Importantly, the performance of an algorithm directly relates to the breadth and complexity of the datasets it learns from. The more novel cases and rare ‘edge’ cases in the dataset, the more the algorithm will learn how to predict similar scenarios in the future. In this way, machine learning can be used to make accurate predictions on a variety of tasks. In general, the algorithm benefits from more data but it is also important to recognize that curated data may be valuable for developing specialized algorithms. For example, if a machine learning algorithm is used specifically to predict cancer epitopes, it should be trained on a specialized cancer database, such as CEDAR ( Table 1 ) (23). Adding data for the prediction of different types of epitopes may improve a machine learning algorithm’s generalizability but could also mask biological patterns unique to cancer epitopes. Additionally, data that are very similar to existing training data do not ‘teach’ an algorithm anything new.

Many publicly accessible databases are available as sources of datasets for either T-cell or B-cell epitopes ( Tables 1 , 2 ). These databases provide large, specialized sources of data that algorithms can study and use for predictions. RCSB PDB, for example, provides primarily 3D structures of various proteins and multimolecular complexes, and information relating to their molecular composition, position, length, chains, etc. AntigenDB, on the other hand, provides sequence, structure, classifications, etc., of various experimentally validated antigens. AntiJen contains data on many topics but, notably, it is a valuable source of data for both B-cell and T-cell epitopes.

In essence, a machine learning algorithm studies databases to find patterns that inform the analysis of novel data. The quality of the algorithm requires careful selection of the appropriate database. For example, EPSVR is a B-cell epitope prediction algorithm focused on conformational epitopes and the type of data needed for accurate conformational epitope prediction is structural (24). Their training set consisted of structures of curated antigen-antibody complexes and was tested on a sampling of structures from the Conformation Epitope Database (CED) (25, 26). By limiting their datasets to a particular type of epitope, the algorithm becomes much more accurate at predicting that type of epitope. In the case of EPSVR, linear epitopes were excluded because of their poor correlation to structural data. If linear epitopes were included, they may affect how the algorithm evaluates structural data and decrease prediction accuracy on conformational epitopes. For LBTOPE, a model focused on linear epitope prediction, only the primary sequences of epitopes for B-cell receptors and non B-cell epitopes are needed (27). Since these data are more available and less complex, it is easier to generate a dataset of tens of thousands of entries. In another example, the BCIpep database contains many linear epitopes and was used to develop ABCpred to support peptide-based vaccine design and allergy research (28, 29). Training the dataset specifically on linear epitopes resulted in a model that can strongly predict these types of epitopes.

Finally, datasets are further divided into ‘training sets’, ‘validation sets’, and ‘test sets’. Machine learning models are built from training sets; a validation sets help determine the optimal parameters or models for a given problem; and test sets are used to evaluate the machine learning models. These sets are mutually exclusive.

Even when an algorithm is provided a database robust enough to characterize the different patterns present in both linear and conformational epitopes, it still requires the tools to make use of those data. These databases are assembled into datasets with each datapoint referred to by the term ‘example’. Within each example there is information relating to ‘labels’ and ‘features’. Labels and features are the tools an algorithm uses to analyze examples.

Labels are what the algorithm learns to predict (the outcome). For instance, the label may refer to the strength of peptide-antibody binding. Machine learning can predict labels based on classification or regression analysis (30); for B-cell epitope prediction, classification involves classifying a peptide as a B-cell epitope or not, while regression analysis would assign a continuous value to the likelihood of the binding.

Features contain descriptive information about the peptide (sequence, structure, physico-chemical properties, etc.) or the parent protein it is derived from. The learning algorithm effectively learns patterns in the features that relate to a specific label value. More specifically, several algorithms use surface accessibility as the feature to predict whether an epitope is likely to be recognized by B-cells as this feature correlates well with binding strength (the label) (31, 32). Developing accurate machine learning models begins with selecting features that correlate strongly to a label (30). The features that are important depend entirely on the nature of the question being asked and the success of an algorithm depends on the design of features ( Figure 1 ). A good feature set uses all data that correlate well with the label (what is being predicted). However, there are many ways to encode the same feature that are more effective for specific machine learning algorithms ( Table 3 ).

For example, algorithms EPITOPIA and CBTOPE both use Grantham polarity and Ponnuswami polarity index to calculate polarity as a feature, while LBtope cites CBTOPE’s feature set but excludes polarity entirely (35–37). This may be because LBtope is trying to predict B-cell epitopes of variable lengths and applies a feature set called Composition-Transition-Distribution (CTD) which allows for the comparison of peptides of variable length by simplifying the residues into categories (38). The CTD divisions used by LBtope is based on a set proposed by Chinnasamy et al. that places each amino acid into groups (either 1, 2, or 3) based on certain physicochemical properties: hydrophobicity, polarizability, polarity, and Van der Waal’s volume (39). CTD characterizes the percent frequency of these groups (Composition), their spatial relationship to one another (Transition), and distribution of each group across the peptide (Distribution). The CTD feature set exemplifies an alternate way to encode certain physical features to allow for expanded function of an algorithm by enabling predictions on variable length peptides.

However, the difference in encoding methodologies between features can be much simpler than CTD; hydrophilicity can be calculated using a hydropathy table or a hydrophobicity index for proximal amino acids (40, 41). Hydropathy plots are frequently used to estimate hydrophobic and hydrophilic properties over a 20 amino acid window and, as such, may be an appropriate feature when an algorithm is focusing on predictions of peptides of a similar length (42, 43). Further, there is some variability between experimentally determined hydropathy tables, and pH will affect the hydrophilicity calculations of some amino acids.

Features that describe protein structure may be better at predicting conformational epitopes but may be unimportant when the algorithm is predicting a linear epitope. For example, surface accessibility and flexibility are more impactful features for B-cell epitope prediction software specializing in conformational epitopes. Sequence-based features, on the other hand, may be more beneficial in predicting linear epitopes. Physicochemical features such as hydrophobicity and polarity would be beneficial to both types of predictions and may not benefit one type of prediction more than another. In general, adding more features that provide new information about the data will improve accuracy but with certain limitations. If every feature represents a decision that can be made by a machine learning algorithm, adding noisy features increases the likelihood that a machine learning model uses a feature that appears effective in the training set but is inaccurate once extrapolated to the test set or new cases. This is because every machine learning model assigns a relative weight to different features. When a poorly correlative feature is used to make a prediction, it will likely increase the error. Typically, this is not a major concern unless the dataset is too small; too many features relative to the size of the dataset may impact the predictive accuracy of a model. Further, information provided by databases may limit the features that will be used to predict the label.

Another equally important aspect of a machine learning algorithm is selecting the algorithm itself. While algorithms can either be classifiers or regressors, there is much more variety to algorithms than just this single trait. We consider four common machine learning algorithms: support vector machine (SVM), neural networks, decision trees, and language models.

SVM or support vector machine is a classifier model that is accurate, simple, common, and elegant; it delineates a boundary in N-dimensional space in which data on one side of the boundary fall into one classification while data on the other side fall into a different classification. For example, if flexibility and net charge are the features for this model, every datapoint is plotted onto a grid using the flexibility score on one axis and net charge on the other. After training on the dataset, SVM algorithms will produce a line that bisects the data. If the algorithm is intended to predict B-cell epitopes, the datapoints would be classified as either “B-cell epitopes” or “non B-cell epitopes” based on this line. SVM is relatively simple computationally but is sensitive to noisy data and too many features relative to the dataset size (44, 45).

Neural networks are another common machine learning algorithm. They process data in a way that mimics how the human brain functions. Neural networks consist of a series of layers: input layer, hidden layer(s), and output layer. The input layer consists of neurons equal to the number of features and the output layer is the prediction made by the algorithm. The hidden layer(s) is defined by the programmer and is characterized by a series of weight values. The sum of all inputs multiplied by the weight values will determine whether the ‘neuron’ in the hidden layer will activate. Activated neurons will propagate data to the output layer in the same process; the hidden layer inputs are multiplied by weight values between the hidden layer and output layer. The activations in the output layer will determine the prediction of the algorithm. During training, a neural network adjusts weight values between layers to strengthen some connections and weaken others similar to how neural pathways in our brain can be reinforced or weakened. Parameters like the optimal size of the hidden layer and number of hidden layers are generally found by trail-and-error. The reason why it is difficult to predict the correct parameters is there is no way to interpret the analysis performed in these hidden layers, a black box. Two common neural networks in biological research are deep neural networks (neural networks with more than four layers) and convolutional neural networks (a neural network architecture that is effective at image processing). Convolutional neural networks can be ‘deep’. Also, neural networks rely on very large datasets and as such, trial-and-error optimization of hidden layers is very demanding computationally. While neural networks struggle with interpretability, they are highly accurate once optimized (46).

Unlike neural networks, decision trees are defined by their interpretability (47). Decision trees involve a series of nodes that each split into two ‘child’ nodes continually until they reach their pre-defined depth limit forming a pyramid shape. At each node, the algorithm imposes a criterion that splits the data into one of the two subsequent nodes. For example, if a peptide is being evaluated by the algorithm, a node could be “net charge ≥3” where all the datapoints that satisfy this criterion go to one of the child nodes and all the other datapoints go to the other child node. During training, the features at each node and the criterion are adjusted to sort the data accurately. Commonly, decision trees are used in random forests that consist of many decision trees where the output is a consensus of a majority of trees (classification), or the mean of tree outputs (regression).

Lastly, language models function by studying patterns in speech and language to predict what comes next probabilistically. These language models are also applied to analyze patterns in protein and genetic sequences to find patterns common to specific types of epitopes and produce prediction tools. Language models accurately represent the data on which they are trained but are very computationally demanding and require large datasets (48).

Since the early 2000s, the integration of machine learning into epitope prediction has increased the accuracy of those predictions and, currently, most modern algorithms use machine learning methods ( Tables 4 , 5 ). These algorithms are provided with sequence and/or structural information about a protein, and they use machine learning to predict which epitopes will be recognized by receptors on B-cells or T-cells (49).

Non-machine learning methods use scoring methods to predict epitopes. In the case of EpiJen, a 4-step process produces a score and eliminates datapoints in a stepwise manner using quantitative matrices. It first determines whether proteasome cleavage would occur, then whether TAP binding would occur, then whether MHC binding would occur, and finally whether T-cell recognition would occur. The output would be a small subsection of the data.

The early B-cell epitope prediction tools focused primarily on the prediction of linear epitopes despite linear epitopes making up only a small portion of B-cell epitopes (10). Rapidly, new tools were developed that boasted higher accuracy on a benchmark dataset or pioneered new methods of analysis; for example, ABCpred was the first B-cell prediction server based on recurrent neural networks (29). Commonly, tools specialized for linear prediction take sequence files as their input and conformational predictions require structural inputs, but more recently, tools focus on predicting both linear and conformational epitopes. In 2017, BepiPred 2.0 became a benchmark algorithm for the comparison of newer ones based on its accuracy (for 2017 standards) and capability to predict both conformational and linear epitopes (50). Other tools were released since that predict both types of epitopes, including BepiPred 3.0, SEMA-1D, and SEMA-3D (51, 52). Ideally, the best tool is one that is specialized to address a given problem and trained on a dataset containing relevant cases.

The value in predicting T-cell epitopes lies in the characterization of potential immune responses prior to treatment to avoid side effects against self-peptides (53). In T-cell epitope prediction, we encounter the dichotomy of algorithms focused primarily on either MHC class I or MHC class II binding prediction ( Table 5 ). The open binding groove of MHC class II molecules makes predictions much more complex compared to MHC class I. Specifically, this means that the ligand does not fit cleanly inside the binding groove; there is a structural element introduced and MHC class II predictive algorithms have lower accuracy compared to MHC class I algorithms. There exist several accessible online machine learning tools for predicting T-cell epitopes and/or MHC class I and class II peptide binding ( Table 5 ). The review by Peters et al. (54) explores T-cell epitope prediction and algorithm design more thoroughly.

Since features, datasets, and algorithm selection all define the specificity of a machine learning model and the type of data it is predicting, it is difficult to meaningfully compare different T-cell and B-cell epitope prediction tools. These tools will often compare themselves to similar contemporary tools based on performance to benchmark datasets, but there are issues with generalizing the accuracy of these tools. The first issue is that the test sets used to compare algorithms to one another are not the same and often tailored to the specific comparison. For example, comparing linear B-cell epitope prediction software to conformational software would be affected by whether the benchmark dataset contained primarily linear or conformational data. If the benchmark set is small, neural networks may perform worse and if it is noisy then SVM algorithms will be disadvantaged. While this makes it difficult to easily select a single ‘best’ tool, it is important to appreciate that certain tools are highly specialized to address particular problems, and the breadth of tools available should be considered in selecting which are most appropriate.

A significant advance in epitope mapping is the integration of machine learning to improve the analysis of peptide arrays ( Table 6 ). For example, Xue et al. demonstrated the utility of machine learning algorithms to address a common issue for peptide arrays: certain peptides result in high noise relative to signal in the output data (low signal-to-noise ratio). This noise ratio results in difficulties for interpreting array data and affects its utility in defining the boundaries of an epitope. Often, array data is simplified to streamline interpretation; a threshold-based approach is used to convert data into a binary compatible with classification-type algorithms (20, 55–57). The data are usually sorted into either binding or not binding. However, machine learning can predict which peptides will result in a low signal to noise ratio. By training an algorithm on a small subset of the intended peptides, Xue et al. demonstrated that their machine learning program succeeded in accurately predicting which peptides in the larger peptide set would result in a low signal-to-noise ratio (58). There are several possible ways to consider predictions of which peptides will result in significant noise: ‘noisy’ peptides could be eliminated from training sets for future algorithm development, excluded from testing all together, or the predictions could be used in interpreting results.

Machine learning is also used in docking software to predict the discrete interactions between a monoclonal antibody and a protein. Non-machine learning algorithms like Zdock and Rosetta are relatively effective and accurate antibody docking software programs, but they specialize in “local docking” which necessitates partial epitope knowledge, like relative location of the epitope within approximately 8 Å (59, 60). Machine learning allows programs like Mabtope to analyze millions of docking poses and identify those that are optimal. When compared to other methods like FRODOCK (another non-machine learning method), Mabtope reports are more accurate at predicting whether a specific residue participates in an epitope (80% accuracy compared to 35% accuracy) (61, 62). The impacts of machine learning on protein-ligand docking are discussed in a review by Yang et al. (63). Docking-based approaches provide the ability to explore specific antibody-protein interactions in silico.

In X-ray crystallography, machine learning can address the difficulty with performing this method on large proteins. When analyzing large proteins, a pre-experimental step can use B-cell mapping algorithms to focus on specific regions of a protein; testing a smaller discrete region of a protein would help address the size issue and improve the throughput. Machine learning can also predict protein crystallizability which saves time and resources (64, 65). Moreover, the study of highly variable regions of a protein is facilitated by algorithms that can predict ‘correctness’ scores (a per-residue estimate of how reliable the crystal structure is) for residue main chains and side chains to improve interpretability of crystallographic protein structures ( Table 6 ).

Cryo-EM depends on machine learning integration where a cryogenically frozen molecule is subjected to electron microscopy to produce and assemble a 3D reconstruction. Machine learning has automated particle selection, improved post-processing resolution, and map reconstruction ( Table 6 ).