A deep learning model for DNA enhancer prediction based on nucleotide position aware feature encoding.

Enhancers, genomic DNA elements, regulate neighboring gene expression crucial for biological processes like cell differentiation and stress response. However, current machine learning methods for predicting DNA enhancers often underutilize hidden features in gene sequences, limiting model accuracy. Hence, this article proposes the PDCNN model, a deep learning-based enhancer prediction method. PDCNN extracts statistical nucleotide representations from gene sequences, discerning positional distribution information of nucleotides in modifier-like DNA sequences. With a convolutional neural network structure, PDCNN employs dual convolutional and fully connected layers. The cross-entropy loss function iteratively updates using a gradient descent algorithm, enhancing prediction accuracy. Model parameters are fine-tuned to select optimal combinations for training, achieving over 95% accuracy. Comparative analysis with traditional methods and existing models demonstrates PDCNN's robust feature extraction capability. It outperforms advanced machine learning methods in identifying DNA enhancers, presenting an effective method with broad implications for genomics, biology, and medical research.

SpatialPPI: Three-dimensional space protein-protein interaction prediction with AlphaFold Multimer.

Rapid advancements in protein sequencing technology have resulted in gaps between proteins with identified sequences and those with mapped structures. Although sequence-based predictions offer insights, they can be incomplete due to the absence of structural details. Conversely, structure-based methods face challenges with respect to newly sequenced proteins. The AlphaFold Multimer has remarkable accuracy in predicting the structure of protein complexes. However, it cannot distinguish whether the input protein sequences can interact. Nonetheless, by analyzing the information in the models predicted by the AlphaFold Multimer, we propose a highly accurate method for predicting protein interactions. This study focuses on the use of deep neural networks, specifically to analyze protein complex structures predicted by the AlphaFold Multimer. By transforming atomic coordinates and utilizing sophisticated image-processing techniques, vital 3D structural details were extracted from protein complexes. Recognizing the significance of evaluating residue distances in protein interactions, this study leveraged image recognition approaches by integrating Densely Connected Convolutional Networks (DenseNet) and Deep Residual Network (ResNet) within 3D convolutional networks for protein 3D structure analysis. When benchmarked against leading protein-protein interaction prediction methods, such as SpeedPPI, D-script, DeepTrio, and PEPPI, our proposed method, named SpatialPPI, exhibited notable efficacy, emphasizing the promising role of 3D spatial processing in advancing the realm of structural biology. The SpatialPPI code is available at: https://github.com/ohuelab/SpatialPPI.

Essential genes identification model based on sequence feature map and graph convolutional neural network.

BACKGROUND: Essential genes encode functions that play a vital role in the life activities of organisms, encompassing growth, development, immune system functioning, and cell structure maintenance. Conventional experimental techniques for identifying essential genes are resource-intensive and time-consuming, and the accuracy of current machine learning models needs further enhancement. Therefore, it is crucial to develop a robust computational model to accurately predict essential genes. RESULTS: In this study, we introduce GCNN-SFM, a computational model for identifying essential genes in organisms, based on graph convolutional neural networks (GCNN). GCNN-SFM integrates a graph convolutional layer, a convolutional layer, and a fully connected layer to model and extract features from gene sequences of essential genes. Initially, the gene sequence is transformed into a feature map using coding techniques. Subsequently, a multi-layer GCN is employed to perform graph convolution operations, effectively capturing both local and global features of the gene sequence. Further feature extraction is performed, followed by integrating convolution and fully-connected layers to generate prediction results for essential genes. The gradient descent algorithm is utilized to iteratively update the cross-entropy loss function, thereby enhancing the accuracy of the prediction results. Meanwhile, model parameters are tuned to determine the optimal parameter combination that yields the best prediction performance during training. CONCLUSIONS: Experimental evaluation demonstrates that GCNN-SFM surpasses various advanced essential gene prediction models and achieves an average accuracy of 94.53%. This study presents a novel and effective approach for identifying essential genes, which has significant implications for biology and genomics research.

Monitoring of Single-Track Melting States Based on Photodiode Signal during Laser Powder Bed Fusion.

Single track is the basis for the melt pool modeling and physics work in laser powder bed fusion (LPBF). The melting state of a single track is closely related to defects such as porosity, lack of fusion, and balling, which have a significant impact on the mechanical properties of an LPBF-created part. To ensure the reliability of part quality and repeatability, process monitoring and feedback control are emerging to improve the melting states, which is becoming a hot topic in both the industrial and academic communities. In this research, a simple and low-cost off-axial photodiode signal monitoring system was established to monitor the melting pools of single tracks. Nine groups of single-track experiments with different process parameter combinations were carried out four times and then thirty-six LPBF tracks were obtained. The melting states were classified into three classes according to the morphologies of the tracks. A convolutional neural network (CNN) model was developed to extract the characteristics and identify the melting states. The raw one-dimensional photodiode signal data were converted into two-dimensional grayscale images. The average identification accuracy reached 95.81% and the computation time was 15 ms for each sample, which was promising for engineering applications. Compared with some classic deep learning models, the proposed CNN could distinguish the melting states with higher classification accuracy and efficiency. This work contributes to real-time multiple-sensor monitoring and feedback control.

Deep learning with explainability for characterizing age-related intrinsic differences in dynamic brain functional connectivity.

Although many deep learning models-based medical applications are performance-driven, i.e., accuracy-oriented, their explainability is more critical. This is especially the case with neuroimaging, where we are often interested in identifying biomarkers underlying brain development or disorders. Herein we propose an explainable deep learning approach by elucidating the information transmission mechanism between two layers of a deep network with a joint feature selection strategy that considers several shallow-layer explainable machine learning models and sparse learning of the deep network. At the end, we apply and validate the proposed approach to the analysis of dynamic brain functional connectivity (FC) from fMRI in a brain development study. Our approach can identify the differences within and between functional brain networks over age during development. The results indicate that the brain network transits from undifferentiated structures to more specialized and organized ones, and the information processing ability becomes more efficient as age increases. In addition, we detect two developmental patterns in the brain network: the FCs in regions related to visual and sound processing and mental regulation become weakened, while those between regions corresponding to emotional processing and cognitive activities are enhanced.

Prediction of DNA Methylation based on Multi-dimensional feature encoding and double convolutional fully connected convolutional neural network.

DNA methylation takes on critical significance to the regulation of gene expression by affecting the stability of DNA and changing the structure of chromosomes. DNA methylation modification sites should be identified, which lays a solid basis for gaining more insights into their biological functions. Existing machine learning-based methods of predicting DNA methylation have not fully exploited the hidden multidimensional information in DNA gene sequences, such that the prediction accuracy of models is significantly limited. Besides, most models have been built in terms of a single methylation type. To address the above-mentioned issues, a deep learning-based method was proposed in this study for DNA methylation site prediction, termed the MEDCNN model. The MEDCNN model is capable of extracting feature information from gene sequences in three dimensions (i.e., positional information, biological information, and chemical information). Moreover, the proposed method employs a convolutional neural network model with double convolutional layers and double fully connected layers while iteratively updating the gradient descent algorithm using the cross-entropy loss function to increase the prediction accuracy of the model. Besides, the MEDCNN model can predict different types of DNA methylation sites. As indicated by the experimental results,the deep learning method based on coding from multiple dimensions outperformed single coding methods, and the MEDCNN model was highly applicable and outperformed existing models in predicting DNA methylation between different species. As revealed by the above-described findings, the MEDCNN model can be effective in predicting DNA methylation sites.

Evaluation of a Series of Turnip Mosaic Virus Chimeric Clones Reveals Two Amino Acid Sites Critical for Systemic Infection in Chinese Cabbage.

Two infectious clones of turnip mosaic virus (TuMV), pKBC-1 and pKBC-8, with differential infectivity in Chinese cabbage (Brassica rapa subsp. pekinensis), were obtained. Both infected Nicotiana benthamiana systemically, inducing similar symptoms, whereas only virus KBC-8 infected Chinese cabbage systemically. To identify the determinants affecting infectivity on Chinese cabbage, chimeric clones were constructed by restriction fragment exchange between the parental clones and tested on several Chinese cabbage cultivars. Chimeric clones p1N8C and p8N1C demonstrated that the C-terminal portion of the polyprotein determines systemic infection of Chinese cabbage despite only three amino acid differences in this region, in the cylindrical inclusion (CI), viral protein genome-linked (VPg), and coat protein (CP). A second pair of hybrid constructs, pHindIII-1N8C and pHindIII-8N1C, failed to infect cultivars CR Victory and Jinseonnorang systemically, yet pHindIII-1N8C caused hypersensitive response-like lesions on inoculated leaves of these cultivars, and could systemically infect cultivars CR Chusarang and Jeongsang; this suggests that R genes effective against TuMV may exist in the first two cultivars but not the latter two. Constructs with single amino acid changes in both VPg (K2045E) and CP (Y3095H) failed to infect Chinese cabbage, implying that at least one of these two amino acid substitutions is essential for successful infection on Chinese cabbage. Successful infection by mutant KBC-8-CP-H and delayed infection with mutant HJY1-VPg-E following mutation or reversion suggested that VPg (2045K) is the residue required for infection of Chinese cabbage and involved in the interaction between VPg and eukaryotic initiation factor eIF(iso)4E, confirmed by yeast two-hybrid assay.

scRNASequest: an ecosystem of scRNA-seq analysis, visualization, and publishing.

BACKGROUND: Single-cell RNA sequencing is a state-of-the-art technology to understand gene expression in complex tissues. With the growing amount of data being generated, the standardization and automation of data analysis are critical to generating hypotheses and discovering biological insights. RESULTS: Here, we present scRNASequest, a semi-automated single-cell RNA-seq (scRNA-seq) data analysis workflow which allows (1) preprocessing from raw UMI count data, (2) harmonization by one or multiple methods, (3) reference-dataset-based cell type label transfer and embedding projection, (4) multi-sample, multi-condition single-cell level differential gene expression analysis, and (5) seamless integration with cellxgene VIP for visualization and with CellDepot for data hosting and sharing by generating compatible h5ad files. CONCLUSIONS: We developed scRNASequest, an end-to-end pipeline for single-cell RNA-seq data analysis, visualization, and publishing. The source code under MIT open-source license is provided at https://github.com/interactivereport/scRNASequest . We also prepared a bookdown tutorial for the installation and detailed usage of the pipeline: https://interactivereport.github.io/scRNAsequest/tutorial/docs/ . Users have the option to run it on a local computer with a Linux/Unix system including MacOS, or interact with SGE/Slurm schedulers on high-performance computing (HPC) clusters.

RNASequest: An End-to-End Reproducible RNAseq Data Analysis and Publishing Framework.

We present RNASequest, a customizable RNA sequencing (RNAseq) analysis, app management, and result publishing framework. Its three-in-one RNAseq data analysis ecosystem consists of (1) a reproducible, configurable expression analysis (EA) module, (2) multi-faceted result presentation in R Shiny, a Bookdown document and an online slide deck, and (3) a centralized data management system. In principle, following up our well-received omics data visualization tool Quickomics, RNASequest automates the differential gene expression analysis step, eases statistical model design by built-in covariates testing module, and further provides a web-based tool, ShinyOne, to manage apps powered by Quickomics and reports generated by running the pipeline on multiple projects in one place. Researchers can experience the functionalities by exploring demo data sets hosted at http://shinyone.bxgenomics.com or following the tutorial, https://interactivereport.github.io/RNASequest/tutorial/docs/introduction.html to set up the framework locally to process private RNAseq datasets. The source code released under MIT open-source license is provided at https://github.com/interactivereport/RNASequest.

Voltage-dependent anion channel 2 (VDAC2) facilitates the accumulation of rice stripe virus in the vector Laodelphax striatellus.

Rice stripe virus (RSV) causes enormous losses in rice production and is transmitted by the small brown planthopper, Laodelphax striatellus, in a persistent-propagative manner. RSV accumulation within the gut lumen of the vector is indispensable for the successful transmission to rice and insects. In this study, we obtained a 1464 bp full-length cDNA of a voltage-dependent anion channel 2 from L. striatellus (LsVDAC2), which encodes a 283 amino acid protein. RSV infection increased the expression of LsVDAC2 in the midguts and ovaries of L. striatellus by 260% and 228%, respectively. Silencing of LsVDAC2 resulted in a 88% reduction of RSV loads at 24 h after RNAi, indicating that LsVDAC2 facilitates RSV accumulation in the vector. Yeast two-hybrid and GST pulldown assays demonstrated that LsVDAC2 interacted with RSV RNA-dependent RNA polymerase, RdRp. Furthermore, experiments in vivo and in vitro showed that LsVDAC2 induced the apoptotic response in RSV-infected insects and tissues. Silencing of LsVDAC2 via RNAi significantly reduced the expression of genes for apoptosis-related caspases 1a and 1c by 62% and 78%, respectively, in RSV-infected vectors. Whether LsVDAC2-induced RSV accumulation is related to RSV RdRp and LsVDAC2-induced cell apoptosis deserves further investigation.

Neutrophil extracellular traps facilitate cancer metastasis: cellular mechanisms and therapeutic strategies.

BACKGROUND: The formation of neutrophil extracellular traps (NETs) was initially discovered as a novel immune response against pathogens. Recent studies have also suggested that NETs play an important role in tumor progression. This review summarizes the cellular mechanisms by which NETs promote distant metastasis and discusses the possible clinical applications targeting NETs. METHOD: The relevant literature from PubMed and Google Scholar (2001-2021) have been reviewed for this article. RESULTS: The presence of NETs has been detected in various primary tumors and metastatic sites. NET-associated interactions have been observed throughout the different stages of metastasis, including initial tumor cell detachment, intravasation and extravasation, the survival of circulating tumor cells, the settlement and the growth of metastatic tumor cells. Several in vitro and in vivo studies proved that inhibiting NET formation resulted in anti-cancer effects. The biosafety and efficacy of some NET inhibitors have also been demonstrated in early phase clinical trials. CONCLUSIONS: Considering the role of NETs in tumor progression, NETs could be a promising diagnostic and therapeutic target for cancer management. However, current evidence is mostly derived from experimental models and as such more clinical studies are still needed to verify the clinical significance of NETs in oncological settings.

Parsing Netlists of Integrated Circuits from Images via Graph Attention Network.

A massive number of paper documents that include important information such as circuit schematics can be converted into digital documents by optical sensors like scanners or digital cameras. However, extracting the netlists of analog circuits from digital documents is an exceptionally challenging task. This process aids enterprises in digitizing paper-based circuit diagrams, enabling the reuse of analog circuit designs and the automatic generation of datasets required for intelligent design models in this domain. This paper introduces a bottom-up graph encoding model aimed at automatically parsing the circuit topology of analog integrated circuits from images. The model comprises an improved electronic component detection network based on the Swin Transformer, an algorithm for component port localization, and a graph encoding model. The objective of the detection network is to accurately identify component positions and types, followed by automatic dataset generation through port localization, and finally, utilizing the graph encoding model to predict potential connections between circuit components. To validate the model's performance, we annotated an electronic component detection dataset and a circuit diagram dataset, comprising 1200 and 3552 training samples, respectively. Detailed experimentation results demonstrate the superiority of our proposed enhanced algorithm over comparative algorithms across custom and public datasets. Furthermore, our proposed port localization algorithm significantly accelerates the annotation speed of circuit diagram datasets.

Recommendations of scRNA-seq Differential Gene Expression Analysis Based on Comprehensive Benchmarking.

To guide analysts to select the right tool and parameters in differential gene expression analyses of single-cell RNA sequencing (scRNA-seq) data, we developed a novel simulator that recapitulates the data characteristics of real scRNA-seq datasets while accounting for all the relevant sources of variation in a multi-subject, multi-condition scRNA-seq experiment: the cell-to-cell variation within a subject, the variation across subjects, the variability across cell types, the mean/variance relationship of gene expression across genes, library size effects, group effects, and covariate effects. By applying it to benchmark 12 differential gene expression analysis methods (including cell-level and pseudo-bulk methods) on simulated multi-condition, multi-subject data of the 10x Genomics platform, we demonstrated that methods originating from the negative binomial mixed model such as glmmTMB and NEBULA-HL outperformed other methods. Utilizing NEBULA-HL in a statistical analysis pipeline for single-cell analysis will enable scientists to better understand the cell-type-specific transcriptomic response to disease or treatment effects and to discover new drug targets. Further, application to two real datasets showed the outperformance of our differential expression (DE) pipeline, with unified findings of differentially expressed genes (DEG) and a pseudo-time trajectory transcriptomic result. In the end, we made recommendations for filtering strategies of cells and genes based on simulation results to achieve optimal experimental goals.

Construction of full-length infectious cDNA clones of two Korean isolates of turnip mosaic virus breaking resistance in Brassica napus.

In this work, two new turnip mosaic virus (TuMV) strains (Canola-12 and Canola-14) overcoming resistance in canola (Brassica napus) were isolated from a B. napus sample that showed typical TuMV-like symptoms and was collected in the city of Gimcheon, South Korea, in 2020. The complete genome sequence was determined and an infectious clone was made for each isolate. Phylogenetic analysis indicated that the strains isolated from canola belonged to the World-B group. Both infectious clones, which used 35S and T7 promoters to drive expression, induced systemic symptoms in Nicotiana benthamiana and B. napus. To our knowledge, this is the first report of TuMV infecting B. napus in South Korea.

Construction of full-length infectious clones of turnip mosaic virus isolates infecting Perilla frutescens and genetic analysis of recently emerged strains in Korea.

Perilla is an annual herb with a unique aroma and taste that has been cultivated in Korea for hundreds of years. It has been widely cultivated in many Asian and European countries as a food and medicinal crop. Recently, several viruses have been reported to cause diseases in perilla in Korea, including turnip mosaic virus (TuMV), which is known as a brassica pathogen due to its significant damage to brassica crops. In this study, we determined the complete genome sequences of two new TuMV isolates originating from perilla in Korea. Full-length infectious cDNA clones of these two isolates were constructed, and their infectivity was tested by agroinfiltration of Nicotiana benthamiana and sap inoculation of Chinese cabbage and radish plants. In addition, we analyzed the phylogenetic relationship of six new Korean TuMV isolates to members of the four major groups. We also used RDP4 software to conduct recombination analysis of recent isolates from Korea, which provided new insight into the evolutionary relationships of Korean isolates of TuMV.

Reassortment of Infectious Clones of Radish Mosaic Virus Shows that Systemic Necrosis in Nicotiana benthamiana Is Determined by RNA1.

Three infectious clones of radish mosaic virus (RaMV) were generated from isolates collected in mainland Korea (RaMV-Gg) and Jeju Island (RaMV-Aa and RaMV-Bb). These isolates differed in sequences and pathogenicity. Examination of the wild-type isolates and reassortants between the genomic RNA1 and RNA2 of these three isolates revealed that severe symptoms were associated with RNA1 of isolates Aa or Gg causing systemic necrosis in Nicotiana benthamiana, or with RNA1 of isolate Bb for induction of veinal necrosis and severe mosaic symptoms in radish. Reverse transcription, followed by quantitative real-time PCR (Q-RT-PCR), results from infected N. benthamiana confirmed that viral RNA2 accumulation level was correlated to RaMV necrosis-inducing ability, and that the RNA2 accumulation level was mostly dependent on the origin of RNA1. However, in radish, Q-RT-PCR results showed more similar viral RNA2 accumulation levels regardless of the ability of the isolate to induce necrosis. Phylogenetic analysis of genomic RNAs sequence including previously characterized isolates from North America, Europe, and Asia suggest possible recombination within RNA1, while analysis of concatenated RNA1+RNA2 sequences indicates that reassortment of RNA1 and RNA2 has been more important in the evolution of RaMV isolates than recombination. Korean isolate Aa is a potential reassortant between isolates RaMV-J and RaMV-TW, while isolate Bb might have evolved from reassortment between isolates RaMV-CA and RaMV-J. The Korean isolates were shown to also be able to infect Chinese cabbage, raising concerns that RaMV may spread from radish fields to the Chinese cabbage crop in Korea, causing further economic losses.

A multi-task two-path deep learning system for predicting the invasiveness of craniopharyngioma.

BACKGROUND AND OBJECTIVE: Craniopharyngioma is a kind of benign brain tumor in histography. However, it might be clinically aggressive and have severe manifestations, such as increased intracranial pressure, hypothalamic-pituitary dysfunction, and visual impairment. It is considered challenging for radiologists to predict the invasiveness of craniopharyngioma through MRI images. Therefore, developing a non-invasive method that can predict the invasiveness and boundary of CP as a reference before surgery is of clinical value for making more appropriate and individualized treatment decisions and reducing the occurrence of inappropriate surgical plan choices. METHODS: The MT-Brain system has consisted of two pathways, a sub-path based on 2D CNN for capturing the features from each slice of MRI images, and a 3D sub-network for capturing additional context information between slices. By introducing the two-path architecture, our system can make full use of the fusion of the above 2D and 3D features for classification. Furthermore, position encoding and mask-guided attention also have been introduced to improve the segmentation and diagnosis performance. To verify the performance of the MT-Brain system, we have enrolled 1032 patients with craniopharyngioma (302 invasion and 730 non-invasion patients), segmented the tumors on postcontrast coronal T1WI and randomized them into a training dataset and a testing dataset at a ratio of 8:2. RESULTS: The MT-Brain system achieved a remarkable performance in diagnosing the invasiveness of craniopharyngioma with the AUC of 83.84%, the accuracy of 77.94%, the sensitivity of 70.97%, and the specificity of 80.99%. In the lesion segmentation task, the predicted boundaries of lesions were similar to those labeled by radiologists with the dice of 66.36%. In addition, some explorations also have been made on the interpretability of deep learning models, illustrating the reliability of the model. CONCLUSIONS: To the best of our knowledge, this study is the first to develop an integrated deep learning model to predict the invasiveness of craniopharyngioma preoperatively and locate the lesion boundary synchronously on MRI. The excellent performances indicate that the MT-Brain system has great potential in real-world clinical applications.

Brain Functional Connectivity Analysis via Graphical Deep Learning.

OBJECTIVE: Graphical deep learning models provide a desirable way for brain functional connectivity analysis. However, the application of current graph deep learning models to brain network analysis is challenging due to the limited sample size and complex relationships between different brain regions. METHOD: In this work, a graph convolutional network (GCN) based framework is proposed by exploiting the information from both region-to-region connectivities of the brain and subject-subject relationships. We first construct an affinity subject-subject graph followed by GCN analysis. A Laplacian regularization term is introduced in our model to tackle the overfitting problem. We apply and validate the proposed model to the Philadelphia Neurodevelopmental Cohort for the brain cognition study. RESULTS: Experimental analysis shows that our proposed framework outperforms other competing models in classifying groups with low and high Wide Range Achievement Test (WRAT) scores. Moreover, to examine each brain region's contribution to cognitive function, we use the occlusion sensitivity analysis method to identify cognition-related brain functional networks. The results are consistent with previous research yet yield new findings. CONCLUSION AND SIGNIFICANCE: Our study demonstrates that GCN incorporating prior knowledge about brain networks offers a powerful way to detect important brain networks and regions associated with cognitive functions.

Prediction of Antibody-Antigen Binding via Machine Learning: Development of Data Sets and Evaluation of Methods.

BACKGROUND: The mammalian immune system is able to generate antibodies against a huge variety of antigens, including bacteria, viruses, and toxins. The ultradeep DNA sequencing of rearranged immunoglobulin genes has considerable potential in furthering our understanding of the immune response, but it is limited by the lack of a high-throughput, sequence-based method for predicting the antigen(s) that a given immunoglobulin recognizes. OBJECTIVE: As a step toward the prediction of antibody-antigen binding from sequence data alone, we aimed to compare a range of machine learning approaches that were applied to a collated data set of antibody-antigen pairs in order to predict antibody-antigen binding from sequence data. METHODS: Data for training and testing were extracted from the Protein Data Bank and the Coronavirus Antibody Database, and additional antibody-antigen pair data were generated by using a molecular docking protocol. Several machine learning methods, including the weighted nearest neighbor method, the nearest neighbor method with the BLOSUM62 matrix, and the random forest method, were applied to the problem. RESULTS: The final data set contained 1157 antibodies and 57 antigens that were combined in 5041 antibody-antigen pairs. The best performance for the prediction of interactions was obtained by using the nearest neighbor method with the BLOSUM62 matrix, which resulted in around 82% accuracy on the full data set. These results provide a useful frame of reference, as well as protocols and considerations, for machine learning and data set creation in the prediction of antibody-antigen binding. CONCLUSIONS: Several machine learning approaches were compared to predict antibody-antigen interaction from protein sequences. Both the data set (in CSV format) and the machine learning program (coded in Python) are freely available for download on GitHub.

Ensemble Manifold Regularized Multi-Modal Graph Convolutional Network for Cognitive Ability Prediction.

OBJECTIVE: Multi-modal functional magnetic resonance imaging (fMRI) can be used to make predictions about individual behavioral and cognitive traits based on brain connectivity networks. METHODS: To take advantage of complementary information from multi-modal fMRI, we propose an interpretable multi-modal graph convolutional network (MGCN) model, incorporating both fMRI time series and functional connectivity (FC) between each pair of brain regions. Specifically, our model learns a graph embedding from individual brain networks derived from multi-modal data. A manifold-based regularization term is enforced to consider the relationships of subjects both within and between modalities. Furthermore, we propose the gradient-weighted regression activation mapping (Grad-RAM) and the edge mask learning to interpret the model, which is then used to identify significant cognition-related biomarkers. RESULTS: We validate our MGCN model on the Philadelphia Neurodevelopmental Cohort to predict individual wide range achievement test (WRAT) score. Our model obtains superior predictive performance over GCN with a single modality and other competing approaches. The identified biomarkers are cross-validated from different approaches. CONCLUSION AND SIGNIFICANCE: This paper develops a new interpretable graph deep learning framework for cognition prediction, with the potential to overcome the limitations of several current data-fusion models. The results demonstrate the power of MGCN in analyzing multi-modal fMRI and discovering significant biomarkers for human brain studies.