Predictive Immunology: Building AI Models of the Human Immune System | Aly Azeem Khan

Predictive Immunology: Building AI Models of the Human Immune System | Aly Azeem Khan

Title history
  • Predictive Immunology: Building AI Models of the Human Immune System | Aly Azeem Khan — observed on 20261104-0952 UTC (YouTube)

by Aly Azeem Khan (University of Chicago, Biohub Chicago) · View the channel on PSI

Life & Natural Systems Biology PSALife sciences: general issuesPSAXComputational biology
🎙 Aly Azeem Khan (University of Chicago, Biohub Chicago) 👥 879 📅 November 4, 2026 ⏱ 41 min 👁 2 📄 original study 🧭 2026-11-04
Available in: English (current) Français

Keywords

immune systemAI modelsT-cell receptorgeneticsmultitask learning

Summary

The talk by Professor Aly Azeem Khan presents recent work on applying AI to immunology, focusing on building predictive models of the human immune system. He begins by drawing parallels between breakthroughs in AI and immunology over the past decade, highlighting advances like transformers and checkpoint immunotherapy. The core of the talk describes a multitask neural network model trained on genomic data from over 3 million individuals across 18 autoimmune and inflammatory diseases. This model uses neighborhood attention to capture interactions within DNA and improves disease risk prediction compared to classical linear approaches, with notable scaling performance. Khan explains that multitask learning across diseases yields better predictions for individual conditions, suggesting shared biological mechanisms. He then discusses the fundamental challenge of predicting T-cell receptor (TCR) and peptide-MHC interactions, crucial for understanding autoimmunity, cancer immunity, and antiviral responses. He emphasizes the immense combinatorial complexity of these interactions and proposes using protein language models and large datasets to predict TCR recognition. Finally, he outlines a closed-loop approach where AI proposes experiments, results refine models, driving both prediction and design of immune function, with implications for therapies and diagnostics.

186 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides high-value information by presenting concrete, published results on AI-driven immunology, including a novel multitask neural network for genetic risk prediction and early efforts on TCR-antigen recognition. The argumentation is solid, grounded in peer-reviewed studies, and clearly explains the rationale for using modern AI architectures to overcome limitations of classical statistical genetics. The speaker supports his claims with empirical scaling plots and discusses both successes and open problems, demonstrating intellectual honesty and rigorous scientific reasoning.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates strong scientific rigour, referencing established works like the transformer paper ‘Attention is All You Need’ and GWAS studies, while the speaker’s own publication (about a year prior) is presented with technical detail. The description provides the ADIA Lab symposium page as a supplementary source. The title accurately matches the content, as the talk is entirely focused on building AI models for immune system prediction and design. No discrepancies or overclaims are apparent.

167 words

Title / Content Match

The title accurately reflects the content, as the talk focuses on building AI models of the human immune system, with specific examples and results.

Quality & Reliability

8/10

Speaker is a recognized academic, presenting recent peer-reviewed work from his group; the talk includes technical details and empirical results, demonstrating high rigour.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Contribution & Novelties

The contribution of this talk lies in demonstrating the application of modern AI architectures (transformers, multitask learning, protein language models) to fundamental problems in immunology, such as genetic risk prediction and TCR-antigen recognition. The speaker presents recent advances from his group, including a scalable neural network that outperforms classical polygenic risk scores and a framework for predicting and designing immune receptors. This represents a novel integration of computational and experimental biology, pushing toward programmable immune systems.

Pour aller plus loin :

  • T-cell receptor — Central molecule in adaptive immunity, directly relevant to the talk’s focus on TCR-antigen interactions.
  • HLA (Human Leukocyte Antigen) — Key genomic region discussed in the talk, strongly associated with autoimmune diseases.
  • Polygenic risk score — Classical method compared against the AI models, providing context for the improvements.

131 words

Radar Profile

The radar profile shows balanced scores across the dimensions, with slightly lower technical depth compared to information quantity and reliability, reflecting the talk's emphasis on conveying concrete results and implications rather than deep algorithmic details.

Reliability 8/10