Summary / Even high-performing machine learning models can have problems when deployed in a real-world setting if the data used to train and test the model contains biases. TCR–epitope binding prediction for novel epitopes is a very important but yet unsolved problem in immunology. In this article, we describe how the technique used to create negative data for the TCR–epitope interaction prediction task can lead to a strong bias and makes that the performance drops to random when tested in a more realistic scenario.
Background The recognition of an epitope by a T-cell receptor (TCR) is crucial for eliminating pathogens and establishing immunological memory. Prediction of the binding of any TCR–epitope pair is still a challenging task, especially with novel epitopes, because the underlying patterns that drive the recognition are still largely unknown to both domain experts and machine learning models. Results The binding of a TCR and epitope sequence can only occur when amino acids from both sequences are in close contact with each other. We analyze the distance between interacting molecules of the TCR and epitope sequences and compare this to the amino acids that are important for TCR–epitope prediction models. Important residues are determined by using interpretable deep learning techniques or, more specifically, feature attribution extraction methods, on two state-of-the-art TCR–epitope prediction models: ImRex and TITAN. Highlighting feature attributions on the molecular complex reveals additional insights to the domain expert about why the prediction was made and can offer novel insights into the factors that determine TCR affinity on a molecular level. We also show which residues of the TCR and epitope sequences determine binding prediction for ImRex and TITAN and use those to explain model performance. Conclusions Extracting feature attributions is a useful way to verify your model and data for challenging problems where small hard-to-detect problems can accumulate to inaccurate results.
Mass spectrometry-based proteomics relies on accurate peptide property prediction models to enhance peptide identification and characterization, especially when dealing with peptidoforms. However, current approaches are limited in their ability to generalize to peptides with novel post-translational modifications (PTMs) due to insufficient training data. To address this challenge, we introduce MoSTERT (Molecular Structure Transformer Encoder for Retention Time prediction) and its enhanced variant, MoSTERT-2S, two transformer-based models designed for zero-shot prediction of retention times of peptides with unseen PTMs. Unlike conventional models, MoSTERT encodes peptide residues at the molecular structure level, allowing it to handle diverse PTMs. MoSTERT-2S further improves accuracy by employing a two-step strategy: first predicting the retention time of the unmodified peptide, then estimating the retention time shift induced by the PTMs. Evaluation on an external dataset demonstrates that MoSTERT-2S achieves state-of-the-art performance, reducing prediction errors compared to existing methods. Its ability to accurately predict retention times for peptides with a wide variety of PTMs not seen during training highlights its potential for advancing proteomic workflows analyzing proteoforms and protein modifications.