Aly Azeem Khan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Aly Azeem Khan has a total of 183 co-authors. Here are the top co-authors by number of shared publications:
- Derek Reiman
Toyota Technological Institute at Chicago
9 shared publications - Robert L. Grossman
Center for Translational Data Science, University of Chicago, Chicago, USA
8 shared publications - Steven Song
Department of Computer Science, University of Chicago, Chicago, Illinois
6 shared publications - Christopher R. Shea
Department of Medicine Section of Dermatology, University of Chicago Chicago Illinois USA
5 shared publications - Habibul Ahsan
Institute for Population and Precision Health (IPPH), University of Chicago, Chicago, IL 60637, USA
5 shared publications - Maria Argos
Boston University
5 shared publications
Latest Publications
- Open Access
- 13 May 2026
- Published by Springer Nature in Nature Biotechnology
Deep peptide recognition profiling decodes TCR specificity and enables disease-associated antigen discovery
Predicting T cell receptor (TCR) specificity on the basis of sequence is challenging because TCRs of similar sequence can recognize entirely different antigens, whereas TCRs of different sequence can recognize the same antigens. Here we present a system that integrates high-throughput yeast display with fine-tuned protein language models (pLMs) to generate deep peptide recognition profiles (PRPs) for individual TCRs, each detailing binding against millions of peptides. We provide detailed PRPs for a panel of HLA-B*27:05-restricted TCRs from persons with ankylosing spondylitis and acute anterior uveitis that almost exclusively recognize peptides through CDR3β. pLMs trained on these PRPs outperform AlphaFold3 and tFold-TCR in predicting T cell activation. We discover and validate novel candidate autoantigens, demonstrate that model generalization to new TCRs correlates with functional distance (PRP divergence) rather than sequence similarity and introduce a model-intrinsic uncertainty metric to quantify prediction confidence. This system and its associated PRP datasets offer a scalable approach to mapping TCR recognition, accelerating antigen discovery and guiding TCR engineering.
- 13 May 2026
- Published by Springer Nature in Nature Biotechnology
Open AccessDeep peptide recognition profiling decodes TCR specificity and enables disease-associated antigen discovery
Predicting T cell receptor (TCR) specificity on the basis of sequence is challenging because TCRs of similar sequence can recognize entirely different antigens, whereas TCRs of different sequence can recognize the same antigens. Here we present a system that integrates high-throughput yeast display with fine-tuned protein language models (pLMs) to generate deep peptide recognition profiles (PRPs) for individual TCRs, each detailing binding against millions of peptides. We provide detailed PRPs for a panel of HLA-B*27:05-restricted TCRs from persons with ankylosing spondylitis and acute anterior uveitis that almost exclusively recognize peptides through CDR3β. pLMs trained on these PRPs outperform AlphaFold3 and tFold-TCR in predicting T cell activation. We discover and validate novel candidate autoantigens, demonstrate that model generalization to new TCRs correlates with functional distance (PRP divergence) rather than sequence similarity and introduce a model-intrinsic uncertainty metric to quantify prediction confidence. This system and its associated PRP datasets offer a scalable approach to mapping TCR recognition, accelerating antigen discovery and guiding TCR engineering.
- Top Cited PapersOpen Access
Virtual Cells Need Context, Not Just Scale
The intersection of AI and biology has entered a phase of explosive growth, driven by the ambition to build “Virtual Cells” or computational models capable of predicting cellular responses to any perturbation. Following the success of structural biology (e.g., AlphaFold) and large language models, the field has converged on training massive, high-capacity models on large-scale single-cell data. This position paper argues that scaling model capacity is insufficient to solve the Virtual Cell problem because the primary failure mode is a lack of adequate coverage over diverse biological contexts , not insufficient model expressivity. We support this claim by reviewing recent studies showing that simple baselines perform on par with sophisticated architectures within a given biological context, and current models fail to consistently generalize across contexts. We connect this finding to the causal inference literature on transportability and contrast it with domains where scaling has succeeded. We substantiate our argument through analysis of a state-of-the-art model on a 22-million-cell immunology dataset. We conclude that the community faces a causal transport problem that cannot be solved by accumulating more data from the same distributions. Instead, we argue that contextual diversity and causal representation learning deserve increased emphasis, complementing ongoing scaling of model capacity and data volume.
Top Cited PapersOpen AccessVirtual Cells Need Context, Not Just Scale
The intersection of AI and biology has entered a phase of explosive growth, driven by the ambition to build “Virtual Cells” or computational models capable of predicting cellular responses to any perturbation. Following the success of structural biology (e.g., AlphaFold) and large language models, the field has converged on training massive, high-capacity models on large-scale single-cell data. This position paper argues that scaling model capacity is insufficient to solve the Virtual Cell problem because the primary failure mode is a lack of adequate coverage over diverse biological contexts , not insufficient model expressivity. We support this claim by reviewing recent studies showing that simple baselines perform on par with sophisticated architectures within a given biological context, and current models fail to consistently generalize across contexts. We connect this finding to the causal inference literature on transportability and contrast it with domains where scaling has succeeded. We substantiate our argument through analysis of a state-of-the-art model on a 22-million-cell immunology dataset. We conclude that the community faces a causal transport problem that cannot be solved by accumulating more data from the same distributions. Instead, we argue that contextual diversity and causal representation learning deserve increased emphasis, complementing ongoing scaling of model capacity and data volume.
- Open Access
DecoderTCR: Compositional Pretraining and Entropy-Guided Decoding for TCR-pMHC Interactions
Modeling recognition between T-cell receptors (TCRs) and peptide-MHC (pMHC) complexes is a fundamental challenge in computational immunology, constrained by sparse paired interaction data relative to abundant unpaired sequences. We introduce DecoderTCR, a masked language model framework that addresses this through two contributions: (1) a compositional continual pre-training curriculum that learns component representations from marginal data before refining cross-chain dependencies from limited pairs, and (2) Iterative Entropy-Guided Refinement (IEGR), a non-autoregressive decoding algorithm that resolves high-confidence positions first to provide context for uncertain regions. On held-out benchmarks, DecoderTCR achieves 0.96 AUROC for zero-shot pMHC binding prediction and 0.76 AUROC for epitope-specific TCR recognition, approaching supervised baselines without epitope-specific training. Learned representations recover structural contacts without coordinate supervision, and generated sequences exhibit realistic recombination statistics. Experimental validation reveals a prediction-generation gap: strong discrimination does not yet yield reliable generation, highlighting an open challenge for the field.
Open AccessDecoderTCR: Compositional Pretraining and Entropy-Guided Decoding for TCR-pMHC Interactions
Modeling recognition between T-cell receptors (TCRs) and peptide-MHC (pMHC) complexes is a fundamental challenge in computational immunology, constrained by sparse paired interaction data relative to abundant unpaired sequences. We introduce DecoderTCR, a masked language model framework that addresses this through two contributions: (1) a compositional continual pre-training curriculum that learns component representations from marginal data before refining cross-chain dependencies from limited pairs, and (2) Iterative Entropy-Guided Refinement (IEGR), a non-autoregressive decoding algorithm that resolves high-confidence positions first to provide context for uncertain regions. On held-out benchmarks, DecoderTCR achieves 0.96 AUROC for zero-shot pMHC binding prediction and 0.76 AUROC for epitope-specific TCR recognition, approaching supervised baselines without epitope-specific training. Learned representations recover structural contacts without coordinate supervision, and generated sequences exhibit realistic recombination statistics. Experimental validation reveals a prediction-generation gap: strong discrimination does not yet yield reliable generation, highlighting an open challenge for the field.
- 1 January 2026
- Published by American Society of Clinical Oncology (ASCO) in JCO Clinical Cancer Informatics
- Vol. 10 (1) , e2500041
Hybrid Computer Vision Model to Predict Lung Cancer in Diverse Populations
PURPOSE: Disparities in lung cancer incidence exist in Black populations, and screening criteria underserve Black populations due to disparately elevated risk in the screening-eligible population. Prediction models that integrate clinical and imaging-based features to individualize lung cancer risk are a potential means to mitigate these disparities. METHODS: This multicenter (National Lung Screening Trial [NLST]) and catchment population–based (University of Illinois Health [UIH], urban and suburban Cook County) cross-sectional study used participants at risk of lung cancer with available lung computed tomography (CT) imaging and follow-up between the years 2015 and 2024. In all, 53,452 in NLST and 11,654 in UIH were included on the basis of age and tobacco use–based risk factors for lung cancer. Cohorts were used for training and testing of deep and machine learning models using clinical features alone or combined with CT image features (hybrid computer vision). RESULTS: An optimized seven-feature clinical model achieved receiver operating characteristic (ROC)-AUC values ranging from 0.64 to 0.67 in NLST and 0.60 to 0.65 in UIH cohorts across multiple years. Incorporation of imaging features to form a hybrid computer vision model significantly improved ROC-AUC values to 0.78-0.91 in NLST but deteriorated in UIH with ROC-AUC values of 0.68-0.80, attributable to Black participants where ROC-AUC values ranged from 0.63 to 0.72 across multiple years. Retraining the hybrid computer vision model by incorporating Black and other participants from the UIH cohort improved performance with ROC-AUC values of 0.70-0.87 in a held-out UIH test set. CONCLUSION: Hybrid computer vision predicted risk with improved accuracy compared with clinical risk models alone. However, potential biases in image training data reduced model generalizability in Black participants. Performance was improved upon retraining with a subset of the UIH cohort, suggesting that inclusive training and validation data sets can minimize racial disparities. Future studies incorporating vision models trained on representative data sets may demonstrate improved health equity upon clinical use.
- 1 January 2026
- Published by American Society of Clinical Oncology (ASCO) in JCO Clinical Cancer Informatics
- Vol. 10 (1) , e2500041
Hybrid Computer Vision Model to Predict Lung Cancer in Diverse Populations
PURPOSE: Disparities in lung cancer incidence exist in Black populations, and screening criteria underserve Black populations due to disparately elevated risk in the screening-eligible population. Prediction models that integrate clinical and imaging-based features to individualize lung cancer risk are a potential means to mitigate these disparities. METHODS: This multicenter (National Lung Screening Trial [NLST]) and catchment population–based (University of Illinois Health [UIH], urban and suburban Cook County) cross-sectional study used participants at risk of lung cancer with available lung computed tomography (CT) imaging and follow-up between the years 2015 and 2024. In all, 53,452 in NLST and 11,654 in UIH were included on the basis of age and tobacco use–based risk factors for lung cancer. Cohorts were used for training and testing of deep and machine learning models using clinical features alone or combined with CT image features (hybrid computer vision). RESULTS: An optimized seven-feature clinical model achieved receiver operating characteristic (ROC)-AUC values ranging from 0.64 to 0.67 in NLST and 0.60 to 0.65 in UIH cohorts across multiple years. Incorporation of imaging features to form a hybrid computer vision model significantly improved ROC-AUC values to 0.78-0.91 in NLST but deteriorated in UIH with ROC-AUC values of 0.68-0.80, attributable to Black participants where ROC-AUC values ranged from 0.63 to 0.72 across multiple years. Retraining the hybrid computer vision model by incorporating Black and other participants from the UIH cohort improved performance with ROC-AUC values of 0.70-0.87 in a held-out UIH test set. CONCLUSION: Hybrid computer vision predicted risk with improved accuracy compared with clinical risk models alone. However, potential biases in image training data reduced model generalizability in Black participants. Performance was improved upon retraining with a subset of the UIH cohort, suggesting that inclusive training and validation data sets can minimize racial disparities. Future studies incorporating vision models trained on representative data sets may demonstrate improved health equity upon clinical use.
- 14 December 2025
- Published by World Scientific Pub Co Pte Ltd in Biocomputing 2020
- Vol. 31, 480-493
Large language models identify causal genes in complex trait GWAS
Pinpointing causal genes at genome-wide association study (GWAS) loci remains a major bottleneck. Existing literature-mining approaches are often limited in accuracy and scalability. We show that large language models (LLMs) can accurately prioritize likely causal genes at GWAS loci. We systematically evaluated several widely available general-purpose LLMs against benchmark datasets of high-confidence causal genes, including a unique set from 23 unpublished GWAS. Our results demonstrate that LLMs outperform or match current state-of-the-art methods and, crucially, exhibit robust performance on novel loci not previously linked to traits, underscoring their generalizability. Moreover, when integrated with existing methods, LLMs substantially enhance overall performance. This work establishes LLMs as an accurate, scalable, and broadly generalizable approach to accelerate causal gene identification in complex traits.
- 14 December 2025
- Published by World Scientific Pub Co Pte Ltd in Biocomputing 2020
- Vol. 31, 480-493
Large language models identify causal genes in complex trait GWAS
Pinpointing causal genes at genome-wide association study (GWAS) loci remains a major bottleneck. Existing literature-mining approaches are often limited in accuracy and scalability. We show that large language models (LLMs) can accurately prioritize likely causal genes at GWAS loci. We systematically evaluated several widely available general-purpose LLMs against benchmark datasets of high-confidence causal genes, including a unique set from 23 unpublished GWAS. Our results demonstrate that LLMs outperform or match current state-of-the-art methods and, crucially, exhibit robust performance on novel loci not previously linked to traits, underscoring their generalizability. Moreover, when integrated with existing methods, LLMs substantially enhance overall performance. This work establishes LLMs as an accurate, scalable, and broadly generalizable approach to accelerate causal gene identification in complex traits.
- 14 December 2025
- Published by World Scientific Pub Co Pte Ltd in Biocomputing 2020
- Vol. 31, 600-611
Integrating Polygenic Risk Improves Generative Forecasting of Disease Trajectories
Predicting the longitudinal sequence of diseases an individual will develop over their lifetime is a central challenge in medicine. While recent AI models can process health histories, they have been limited by cohort size and the omission of genetic data. Here we introduce the Next Health Event (NHE) model, a generative transformer trained on the health trajectories of 7.1 million research participants. By using a transformer architecture to integrate demographic data, longitudinal BMI, and polygenic risk scores (PRS) for 297 traits with sequential health history, NHE significantly outperforms baseline models, including XGBoost with the same inputs, in predicting the next diagnosis across 129 conditions (Top-1 accuracy 25.5% vs. 22.3%). Systematic ablation studies reveal that both PRS and longitudinal BMI provide substantial, non-redundant predictive power, whereas self-reported lifestyle information offers limited additional value. The model's predictive accuracy is the same when forecasting prospectively reported incident outcomes vs. combined prospectively and retrospectively reported outcomes (AUROC 0.917), demonstrating its utility for real-world risk assessment. By uniting large-scale health histories with genetics, our work establishes a new framework for predictive health and demonstrates that generative models can effectively forecast individual disease pathways.
- 14 December 2025
- Published by World Scientific Pub Co Pte Ltd in Biocomputing 2020
- Vol. 31, 600-611
Integrating Polygenic Risk Improves Generative Forecasting of Disease Trajectories
Predicting the longitudinal sequence of diseases an individual will develop over their lifetime is a central challenge in medicine. While recent AI models can process health histories, they have been limited by cohort size and the omission of genetic data. Here we introduce the Next Health Event (NHE) model, a generative transformer trained on the health trajectories of 7.1 million research participants. By using a transformer architecture to integrate demographic data, longitudinal BMI, and polygenic risk scores (PRS) for 297 traits with sequential health history, NHE significantly outperforms baseline models, including XGBoost with the same inputs, in predicting the next diagnosis across 129 conditions (Top-1 accuracy 25.5% vs. 22.3%). Systematic ablation studies reveal that both PRS and longitudinal BMI provide substantial, non-redundant predictive power, whereas self-reported lifestyle information offers limited additional value. The model's predictive accuracy is the same when forecasting prospectively reported incident outcomes vs. combined prospectively and retrospectively reported outcomes (AUROC 0.917), demonstrating its utility for real-world risk assessment. By uniting large-scale health histories with genetics, our work establishes a new framework for predictive health and demonstrates that generative models can effectively forecast individual disease pathways.
- Open Access
- 16 November 2025
- Published by F1000 Research Ltd
- Vol. 14
Large language models identify causal genes in complex trait GWAS
Pinpointing causal genes at genome-wide association study (GWAS) loci remains a major bottleneck. Existing literature-mining approaches are often limited in accuracy and scalability. We show that large language models (LLMs) can accurately prioritize likely causal genes at GWAS loci. We systematically evaluated several widely available general-purpose LLMs against benchmark datasets of high-confidence causal genes, including a unique set from 23 unpublished GWAS. Our results demonstrate that LLMs outperform or match current state-of-the-art methods and, crucially, exhibit robust performance on novel loci not previously linked to traits, underscoring their generalizability. Moreover, when integrated with existing methods, LLMs substantially enhance overall performance. This work establishes LLMs as an accurate, scalable, and broadly generalizable approach to accelerate causal gene identification in complex traits.
- 16 November 2025
- Published by F1000 Research Ltd
- Vol. 14
Open AccessLarge language models identify causal genes in complex trait GWAS
Pinpointing causal genes at genome-wide association study (GWAS) loci remains a major bottleneck. Existing literature-mining approaches are often limited in accuracy and scalability. We show that large language models (LLMs) can accurately prioritize likely causal genes at GWAS loci. We systematically evaluated several widely available general-purpose LLMs against benchmark datasets of high-confidence causal genes, including a unique set from 23 unpublished GWAS. Our results demonstrate that LLMs outperform or match current state-of-the-art methods and, crucially, exhibit robust performance on novel loci not previously linked to traits, underscoring their generalizability. Moreover, when integrated with existing methods, LLMs substantially enhance overall performance. This work establishes LLMs as an accurate, scalable, and broadly generalizable approach to accelerate causal gene identification in complex traits.
- Open Access
Scalable Single-Cell Gene Expression Generation with Latent Diffusion Models
Computational modeling of single-cell gene expression is crucial for understanding cellular processes, but generating realistic expression profiles remains a major challenge. This difficulty arises from the count nature of gene expression data and complex latent dependencies among genes. Existing generative models often impose artificial gene orderings or rely on shallow neural network architectures. We introduce a scalable latent diffusion model for single-cell gene expression data, which we refer to as scLDM, that respects the fundamental exchangeability property of the data. Our VAE uses fixed-size latent variables leveraging a unified Multi-head Cross-Attention Block (MCAB) architecture, which serves dual roles: permutation-invariant pooling in the encoder and permutation-equivariant unpooling in the decoder. We enhance this framework by replacing the Gaussian prior with a latent diffusion model using Diffusion Transformers and linear interpolants, enabling high-quality generation with multi-conditional classifier-free guidance. We show its superior performance in a variety of experiments for both observational and perturbational single-cell data, as well as downstream tasks like cell-level classification.
Open AccessScalable Single-Cell Gene Expression Generation with Latent Diffusion Models
Computational modeling of single-cell gene expression is crucial for understanding cellular processes, but generating realistic expression profiles remains a major challenge. This difficulty arises from the count nature of gene expression data and complex latent dependencies among genes. Existing generative models often impose artificial gene orderings or rely on shallow neural network architectures. We introduce a scalable latent diffusion model for single-cell gene expression data, which we refer to as scLDM, that respects the fundamental exchangeability property of the data. Our VAE uses fixed-size latent variables leveraging a unified Multi-head Cross-Attention Block (MCAB) architecture, which serves dual roles: permutation-invariant pooling in the encoder and permutation-equivariant unpooling in the decoder. We enhance this framework by replacing the Gaussian prior with a latent diffusion model using Diffusion Transformers and linear interpolants, enabling high-quality generation with multi-conditional classifier-free guidance. We show its superior performance in a variety of experiments for both observational and perturbational single-cell data, as well as downstream tasks like cell-level classification.
- Open Access
- 1 July 2025
- Published by American Association for Cancer Research (AACR)
Table 1 from AI-assisted Diagnosis of Nonmelanoma Skin Cancer in Resource-Limited Settings
Important concepts and terms used in this study.
- 1 July 2025
- Published by American Association for Cancer Research (AACR)
Open AccessTable 1 from AI-assisted Diagnosis of Nonmelanoma Skin Cancer in Resource-Limited Settings
Important concepts and terms used in this study.
- Open Access
- 1 July 2025
- Published by American Association for Cancer Research (AACR)
Table 2 from AI-assisted Diagnosis of Nonmelanoma Skin Cancer in Resource-Limited Settings
Dataset characteristics broken down by NMSC subtype at the individual patient level and the biopsy level.
- 1 July 2025
- Published by American Association for Cancer Research (AACR)
Open AccessTable 2 from AI-assisted Diagnosis of Nonmelanoma Skin Cancer in Resource-Limited Settings
Dataset characteristics broken down by NMSC subtype at the individual patient level and the biopsy level.