporal Dual-pathway Network with Multi-scale Feature Fusion (SDMFF) framework to enhance the performance of PD diagnosis. Specifically, we develop a spatial-temporal dual-pathway network capable of effectively extracting representations from sMRI and fMRI. For sMRI, we design a spatial CNN-transformer module to extract both local and global structural features. For fMRI, we design a spatial-temporal Transformer to capture dynamic spatial-temporal features. To effectively fuse the information from both sMRI and fMRI, we design a multi-scale convolutional attention feature fusion module, which fully integrates multi-scale feature information of sMRI and fMRI. Extensive experimental results demonstrate that our proposed SDMFF achieves state-of-the-art performance on both public and private datasets, with accuracy of 0.926 and 0.858, respectively.
Viral mutations and waning immunity play a key role in the spread of infectious diseases. Traditional compartmental models often assume that individuals acquire complete but temporary immunity after their infection, overlooking the complex dynamics of immune escape by emerging variants. To address this oversight, we built an agent-based model using empirical data from COVID-19 to investigate how viral evolution, mutation, and immune escape reinfections shape transmission dynamics. For the pathogen characteristics examined here, simulations with and without immune escape reinfections revealed that immune escape reinfections cause approximately 30
