Professor David Gfeller discusses interpretable AI, TCR specificity profiles, cross-reactivity, chain pairing, and why understanding T cell recognition may matter more than black-box prediction.
In this episode of Immunity by Design, Prof Hashem Koohy speaks with Professor David Gfeller (University of Lausanne & Ludwig Institute for Cancer Research) about the emerging rules governing T cell recognition and how interpretable AI may reshape computational immunology.
The conversation explores three recent studies from David Gfeller’s group focused on TCR specificity profiles (TSPs), probabilistic modelling of TCR recognition, and the role of alpha-beta chain pairing in T cell specificity prediction.
Together, these studies challenge several widely held assumptions in the field, including the dominance of black-box sequence prediction models, the centrality of paired TCR sequencing for prediction performance, and the underappreciated role of V and J gene usage in shaping antigen recognition.
Key topics discussed include:
Why sequence-based TCR prediction struggles to generalize across unseen epitopes
o The biological and statistical foundations of TCR specificity profiles (TSPs)
o TEMPO: an interpretable probabilistic framework for TCR epitope prediction
o The importance of baseline repertoires in modelling T cell specificity
o Structural insights into V/J gene usage and TCR recognition
o Why alpha-beta chain pairing may contribute less predictive information than expected o Cross-reactivity, neoantigen recognition, and off-target toxicity prediction
o The translational implications for TCR-based immunotherapies
o Open science, preprints, and the future of AI-driven immunology research
o The importance of interpretability in the era of increasingly powerful AI systems
The episode also explores broader questions around multidisciplinary research culture, scientific training, and how computational and experimental scientists can collaborate more effectively.
Music: “Minimal” by paulyudin, sourced from Pixabay and used under the Pixabay Content License (royalty-free for commercial use).
