A multimodal model for protein function prediction

MotivationProtein function prediction is important for drug development and disease treatment. Recently, deep learning methods have leveraged protein sequence and structural information, achieving remarkable progress in the field of protein function prediction. However, existing methods ignore the complex multimodal interaction information between sequence and structural features. Since protein sequence and structural information reveal the functional characteristics of proteins from different perspectives, it is challenging to effectively fuse the information from these two modalities to portray protein functions more comprehensively. In addition, current methods have difficulty in effectively capturing long-range dependencies and global contextual information in protein sequences during feature extraction, thus limiting the ability of the model to recognize critical functional residues.ResultsIn this study, we propose a novel framework termed Multi-stage Attention-based Extraction and Fusion model for GO prediction (MAEF-GO) based on a multistage attention mechanism to predict protein functions. MAEF-GO innovatively integrates the graph convolutional network and the graph attention network to extract protein structural features. To address the issue of modeling long-range dependencies within protein sequences, we introduce a frequency-domain attention mechanism capable of extracting global contextual relationships. Additionally, a cross-attention module is implemented to facilitate interactive fusion between protein sequence and structural modalities. Experimental evaluations demonstrate that MAEF-GO achieves superior performance compared to several state-of-the-art baseline models across standard benchmarks. Furthermore, analysis of the cross-attention weight distributions demonstrates MAEF-GO’s interpretability. It can effectively identify critical functional residues of proteins.Availability and implementationThe MAEF-GO source code can be found at https://github.com/nebstudio/MAEF-GO, an archived snapshot of the code used in this study is also available via Zenodo at https://doi.org/10.5281/zenodo.15422392.

Completely sequenced genomes and other types of genomics data provide us with new information to predict protein function. While classical, homology-based function prediction provides information about a proteins’ molecular function (what does the protein do at a molecular scale?), the analysis of the sequence in the context of its genome or in other types of genomics data provides information about its functional context (what are the proteins’ interaction partners, and in which biological process does it play a role?) Genomic context data are however inherently noisy. Only by combining different types of genomic(s) context data (vertical comparative genomics) or by combining the same type of genomics data from different species (horizontal comparative genomics) do they become sufficiently reliable to be used for protein function prediction. Homology-based function prediction and context-based function prediction provide complementary information about a protein’s function and can becombined to make predictions that are specific enough for experimental testing. Here we discuss the genomic coverage and reliability of combining genomics data for protein function prediction and survey predictions that have actually led to experimental confirmation. Using a number of examples we illustrate how combining the information from various types of genomics data can lead to specific protein function predictions. These include the prediction that the Ribonuclease L inhibitor (RLI) is involved in the maturation of ribosomal RNA.

AlphaFold2 (AF2) is an artificial intelligence (AI) system developed by DeepMind that can predict three-dimensional (3D) structures of proteins from amino acid sequences with atomic-level accuracy. Protein structure prediction is one of the most challenging problems in computational biology and chemistry, and has puzzled scientists for 50 years. The advent of AF2 presents an unprecedented progress in protein structure prediction and has attracted much attention. Subsequent release of structures of more than 200 million proteins predicted by AF2 further aroused great enthusiasm in the science community, especially in the fields of biology and medicine. AF2 is thought to have a significant impact on structural biology and research areas that need protein structure information, such as drug discovery, protein design, prediction of protein function, et al. Though the time is not long since AF2 was developed, there are already quite a few application studies of AF2 in the fields of biology and medicine, with many of them having preliminarily proved the potential of AF2. To better understand AF2 and promote its applications, we will in this article summarize the principle and system architecture of AF2 as well as the recipe of its success, and particularly focus on reviewing its applications in the fields of biology and medicine. Limitations of current AF2 prediction will also be discussed.

Protein sequences are symbols generally different characters representing the 20 amino acids used in human proteins those sequences can range from the very sort to the very long. There are many proteins database for the sequences are known but the function and functional annotation is not. Protein function prediction (PFP) as well as functional annotation (FA) from its structure or sequence is a major field of bioinformatics at the same time how to judge how well perform these algorithms. We proposed the novel method that converts the protein function problem into a language translation problem by a new proposed protein sequence language encoded to the protein function language decoded and build a recurrent neural machine encoding decoding translator (RNNEDT) based on the recurrent neural networks model. The excellent acting on training, testing datasets exhibits the proposed system as an improving direction for PFP. The proposed system alters the PFP matter to a language translation issue as well as applies a recurrent neural network machine version model for PFP, and visualizes the annotation of biological process (BP), molecular function (MF), as well as cellular component (CP).

The iterative threading assembly refinement (I-TASSER) server is an integrated platform for automated protein structure and function prediction based on the sequence-to-structure-to-function paradigm. Starting from an amino acid sequence, I-TASSER first generates three-dimensional (3D) atomic models from multiple threading alignments and iterative structural assembly simulations. The function of the protein is then inferred by structurally matching the 3D models with other known proteins. The output from a typical server run contains full-length secondary and tertiary structure predictions, and functional annotations on ligand-binding sites, Enzyme Commission numbers and Gene Ontology terms. An estimate of accuracy of the predictions is provided based on the confidence score of the modeling. This protocol provides new insights and guidelines for designing of online server systems for the state-of-the-art protein structure and function predictions. The server is available at http://zhanglab.ccmb.med.umich.edu/I-TASSER.

Proteins are considered the central compound necessary for life, as they play a crucial role in governing several life processes by performing the most essential biological and chemical functions in every living cell. Understanding protein structures and functions will lead to a significant advance in life science and biology. Such knowledge is vital for various fields such as drug development and synthetic biofuels production. Most proteins have definite shapes that they fold into, which are the most stable state they can adopt. Due to the fact that the protein structure information provides important insight into its functions, many research efforts have been conducted to determine the protein 3-dimensional structure from its sequence. The experimental methods for protein 3-dimensional structure determination are often time-consuming, costly, and even not feasible for some proteins. Accordingly, recent research efforts focus more and more on computational approaches to predict protein 3-dimensional structures. Template-based modeling is considered one of the most accurate protein structure prediction methods. The success of template-based modeling relies on correctly identifying one or a few experimentally determined protein structures as structural templates that are likely to resemble the structure of the target sequence as well as accurately producing a sequence alignment that maps the residues in the target sequence to those in the template. In this work, we aim at improving the template-based protein structure modeling by enhancing the correctness of identifying the most appropriate templates and precisely aligning the target and template sequences. Firstly, we investigate employing inter-residue contact score to measure the favorability of a target sequence fitting in the folding topology of a certain template. Secondly, we design a multi-objective alignment algorithm extending the famous Needleman-Wunsch algorithm to obtain a complete set of alignments yielding Pareto optimality. Then, we use protein sequence and structural information as objectives and generate the complete Pareto optimal front of alignments between target sequence and template. The alignments obtained enable one to analyze the trade-offs between the potentially conflicting objectives. These approaches lead to accuracy enhancement in template-based protein structure modeling.

Well understanding protein function and structure in computational biology helps in the understanding of human beings. To face the limited proteins that are annotated structurally and functionally, the scientific community embraces the self-supervised pre-training methods from large amounts of unlabeled protein sequences for protein embedding learning. However, the protein is usually represented by individual amino acids with limited vocabulary size (e.g. 20 type proteins), without considering the strong local semantics existing in protein sequences. In this work, we propose a novel pre-training modeling approach SPRoBERTa. We first present an unsupervised protein tokenizer to learn protein representations with local fragment pattern. Then, a novel framework for deep pre-training model is introduced to learn protein embeddings. After pre-training, our method can be easily fine-tuned for different protein tasks, including amino acid-level prediction task (e.g. secondary structure prediction), amino acid pair-level prediction task (e.g. contact prediction) and also protein-level prediction task (remote homology prediction, protein function prediction). Experiments show that our approach achieves significant improvements in all tasks and outperforms the previous methods. We also provide detailed ablation studies and analysis for our protein tokenizer and training framework.

The large amounts of available protein structures emerges the need for computational methods for protein function prediction. Predicting protein function is mainly based on finding similarities between proteins with unknown function with already annotated proteins. This may be achieved using different protein characteristics: sequences, interactions, localization, structure and or psychochemical. A lot of review papers mainly focus on sequence and psychochemical features- based methods. This is because sequence and psychochemical data are easy to deal with and to interpret the results, and much available compared to protein structures. However, structure-based computational methods provide additional accuracy and reliability of protein function prediction. Therefore, unlike many review papers, this paper presents an up-to-date review on the structure-based protein function prediction. The aim was to provide a recent and comprehensive review of protein structure related topics: function aspects, structural classification, databases, tools and methods.

Sequence-based prediction of one dimensional structural properties of proteins has been a long-standing subproblem of protein structure prediction. Recently, prediction accuracy has been significantly improved due to the rapid expansion of protein sequence and structure libraries and advances in deep learning techniques, such as residual convolutional networks (ResNets) and Long-Short-Term Memory Cells in Bidirectional Recurrent Neural Networks (LSTM-BRNNs). Here we leverage an ensemble of LSTM-BRNN and ResNet models, together with predicted residue-residue contact maps, to continue the push towards the attainable limit of prediction for 3- and 8-state secondary structure, backbone angles (θ, τ, ϕ and ψ), half-sphere exposure, contact numbers and solvent accessible surface area (ASA). The new method, named SPOT-1D, achieves similar, high performance on a large validation set and test set (≈1000 proteins in each set), suggesting robust performance for unseen data. For the large test set, it achieves 87% and 77% in 3- and 8-state secondary structure prediction and 0.82 and 0.86 in correlation coefficients between predicted and measured ASA and contact numbers, respectively. Comparison to current state-of-the-art techniques reveals substantial improvement in secondary structure and backbone angle prediction. In particular, 44% of 40-residue fragment structures constructed from predicted backbone Cα-based θ and τ angles are less than 6 Å root-mean-squared-distance from their native conformations, nearly 20% better than the next best. The method is expected to be useful for advancing protein structure and function prediction. SPOT-1D and its data is available at: http://sparks-lab.org/. Supplementary data are available at Bioinformatics online.

MotivationIn recent years, protein function prediction has broken through the bottleneck of sequence features, significantly improving prediction accuracy using high-precision protein structures predicted by AlphaFold2. While single-species protein function prediction methods have achieved remarkable success, multi-species approaches still face challenges such as difficulties in multi-source data integration and insufficient knowledge transfer between distantly-related species. How to integrate large-scale data and provide effective cross-species label propagation for species with sparse protein annotations remains a critical and unresolved challenge. To address this problem, we propose the MSNGO (Multi-species protein Structures and Network to predict GO terms) model, which integrates structural features and network propagation methods. Our validation shows that using structural features can significantly improve the accuracy of multi-species protein function prediction.ResultsWe employ graph representation learning techniques to extract amino acid representations from protein structure contact maps and train a structural model using a graph convolution pooling module to derive protein-level structural features. After incorporating the sequence features from ESM-2, we apply a network propagation algorithm to aggregate information and update node representations within a heterogeneous network. The results demonstrate that MSNGO outperforms previous multi-species protein function prediction methods that rely on sequence features and protein-protein networks.Availability and implementationhttps://github.com/blingbell/MSNGO.

Proteins are the main, building cell blocks, responsible for the existing cell biological processes. Therefore, precise knowledge of protein function is of great significance. There are a lot of methods which are used for protein comparison and for determining protein function. Some of them use structure alignment, others use sequence alignment, while some use protein descriptors. Here, we use two protein descriptors: Voxel and Ray-based descriptors to encode the structural and biological features of proteins. In biology there is a trend to hierarchically organize the things, like protein functions, cell components and the whole living world. There are a lot of classification systems which generate proteins in tree structure. However, due to the fact that it often happens that one protein has more than one parent, the Directed Acyclic Graph (DAG) hierarchy is used. Gene Ontology (GO) is a system for structural and hierarchical representation of proteins and gene products which support DAG hierarchy. CLUS, however, is a system which deals with hierarchical data. In this paper, we present a comparison between the two previously mentioned protein descriptors for predicting protein function. Firstly, protein descriptors are extracted from the structural coordinates found in the Protein Data Bank (PDB) and proteins backbone, appropriately. Afterwards, GO class hierarchy is added to each protein which has descriptor data. This created file is used as an input to the CLUS system. CLUS generates a decision tree model which is trained from the protein structure. The results from this system are the GO classes in which the protein belongs. The generated output shows that the predicting protein function with the Voxel protein descriptor gives better results instead of predicting protein function with the Ray protein descriptor.

Emergence of in-silico techniques have revolutionized the traditional methods of protein structure and function prediction of proteins. These computational tools encouraged the development of protein models in cumin (Cuminum cyminum L.) and have opened up new vistas in the area of research. In the present study, a high quality 3D structure and function of cumin RZ 19 protein have been predicted for the hypothetical amino acid sequence which showed homology with the protein domain of humans and E. coli illustrating that the database available on Apiaceae family is very low. The estimated molecular weight of identified cumin RZ 19 protein was 55028.6 and was predicted as an acidic protein with pI 5.14. Its functions are broadly classified into 2 categories: 1) Biological process: oxidationreduction and ion transport with a probability of 98.9% and 97.8% respectively and 2) Molecular function: structural constituent of ribosome with 99.0% and oxidoreductase activity with 98.7% probability. We have also identified a channel in cumin transmembrane protein, through which a ligand (an ion or a small molecule) might pass. The present finding may be a valuable addition to the proteomic information available on cumin. Further validation can be performed using wet lab experiments.

Inter-residue interactions play an essential role in driving protein folding, and analysis of these interactions increases our understanding of protein folding and stability and facilitates the development of tools for protein structure and function prediction. In this work, we systematically characterized the change of inter-residue interactions at various sequence separation cutoffs using two protein datasets. The first set included 100 diverse, nonredundant and high-resolution soluble protein structures, covering all four major structural classes, all-alpha, alpha/beta, alpha+beta, and all-beta; and the second set included 20 diverse, nonredundant and high-resolution membrane protein structures, representing 19 unique superfamilies. It was shown that the average number of inter-residue interactions in structures of both datasets displays the power-law behavior. Fitting parameters of the power-law function are directly related to the structural classes analyzed. These findings provided further insight into the distribution of short-, medium-, and long-range inter-residue interactions in both soluble and membrane proteins and could be used for protein structure prediction.

Protein function prediction plays an important role in the field of biology. With the wide application of deep learning in the field of bioinformatics, more and more natural language processing (NLP) technologies are applied to the downstream tasks in the field of bioinformatics, and it has also shown excellent performance in protein function prediction. Protein-protein interaction (PPI) networks and other biological attributes contain rich information critical for annotating protein functions. However, existing deep learning networks still suffer from overfitting and noise issues, resulting in low accuracy in protein function prediction. Consequently, developing efficient models for protein function prediction remains a popular and challenging topic in the application of NLP in bioinformatics. In this study, we propose a novel protein function prediction model based on attention mechanisms, termed IMAGO. This model employs the Transformer pre-training process, integrating multi-head attention mechanisms and regularization techniques, and optimizes the loss function to effectively reduce overfitting and noise issues during training. It generates more robust embeddings, ultimately improving the accuracy of protein function prediction. Experimental results on human and mouse datasets indicate that our model surpasses other protein function prediction models across multiple metrics. Thus, this efficient, stable, and accurate deep learning model holds significant promise for protein function prediction.

To compare the diagnostic efficiencies of deep learning single-modal and multi-modal for the classification of benign and malignant breast mass lesions. We retrospectively collected data from 203 patients (207 lesions, 101 benign and 106 malignant) with breast tumors who underwent breast magnetic resonance imaging (MRI) before surgery or biopsy between January 2014 and October 2020. Mass segmentation was performed based on the three dimensions-region of interest (3D-ROI) minimum bounding cube at the edge of the lesion. We established single-modal models based on a convolutional neural network (CNN) including T2WI and non-fs T1WI, the dynamic contrast-enhanced (DCE-MRI) first phase was pre-contrast T1WI (d1), and Phases 2, 4, and 6 were post-contrast T1WI (d2, d4, d6); and Multi-modal fusion models with a Sobel operator (four_mods:T2WI, non-fs-T1WI, d1, d2). Training set (n=145), validation set (n=22), and test set (n=40). Five-fold cross validation was performed. Accuracy, sensitivity, specificity, negative predictive value, positive predictive value, and area under the ROC curve (AUC) were used as evaluation indicators. Delong's test compared the diagnostic performance of the multi-modal and single-modal models. All models showed good performance, and the AUC values were all greater than 0.750. Among the single-modal models, T2WI, non-fs-T1WI, d1, and d2 had specificities of 77.1%, 77.2%, 80.2%, and 78.2%, respectively. d2 had the highest accuracy of 78.5% and showed the best diagnostic performance with an AUC of 0.827. The multi-modal model with the Sobel operator performed better than single-modal models, with an AUC of 0.887, sensitivity of 79.8%, specificity of 86.1%, and positive prediction value of 85.6%. Delong's test showed that the diagnostic performance of the multi-modal fusion models was higher than that of the six single-modal models (T2WI, non-fs-T1WI, d1, d2, d4, d6); the difference was statistically significant (p = 0.043, 0.017, 0.006, 0.017, 0.020, 0.004, all were greater than 0.05). Multi-modal fusion deep learning models with a Sobel operator had excellent diagnostic value in the classification of breast masses, and further increase the efficiency of diagnosis.