High-accuracy protein model quality assessment using attention graph neural networks.

Great improvement has been brought to protein tertiary structure prediction through deep learning. It is important but very challenging to accurately rank and score decoy structures predicted by different models. CASP14 results show that existing quality assessment (QA) approaches lag behind the development of protein structure prediction methods, where almost all existing QA models degrade in accuracy when the target is a decoy of high quality. How to give an accurate assessment to high-accuracy decoys is particularly useful with the available of accurate structure prediction methods. Here we propose a fast and effective single-model QA method, QATEN, which can evaluate decoys only by their topological characteristics and atomic types. Our model uses graph neural networks and attention mechanisms to evaluate global and amino acid level scores, and uses specific loss functions to constrain the network to focus more on high-precision decoys and protein domains. On the CASP14 evaluation decoys, QATEN performs better than other QA models under all correlation coefficients when targeting average LDDT. QATEN shows promising performance when considering only high-accuracy decoys. Compared to the embedded evaluation modules of predicted ${C}_{\alpha^{-}} RMSD$ (pRMSD) in RosettaFold and predicted LDDT (pLDDT) in AlphaFold2, QATEN is complementary and capable of achieving better evaluation on some decoy structures generated by AlphaFold2 and RosettaFold. These results suggest that the new QATEN approach can be used as a reliable independent assessment algorithm for high-accuracy protein structure decoys.

Leveraging scaffold information to predict protein-ligand binding affinity with an empirical graph neural network.

Protein-ligand binding affinity prediction is an important task in structural bioinformatics for drug discovery and design. Although various scoring functions (SFs) have been proposed, it remains challenging to accurately evaluate the binding affinity of a protein-ligand complex with the known bound structure because of the potential preference of scoring system. In recent years, deep learning (DL) techniques have been applied to SFs without sophisticated feature engineering. Nevertheless, existing methods cannot model the differential contribution of atoms in various regions of proteins, and the relationship between atom properties and intermolecular distance is also not fully explored. We propose a novel empirical graph neural network for accurate protein-ligand binding affinity prediction (EGNA). Graphs of protein, ligand and their interactions are constructed based on different regions of each bound complex. Proteins and ligands are effectively represented by graph convolutional layers, enabling the EGNA to capture interaction patterns precisely by simulating empirical SFs. The contributions of different factors on binding affinity can thus be transparently investigated. EGNA is compared with the state-of-the-art machine learning-based SFs on two widely used benchmark data sets. The results demonstrate the superiority of EGNA and its good generalization capability.

MINDG: a drug-target interaction prediction method based on an integrated learning algorithm.

MOTIVATION: Drug-target interaction (DTI) prediction refers to the prediction of whether a given drug molecule will bind to a specific target and thus exert a targeted therapeutic effect. Although intelligent computational approaches for drug target prediction have received much attention and made many advances, they are still a challenging task that requires further research. The main challenges are manifested as follows: (i) most graph neural network-based methods only consider the information of the first-order neighboring nodes (drug and target) in the graph, without learning deeper and richer structural features from the higher-order neighboring nodes. (ii) Existing methods do not consider both the sequence and structural features of drugs and targets, and each method is independent of each other, and cannot combine the advantages of sequence and structural features to improve the interactive learning effect. RESULTS: To address the above challenges, a Multi-view Integrated learning Network that integrates Deep learning and Graph Learning (MINDG) is proposed in this study, which consists of the following parts: (i) a mixed deep network is used to extract sequence features of drugs and targets, (ii) a higher-order graph attention convolutional network is proposed to better extract and capture structural features, and (iii) a multi-view adaptive integrated decision module is used to improve and complement the initial prediction results of the above two networks to enhance the prediction performance. We evaluate MINDG on two dataset and show it improved DTI prediction performance compared to state-of-the-art baselines. AVAILABILITY AND IMPLEMENTATION: https://github.com/jnuaipr/MINDG.

DeepQs: Local quality assessment of cryo-EM density map by deep learning map-model fit score.

Cryogenic electron microscopy maps are valuable for determining macromolecule structures. A proper quality assessment method is essential for cryo-EM map selection or revision. This article presents DeepQs, a novel approach to estimate local quality for 3D cryo-EM density maps, using a deep-learning algorithm based on map-model fit score. DeepQs is a parameter-free method for users and incorporates structural information between map and its related atomic model into well-trained models by deep learning. More specifically, the DeepQs approach leverages the interplay between map and atomic model through predefined map-model fit score, Q-score. DeepQs can get close results to the ground truth map-model fit scores with only cryo-EM map as input. In experiments, DeepQs demonstrates the lowest root mean square error with standard method Fourier shell correlation metric and high correlation with map-model fit score, Q-score, when compared with other local quality estimation methods in high-resolution dataset (<=5 Å). DeepQs can also be applied to evaluate the quality of the post-processed maps. In both cases, DeepQs runs faster by using GPU acceleration. Our program is available at http://www.csbio.sjtu.edu.cn/bioinf/DeepQs for academic use.

SIFLoc: a self-supervised pre-training method for enhancing the recognition of protein subcellular localization in immunofluorescence microscopic images.

With the rapid growth of high-resolution microscopy imaging data, revealing the subcellular map of human proteins has become a central task in the spatial proteome. The cell atlas of the Human Protein Atlas (HPA) provides precious resources for recognizing subcellular localization patterns at the cell level, and the large-scale annotated data enable learning via advanced deep neural networks. However, the existing predictors still suffer from the imbalanced class distribution and the lack of labeled data for minor classes. Thus, it is necessary to develop new methods for coping with these issues. We leverage the self-supervised learning protocol to address these problems. Especially, we propose a pre-training scheme to enhance the conventional supervised learning framework called SIFLoc. The pre-training is featured by a hybrid data augmentation method and a modified contrastive loss function, aiming to learn good feature representations from microscopic images. The experiments are performed on a large-scale immunofluorescence microscopic image dataset collected from the HPA database. Using the same deep neural networks as the classifier, the model pre-trained via SIFLoc not only outperforms the model without pre-training by a large margin but also shows advantages over the state-of-the-art self-supervised learning methods. Especially, SIFLoc improves the prediction accuracy for minor organelles significantly.

Learning protein subcellular localization multi-view patterns from heterogeneous data of imaging, sequence and networks.

Location proteomics seeks to provide automated high-resolution descriptions of protein location patterns within cells. Many efforts have been undertaken in location proteomics over the past decades, thereby producing plenty of automated predictors for protein subcellular localization. However, most of these predictors are trained solely from high-throughput microscopic images or protein amino acid sequences alone. Unifying heterogeneous protein data sources has yet to be exploited. In this paper, we present a pipeline called sequence, image, network-based protein subcellular locator (SIN-Locator) that constructs a multi-view description of proteins by integrating multiple data types including images of protein expression in cells or tissues, amino acid sequences and protein-protein interaction networks, to classify the patterns of protein subcellular locations. Proteins were encoded by both handcrafted features and deep learning features, and multiple combining methods were implemented. Our experimental results indicated that optimal integrations can considerately enhance the classification accuracy, and the utility of SIN-Locator has been demonstrated through applying to new released proteins in the human protein atlas. Furthermore, we also investigate the contribution of different data sources and influence of partial absence of data. This work is anticipated to provide clues for reconciliation and combination of multi-source data for protein location analysis.

Accurate flexible refinement for atomic-level protein structure using cryo-EM density maps and deep learning.

With the rapid progress of deep learning in cryo-electron microscopy and protein structure prediction, improving the accuracy of the protein structure model by using a density map and predicted contact/distance map through deep learning has become an urgent need for robust methods. Thus, designing an effective protein structure optimization strategy based on the density map and predicted contact/distance map is critical to improving the accuracy of structure refinement. In this article, a protein structure optimization method based on the density map and predicted contact/distance map by deep-learning technology was proposed in accordance with the result of matching between the density map and the initial model. Physics- and knowledge-based energy functions, integrated with Cryo-EM density map data and deep-learning data, were used to optimize the protein structure in the simulation. The dynamic confidence score was introduced to the iterative process for choosing whether it is a density map or a contact/distance map to dominate the movement in the simulation to improve the accuracy of refinement. The protocol was tested on a large set of 224 non-homologous membrane proteins and generated 214 structural models with correct folds, where 4.5% of structural models were generated from structural models with incorrect folds. Compared with other state-of-the-art methods, the major advantage of the proposed methods lies in the skills for using density map and contact/distance map in the simulation, as well as the new energy function in the re-assembly simulations. Overall, the results demonstrated that this strategy is a valuable approach and ready to use for atomic-level structure refinement using cryo-EM density map and predicted contact/distance map.

circRNA-binding protein site prediction based on multi-view deep learning, subspace learning and multi-view classifier.

Circular RNAs (circRNAs) generally bind to RNA-binding proteins (RBPs) to play an important role in the regulation of autoimmune diseases. Thus, it is crucial to study the binding sites of RBPs on circRNAs. Although many methods, including traditional machine learning and deep learning, have been developed to predict the interactions between RNAs and RBPs, and most of them are focused on linear RNAs. At present, few studies have been done on the binding relationships between circRNAs and RBPs. Thus, in-depth research is urgently needed. In the existing circRNA-RBP binding site prediction methods, circRNA sequences are the main research subjects, but the relevant characteristics of circRNAs have not been fully exploited, such as the structure and composition information of circRNA sequences. Some methods have extracted different views to construct recognition models, but how to efficiently use the multi-view data to construct recognition models is still not well studied. Considering the above problems, this paper proposes a multi-view classification method called DMSK based on multi-view deep learning, subspace learning and multi-view classifier for the identification of circRNA-RBP interaction sites. In the DMSK method, first, we converted circRNA sequences into pseudo-amino acid sequences and pseudo-dipeptide components for extracting high-dimensional sequence features and component features of circRNAs, respectively. Then, the structure prediction method RNAfold was used to predict the secondary structure of the RNA sequences, and the sequence embedding model was used to extract the context-dependent features. Next, we fed the above four views' raw features to a hybrid network, which is composed of a convolutional neural network and a long short-term memory network, to obtain the deep features of circRNAs. Furthermore, we used view-weighted generalized canonical correlation analysis to extract four views' common features by subspace learning. Finally, the learned subspace common features and multi-view deep features were fed to train the downstream multi-view TSK fuzzy system to construct a fuzzy rule and fuzzy inference-based multi-view classifier. The trained classifier was used to predict the specific positions of the RBP binding sites on the circRNAs. The experiments show that the prediction performance of the proposed method DMSK has been improved compared with the existing methods. The code and dataset of this study are available at https://github.com/Rebecca3150/DMSK.

MDGF-MCEC: a multi-view dual attention embedding model with cooperative ensemble learning for CircRNA-disease association prediction.

Circular RNA (circRNA) is closely involved in physiological and pathological processes of many diseases. Discovering the associations between circRNAs and diseases is of great significance. Due to the high-cost to verify the circRNA-disease associations by wet-lab experiments, computational approaches for predicting the associations become a promising research direction. In this paper, we propose a method, MDGF-MCEC, based on multi-view dual attention graph convolution network (GCN) with cooperative ensemble learning to predict circRNA-disease associations. First, MDGF-MCEC constructs two disease relation graphs and two circRNA relation graphs based on different similarities. Then, the relation graphs are fed into a multi-view GCN for representation learning. In order to learn high discriminative features, a dual-attention mechanism is introduced to adjust the contribution weights, at both channel level and spatial level, of different features. Based on the learned embedding features of diseases and circRNAs, nine different feature combinations between diseases and circRNAs are treated as new multi-view data. Finally, we construct a multi-view cooperative ensemble classifier to predict the associations between circRNAs and diseases. Experiments conducted on the CircR2Disease database demonstrate that the proposed MDGF-MCEC model achieves a high area under curve of 0.9744 and outperforms the state-of-the-art methods. Promising results are also obtained from experiments on the circ2Disease and circRNADisease databases. Furthermore, the predicted associated circRNAs for hepatocellular carcinoma and gastric cancer are supported by the literature. The code and dataset of this study are available at https://github.com/ABard0/MDGF-MCEC.

De novo drug design by iterative multiobjective deep reinforcement learning with graph-based molecular quality assessment.

MOTIVATION: Generating molecules of high quality and drug-likeness in the vast chemical space is a big challenge in the drug discovery. Most existing molecule generative methods focus on diversity and novelty of molecules, but ignoring drug potentials of the generated molecules during the generation process. RESULTS: In this study, we present a novel de novo multiobjective quality assessment-based drug design approach (QADD), which integrates an iterative refinement framework with a novel graph-based molecular quality assessment model on drug potentials. QADD designs a multiobjective deep reinforcement learning pipeline to generate molecules with multiple desired properties iteratively, where a graph neural network-based model for accurate molecular quality assessment on drug potentials is introduced to guide molecule generation. Experimental results show that QADD can jointly optimize multiple molecular properties with a promising performance and the quality assessment module is capable of guiding the generated molecules with high drug potentials. Furthermore, applying QADD to generate novel molecules binding to a biological target protein DRD2 also demonstrates the algorithm's efficacy. AVAILABILITY AND IMPLEMENTATION: QADD is freely available online for academic use at https://github.com/yifang000/QADD or http://www.csbio.sjtu.edu.cn/bioinf/QADD.

MLNGCF: circRNA-disease associations prediction with multilayer attention neural graph-based collaborative filtering.

MOTIVATION: CircRNAs play a critical regulatory role in physiological processes, and the abnormal expression of circRNAs can mediate the processes of diseases. Therefore, exploring circRNAs-disease associations is gradually becoming an important area of research. Due to the high cost of validating circRNA-disease associations using traditional wet-lab experiments, novel computational methods based on machine learning are gaining more and more attention in this field. However, current computational methods suffer to insufficient consideration of latent features in circRNA-disease interactions. RESULTS: In this study, a multilayer attention neural graph-based collaborative filtering (MLNGCF) is proposed. MLNGCF first enhances multiple biological information with autoencoder as the initial features of circRNAs and diseases. Then, by constructing a central network of different diseases and circRNAs, a multilayer cooperative attention-based message propagation is performed on the central network to obtain the high-order features of circRNAs and diseases. A neural network-based collaborative filtering is constructed to predict the unknown circRNA-disease associations and update the model parameters. Experiments on the benchmark datasets demonstrate that MLNGCF outperforms state-of-the-art methods, and the prediction results are supported by the literature in the case studies. AVAILABILITY AND IMPLEMENTATION: The source codes and benchmark datasets of MLNGCF are available at https://github.com/ABard0/MLNGCF.

Cryo-EM image alignment: From pair-wise to joint with deep unsupervised difference learning.

Cryo-electron microscopy (cryo-EM) single-particle analysis is a revolutionary imaging technique to resolve and visualize biomacromolecules. Image alignment in cryo-EM is an important and basic step to improve the precision of the image distance calculation. However, it is a very challenging task due to high noise and low signal-to-noise ratio. Therefore, we propose a new deep unsupervised difference learning (UDL) strategy with novel pseudo-label guided learning network architecture and apply it to pair-wise image alignment in cryo-EM. The training framework is fully unsupervised. Furthermore, a variant of UDL called joint UDL (JUDL), is also proposed, which is capable of utilizing the similarity information of the whole dataset and thus further increase the alignment precision. Assessments on both real-world and synthetic cryo-EM single-particle image datasets suggest the new unsupervised joint alignment method can achieve more accurate alignment results. Our method is highly efficient by taking advantages of GPU devices. The source code of our methods is publicly available at "http://www.csbio.sjtu.edu.cn/bioinf/JointUDL/" for academic use.

Recognizing binding sites of poorly characterized RNA-binding proteins on circular RNAs using attention Siamese network.

Circular RNAs (circRNAs) interact with RNA-binding proteins (RBPs) to play crucial roles in gene regulation and disease development. Computational approaches have attracted much attention to quickly predict highly potential RBP binding sites on circRNAs using the sequence or structure statistical binding knowledge. Deep learning is one of the popular learning models in this area but usually requires a lot of labeled training data. It would perform unsatisfactorily for the less characterized RBPs with a limited number of known target circRNAs. How to improve the prediction performance for such small-size labeled characterized RBPs is a challenging task for deep learning-based models. In this study, we propose an RBP-specific method iDeepC for predicting RBP binding sites on circRNAs from sequences. It adopts a Siamese neural network consisting of a lightweight attention module and a metric module. We have found that Siamese neural network effectively enhances the network capability of capturing mutual information between circRNAs with pairwise metric learning. To further deal with the small-sample size problem, we have performed the pretraining using available labeled data from other RBPs and also demonstrate the efficacy of this transfer-learning pipeline. We comprehensively evaluated iDeepC on the benchmark datasets of RBP-binding circRNAs, and the results suggest iDeepC achieving promising results on the poorly characterized RBPs. The source code is available at https://github.com/hehew321/iDeepC.

RNA-binding protein recognition based on multi-view deep feature and multi-label learning.

RNA-binding protein (RBP) is a class of proteins that bind to and accompany RNAs in regulating biological processes. An RBP may have multiple target RNAs, and its aberrant expression can cause multiple diseases. Methods have been designed to predict whether a specific RBP can bind to an RNA and the position of the binding site using binary classification model. However, most of the existing methods do not take into account the binding similarity and correlation between different RBPs. While methods employing multiple labels and Long Short Term Memory Network (LSTM) are proposed to consider binding similarity between different RBPs, the accuracy remains low due to insufficient feature learning and multi-label learning on RNA sequences. In response to this challenge, the concept of RNA-RBP Binding Network (RRBN) is proposed in this paper to provide theoretical support for multi-label learning to identify RBPs that can bind to RNAs. It is experimentally shown that the RRBN information can significantly improve the prediction of unknown RNA-RBP interactions. To further improve the prediction accuracy, we present the novel computational method iDeepMV which integrates multi-view deep learning technology under the multi-label learning framework. iDeepMV first extracts data from the views of amino acid sequence and dipeptide component based on the RNA sequences as the original view. Deep neural network models are then designed for the respective views to perform deep feature learning. The extracted deep features are fed into multi-label classifiers which are trained with the RNA-RBP interaction information for the three views. Finally, a voting mechanism is designed to make comprehensive decision on the results of the multi-label classifiers. Our experimental results show that the prediction performance of iDeepMV, which combines multi-view deep feature learning models with RNA-RBP interaction information, is significantly better than that of the state-of-the-art methods. iDeepMV is freely available at http://www.csbio.sjtu.edu.cn/bioinf/iDeepMV for academic use. The code is freely available at http://github.com/uchihayht/iDeepMV.

Accurate inference of gene regulatory interactions from spatial gene expression with deep contrastive learning.

MOTIVATION: Reverse engineering of gene regulatory networks (GRNs) has long been an attractive research topic in system biology. Computational prediction of gene regulatory interactions has remained a challenging problem due to the complexity of gene expression and scarce information resources. The high-throughput spatial gene expression data, like in situ hybridization images that exhibit temporal and spatial expression patterns, has provided abundant and reliable information for the inference of GRNs. However, computational tools for analyzing the spatial gene expression data are highly underdeveloped. RESULTS: In this study, we develop a new method for identifying gene regulatory interactions from gene expression images, called ConGRI. The method is featured by a contrastive learning scheme and deep Siamese convolutional neural network architecture, which automatically learns high-level feature embeddings for the expression images and then feeds the embeddings to an artificial neural network to determine whether or not the interaction exists. We apply the method to a Drosophila embryogenesis dataset and identify GRNs of eye development and mesoderm development. Experimental results show that ConGRI outperforms previous traditional and deep learning methods by a large margin, which achieves accuracies of 76.7% and 68.7% for the GRNs of early eye development and mesoderm development, respectively. It also reveals some master regulators for Drosophila eye development. AVAILABILITYAND IMPLEMENTATION: https://github.com/lugimzheng/ConGRI. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

CoCoPRED: coiled-coil protein structural feature prediction from amino acid sequence using deep neural networks.

MOTIVATION: Coiled-coil is composed of two or more helices that are wound around each other. It widely exists in proteins and has been discovered to play a variety of critical roles in biology processes. Generally, there are three types of structural features in coiled-coil: coiled-coil domain (CCD), oligomeric state and register. However, most of the existing computational tools only focus on one of them. RESULTS: Here, we describe a new deep learning model, CoCoPRED, which is based on convolutional layers, bidirectional long short-term memory, and attention mechanism. It has three networks, i.e. CCD network, oligomeric state network, and register network, corresponding to the three types of structural features in coiled-coil. This means CoCoPRED has the ability of fulfilling comprehensive prediction for coiled-coil proteins. Through the 5-fold cross-validation experiment, we demonstrate that CoCoPRED can achieve better performance than the state-of-the-art models on both CCD prediction and oligomeric state prediction. Further analysis suggests the CCD prediction may be a performance indicator of the oligomeric state prediction in CoCoPRED. The attention heads in CoCoPRED indicate that registers a, b and e are more crucial for the oligomeric state prediction. AVAILABILITY AND IMPLEMENTATION: CoCoPRED is available at http://www.csbio.sjtu.edu.cn/bioinf/CoCoPRED. The datasets used in this research can also be downloaded from the website. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

GraphLoc: a graph neural network model for predicting protein subcellular localization from immunohistochemistry images.

MOTIVATION: Recognition of protein subcellular distribution patterns and identification of location biomarker proteins in cancer tissues are important for understanding protein functions and related diseases. Immunohistochemical (IHC) images enable visualizing the distribution of proteins at the tissue level, providing an important resource for the protein localization studies. In the past decades, several image-based protein subcellular location prediction methods have been developed, but the prediction accuracies still have much space to improve due to the complexity of protein patterns resulting from multi-label proteins and the variation of location patterns across cell types or states. RESULTS: Here, we propose a multi-label multi-instance model based on deep graph convolutional neural networks, GraphLoc, to recognize protein subcellular location patterns. GraphLoc builds a graph of multiple IHC images for one protein, learns protein-level representations by graph convolutions and predicts multi-label information by a dynamic threshold method. Our results show that GraphLoc is a promising model for image-based protein subcellular location prediction with model interpretability. Furthermore, we apply GraphLoc to the identification of candidate location biomarkers and potential members for protein networks. A large portion of the predicted results have supporting evidence from the existing literatures and the new candidates also provide guidance for further experimental screening. AVAILABILITY AND IMPLEMENTATION: The dataset and code are available at: www.csbio.sjtu.edu.cn/bioinf/GraphLoc. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Fast protein structure comparison through effective representation learning with contrastive graph neural networks.

Protein structure alignment algorithms are often time-consuming, resulting in challenges for large-scale protein structure similarity-based retrieval. There is an urgent need for more efficient structure comparison approaches as the number of protein structures increases rapidly. In this paper, we propose an effective graph-based protein structure representation learning method, GraSR, for fast and accurate structure comparison. In GraSR, a graph is constructed based on the intra-residue distance derived from the tertiary structure. Then, deep graph neural networks (GNNs) with a short-cut connection learn graph representations of the tertiary structures under a contrastive learning framework. To further improve GraSR, a novel dynamic training data partition strategy and length-scaling cosine distance are introduced. We objectively evaluate our method GraSR on SCOPe v2.07 and a new released independent test set from PDB database with a designed comprehensive performance metric. Compared with other state-of-the-art methods, GraSR achieves about 7%-10% improvement on two benchmark datasets. GraSR is also much faster than alignment-based methods. We dig into the model and observe that the superiority of GraSR is mainly brought by the learned discriminative residue-level and global descriptors. The web-server and source code of GraSR are freely available at www.csbio.sjtu.edu.cn/bioinf/GraSR/ for academic use.

GraphBind: protein structural context embedded rules learned by hierarchical graph neural networks for recognizing nucleic-acid-binding residues.

Knowledge of the interactions between proteins and nucleic acids is the basis of understanding various biological activities and designing new drugs. How to accurately identify the nucleic-acid-binding residues remains a challenging task. In this paper, we propose an accurate predictor, GraphBind, for identifying nucleic-acid-binding residues on proteins based on an end-to-end graph neural network. Considering that binding sites often behave in highly conservative patterns on local tertiary structures, we first construct graphs based on the structural contexts of target residues and their spatial neighborhood. Then, hierarchical graph neural networks (HGNNs) are used to embed the latent local patterns of structural and bio-physicochemical characteristics for binding residue recognition. We comprehensively evaluate GraphBind on DNA/RNA benchmark datasets. The results demonstrate the superior performance of GraphBind than state-of-the-art methods. Moreover, GraphBind is extended to other ligand-binding residue prediction to verify its generalization capability. Web server of GraphBind is freely available at http://www.csbio.sjtu.edu.cn/bioinf/GraphBind/.

Incorporating label correlations into deep neural networks to classify protein subcellular location patterns in immunohistochemistry images.

Analysis of protein subcellular localization is a critical part of proteomics. In recent years, as both the number and quality of microscopic images are increasing rapidly, many automated methods, especially convolutional neural networks (CNN), have been developed to predict protein subcellular location(s) based on bioimages, but their performance always suffers from some inherent properties of the problem. First, many microscopic images have non-informative or noisy sections, like unstained stroma and unspecific background, which affect the extraction of protein expression information. Second, the patterns of protein subcellular localization are very complex, as a lot of proteins locate in more than one compartment. In this study, we propose a new label-correlation enhanced deep neural network, laceDNN, to classify the subcellular locations of multi-label proteins from immunohistochemistry images. The model uses small representative patches as input to alleviate the image noise issue, and its backbone is a hybrid architecture of CNN and recurrent neural network, where the former network extracts representative image features and the latter learns the organelle dependency relationships. Our experimental results indicate that the proposed model can improve the performance of multi-label protein subcellular classification.