3.1. Sequence Feature-Based Method

Sequence features are commonly used in localization prediction since some differences in the sequence features are empirically known to be correlated with different localization sites. Nishikawa and Ooi ^[18] [18] first noted the correlation of amino acid composition to its biological and functional character in 1982. After that in 1983 Nakashima developed the first sequence-based method for subcellular localization [19]. They used amino acid composition to discriminate between intracellular and extracellular proteins. Later several research groups successfully used amino acid composition as a tool for subcellular localization predictions ^[16][20][21][16,20,21].

In sequence feature-based methods, the complete sequence of proteins is transformed into a numerical feature vector, which is then used to predict the subcellular location. There are different types of sequence feature-based methods available: (i) amino acid composition based method, in which the frequency of 20 different amino acids is calculated but it ignores the sequence order information of each residue. (ii) Chou’s Pseudo amino acid composition (PseAAC) [22], which considers the amino acid composition along with the potential interaction among the adjacent residues. This can be further categorized into different modes of PseAAC such as the gene ontology mode, functional domain mode and sequential evolution mode. (iii) Hybrid method, which allows the integration of different parameters or features for the prediction and usually results in an increased the prediction performance [4].

3.2. Homology Based Method

This is the most common way to predict the uncharacterized protein on the basis of the presence of homologous sequences of known function with an assumption that function is evolutionarily conserved [23]. This approach first identifies for homologous sequences in the proteins with known subcellular location and then extrapolates to predict the location of unknown proteins, hence this approach is also known as “Annotation by Homology Transfer”. Homology is a qualitative term, which attributes evolutionary relationships among different protein sequences. Orthologous proteins also typically have similar sequences and thus similar subcellular localization patterns. Proteins with a highly similar sequence correlate well with the cellular localization site while those with dissimilar sequences indicate that they are distant and may or may not be colocalized. In 2002, Nair and Rost ^[24] [24] showed the correlation between sequence similarity with subcellular localization. They considered 11 different compartments and observed sequence conservation among the major compartments. BLAST, PSI-BLAST, and hidden Markov models (HMM) are routinely used for searching homologous sequences. The limitation of homology-based methods is more pronounced in cases where no homology is found between the query sequence and the annotated proteins sequence. Additionally, it is known that a single amino acid substitution in localization signals can change the localization of a protein ^[25][26][27][25,26,27]. Thus, sequence homology is a noncausal feature for the localization prediction and should be used with caution when applied to nonnative sequences or in case when homology is less [28].

3.3. Functional Motifs, Domains, and other Signatures Based Method

Proteins have evolved in different compartments, which limit their interactions with other proteins and ultimately impact their functions. Some of these proteins preserved some sequential or structural patterns or motifs. Though not all of these motifs and domains are specific to subcellular localization, many preferentially occur in some specific compartments and such domains can be used to predict the localization of any proteins. Studying proteins at a domain/motif level allows more accurate functional inference [29]. In 2002, Mott et al. ^[30] [30] first used 300 Simple Modular Architecture Research Tool (SMART) domains to predict three subcellular locations viz secreted, cytoplasm, and nucleus. After that, several works have been used for the protein motif and domains as features for protein localization predictions ^[31][32][31,32].

These motifs are not just limited to sequence patterns, but also extended to the structural information. There are a couple of tools such as PROSITE ^[33] [33] and MEME [34] that employed this feature to use for protein localization. While the structure is not available for a big chunk of protein sequences, this gap is filled by several proteins structure predictions servers, like I-TASSER and C-I-TASSER servers ^[35].[35]

3.4. Signal Peptide Based Method

Signal peptides are short amino acid sequences in the amino terminus of the newly synthesized proteins and are found in all organisms including bacteria, archaea, and eukaryotes. The function of the signal peptide is to enable the transport machinery to translocate the proteins to different subcellular locations. They are present in secretory proteins and in transmembrane proteins and the protein residing in different eukaryotic organelles have different types of signal peptide sequences [36]. The signal peptide is followed by a stretch of amino acids that form the cleavage site recognized by peptidases and the signal peptide is removed after translocation, except in the case of transmembrane proteins. In case of transmembrane proteins, this signal peptide serves as signal anchor sequences. The importance of various signal peptide sequences in proteins in their subcellular localization has led to attempts to predict the subcellular location on the basis of the signal peptide present in proteins. The prediction of the signal peptide involves two main tasks: (1) discriminating between the signal peptide and signal anchor sequences and (2) also predicting the position of the signal peptide cleavage site [37]. The major challenge in signal peptide prediction is discriminating between true signal sequences and other hydrophobic regions. In addition to it, the accurate prediction of the cleavage site is also very important due to the high variability of the signal sequence length and the absence of sequence motifs that unambiguously mark the position of the cutting site ^[37][38]. A number of prediction methods are available that recognize and predict the subcellular location on the basis of signal peptides (Table 1). SignalP was the first publicly available method ^[15] [15] and there are many versions available, which were developed based on different methods. Version-1 [39] was based on artificial neural networks, while version-2 [40] was based on hidden Markow models, version-3 ^[41] [41] has an improved cleavage site prediction, version-4 [42] has improved discrimination of signal peptides and TM helices, and version-5 ^[38] [38] is a deep neural network-based method combined with a conditional random field classification and an optimized transfer learning for improved signal peptide prediction.

The signal peptide-based method is a good approach to predict the proteins that contain the signal peptide, but it has some drawbacks, which make these methods not able to be applied for proteome scale prediction. (i) Not all proteins contain signal peptides. There are many proteins that do not have any reported signal peptide sequence and despite this are still translocated to their respective subcellular location. (ii) Many proteins follow the “piggyback import” mechanism during protein translocation, which means these proteins do not have any specific signal peptide for the localization, but they interact and bind to different proteins that have a signal peptide for translocation and then are co-imported to specific target locations ^[43][44][43,44].

3.5. Non-Sequence Derived Features

A variety of non-sequence derived features have been used to predict subcellular localization. For example, LOC3D [45], which used the structural information for identification and prediction of proteins subcellular locations. There are a number of non-sequence derived features that have been used in an automated classifiers including immunohistochemistry ^[3][46][47][3,46,47], fluorescence microscopy image ^[48][49][48,49], protein–protein interaction (PPI) data [50], expression data [51], and recommendation systems [52].

3.6. Integrated Method

The different strategies for predicting protein localization have their own strengths and weaknesses. To enhance the performance of prediction methods, it is important to combine multi-characteristic strategies, which give more complete information to understand the relationship between protein localization with its sequence, structure, physicochemical properties, and function. Hence a combination of different input vectors and different tools will be the successful strategy in protein subcellular localization prediction. Many methods have successfully utilized the combination of protein features to enhance the performance of protein subcellular localization predictions. The Protein Subcellular Localization Prediction Tool (PSORT) family method is one of the integrated methods, which contains several tools for localization prediction. The family includes a number of tools: (i) PSORT [53], the first integrated method of the PSORT family (http://psort.org) for the plant and bacterial protein, (ii) PSORT II ^[53] [54] for yeast and animal proteins, (iii) iPSORT ^[55] [55] for N-terminal sorting signals for plant or non-plants; (iv) PSORTb ^[56][57][58] [56,57,58] for bacterial and archaeal proteins, and (v) WoLF PSORT ^[59] [59] for eukaryotic proteins including plants, animals, and fungi.

A similar approach was taken by many researchers where they integrated biological or empirical sequence features correlated with subcellular location with a variety of machine-learning algorithm such as KNN, SVM, and deep learning: MultiLoc, integration of the phylogenetic profile and GO terms of retrieved homologues such as MultiLoc2, CELLO2.5, SherLoc2, YLoc, iLoc-Euk, Loctree3, DeepLoc, etc. People are also integrating different computational tools for predicting subcellular localization. The Bologna Unified Subcellular Component Annotator (BUSCA) ^[60] [60] is an example of such an integrated tool where the author combines methods for identifying signal and transit peptides (DeepSig and TPpred3), GPI-anchoes (PredGPI), and transmembrane domains (ENSEMBLE3.0 and BetAware) with tools for discriminating subcellular localization of both globular and membrane proteins (BaCelLo, MemLoci, and SChloro). This integrated method performs better than the other methods based on single feature approaches. There are a number of recently developed subcellular localization methods available, which are used by a wide range of researchers (Table 2 and Table S1).