Elucidating the functional roles of prokaryotic proteins using big data and artificial intelligence

Abstract

Annotating protein sequences according to their biological functions is one of the key steps in understanding microbial diversity, metabolic potentials, and evolutionary histories. However, even in the best-studied prokaryotic genomes, not all proteins can be characterized by classical in vivo, in vitro, and/or in silico methods—a challenge rapidly growing alongside the advent of next-generation sequencing technologies and their enormous extension of ‘omics’ data in public databases. These so-called hypothetical proteins (HPs) represent a huge knowledge gap and hidden potential for biotechnological applications. Opportunities for leveraging the available ‘Big Data’ have recently proliferated with the use of artificial intelligence (AI). Here, we review the aims and methods of protein annotation and explain the different principles behind machine and deep learning algorithms including recent research examples, in order to assist both biologists wishing to apply AI tools in developing comprehensive genome annotations and computer scientists who want to contribute to this leading edge of biological research.

Different principles and methods to functionally characterize and annotate prokaryotic proteins are reviewed, with a special emphasis on machine learning and deep learning algorithms.

Introduction

Bacteria and archaea are the oldest, most abundant, and most diverse forms of life on Earth (Eme and Doolittle 2015, Louca et al. 2019). They dominate many functions of the biosphere and harbour a huge potential for biotechnological applications (Singh et al. 2020, Pfeifer et al. 2021). However, the task of fully characterizing microbial diversity is almost incomprehensibly vast. The ‘known unknowns’ (Logan 2009), i.e. the diversity we know is there, but which we have not characterized yet, have become increasingly apparent in the recent years, while the extent of ‘unknown unknowns’, i.e. the diversity which remains completely undiscovered, is still debated. With some approximates suggesting billions or even trillions (Locey and Lennon 2016, Larsen et al. 2017), another study estimated 0.8–1.6 million prokaryotic species (Louca et al. 2019) on Earth. Of those, only ∼2% have whole or at least partial genome sequences (Zhang et al. 2020), many with only the prokaryotic marker gene 16S rRNA known (Hugenholtz et al. 2021). Approximately, 70% belong to the so-called candidate phyla radiation (CPR), which—with very few exceptions—have no cultured representatives (Hug et al. 2016). At this point it is, therefore, it is safe to say that ∼99% of all microbial species from the environment remain uncharacterized. They are, therefore, referred to as Microbial Dark Matter (Bernard et al. 2018).

Strikingly, even todays’ best-studied microorganisms have not been fully functionally characterized yet. For instance, in the ‘favorite pet’ of microbiologists for over 100 years—the model organism Escherichia coli—the function of more than 30% of proteins has not been determined experimentally and more than 2% of protein-encoding genes have no characterization at all (Ghatak et al. 2019). These so-called hypothetical proteins (HPs), which have been referred to as ‘functional dark matter’ (Escudeiro et al. 2022), are found across all microbial species and represent an enormous gap in knowledge. In addition to their importance for the fundamental understanding of biology and evolution, these proteins might also provide novel solutions for medical treatments, bioremediation, or bioenergy production, to help solve 21st century challenges (Rehman et al. 2021, Arslan et al. 2022). However, recent analysis suggests that only 3% of the biochemical potential of bacterial genomes has been discovered (Gavriilidou et al. 2021). Genomes of taxa with no cultured representative, archaea and relatively large bacteria with a complex lifestyle, have a particularly high percentage of HPs (Makarova et al. 2011, Lobb et al. 2020) (Fig. 1). While some uncharacterized genes are conserved across species or genera, many are taxonomically restricted (Yu and Stoltzfus 2012, Tatarinova et al. 2016). Hence, since the majority of microorganisms, proteins, and products are still uncharacterized, their use for future applications is heavily constrained in comparison to what is possible (Kalkreuter et al. 2020).

Figure 1.

(A) Circular tree of archaeal and bacterial phyla, showing the percentage of uncultured genomes in the Genome Taxonomy Database (GTDB) (archaea = green, bacteria = blue) (Parks et al. 2022) and median percentage of HPs according to AnnoTree data (orange) (Mendler et al. 2019) with at least 10 (bacteria) or five genomes (archaea); phyla were renamed according to the List of Prokaryotic names with Standing in Nomenclature (LSPN) and visualized with GraPhlAn (Asnicar et al. 2015). (B) Numbers of characterized and uncharacterized (hypothetical) proteins in bacteria and archaea. Numbers after clustering to an identity of <50% are ‘unique’ proteins. Among unique proteins HPs include more sequence diversity than annotated proteins.

This biological problem can now be addressed with the help of AI-based tools, using the large quantities of biological data available in public databases from high-throughput experiments. A number of reviews on AI in relation to protein function prediction, genomics, or biology more generally, have recently been published (Greener et al. 2022, Whalen et al. 2022), however, all have focused on human proteins or eukaryotic model organisms (Bonetta and Valentino 2020, Mahood et al. 2020, Ofer et al. 2021). Our review particularly concentrates on the functions of proteins in bacteria and archaea, the data and algorithms available, and the difficult conceptual issues underlying the task of predicting protein function.

Functional analysis and categorization of proteins

Annotating protein-encoding genes is the process of assigning functional labels to protein products. This can be approached from different biological perspectives and the best approach to use is often simply taken for granted according to the norms of a biological subdiscipline. The concept of function is, however, multifaceted, and each aspect of function has become associated with particular experimental methods, whether in vivo, in vitro, in silico, or in combination. While different approaches can be complementary and integrated, they are distinct and need to be understood.

Biochemical and biophysical phenotypes

The classic microbiological account of function pertains to the phenotype of a gene. Here, researchers can either study what happens if the gene is taken away from or added to the genome, at the level of DNA or the level of protein expression. In addition, one can also study the properties of a protein in isolation. This is a subset of the causal role view, assessing the difference that the presence or absence of the protein makes to the operation of the cell, as measured for instance by population growth under different conditions, or in terms of biochemical reactions in which the protein is involved. In early molecular genetics research, gene knockouts were often achieved by transposon mutagenesis, leading to distinct phenotypes (Handfield and Levesque 1999) and multiple high throughput versions of this technique have now been developed (Cain et al. 2020). Other new technologies include CRISPR gene editing (Liu et al. 2020) and gene silencing with CRISPR interference (Zhang et al. 2021). Other traditional approaches for characterizing proteins include 2D electrophoresis for separating proteins by mass and isoelectric point, enzyme assays to determine catalytic abilities, and analysis of the 3D structure using X-rays or electron microscopy, which can help to clarify mechanisms of action (Aslam et al. 2017). These classical methods are often experimentally challenging and very time-consuming since they require bacterial cultures with sufficient biomass or means for overexpression of the protein of interest. This can be complicated or even impossible due to the aforementioned problem of not having cultures available for the majority of microbial phylotypes. New technologies using membrane diffusion, cell-sorting, or microfluidics and the use of cocultures to cultivate interacting microbes in combination are trying to circumvent some of the problems (Lewis et al. 2021). There are also ongoing developments in laboratory assays, e.g. using a high throughput system to assess the kinetic effects of more than 1000 mutations in an esterase enzyme (Markin et al. 2021).

Gene expression and regulation

Another ‘causal role’ approach to function, is determining under what conditions a gene is expressed. This approach assumes that regulation in relation to an environmental condition or stimulus implies a functional role in that condition. Condition-dependent protein expression and interactions between proteins capture different dimensions of protein function (Morcinek-Orłowska et al. 2021). The transcriptional regulatory network to a large extent determines the gene expression responses to environmental and cellular conditions. There are multiple levels of control of expression, including post-transcriptional and post-translational factors. In general, correlations between expression of genes indicate coregulation, which may imply a functional relationship (Junier and Rivoire 2016). The genes expressed under different environmental conditions are largely determined by transcription factor binding sites. These sites can be discovered either through experimental detection of binding or computational comparisons for the specific motifs associated with particular transcription factors (Gao et al. 2018). Direct interactions between proteins similarly imply a functional relationship, with the underlying implicit assumption that interactions are specific, as nonspecific interactions tend to be toxic (Bhattacharyya et al. 2016). Protein–protein interactions can be detected with methods such as affinity purification mass spectrometry (Morris et al. 2014).

Evolution

Evolutionary considerations have a complex relationship with function. Some have proposed that to be functional simply is to be (or have recently been) a subject of natural selection (Neander 1991, Graur et al. 2013). Evolutionary biologists are interested in questions regarding historical or population dynamics processes. This gives an account of what is functional and what is not if one can detect the effects of selection; however, given our general lack of access to historical evolutionary forces, it is of little use determining what a biological entity’s particular function was or is. More relevant to assigning a specific function is protein conservation across species, where it is typically assumed that the function is likewise conserved. Technically, homology—as originally used by Richard Owen (Cooper and Owen 1843), and in evolutionary theory since—refers only to a biological structure, such as a protein sequence, but it is now widely applied to functions as well (Love 2007). A related source of microbial functional insights lies in pangenomes—i.e. the patterns of gene diversity across strains within a species, genus, or other monophyletic clade (Golicz et al. 2020). This approach overlaps with the classic phenotype method—again gene presence/absence is used to infer function, but via presence/absence across natural strains rather than genetic manipulation in the laboratory. This kind of analysis presumes that patterns of evolution in functionally related versus functionally unrelated genes will differ. The extent to which this is the case will depend on the role of natural selection in shaping the pangenome. This topic is controversial, but there have been many cases where functional relationships between genes have been shown to leave genomic footprints (Chen et al. 2013). Other approaches integrate some of these diverse information sources to find functional clusters of genes, for instance those relating to specific metabolic processes (Psomopoulos et al. 2020).

Categorization labels

The functional concepts described above do not, however, map directly onto useful labels or gene categories for scientists to work with. Given that the concept of function is contentious and plural, many kinds of categories or labels can be assigned to proteins, each with various advantages and disadvantages, based on information from diverse sources including high throughput gene knockout or silencing, biochemical assays or biophysical studies, ‘omics’ expression studies, and/or evolutionary analyses. Perhaps the most popular hierarchical system for gene function annotation with a controlled vocabulary is the use of ‘gene ontology’ (GO) terms (Ashburner et al. 2000). The GO terms comprise three different classification systems: cellular components (e.g. an organelle), molecular functions (e.g. a particular enzyme activity), and biological processes (e.g. mismatch repair). Methods for automated function prediction using GO terms have recently been reviewed (Makrodimitris et al. 2020). Another widely used comprehensive hierarchical system is the system of enzyme commission (EC) numbers, a numerical classification scheme for enzymes with biochemical evidence (Bairoch 2000), which classify proteins into seven major types of enzymes and 6646 entries of different catalyzed reactions (as of March 2022) (Gasteiger et al. 2003). These EC numbers are very useful for well-understood enzyme families and other genes with characterized homologs. Every EC number is associated with a recommended name for the respective enzyme. If different enzymes (for instance from different organisms) catalyze the same reaction, then they receive the same EC number (Fleischmann et al. 2004). Furthermore, through convergent evolution, completely different enzymes can catalyze an identical reaction and, therefore, would be assigned an identical EC number (Omelchenko et al. 2010). Protein structures are grouped into families or folds in databases such as Pfam, SCOP, and CATH (Fox et al. 2015). Typically, members of the same protein family perform similar functions, but some ancient families or super families have significantly diverged in sequence and/or function (Jaroszewski et al. 2009).

Standard bioinformatics methods for function prediction

Many in silico methods have been developed for functional prediction of proteins using bioinformatics tools for classification and annotation. However, they are not high-throughput and require quite extensive computing power and time, can often only find what is already known, and do not make use of the full metadata available, such as linking gene expression data with environmental parameters. In addition, the predicted functions are usually not experimentally verified, raising the possibility of untrue annotations.

Sequence-based annotations

Comparative genomics, based on evolutionary theory, allows for the propagation of annotations across genomes (König et al. 2018). The use of ‘BLAST’ (Basic Local Alignment Sequence Tool) (Altschul et al. 1990) is nearly ubiquitous in biology, and for some synonymous with bioinformatics. This approach is so powerful because only a few model organisms have been probed in depth in laboratory studies, but many proteins are conserved across diverse taxonomic groups. Homology inference is applied in many annotation tools (Mahlich et al. 2018), which can also be used to identify associations between protein domains and functions (Rojano et al. 2022). However, even when a conserved domain is predicted, the function itself may not be conserved (Punta and Ofran 2008). On the other hand, proteins with different sequences/domains might be able to catalyze the same chemical reaction. In addition, characterized domains are often simply not found in HPs (Goodacre et al. 2013). Further, it is a known problem that many annotations in databases are simply wrong or suffer from ‘over’-annotation (Moreno-Hagelsieb and Hudy-Yuffa 2014). Any errors in the initial assignment of function is, therefore, propagated outwards across genomes with no ‘proofreading’ if only the criterion of homology is used.

Structure-based annotations

Similarly, a protein’s structure may be highly similar even when there is no trace of higher-than-chance similarity at the sequence level (Rost 1999). The function of a putative protein can be predicted by so called ‘homology modelling’ (David and Andrej 2001), in which the protein has to align with a known protein sequence whose 3D structure is known or using protein signatures, which classify proteins into families and domains based on sequence models such as hidden Markov models, with various confidence levels (Zohra Smaili et al. 2021). Further approaches include determination of protein 3D structure by structural genomics initiatives, understanding the nature and mode of prosthetic groups or metal ion binding, fold similarity with other proteins of known functions and annotating possible catalytic and regulatory sites (Myers et al. 2015). Another promising approach is structure prediction followed by biochemical function assessment by in silico screening for various substrates (Mills et al. 2015).

Transcription-based annotations

The transcriptional network within which a gene is located is informative of its function within the cell. These networks are currently only available in databases for a few model organisms, but in principle can now be predicted for a wider range of organisms, using high throughput data on RNA expression and transcription factor binding. The main functional units here are the operon and the regulon. The operon has been classically thought of as a cluster of colocated genes on the same strand controlled by a regulatory region (Jacob and Monod 1961), which may include positive or negative regulation. There are many additional complexities that have since been discovered, for instance recently an example of a noncontiguous operon has been proposed, where genes situated in antisense to each other are coregulated (Sáenz-Lahoya et al. 2019). Functional groupings of genes (operons and higher-level groupings) have been inferred from correlations between the expression levels of different genes in bacterial populations grown under diverse environmental conditions (Chen et al. 2018). Multiple operons can together be coregulated, and together grouped into a regulon, which can coceivably be inferred from either transcriptomic or pangenomic data. Because operons and regulons are relatively discrete functional units (Sastry et al. 2019), they may avoid some of the common pitfalls which other approaches such as GO classes face due to their complex hierarchical relations (Gaudet and Dessimoz 2017).

Pangenome-based annotations

New evolutionary or comparative genomic analyses have also been developed on genome data available across related strains and species. Differing gene content across strains mean that there is a large ‘pangenome’ for each species (Brockhurst et al. 2019). Studying patterns of copresence and coabsence allows for inferring functional networks of genes, which have been termed ‘components’ in this kind of analysis (Hall et al. 2021). Information on gene–gene relationships and combining evolutionary information with transcriptomic/translatomic data has the potential to greatly increase our knowledge of regulatory networks and gene functions.

The development of next-generation and third-generation ‘long read’ sequencing technologies has resulted in huge amounts of data (Fig. 1), and currently almost 200 million HPs are listed in public databases. The aforementioned in vivo, in vitro, and/or in silico methods of assigning functions to sequences are thus no longer able to catch up with the exponentially growing number of sequencing data and, hence, the number of HPs, leaving an enormous potential for biotechnological application uncovered (Ijaq et al. 2015). The amount of information nowadays available in public databases, however, creates opportunities for new ‘Big Data’-based approaches beyond traditional analyses.

The era of Big Data

‘Omics’ datasets have become one of the main examples of ‘Big Data’ in the last few years. In general, the characteristics of Big Data can be summarized by the 6 Vs: Value, Volume, Velocity, Variety, Veracity, and Variability (Emmanuel and Stanier 2016). Aside from the possibility to obtain genomes, transcriptomes and proteomes from individual species in cultures or single cells (Dam et al. 2020, Kaster and Sobol 2020, Wiegand et al. 2021) shotgun sequencing approaches such as metagenomics and metatranscriptomics have now become the methods of choice for studying microbes from various habitats and researching phylogenetic groups that currently lack cultured representatives (Aguiar-Pulido et al. 2016). Shotgun proteomics can also be applied to whole microbial communities for metaproteomics (Karaduta et al. 2021), and the translated portion of RNA has begun to be examined via metatranslatomics (Fremin et al. 2020). All this data could in theory be used to improve gene annotation, but it has been argued that the application of methods designed for smaller datasets to this new wealth of data has in fact reduced the quality of annotations (Salzberg 2019).

Application of Big Data in biology therefore requires integration and analysis of complex heterogeneous data, including metadata for each dataset (Subramanian et al. 2020). This demands cost-effective, innovative forms of information processing. The quality and reliability of data are important and vary significantly across datasets, which is why maintaining curated public databases is paramount. The most widely used database for depositing sequencing and metadata is the National Center for Biotechnology Information database (NCBI), which includes GenBank (Benson et al. 2018) and RefSeq (Li et al. 2021). In addition, there are databases from the European Bioinformatics Institute (EBI), including the ENA (European Nucleotide Archive) repository (Park et al. 2017), EggNOG (Huerta-Cepas et al. 2019), and Ensembl Genomes (Howe et al. 2021). Protein sequences are stored in databases such as InterPro (Blum et al. 2021), Pfam (hosted by the Sanger Institute) (Mistry et al. 2021), or UniProtKB/Swiss-Prot (hosted by the Swiss Institute of Bioinformatics) (Bateman et al. 2021). Metabolic pathway and protein function databases include KEGG (maintained by the University of Kyoto) (Kanehisa et al. 2017) or BioCyc (maintained by SRI International) (Karp et al. 2019). The journal Nucleic Acid Research regularly publishes special issues on biological databases and currently, there are over 1640 databases listed with different purposes, some being complementary (Rigden et al. 2021). It has been estimated that the amount of omics data will double every 9–12 months (Stephens et al. 2015). Large-scale comparisons of genomes and transcriptomes have already been used for diverse analyses such as biosynthetic gene clusters (Navarro-Muñoz et al. 2020), the relative expression of different mRNAs and noncoding RNAs (Ireland et al. 2020) or ‘ribosome profiling’ (Steizt 1969). CHiPseq (Furey 2012) can be used to discover transcription factor binding sites, and thereby bacterial regulons (Myers et al. 2015) and a modification of IPOD-HR gives an overview of all protein binding to DNA, giving a fuller picture of gene regulation and functional components (Freddolino et al. 2021).

The next step: using artificial intelligence to characterize the proteome

With the breakthrough of technological advances over the last decades, the ‘educated guesses’ that had been previously used for creating specific scientific hypotheses are rapidly being replaced by the knowledge provided through untargeted high-throughput methods. Artificial intelligence (AI) provides novel opportunities to use large quantities of high-quality data on another level with a wide range of rational processes, including reasoning, learning, decisions, language processing, and perception (Oliveira 2019). We can think of AI, machine learning (ML) and deep learning (DL) as a set of concentric circles where DL is a subset for ML and ML is a subset for AI.

The massive, complex, and rapid evolution of datasets as well as computational mathematics make it now possible for AI applications to learn, make intelligent decisions, and improve pattern recognition capabilities (Perakakis et al. 2018). Manual intervention in data management and analytics are still needed, but processes that might take days or weeks (or longer) or which were not humanly possible, are now quickly achieved (Serres et al. 2001) (Fig. 2). Table 1 provides a glossary of generally used terms regarding AI, to help guide the reader through the next sections.

Figure 2.

Annotation of data using classical methods and AI. AI surpasses the traditional methods from data curation to data annotation. ML-based techniques have now paved the way for DL neural networks. AI can be used directly on data acquired from experimental or publicly available repositories and extract meaningful features, ensuring a reduction in total number of dimensions, which translates to reduction in the number of input variables for training datasets, hence, reducing computational load.

Table 1.

A glossary of generally used terms regarding AI.

Term	Description
Activation function	A function which is applied at each neuron on its summed input and which determines its output to the next layer. Can be linear, i.e. multiplying the neuron’s input by a constant factor, or nonlinear. Nonlinear activation functions give neural networks the ability to model complex problems
Categorization	Assigning labels to elements of a dataset (such as functional classes to protein sequences), where classes can be hierarchical, i.e. related to each other, and one element can belong to multiple classes (e.g. GO terms)
Class	A discrete value returned by a classifier, which is mutually exclusive from all the other values obtained for all the data points of a dataset (such as EC numbers for protein sequences)
Classification	Assigning classes to elements of a dataset, where classes are mutually exclusive, i.e. one element can only belong to one class
Classifier	An algorithm used for the classification of input data into specific classes
Clustering	Divides the elements of a dataset into several groups based on similarities in their features/attributes, without defining these groups a priori, i.e. learning without any labels
Collaborative filtering	Filtering for patterns of features across large datasets. Similarities in feature patterns across different data points can be used to make predictions (‘filtering’) about missing data
Convolutional layer	Layer type that gives convolutional neural networks (CNNs) their name; it extracts Feature from input data, such as DNA or protein sequences in their matrix form by applying several filters; these decompose the input into smaller matrices (feature Maps), which are passed to the next layer (e.g. a pooling layer)
DL	Use of artificial neural networks comprising multiple (more than three) layers of neurons to identify complex patterns over large datasets
Dimension	The number of features/attributes of a data point/dataset (e.g. the amino acid frequencies of a protein sequence would be a 20-dimensional vector, describing/containing 20 attributes)
Ensemble classification	Classification based on the predictions of several independent classifiers. Ensemble methods usually perform better than their individual constituent classifiers. The classifiers can have different algorithms, training sets and hyperparameters and are all independently trained. Two methods by which a Classification may be reached by the ensemble are unweighted and weighted voting.
Feature	A single representation (typically numeric) that combines/summarizes multiple attributes/metrics of a data point, e.g. combining the frequencies of R (Arginine), H (Histidine), and K (Lysine) (three attributes) in a protein into the frequency of positively charged amino acids (one feature), or computing a feature map on an input matrix in a convolutional neural network
Feature extraction	Defining a set of features for a dataset (usually done by dedicated algorithms), each of which combines several attributes into one numerical representation. Effectively reduces the dimensionality of the dataset and thus speeds up computations
Feature map	The activation pattern of neurons for a particular filter across the input matrix in a convolutional layer.
Feature selection	Manually or algorithmically selecting a subset of features that best represents the structure of the dataset. Further reduces the number of dimensions of the dataset and thus also computational complexity
Filter	Filters define features in convolutional neural networks. Each one is a small matrix of weights, which is slid across the input matrix and at each position may or may not activate a neuron. This results in a smaller matrix of activations, called a feature map. Effectively, each filter acts as a template or a pattern for which the input is scanned
Fully connected layer	Layer type in a neural network, where each neuron is connected to the next one of the subsequent layer; in CNN-based classifiers, the classification process happens in the fully connected layers
Feedforward neural network	A simple network where the output of a neuron layer is the input of the layer below; i.e. information flows linearly from input-to-hidden-to-output layers, unlike in recurrent neural networks or Boltzmann machines
Hidden layer	Layer type in neural networks, i.e. ‘hidden’ between the input and the output layers. Here, from layer to layer, the input is transformed through mathematical functions before being passed to the output layer. In effect, the learning takes place here. The number of hidden layers determines the model complexity and in CNNs, how well higher-level features can be derived from the input features
Hyperparameter	A parameter, i.e. not changed/adapted during model training, but which describes the model or training itself. For example, this can be the number of hidden layers in the network, or the number of neurons in each layer
Inference data	Data for which the true labels are unknown in a supervised learning setting, and for which the model(s) are trained in order to classify them
Input layer	Layer type in a neural network that takes in raw information or data from the user. No computation is performed at this layer
Kernel functions	When high-dimensional data are not linearly separable, kernel functions (such as polynomial or radial basis functions) are applied to project them into a higher-dimensional space where they become linearly separable; kernel functions thus facilitate classification tasks in ML
Label	A value returned by a classifier which is not mutually exclusive from other values obtained for all elements of a dataset (such as functional classes to protein sequences using GO terms)
Labelled data	Data for which each element's true label is known, often split 80:20 into training data, to train a model and test data with which final model accuracy is tested
Layer	Highest-level building block in neural networks; effectively a group of neurons, which are not connected to each other and which receive input (either user-provided data, or the output from the preceding layer of neurons), transform it using mathematical functions, and pass it to the next layer
ML	Field of study in which machines learn from existing data via dedicated algorithms to perform tasks or make accurate predictions without human intervention
Model	Pertaining to ML/DL, a model is the result of an algorithm being trained on data in order to make predictions about unknown data or to find specific patterns
Natural language processing (NLP)	Branch of AI, which teaches machines to understand, utilize, and generate human language with statistical, ML or DL algorithms
Neural network	Collection of artificial neurons that mimic a group of biological neurons; the neurons are connected to each other in patterns that generally influence the behaviour and capabilities of the algorithm
Neuron/node	The individual unit of neural networks, organized in layers of several neurons; each neuron receives numerical input, e.g. from neurons of the preceding layer, and produces an output via a mathematical function, i.e. applied on the sum of all its inputs, which it then passes on to all its downstream connected neurons
Neural depth	The number of neural layers in a neural network
Output layer	Layer type in neural networks, which brings the information learned through the hidden layer(s) into a form, i.e. interpretable by the user; e.g. in case of a simple classifier, the output layer contains as many neurons as there are classes (each neuron corresponding to a specific class) and the neuron with the highest output corresponds to the class predicted by the model for that input
Parameter	A variable, i.e. internal to the algorithm and optimized during training, e.g. weights and biases
Pooling layer	Layer type in a neural network, which down-samples the output feature maps from a convolutional layer (usually to half their original size). Among others, this helps in reducing the number of parameters and thus computational cost during training
Predicted data	It is the output generated from a model
Regression	Statistical method(s) for understanding and describing the relationship between two features or variables
Reinforcement learning	Field of ML in which learning is based on a trial-and-error system. Instead of training on labelled or unlabelled data, the algorithm (or ‘agent’) learns the cost or reward associated with the outcomes of its different choices and then tries to maximize the reward. It can be considered a separate paradigm from supervised and unsupervised learning
Supervised learning	Field of ML, which requires labelled data. In this type of learning, the user defines the classes and their corresponding labels, and the algorithm learns to associate data/attribute/feature patterns with these labels
Semisupervised learning	Field of ML where learning takes place by using a combination of small amount of labelled data and a large amount of unlabeled data
Test data	Subset of the labelled data, used to test the accuracy of the model, after its parameters have been optimized in several rounds of training and validation, and before it is fed the inference data
Training data	Subset of the labelled data, which is used to train an algorithm/classifier, i.e. teach the algorithm to associate (patterns of) features with the different classes or class labels. Part of the training data (usually 20%) is held back as validation data and not used in adjusting algorithm parameters
Training Iteration	After each training iteration, algorithm parameters are updated. The total number of training iterations depends on how much of the training data the algorithm sees in each iteration, as well as how often it should go over the training data to optimize its parameters
Unsupervised learning	Field of ML that deals with unlabelled data. The algorithm learns previously undetected patterns in a dataset, without relying on user-provided associations between data points and classes/labels
Unweighted voting	Assigning a class/label to an Inference data point via an ensemble method, using the label, which was predicted by the majority of classifiers
Validation data	A subset of the labelled training data, not used in algorithm parameter tuning, but retained to check algorithm accuracy. Each training iteration predicts classes or labels for the validation data
Weights and biases	Learnable parameters of a neural network, i.e. those that are optimized during training; weights signify the strength of a connection between two neurons. Biases are constant values that are added at each neuron and influence its output; unlike weights, they are not influenced by a neuron’s incoming connections
Weighted voting	Assigning a class/label to an inference data point via an ensemble method by weighting the prediction of each classifier where the weights are adjusted over the learning period. The algorithms emit predictions, followed by the input of correct labels by the user. For each error made by the best performing classifier, its weight decreases exponentially

Traditional classifiers

ML includes a wide range of methods, which can be divided into supervised, unsupervised, semisupervised, semiunsupervised, and reinforcement-based learning approaches (Alloghani et al. 2020). There are several algorithms that can be used for the task of categorization—assigning functional labels to putative genes, which have already been predicted to encode a protein (Fig. 3). The technique of classification has long been used for function prediction by identifying suitable features using feature engineering and generating numeric vectors to subsequently develop suitable models.

Figure 3.

A workflow for functional annotation of proteins using AI. Specific features from sequence and/or structure data are extracted and used as training data. Apart from classification-based tasks, ML algorithms can also perform tasks pertaining to regression that predict the relationship between two known variables. Other methods involve algorithms that group the input data into specific clusters, the output data. A validation dataset is then used to test the efficiency of the model. Test datasets (in this case HPs) are then fed to the optimized model. An experimental feedback loop could also be used for the validation of correct assignment for function. Partially adapted from (Mahood et al. 2020).

Support vector machine (SVM) (Vapnik 1995) a supervised learning approach, which performs binary classification, with linear or nonlinear functions. It establishes a maximum-margin hyperplane within the n-dimensional space of the data, which separates the data into two classes. The ML algorithm achieves this by determining an appropriate kernel function (e.g. linear, polynomial, or radial basis) (Kulkarni-Kale et al. 2014).

k-nearest-neighbor (KNN) (Altman 1992) is a supervised learning algorithm that tries to classify each data point by locating the nearest ‘k’ neighbors with known labels and subsequently assigns a class label, i.e. determined by a majority vote among neighbours. Traditional KNN-based methods are easy to use but involve higher computation times (Borah et al. 2020).

Decision tree (DT) (Quinlan 1986) is a branch-test-based classifier, supervised algorithm, which recursively partitions the data based on its attributes, until some stopping condition is reached. This recursive partitioning gives rise to a tree-like structure. The route undertaken for classification of data can be traced from the root node to each leaf node in the tree (Schietgat et al. 2010).

Random forest (RF) (Breiman 2001) uses an ensemble of DTs to obtain a majority vote on the correct classification. Classification trees are constructed by randomly selecting from training datasets. Results from each tree can then be gathered to give a prediction for each observation.

Some studies also mention regression-based protein function annotations, which are now used as ensemble methods along with other classifiers (You et al. 2018). Although ML has gained immense popularity over the last decade, DL methods are now increasingly being explored. Unlike traditional ML algorithms that require a lot of domain expertise and human intervention, DL algorithms automatically learn from raw input data, therefore, describing highly nonlinear and complex patterns more effectively (Lv et al. 2019). DL is widely associated with artificial neural network (ANN) architectures having numerous hidden layers for feature extraction. Unlike ML, which is limited to predicting discrete outputs obtained by counting the data or continuous outputs obtained by measuring input data, DL-based methods are able to learn data representations i.e. feature learning (Fig. 3). These models often have the capability to automatically obtain useful information from input datasets and bypass traditional feature engineering and selection processes (Bonetta and Valentino 2020).

ANNs (Mcculloch and Pitts 1943) can be simply described as being similar to biological neurons where the learning process is due to synaptic connections. The data passes through input, hidden, and output layers. The input layer feeds in the data, from which meaningful information is extracted by hidden layers, which lead to the prediction of data for the classification problem in hand. All the layers comprising this network architecture are together known as a Deep neural network (DNN). ANNs can process nonlinear data and handle noisy data but are prone to overfitting (Kulkarni-Kale et al. 2014). ANNs include RNNs, CNNs, and GCNs (see below).

Recurrent neural networks (RNN) (Sperduti and Starita 1997) can be used for supervised learning with artificial neurons having one or more feedback loops. These loops are recurrent cycles over data. Input-target pairs are provided by the user. RNNs are expected to optimize the networks by minimizing the difference between the target-output pairs (Salehinejad et al. 2018).

Convolutional neural networks (CNN) (Lecun et al. 1998) are comprised of convolutional, pooling, and fully connected layers. They can identify relevant features without human supervision and resemble a feedforward NN. The pooling layer is involved in dimension reduction and the results are forwarded into the fully connected layers. Its massive parallelism yields a great amount of computational efficiency (Alzubaidi et al. 2021).

Graph convolutional networks (GCN) (Miller et al. 1989) have numerous spectral or spatial convolutional layers. Input data featurization and elaborate architectures render them suitable for complicated problems. The graphs can extract meaningful information from their own structures (Zhou et al. 2022).

Transformers are novel neural network architectures based on self-attention mechanism, which are most widely used in the field of natural language processing (NLP). When performing text translation tasks, these DL models are not required to process each sentence from the beginning but refer to the context of each word in a sentence, which improves parallelization of the classification task. These, however, suffer from the problem of sequential computation (Vaswani et al. 2017).

Ensemble classifiers (Dietterich 2000) ensembles make individual decisions by different classifiers, which are combined either by weighted or unweighted voting for classification of new instances. Multiple learning algorithms make the classification model more robust. Methods like this increase efficiency but can be biased as performance heavily depends on weights in a weighted voting (Gokalp and Tasci 2019). These can involve a combination of any of the classifiers described above.

Clustering-based methods are capable of exploiting the direct, as well as the indirect, interaction proteins of the unannotated protein to predict functions, and of more effectively interpreting the protein interaction relationships in the prediction process (Hou 2017). One of the oldest and most used techniques is the k-means clustering. This clustering algorithm follows an unsupervised approach and does not require the user to define the output clusters.

Reinforcement learning (RF) or reward-based approaches have often been utilized in DL techniques. The famous AlphaFold2 software (Jumper et al. 2021), developed by DeepMind, has used this approach to identify protein structures, which has aided in deciphering how a protein folds in its natural environment (also see below). RF also finds its use in data augmentation, i.e. generating artificial data points or sequences, which helps in balancing the datasets by compensating the lack of adequate protein sequences as described in a recent study (Eftekhar 2020) for the prediction of subcellular localization of proteins.

By using traditional ML or DL classifiers, it is possible to predict the function of HPs without using homology information (Han et al. 2006). There are several other data clustering algorithms such as DBSCAN (density-based spatial clustering of applications with noise) (Ester et al. 1996), which along with ICA (independent component analysis) (Bell and Sejnowski 1995)—a blind-source separation algorithm—help in computing context-specific activities (e.g. metadata) of gene modules, determining their relative strengths, followed by an unsupervised approach and do not require a predetermined number of clusters to perform classification. As an example, DBSCAN and ICA for clustering gene expression data into so called i-modulons (Sastry et al. 2019). Summary aspects and usage of each classifier is given in Table 2, where also the pros and cons for each type of method are listed.

Table 2.

Aspects and usage of classifiers. All classifiers/methods used for classification or clustering of protein sequence or structure data. Each classifier is listed with pros, cons, and problem-specific usages, to guide researchers in choosing a method, which is best suited to their datasets and biological question.

Classifier/Method	Task	Pros	Cons	Usage
Support Vector Machine (SVM)	Classification/Regression	Efficient when classes are distinctly separable, relatively memory efficient, suited for binary classification and for data with very high dimensions	Not suitable for large datasets or noisy data, data preprocessing required, overfitting risk, computation can be costly, interpretability of output is low	Protein-protein interaction prediction using sequence data (Ma et al. 2020) Identification of intrinsically disordered proteins using sequence data (Kandemir Cavas and Yildirim 2016)
k-nearest-neighbor (k-NN)	Classification/Regression	Easy to implement for multi-class problems, selected Hyperparamerter remains the same, non-parametric algorithm	Slow algorithm, works well with small number of input variables, not ideal for imbalanced data, sensitive to outliers	Protein structure classification into distinct classes by using structure data (Mirceva et al. 2020) Discrimination of Membrane Transporters using sequence data (Zuo et al. 2015)
Decision Tree (DT)	Classification/Regression	Data pre-processing is easy and does not require scaling or normalization, good for visual representation of output, can utilize numerical and categorical features	Higher training times, overfitting, inadequate prediction for large complex datasets, not ideal for imbalanced datasets	Classification of proteins from PDB into families based on surface roughness using structure data (Satpute and Yadav 2019) Protein-protein interaction prediction using different physicochemical properties by utilising protein sequence data (Zhou et al. 2017)
Random Forest (RF)	Classification/Regression	Scaling and transformation of input variables is not required, ideal for working with large number of features, lower overfitting risk, works well with non-linear data	Not easily interpretable, computationally intensive, training is slow, limited performance with regeression while dealing with linear varibales	Prediction of small encoded proteins and their functions in a species-specific context by protein sequence data (Miravet-Verde et al. 2019). Prediction of protean segments from sequence of an Intrinsically disordered protein (Basu et al. 2017)
Recurrent Neural Network (RNN)	Classification/Regression	Dynamic neural network, computationally powerful, useful for non-linear systems	Rough to train long sequences due to gradient vanishing problem and exploding gradient problem, slow and complex training	Antibiotic resistance class prediction using protein sequences (Hamid and Friedberg, 2020) Prediction of essential proteins by integrating information from protein-interaction networks, gene expression profiles and subcellular localization (Zeng et al. 2021)
Convolutional Neural Network (CNN)	Classification/Regression	Unsupervised learning, high accuracy, weight sharing, reduce dimensionality in neural network	Long training period, requires large datasets, fails to encode the position and orientation of objects	Prediction of metal binding sites using structure data (Mohamadi et al. 2022) Identification of efflux proteins in transporters using sequence data (Taju et al. 2018)
Graphical Neural Network (GCN)	Classification/Regression/Clustering	Uses same parameters in the training iteration, inexpensive data storage, adaptively learn the importance of neighbours in a graph-based system	The ‘Black Box‘ problem, which makes algorithm's processes untraceable, computation cost	Graph based prediction of PPIs using raw sequence data (Yang et al. 2020) Protein-protein interactions prediction using structure data (Pancino et al. 2020)
Ensemble Classifier	Classification/Regression/Clustering	Higher predictive accuracy than individual models, can handle both linear and non-linear data, Bias/overfitting can be reduced, less noisy and more stable	Model that is closest to the ‘true’ data generating process beats other methods, lack of interpretability, computationally expensive, memory intensive	Sequence-based classification of antioxidant proteins by RF and SVM (Meng et al. 2019) Protein function prediction using Sequence, taxonomic, structural domains and amino acid index based classification by CNN and RF (Hakala et al. 2022) Protein secondary structure prediction from sequence data by DT and SVM (Afify et al. 2021) Complex data clustering using General adversial networks for assigning proteins to different sub-families by RNN and Clustering (Bitard-Feildel, 2021)
Transformer Model	Classification/Regression/Clustering	Self-attention, universal and flexible architecture, do not provide specific structure to input data (for eg. The need for specifying nodes and edges as in GCN)	Limited to higher level representation to data, limited memory span, overfitting	Utilization of sequence statistics, chemical and biological features to generate biologically active sequences (Rives et al. 2020) Structure, remote homology prediction and protein engineering using sequence data (Rao et al. 2019) Protein Language model for prediction of secondary structure and subcellular localization (Heinzinger et al. 2019)
Clustering Model	k-means Mean shift DBSCAN Gaussian mixture Hierarchical clustering	Fast, few computations No pre-set clusters, intuitive No pre-set clusters, identifies outliers More flexible than k-means Not sensitive to the choice of distance metric, visualize hierarchy	Inconsistent, outliers Selection of a clusters Inconsistent for high dimensions Uses all available data points Lower efficiency, time complexity	Construction of protein networks using sequence data (Keel et al. 2018) Unsupervised clustering based prediction of protein structure and function using relative solvent accessibility of amino acid residues (Teletin et al. 2018) Identification of antibiotic and virulence resistance genes in pathogenic bacteria using genomic sequence data (Li et al. 2018a)

Current computational challenges and resources

One of the biggest challenge in AI is the cost of algorithm training, which scales up with the number of parameters and data points while the computational requirements grow exponentially with number of data points. In order to address the problem emanating with Big Data processing, scientists require dedicated servers with high computing power (Central and Graphics Processing Units) (Thompson et al. 2020).

For research groups with no or very limited funding in terms of computational infrastructure, there are some free services available such as Amazon Web Services (https://aws.amazon.com/machine-learning/ai-services/), which provide processing power for ML-based models but are rather ineffective for training DL models when using the free tier which currently allows 750 hours total computing time for 1 month with two cores, but excludes GPUs required for faster computing. Google cloud (https://cloud.google.com/products/ai) and Azure (https://azure.microsoft.com/en-us/solutions/ai/) are other options but have similar limited capabilities for free accounts. Researchers might also be interested in GPU-based Jupyter notebook servers to train their ML/DL models. These include some freely available platforms like Google Colab (Bisong 2019) and Kaggle (https://www.kaggle.com), again with limitations for storage and processing power. For instance, Collab has a space limitation of 15 GB and allows users to run their notebooks for 12 hours per day after which the user is required to pay for additional hours. Amazon Sagemaker (https://aws.amazon.com/sagemaker/) is another option for large-scale training of ML models. It is time efficient, having built-in algorithms and optimized frameworks making it easier to use, but currently requires an hourly fee of $1.125 to train large datasets. Other considerable platforms are DataCrunch (https://datacrunch.io) and Paperspace (https://www.paperspace.com), which have lower GPU costs of $1.1 and $0.18 per hour, respectively as compared to Sagemaker (https://vitalflux.com/deep-learning-top-5-online-jupyter-notebooks-servers/).

Examples of AI used in protein function prediction of prokaryotes

Table 3 gives an overview of all tools/software that are used for prokaryotic protein function prediction, with the type of input data required by the model and the generated output. Tables 2 and 3 can, therefore, aid researchers in choosing the model best suited for their data and biological question. Based on the most cited tools, we here, discuss a few examples, using methods including KNN, ensemble approaches, DL, and NLP.

Table 3:

Machine and Deep Learning-based tools/software forprotein function prediction. ARGs, antibiotic resistance genes; BGCs, bacterial gene clusters; CS, contribution scores; DBP, DNA binding protein; EC, enzyme commission IDs; emb, embeddings; EP/NEP, essential/non-essential proteins; FD, fingerprint descriptors; func, functional classes; GA, gene annotations; GE, gene expression data; GC, gene cluster; Ge, genome; GO, gene ontology annotations; HbL, haemoglobin; Kace/Non-Kace, lysine acetylation site/non-lysine acetylation site; LBS, ligand binding sites; Nt seq, nucleotide sequences; ORF, open reading frame; PC, protein clusters; Pfams, protein families; PI, protein interactions; PS, probability scores; PSC/PRS, per sequence/per residue scores, RP, ribosomal profiling data; SA, species abundance; Sbfam, subfamily; Seq, protein sequence data; SB; source biome; Sdr, short DNA reads; Strc, protein structure data; SL, subcellular localizations; TA: taxonomic abundance score.

Tool/Software	Classifier /method	Type of Input(s)	Type of Output(s)	References
SVMProt BacHbpred iProEP	Support Vector Machine	Seq, Strc Seq Seq	Pfams HbL Pfams	Li et al. (2016) Krishnan et al. (2016) Lai et al. (2019)
GrAPFI PseAAC	k-Nearest Neighbour	Seq Seq	EC SL	Sarker et al. (2020) Jiang et al. (2019)
CrowdGO	Decision Trees	GO	GO	Reijnders and Waterhouse (2021)
PPI & Gabor filter P2Rank FEAST	Random Forest	Seq Seq TA	PI LBS SB	Zhan et al. (2020) Krivák and Hoksza (2018) Shenhav et al. (2019)
DEEPred UDSMProt ProDec-BLSTM PARROT DeepBGC LookingGlass UniRep ProLanGO	Recurrent Neural Network	Seq Seq Seq Seq Ge Sdr Seq Seq	GO GO, EC, Pfams PS PSC/PRS BGCs EC, Emb Emb GO	Rifaioglu et al. (2019) Strodthoff et al. (2020) Li et. al (2017) Griffith and Holehouse (2021) Hannigan et al. (2019) Hoarfrost et al. (2020) Alley et al. (2019) Cao et al. (2017)
MultiPredGO DeepGOPlus DeepAdd Balrog ProtCNN DeepHiFam SmORFinder DeepEP MaSIF ONN4MST ON4ARG DeeProtGO EXPERT ProteInfer SeqVec	Convolutional Neural Network	Seq, Strc Seq Seq, PI Seq Seq Nt seq PI, GE Str SA Seq, ARGs Seq SA Seq Seq Seq	GO GO GO Genes Pfams CS EP/NEP FD SB GA GO SB EC EC, GO Pfams, Sbfam	Giri et al. (2021) Kulmanov and Hoehndorf (2020) Du et al. (2020) Sommer and Salzberg (2021) Bileschi et al. (2019) Sandaruwan and Wannige (2021) Durrant et al. (2020) Zeng et al. (2019) Gainza et al. (2020) Zha et al. (2020) Zha et al. (2021) Merino et al. (2022) Chong et al. (2021) Sandesron et al. (2021) Heinzinger et al. (2019)
AlphaFold DeepFri PersGNN PANDA2	Graphical Neural Network	Seq, Strc Strc Strc Seq, GO	Strc GO GO GO	Jumper et al. (2021) Gligorijević et al. (2021) Swenson et al. (2020) Zhao et al. (2022)
DEEPre ECPred BioSeq-Analysis DeepRibo LargeGOPred DeepGraphGO GOLabeler NetGO ProPythia DeepMicrobes STALLION Deep_CNN_LSTM_GO PFmulDL	Ensemble classifiers	Seq EC Seq Seq, RP Seq Seq, PI Seq, PI, GO Seq Seq Nt seq Seq Seq Seq	EC EC DBPs ORFs GO GO GO GO EC Emb Kace/ Non-Kace GO GO	Li et al. (2018) Dalkiran et al. (2018) Liu (2019) Clauwaerts et al. (2018) Wang et al. (2020) You et al. (2021) You et al. (2018) You et al. (2019) Sequeira et al. (2022) Liang et al. (2019) Basith et al. (2022) Abdou et al. (2021) Xia et al. (2022)
TALE ProteinBERT	Transformer	Seq Seq, GO	GO GO	Cao and Shen (2021) Brandes (2022)
iModulonDB AGNOSTOS DB PPI-GA	Clustering	GE Nt seq PI	GC GC PC	Rychel et al. (2021) Vanni et al. (2021) Shirmohammady et al. (2019)

NetGO (You et al. 2019) is a GO-term prediction tool, which builds on a previous ensemble learning framework, GOLabeler (You et al. 2018), by adding a module using KNN, a supervised classification algorithm. The protein features used for learning are the networks in which a protein is involved in, as found in the ‘STRING’ database of protein–protein interactions. If the protein is not present in this database, then data for the closest homolog is used, if one is available. GO-terms from the closest neighbours in the (potentially multiple) relevant networks are applied to the new protein by aggregating weights from different networks followed by the kNN approach of plurality voting. Significant limitations of this method are genes, which lack homology to entries in the STRING database, and examples where homologous proteins with low similarity have different functions.

Another ensemble approach, SVM-Prot 2016 (Li et al. 2016), an update to an earlier tool (Cai et al. 2003), aims to predict what it calls ‘functional families’ from protein sequences without relying on detectable homology. Physicochemical features of proteins such as solubility are derived from sequences and used to train a SVM algorithm to classify sequences into functional classes. The functional groupings of proteins for training purposes (labels) are derived from GO and other databases. A negative training set is derived by choosing some Pfam families with no members included in the positive training set.

There is also a growing set of tools for protein structure prediction which use DL approaches. Two of the most prominent are AlphaFold2 from Google’s DeepMind (Jumper et al. 2021) and ESMfold from Meta AI Research (Lin et al. 2022). Alphafold2 uses information obtained from a multiple sequence alignment on evolutionary couplings between amino acid residues to infer pairwise distances between the residues, and from there infers a 3D protein structure. While the earlier version of AlphaFold used statistics calculated from multiple sequence alignment input, AlphaFold2 embeds the full sequence information, and uses a transformer neural network, which is able to take into account relationships between residues and apply an ‘attention’ mechanism to take into account those previously learned from training sets to be most important. The neural network then feeds the information into a structural model neural network, which uses another attention mechanism transformer to take into account the most important pairwise relationships and output a structural model, i.e. the 3D co-ordinates of each of the residues’ atoms (Jones and Thornton 2022). ESMfold, on the other hand, uses a very large natural language model designed for proteins to capture key aspects of an input sequence, which relate to structure as an attention map, predicts pairwise distances from this, and uses both to predict structures (Lin et al. 2022). A protein language model is useful here because structural features are emergent properties of sequences, which can be learned with large input datasets. The overall architecture of the approach is based on Alphafold2, but it replaces the transformer neural network that Alphafold2 uses to process a multiple sequence alignment with a protein language model, which takes a single input sequence. High confidence structures predicted with Alphafold2 are used as part of the training data. ESMfold is less accurate than AlphaFold2, but more than an order of magnitude faster, without the need for building a multiple sequence alignment. Protein structures are not identical to protein functions, but there is often a close relationship between them, which is why structures can then be used by ML models to predict protein function. E.g. DeepFRI is a GCN for predicting functions, which uses protein structures as input along with sequence features derived from a protein language model (Gligorijević et al. 2021).

Major developments have occurred regarding inferring protein structures from sequence data for function prediction. Nevertheless, there are still obstacles, e.g. when applying algorithms for human HPs to microbial HPs (Suravajhala and Sundararajan 2012) since these are based on finding specific domains whilst the majority of the latter have uncharacterized domains. Additionally, almost 20% of the total proteins from 944 bacterial species have no identifiable domains at all (Wang et al.2019). Feed-forward DNN-based modelling approaches for large-scale automated protein function prediction are also probably not a good choice for functional terms with low or moderate number of annotated proteins and it is currently not feasible to carry out a fold-based cross-validation analysis, especially when the number of model training operations are high, since it usually requires extremely high computational power (Sureyya Rifaioglu et al. 2019). Furthermore, many state-of-the-art ML and DL models, have not yet been extensively explored for protein function prediction.

Outlook

Accurate prediction of microbial protein functions has the potential to revolutionize multiple fields of biological research. The accelerating expansion in biological Big Data presents many opportunities but also challenges for researchers, such as computational power and storage limitations and properly integrating diverse data types. DL algorithms have now paved the way for a better and more competent prediction of HPs. Furthermore, new genetic elements and sources of genetic information will open the door to an unexplored world of coding and noncoding complexity in prokaryote genomes (Grainger 2016, Kirchberger et al. 2020). This new world, if it was included in annotation, would further expand the class of ‘hypothetical proteins’. Gene annotation tools, however, have typically excluded short ORFs and overlapping ORFs, largely as a matter of practicality. In recent years, many of both kinds have been discovered to be true protein-coding genes (Storz et al. 2014, Ardern et al. 2020). The number of ‘dual-coding’ RNAs, where there are separate functions for a transcript at the level of RNA and protein structure, remains unknown. Further, the phenomenon of ‘proteoforms’ (Smith et al. 2013), i.e. alternative start sites for genes, has shown the terrain of prokaryotic protein coding to be yet more complex. Nevertheless, direct laboratory analysis will still remain the gold standard of functional annotation for the foreseeable future. An experimental feedback loop is consequently of great importance to further test and train the models for accurate predictions and progress will come from creative and efficient integration of existing data with careful experimental validation. Metaservers or servers, which can be used to input query data simultaneously to extract specific features from the data keep improving with time, hence facilitating developers and users alike, by making the training and testing datasets as well as benchmarking results available in public domain. While there is a need to refine the methods to achieve higher accuracy, there is also a growing need to bring various aspects of protein function prediction under the realm of AI.

One of the key takeaway points from our review of the available resources and methods is that at least something informative can be said about essentially all protein sequences. The label ‘uncharacterized’ or ‘hypothetical’ is completely uninformative, but for most proteins it should be possible to derive at least some characteristics from their general sequence properties (e.g. do they have a transmembrane domain?), whether there is evidence of expression, their possible cellular location, details of homologs, and whether they are predicted or demonstrated to be expressed as part of an operon and/or regulon. New approaches to storing gene annotation data should take these diverse lines of evidence into account.

Authors’ contributions

S.C., Z.A., and A.K.K. wrote the manuscript, F.L. helped with the revision, and A.K.K. had the idea for this review and acquired the funding.

Competing interests

The authors declare no competing interests.

Funding

This work was supported by the Future Fields DigeteLiSe grant of the Karlsruhe Institute of Technology (5207.0011.5096).