Current challenges for unseen-epitope TCR interaction prediction and a new perspective derived from image classification

Abstract

The prediction of epitope recognition by T-cell receptors (TCRs) has seen many advancements in recent years, with several methods now available that can predict recognition for a specific set of epitopes. However, the generic case of evaluating all possible TCR-epitope pairs remains challenging, mainly due to the high diversity of the interacting sequences and the limited amount of currently available training data. In this work, we provide an overview of the current state of this unsolved problem. First, we examine appropriate validation strategies to accurately assess the generalization performance of generic TCR-epitope recognition models when applied to both seen and unseen epitopes. In addition, we present a novel feature representation approach, which we call ImRex (interaction map recognition). This approach is based on the pairwise combination of physicochemical properties of the individual amino acids in the CDR3 and epitope sequences, which provides a convolutional neural network with the combined representation of both sequences. Lastly, we highlight various challenges that are specific to TCR-epitope data and that can adversely affect model performance. These include the issue of selecting negative data, the imbalanced epitope distribution of curated TCR-epitope datasets and the potential exchangeability of TCR alpha and beta chains. Our results indicate that while extrapolation to unseen epitopes remains a difficult challenge, ImRex makes this feasible for a subset of epitopes that are not too dissimilar from the training data. We show that appropriate feature engineering methods and rigorous benchmark standards are required to create and validate TCR-epitope predictive models.

TITAN: T-cell receptor specificity prediction with bimodal attention networks

“…This can be achieved by shuffling the sequences, thereby associating TCRs with epitopes that they have not been shown to bind. Due to the low probability of a randomly drawn TCR binding a specific epitope, this manner of generating negative samples is established in the field ( Fischer et al, 2020 ; Moris et al, 2020 ). It has also been shown to limit overestimation of performances in comparison to adding additional naive TCR sequences from other sources ( Moris et al, 2020 ).…”

“…Due to the low probability of a randomly drawn TCR binding a specific epitope, this manner of generating negative samples is established in the field ( Fischer et al, 2020 ; Moris et al, 2020 ). It has also been shown to limit overestimation of performances in comparison to adding additional naive TCR sequences from other sources ( Moris et al, 2020 ). Furthermore, by shuffling the pairing of TCRs and epitopes, we can match the number of negative examples to that of positive examples for each TCR, avoiding unbalanced datasets.…”

“…These have the potential to predict binding of any TCR–epitope pair, opening the door to the development of models that can generalize to both, unseen TCRs and epitopes. Current models show moderate performance on test data containing epitopes already encountered in training, but cannot extrapolate to unseen epitopes ( Jurtz et al, 2018 ; Moris et al, 2020 ; Springer et al, 2019 ).…”

immuneML: an ecosystem for machine learning analysis of adaptive immune receptor repertoires

“…Together this information shapes the foundation for AIRR-based diagnostics 6,[10][11][12][13] . Similarly, s equence -based prediction of antigen and epitope binding is of fundamental importance for AIR-based therapeutics discovery and engineering [14][15][16][17][18][19][20][21][22][23][24] . In this manuscript, the term AIRR signifies both AIRs and AIRRs (a collection of AIRs) if not specified otherwise.…”

“…Briefly, (i) ~10 8 -10 10 distinct AIRs exist in a given individual at any one time [29][30][31] , with little overlap among individuals, necessitating encodings that allow detection of predictive patterns. These shared patterns may correspond to full-length AIRs 6 or subsequences 15 alternative representations thereof 11,12,16,17,21,[32][33][34] . (ii) In repertoire-based ML, the patterns relevant to any immune state may be as rare as one antigen-binding AIR per million lymphocytes in a repertoire 35 translating into a very low rate of relevant sequences per repertoire (low witness rate) 11,36,37 .…”