In silico protein function prediction: the rise of machine learning-based approaches

Traditional protein representation methods

During the preliminary phase of machine learning-driven protein function prediction, the prevailing constraints encompassing algorithmic limitations, data volume restrictions, and computational hardware barriers frequently necessitated a manual one-step feature extraction process. Instead of solely relying on direct sequence similarity or clustering methodologies, certain algorithms embraced the integration of proteins’ chemical properties as inputs for machine learning models. To exemplify, select studies incorporated amino acid composition, hydrophobicity, solvent-accessible surface area, and polarizability as input features. Then these inputs were combined through a support vector machine (SVM) classifier to solve the binary classification problem of DNA-binding protein or RNA-binding protein [28], [29], [30]. Correspondingly, akin methodologies have found application within studies aiming at enzyme family classification [24].

The recognition has gradually emerged that manual feature summarization alone is insufficient to comprehensively address the intricate challenges inherent in protein function prediction. Acknowledging the potential of machine learning in facilitating feature extraction, the incorporation of comprehensive sequence information of proteins assumes significance. Within this context, the most straightforward approach for encoding amino acid sequences (AAS) involves the sequential arrangement of amino acids, accompanied by the specification of amino acid types at respective positions. Previous studies have demonstrated that combining AAS with certain physical or chemical descriptors can yield informative protein representations. Research groups have developed several servers for computing these descriptors, such as PROFEAT, which calculates 6 feature groups composed of 10 features, including 51 descriptors and 1,447 values [31]. The features calculated by these servers include amino acid composition, dipeptide composition, normalized Moreau–Broto autocorrelation, Moran autocorrelation, Geary autocorrelation, number of sequence-order coupling, quasi-sequence descriptors, and distributions of various structural and chemical properties. This method for protein encoding and feature extraction has been widely used in a variety of downstream tasks related to protein function, such as predicting drug–protein interactions (DPI), anti-hypertensive peptides, and RNA–protein interactions [32], [33], [34].

In order to enhance the feature extraction of protein sequences, various protein encoding methods have been proposed. To better facilitate amino acid alignment and incorporate evolutionary information, substitution matrix representation (SMR) was developed [35]. It calculates the probability that amino acid at each position mutates into another type of amino acid and represents any given protein sequence with length N as an N × 20 substitution matrix, where the sequential similarity depends on the divergence time and substitution rate in the matrix. This approach is often applied to the prediction of interactions between proteins and biomolecules. For example, some studies added discrete cosine transform (DCT) on the basis of SMR for protein interaction prediction in various species, and the average accuracy is up to 96.28, 96.30, and 86.74 % for different species, respectively, which is significantly better than previous methods [36]. Meanwhile, this approach has also been applied to predict drug–protein interaction. For example, a study from Huang et al. encoded protein sequences with SMR descriptors, which achieved more than 80.00 % accuracy on multiple benchmark datasets for the prediction of drug–protein interaction [37].

In order to emphasize the specificity of different positions on the protein sequence, the position-specific scoring matrix (PSSM) method was proposed. This method utilizes PSI-BLAST (Position-specific Iterative BLAST) to calculate the percentage of different residues at each position [38], employing sequence alignment extract evolutionarily relevant feature information. It was applied in the PPlevo algorithm for predicting PPIs [39]. Furthermore, PSSM encoding method has been combined with various classifiers to predict protein function classification in yeast [40]. PSSM can also be integrated with other methods such as orthogonal local preservation projection (OLPP) to encode protein as a feature vector of fixed length and then combined with a RoF classifier to identify non-interacting and interacting protein pairs, with an accuracy of more than 90.00 % in yeast [41]. Autocovariance based on PSSM is another effective sequence-based protein representation method. This method extracts features from PSSMs by considering proximity effects, enabling the highlighting of some specific patterns in the whole sequence, which is also widely employed in protein classification tasks [42], [43], [44], [45], [46]. Following the principles in PSSM, SPRINT (Protein interaction Score) and PIPE (Protein Interaction Prediction Engine) determine the interaction between protein pairs by searching for similar pairwise regions among known protein complexes [47, 48].

The Conjoint Triad Feature (CTF) encapsulates not only the attributes of the target amino acid but also those of its neighboring amino acids. By treating any three consecutive amino acids as an entity, it extracts the intrinsic characteristics of a protein. Consequently, it possesses the capability to encompass both the protein sequence’s compositional information and the interconnected relationships among adjacent amino acids. The application of CTF extends across various domains, encompassing the prediction of PPI, RNA–protein interactions, and enzyme function [49, 50]. For example, Dey et al. used CTF protein representation methods combined with supervised machine learning methods (SVM, KNN, NB) to predict the interaction between the DENV virus and human proteins, as well as further predict the GO and KEGG pathway [51]. Wang et al. combined CTF and chaos game representation (CGR) with a random forest model to predict RNA–protein interactions [52]. Another study developed an SVM-based method to predict Enzyme Commission (EC), which used CTF to represent a given protein sequence [53].

Another notable approach is the multi-scale local descriptor (MLD), which partitions the protein sequence into segments of varying lengths to capture multi-scale local insights. These methodologies have showcased pronounced efficacy in the encoding of protein sequences and have found widespread application across diverse domains connected to protein function prediction [54, 55]. Despite their divergence in processing techniques for protein sequences, these methodologies collectively share a common hallmark: the integration of essential empirical features and the application of artificial feature extraction processes based on AAS. While these methods demonstrate superior performance compared to the mere encoding of protein sequences and amino acid types, the integration of their respective strengths becomes intricate when confronted with the intricate downstream task of protein function prediction.

NLP-based protein representation methods

The original object of NLP is human language, which shares analogous data structures with protein sequences [56]. Both utilize discrete units to construct structures endowed with specific attributes, ultimately express specific semantics or functions from this specific coding method. Experimental and computational biology have provided a large amount of protein-related data. In recent years, drawing inspiration from the paradigms of NLP, pre-training models tailored for protein encoding have emerged. In 2015, Asgari et al. introduced word2vec to the realm of biomolecules, pioneering the protein representation method provec [57]. This representation method focused on the first-order and second-order information in the protein sequences, generated vectors in protein space, and extracted corresponding protein properties in the embedding space. Combined with the SVM classifier, it was used for the classification of protein families.

Over the past two years, several sequence-based protein pre-training models have emerged. For example, Elnaggar et al. used six mature models in the field of NLP, such as Transformer-XL, XLNet, BERT, Albert, Electra, and T5, to train 393 billion amino acids in UniRef [58]. They tried to capture the biophysical features of protein sequences and verified the advantages of these embedded features on downstream tasks such as protein secondary structure prediction and protein subcellular localization prediction [59]. Brandes et al. presented ProteinBERT, which is a deep language model specifically designed for proteins [60]. The framework used in ProteinBERT is smaller and faster to train, and it achieves near-state-of-the-art performance across multiple benchmarks covering a variety of protein properties including protein structure, post-translational modifications, and biophysical properties. Roshan et al. trained millions of protein sequences using a self-supervised protein language model [61], which showed excellent generalization capabilities with parametric efficiency far higher than previous protein language models. It is worth mentioning that the basic architecture of ESM-1b is transformer, which is a common model in the field of NLP.

Protein pre-training models based on sequences or MSA have shown great potential, which indicates the rationality of applying the pre-training model in NLP to the field of bioinformatics. Protein structure information is also one of the determinants of protein function, while a large number of structural protein information has not been well utilized in protein pre-training models. In this case, researchers tried to add structural information to the pre-training model to obtain richer protein embedding information. For example, Gligorijević et al. proposed DeepFRI to predict protein function by extracting features from both sequences and structures [62]. DeepFRI utilized the LSTM-LM architecture combined with a large number of available sequences and 3D structural data in the form of contact maps. And the result of protein function prediction based on DeepFRI outperformed sequence-based methods on several tasks.

In addition to introducing 3D information in the form of contact maps, GearNet attempted to encode structure information by directly introducing geometric 3D representation [63]. GearNet leveraged AlphaFold2 predicted protein structure for pre-training through self-supervised contrastive learning and outperformed the previous baselines with its acquired structural embeddings on some metrics in the prediction of EC number and GO. Recently, an energy-based protein pre-training model was proposed and applied to two downstream tasks: Protein structure quality assessment (QA) and PPI assessment [64].

In summary, empirically-driven protein feature extraction methodologies continue to maintain a significant foothold. And to deal with diverse task, a variety of well-crafted designs have emerged, which include sequential adjacency and evolutionary relationships among sequences. On the other hand, in recent years, protein encoding methods based on pre-training models are gradually showing their advantages. When confronted with vast amounts of data and complex features, pre-training models have robust capabilities for feature integration. Efficient protein representation and feature extraction is the core of protein function prediction. Within this framework, we have introduced a variety of protein encoding methods. These encoding methods need to be combined with various classifiers, including traditional machine learning classifiers and deep neural network classifiers, for specific downstream tasks.

Prediction of protein–protein interaction

The process of protein interaction involves the binding of two or more proteins, which plays a pivotal role in numerous biochemical processes. For example, some signaling molecules transmit extracellular signals into the cell through PPI, which is the basis of many biochemical functions [65]. Another example is that proteins can form complexes through long-term interaction and participate in important biological processes such as transport. Moreover, some transient interactions can add modifications to proteins and regulate their function. Therefore, PPI is the core of cell biochemical reactions, and studies about PPI can enhance the comprehension of the mechanisms behind the disease.

The dataset pertaining to PPIs originates from two primary sources. Firstly, a portion of this data is collected from complexes from the Protein Data Bank (PDB), affording atomic-level insights. Secondly, another segment emerges from PPIs elucidated through high-throughput methodologies, including yeast two-hybrid assays, immunoprecipitation, mass spectrometry-based protein complex identification, and affinity purification. The amalgamation of data from these diverse origins has been curated within the publicly accessible Protein Interaction Database (DIP), encompassing protein interaction records spanning diverse organisms ranging from yeast to humans. This reservoir of data constitutes a valuable resource, furnishing ample material for the application of machine learning techniques in the investigation of PPIs.

Traditional molecular docking algorithms can predict the binding conformations of protein complexes, which are effective approaches to study PPIs [66]. These molecular docking algorithms mainly consist of two steps. The first step involves employing a spatial search algorithm such as FFT algorithm to search the spatial conformation of two proteins bound to each other. The second step is to evaluate the affintity of protein binding through scoring function. These traditional molecular docking algorithms often require the spatial conformational coordinates of the protein as input and the three-dimensional space lattice. The traditional molecular docking algorithm has the advantage of being able to obtain multiple candidates binding conformations for any two protein pairs [67, 68]. These algorithms also exhibit certain limitations. First, the prediction of PPIs using molecular docking algorithms heavily relies on the spatial conformation of proteins, which is hampered by the fact that the number of proteins with experimentally-resolved structures is far less than that of protein sequences. Secondly, the interaction prediction of molecular docking algorithm also depends on the scoring function, which is based on experience, physical and chemical laws [69, 70]. The scoring function itself still has considerable potential for improvement. The inputs of molecular docking algorithms are often the rigid conformations of individual proteins, which may undergo flexible backbone changing during the process of interaction. Therefore, to optimize the docking poses, some molecular docking algorithms have to allocate more computational time by introducing local molecular dynamics simulations or flexible conformational libraries [71, 72]. How to deal with flexible docking remains an unresolved issue in this field. Molecular docking algorithms necessitate performing conformational searches for each input protein pair, and occasionally even require conducting conformational modeling from protein sequences. Consequently, the execution of molecular docking algorithms on a large-scale screening basis could potentially be hampered by computational time constraints.

In recent times, machine learning-based methodologies have in part addressed the limitations inherent in traditional approaches concerning the prediction of PPIs. Noteworthy studies in recent years using machine learning-based methods to predict PPI are listed in Table 1. First, machine learning-based prediction methods can directly treat the target of the task as binary classification, which means that the input protein representation could be more flexible. Beyond characterizing protein sequences and conformations, the integration of effective feature extraction founded on physicochemical priori knowledge can be incorporated into the model. Since 2001, computational methods have been employed in attempts to predict PPIs [73], [74], [75]. Until recent years, various protein representation methods have been applied to this objective. For example, Carlos et al. applied six different new features to represent proteins [76]. Sun et al. applied the Autocovariance method with Stacked autoencoder (SAE) to study sequence-based human PPI predictions [77]. Bryant et al. used multiple sequence alignment (MSA) as input [78]. Beyond augmenting flexibility in protein representation and prediction targets, machine learning-grounded PPI prediction algorithms pivot around a data-driven paradigm rather than relying on prior knowledge. Therefore, the use of extensive PPI datasets significantly enhances the precision of machine learning-based PPI prediction. For example, Hanggara et al. obtained a large number of PPI datasets based on string-DB, and the validation accuracy was close to 90 % [79]. Machine learning-based prediction methods have also contributed to the exploration of fundamental principles underlying PPI. Methods have been devised to discern akin targets within protein interaction networks. For instance, Zhou et al. employed a PPIs network between SARS-CoV-2 and human, constructed through high-throughput yeast experiments and mass spectrometry, to unveil 361 new host factors, including proteins devoid of specific experimental structures such as BAG3—an entity implicated in diverse diseases like heart disease and cancer [80]. Kovács et al. delved into the role of BAG3 in bacterial infections through the lens of a PPI network. These network-based prediction methods even have the potential to challenge the conventional wisdom that interacting proteins are not necessarily similar, and that similar proteins do not necessarily interact with each other [81].

AlphaFold2 greatly facilitated protein structure prediction [2, 3], making it feasible to achieve structure information from mere protein sequences. Since protein spatial structure information is difficult to extract directly from the embedding of protein sequences, integrating the predicted protein spatial structure in the PPI prediction process can effectively improve the accuracy in the post-AlphaFold era. For example, TAGPPI relied solely on sequences and performs PPI prediction end-to-end, without additional input of protein 3D structure [87]. TAGPPI used AlphaFold2 in the algorithm to construct the residue contact map of the protein, which contained precise spatial structure information, thus effectively improving the prediction ability of PPI. For the protein complex prediction task, the DeepMind team retrained AlphaFold-multimer for the protein complexes [88]. They linked multiple proteins into single chains with cross-chain positional encodings as input to AlphaFold2. This approach demonstrated a great improvement in heterologous complex structure prediction. Bryant et al. employed AlphaFold2 to incorporate species-specific multiple sequence alignment (MSA), thereby enhancing the precision of protein complex prediction [78]. AF2Complex enables the structural inference of a polymeric protein complex using a single protein sequence without necessitating retraining of AlphaFold2. In contrast to other approaches, this method integrates MSA regions from diverse proteins by means of sequence alignment. Leveraging these sequence and template features, the AF2Complex model generates a comprehensive complex model by iteratively computing its interface score S for ranking the confidence level [89].

In summary, PPI prediction is a crucial research direction for protein function prediction, with significant implications for comprehending protein interaction network and identifying disease targets. Treating PPI prediction as a classification task or directly predicting protein complex binding conformations are both meaningful prediction targets. With the accumulation of protein data, data-driven machine learning prediction algorithms are playing an increasingly important role in PPI prediction. How to extract features and integrate information more effectively remains a problem.

Prediction of drug–protein interaction

The development of algorithms for predicting small molecule-protein interactions is crucial not only for drug screening, but also for identifying potential drug targets. Additionally, these algorithms can be utilized to predict the interactions between endogenous metabolic molecules and proteins, including sugars, bioactive peptides, endogenous regulatory factors, signaling molecules, etc., thereby shedding light on cellular regulatory mechanisms. Similar to the encoding of macromolecular proteins, representation of small molecular compounds has experienced a paradigm shift from traditional molecular descriptors to machine learning training for embedding.

As shown in Figure 3, we will introduce the following forms of molecular representation: one-dimensional linear inputs (such as SMILES or selfies, inches), structural or path-based fingerprints, and two-dimensional graphical structures (atoms and bonds) that involving topological information.

Figure 3:

Small molecule and protein representation based on machine learning pre-training models. LSTM, long short-term memory.

The characterization of small molecules through string representation is widely employed in scientific research. Molecular structures can be translated into machine-readable string representations that are more suitable for NLP, among which, SMILES is typical molecular string representation. Several deep generative models were developed to learning the distribution of SMILES representation [90], [91], [92]. It is worth noting that the SMILES string is non-unique, often leading to multiple encoded representations for a single molecule. In this context, certain deep generative models have proposed enhancements to the traditional SMILES format. An illustrative example is SELFIES, which serves as an advanced alternative to SMILES. Particularly in the context of Pangu-based models, SELFIES is preferred over SMILES as input. This preference stems from findings that molecules generated using SELFIES exhibit an efficacy level of up to 100 %.

The internal topological structure of the small molecule naturally allows the molecule to be represented as a two-dimensional graph. The atoms of a molecule are mapped to nodes of a graph containing information such as atomic type, chirality, etc. The edges are linked when there exists a covalent bond between two atoms, and the edge attributes include the types of chemical bonds. Such a graph structure is often represented as inputs of GNN. Combined with deep neural networks such as transformer, the topological structure of molecules can be better extracted [93], [94], [95], [96].

Recent molecular characterization methods based on deep learning pre-training models aim to add molecular structure, molecular properties and other information into the training process to generate efficient embedding.

Self-supervised frameworks are frequently employed in small molecule pre-training. Given the substantial data requirements of pre-trained models, leveraging contrastive learning for data augmentation represents an effective strategy. After data augmentation, the consistency between similar inputs is maximized in the feature space and the differences between different classes of data are enlarged. At present, the prevailing approach to enhance small molecules is to randomly mask the atoms, chemical bonds, and subgraphs of molecules. MolCLR, for example, constructed molecular maps using extensive unlabeled data and developed graph neural network encoders to learn molecular properties, which performed impressively in the benchmark test [97]. iMolCLR reduced negative pairs between similar molecules [98]. In addition to directly comparing the degree of similarity between molecules, ATMOL compared the molecular map with masked attention matrices generated by graph attention networks (GAT), which improved the performance on downstream tasks. There are also been some studies trying to combine 3D information of molecules with generative models. GraphMVP, for example, used 2D topologies and 3D geometric views for sub-supervised learning. GraphMVP used accurate 3D molecular conformations from the GEOM dataset to do more discriminating 3D geometric enhancements than the 2D molecular map encoder [99]. The success of denoising in image generation has led to its application in molecular characterization tasks. A recent work utilized denoising-based autoencoders to learn molecular force fields for pre-training, and improved molecular property prediction performance on multiple benchmark datasets [100].

Most small molecular drugs exert their efficacy by interacting with their target proteins, such as enzymes, ion channels, and G-protein-coupled receptors. Therefore, identifying DPI is an important prerequisite for drug discovery, pharmacology, drug side effects, and other studies [101]. Biochemical assays for experimentally undiscovered DPI are costly and time-consuming. In the face of a large number of potential unpaired small molecule compounds and drug target proteins, large-scale virtual screening by computational methods can provide a very valuable reference and guidance for experimental verification.

Similar to the prediction of PPI, there are three main methods to predict DPI, the first of which is based on molecular docking. Conformation search and molecular dynamics simulation are combined to reconstruct the contact relationship between small drug molecules and proteins in 3D space, with the goal to find the best binding pose. The disadvantage of molecular docking-based methods is that they require accurate protein structure as input and are time-consuming [102], [103], [104]. The second approach is to predict interactions based on drug–protein association networks [105, 106]. The underlying principle here is that proteins sharing similar structures and exhibiting close relationships are more likely to interact with the same drug. This methodology typically involves the establishment of a network encompassing existing drugs and proteins, followed by the computation of similarity scores for both drug pairs and protein pairs. However, given the absence of a standardized protein similarity score, a drawback of this approach pertains to the accuracy of similarity scoring, particularly when dealing with rare or novel proteins. It is imperative to acknowledge that network-based methodologies hinge upon assumptions that may not universally hold true, as not all similar proteins necessarily interact with similar drugs. The third approach involves a data-driven method rooted in learning, which operates independently of a priori assumptions but demands substantial data quantity and quality to be effective [107]. Several databases, such as PubChem, ChEMBL, DrugBank, and DUD-E contain a large amount of information on the interaction of ligand molecules with target proteins. The PDB database also contains a large amount of structural data. The collective information within these databases underpins the utilization of machine learning techniques for the prediction of DPI. In this context, machine learning algorithms have been widely used in the field of computer-aided drug design (CADD) to predict DPI.

We have already reviewed the methods of small molecule characterization commonly used in machine learning. Small molecule compounds can be naturally described in computer-readable formats, such as strings, graphs, etc. The Simplified Molecular Input Line Entry System (SMILES) is the most widely used string format. It is worth noting that both small molecules and proteins are essentially composed of atoms and chemical bonds, and are therefore very easy to represent in the form of a connection graph. Representing atoms as nodes and chemical bonds as edges, GNN are natural small molecule machine learning network frameworks. Multiple variants of graph networks achieved state of the art (SOTA) performance in multiple machine learning domains, such as graph convolutional networks (GCNs), graph attention networks (GATs), and graph isomorphic networks (GINs). These network architectures can be efficiently employed for various downstream tasks concerning small molecules. Furthermore, the utilization of pre-training models specifically designed for small molecules to achieve effective embeddings is an emerging concept. This representation approach, rooted in pre-training, is relatively novel and currently lacks comprehensive evaluation on downstream tasks, particularly in the context of DPI.

Typical studies in recent years using machine learning-based methods to predict DPI are listed in Table 2. The prediction of DPI is quite different from the prediction of PPI. DPI prediction requires not only integrating drug molecule and protein databases but also embedding the chemical space and protein space into a unified space. In this case, the deep neural network framework is gradually showing more advantages over traditional machine learning methods. A straightforward idea for embedding small molecules and proteins into the same hidden space is to use simple concatenation to combine protein and small molecule representations. However, this approach has limitations as simple concatenation does not comprehensively capture the intricate interactions between the two compound types. Recent efforts have aimed at achieving more robust information interactions that effectively capture the nuanced connections between small molecules and proteins. For instance, the Perceiver CPI model integrated a cross-attention module to compel the model to discern the impact of compound information on protein information [108]. Other investigations have focused on predicting interactions between previously uncharacterized proteins and unknown small molecules. Yuel, for instance, prominently incorporated a characterized FC layer alongside an attention-based affinity prediction module that employed the outer product to combine protein and small molecule features [109]. Additionally, a noteworthy example emphasizing the prediction of DPI involving new proteins is AttentionSiteDTI, drawing inspiration from sentence classification models [110]. Here, the drug target complex was likened to a sentence with meaningful connections between its biochemical entity (referred to as the protein pocket) and the drug molecule. The authors highlighted that unlike previous studies, this model demonstrated exceptional performance when applied to novel proteins. Furthermore, to address potential information loss when characterizing molecular graphs through graph convolutional networks, SSGraphCPI introduces a comprehensive approach by incorporating both 1D SMILES representations and 2D molecular graphs, thereby incorporating sequence and structural features [111]. In contrast to many other methods for predicting DPI, which primarily emphasize molecular representation, STAMP-DPI places a stronger focus on protein representation and the higher-level relationships between distinct instances [112]. STAMP-DPI employs TAPE encoding to integrate contact maps and GNN as a protein representation and establishes GalaxyDB, an esteemed benchmark dataset specifically designed for DPI prediction. HyperAttentionDTI highlights the incorporation of intricate non-covalent interactions between atoms and amino acids by employing an attention mechanism that assigns an attention vector to each atom and amino acid [113]. HyperAttentionDTI has exhibited noteworthy performance improvements on benchmark datasets. Notably, BridgeDPI merges network-based and learning-based concepts [114]. It constructs a drug–protein association network by introducing a class of virtual nodes designed to bridge the gap between drugs and proteins. Furthermore, it leverages drug molecules and protein sequences as prior knowledge to generate features for interaction prediction.

In the last two years, with the development of protein and small molecule databases, a few researchers tried to build pre-training models of proteins and small molecules for DPI prediction [119, 120]. One such instance is Uni-Mol, a 3D molecular pre-training model. Differing from its counterparts, Uni-Mol directly employs the 3D molecular structure as input to the model, thereby eschewing the use of 1D sequence or 2D graph structure representations. Uni-Mol draws upon its own expansive dataset encompassing 3D structural information of organic small molecules and protein pockets. This model was trained via a unified pre-training framework and strategic tasks on a large-scale distributed cluster. The utilization of 3D information in representation learning empowered Uni-Mol to yield remarkable performance across multiple downstream tasks, while simultaneously facilitating 3D conformation-related endeavors like molecular conformation prediction and protein–ligand binding conformation prediction [112]. To further explore a more comprehensive representation of molecules and proteins, STAMP-DPI employed a pre-training approach to encode the semantic information of small molecules and proteins within an end-to-end deep learning architecture [112]. Protein representation was accomplished via a hybridization of structural topology mapping and Tape Embedding pretraining features, while drug molecules were concurrently represented using molecular mapping and Mol2vec Embedding pretraining features. Leveraging an attention mechanism, STAMP-DPI captured the intricate interaction information between molecules and proteins, ultimately realizing the prediction of DPI.

In recent years, the proliferation of research on DPI prediction has witnessed an escalation in methodological intricacy, paralleled by an augmentation in predictive accuracy. This trend has prompted a thoroughgoing evaluation of this domain. For instance, they highlighted that excessive similarity among samples in the validation set could lead to inflated accuracy levels, while the spurious negative samples could compromise the model’s generalizability. In addition to the network architecture, equal emphasis should be placed on the composition of the training dataset. This is especially significant for big data-driven models which rely heavily on data quality.

Prediction of proteins with specific properties

There is also a need to predict specific functional proteins in certain research scenarios [122], [123], [124]. The prediction of proteins with specific properties relies on specifically collected datasets, which also has considerable value in application.

There are proteins such as transcription factors or RNA-binding proteins that function by binding to DNA or RNA [125]. Recognizing these kinds of proteins is of great significance for understanding transcriptional and translational regulation. Therefore, studies have been devoted to predicting DNA-binding proteins based on machine learning [126], [127], [128], [129], [130]. A number of methods using deep multi-task architectures to predict protein and DNA or RNA binding were published in 2022. DeepDISOBind implemented intrinsically disordered residues (IDR) that predicted the interaction between proteins and DNA as well as RNA. It used common input layers that are followed by different layers that distinguish between DNA and RNA interactions [131]. The classifier architecture mainly consisted of CNN and FNN. Using PSSM, HMM, DSSP and AlphaFold2 predicted structures to jointly construct amino acid features, GraphSite not only improved the performance of predicting protein binding to nucleic acids, but also had the potential to identify binding sites [132]. In predicting proteins of particular functional categories, the utilization of protein structure prediction tools, exemplified by AlphaFold2, offers significant advantages. Huang et al., for instance, implemented a high-throughput protein clustering approach relying on tertiary structural information. They harnessed structural clustering of proteins to identify deaminase functionality. This method facilitated the identification of a deaminase protein strongly amenable to editing in soybean plants, an achievement unattainable through cytosine base editing (CBE) alone. Moreover, their efforts yielded a suite of novel base editing tools endowed with autonomous intellectual property right [133].

There were also studies that attempted to summarize the properties of drug target proteins and identified drug target proteins based on machine learning methods (Table 4). Most of the studies utilized traditional protein representation methods, while in recent years, NLP-based protein representation methods have also been used [134, 135]. Sun et al. evaluated the performance of various combinations of machine learning algorithms for predicting druggable proteins, utilizing Word2Vec to characterize protein sequences and showcasing its potential in this regard [135]. Chen et al. integrated ESM1b, a sequence-based self-supervised pre-trained protein language model, with a graph convolutional neural network classifier to develop an enhanced sequence-based identification method for drug target proteins. The comprehensive model, named QuoteTarget, successfully identified 1,213 potential untapped drug targets when applied to all Homo sapiens proteins. Additionally, the authors employed the gradient-weighted class-activation Mapping (Grad-Cam) algorithm to infer residual binding weights from well-trained networks [136]. In terms of classification algorithms, most drug–target protein prediction algorithms used traditional machine learning algorithms or simple neural networks. And the highest accuracy in these studies was above 90 %. It is worth noting that fewer studies used deep neural network framework in this objective, probably due to the fact that there are not many drug target protein datasets available for training.

Liquid-liquid phase separation is a key principle of intracellular organization in biological systems and has been implicated in a variety of biological processes as well as a range of neurodegenerative diseases. In recent years, there have been many in-depth studies on the LLPS phenomenon of biomolecules [149, 150]. Liquid condensates formed by LLPS are generally thought to be the result of multivalent weak interactions of multiple interacting moieties in multiple folded regions or intrinsically disordered regions (IDRs) [151], [152], [153]. Because of this special property, many traditional protein-coding methods may no longer be suitable. The PLAAC tool is web to retrieve protein sequence domains and extract pertinent information encompassing prion-like amino acid compositions [154]. CatGRANULE is an algorithm for a single species [155]. Based on the previously published phase separate protein database PhaSepDB, Chen et al. divided phase separate proteins into two sets of spontaneous phase separate proteins (hSaPS) and interaction dependent phase separate proteins (hPdPS). The distribution of the two phase isolated protein sets and the background protein sets were significantly different from each other by comparing the multimodal characteristics [156]. Chu et al. has developed PSPredictor, a sequence-based protein prediction tool for liquid-liquid phase separation (LLPS), which integrates compositional and sequence information during the protein embedding stage and employs a machine learning algorithm to yield accurate predictions [157].