AI-driven modelling of the immune system

Prof María Rodríguez Martínez ( Department of Biomedical Informatics and Data Science & Center for Systems and Engineering Immunology, Yale School of Medicine )

In recent years, deep learning models have driven groundbreaking advancements across various areas of computational biology and bioinformatics. However, their "black box" nature often obscures potential data biases, incorrect assumptions, and software errors. In this talk, I will present recent work from my team, where we developed novel deep learning approaches to better understand the immune system, with a particular focus on enhancing the interpretability, trustworthiness, and reliability of predictive modeling.

Focusing first on T cell receptors (TCRs) and the challenge of predicting their binding and specificity, I will introduce TITAN (Tcr epITope bimodal Attention Network), a neural network that jointly encodes TCR sequences and epitopes, allowing us to independently evaluate generalization capabilities to unseen TCRs and/or epitopes. TITAN leverages attention mechanisms to identify critical amino acids for prediction, offering insights into situations where data limitations hinder model generalization.

Next, I will discuss how protein language models (PLMs), including generalist models trained on extensive amino acid datasets and domain-specific models fine-tuned for immune receptor sequences, achieve performance comparable to traditional deep learning methods with significantly reduced data requirements. Despite this, interpretability remains a challenge for transformer-based PLMs. To address this, we developed DECODE, an interpretable pipeline that extracts human-readable binding rules from black-box models, enabling robust TCR specificity predictions.

Finally, I will present recent work on pediatric acute myeloid leukemia (AML), where we developed a novel computational tool to classify malignant cells using single-cell flow cytometry data. Coupling a classifier with an autoencoder for anomaly detection, we achieved 90% accuracy in identifying malignant blasts and gained new insights into their developmental stages. This method revealed significant immunophenotypic changes between diagnosis and relapse, underscoring the transformative potential of AI in advancing immunology and cancer research.

I will close by discussing how the integration of AI and mechanistic models is necessary to tackle many current computational challenges and enable the personalized design of new therapeutic interventions.