Athan Z. Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Athan Z. Li has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
- Ruishan Liu
Department of Computer Science, University of Southern California, Los Angeles, CA 90089, USA
3 shared publications - Yuxuan Du
Department of Electrical Engineering, University of Texas at San Antonio, San Antonio, TX 78249, USA
3 shared publications Chirag Gupta
Waisman Center, University of Wisconsin-Madison, Madison, WI, USA
2 shared publications- Daifeng Wang
Waisman Center, University of Wisconsin-Madison, Madison, USA
2 shared publications - Donghoon Lee
Icahn School of Medicine at Mount Sinai
2 shared publications - Gabriel E. Hoffman
Icahn School of Medicine at Mount Sinai
2 shared publications
Latest Publications
- Open Access
A Taxonomy-Informed Sparse DNA Foundation Model for Microbial Genomics
: Microorganisms are crucial to Earth’s ecosystems, with genomes encoding functions important to agriculture, biotechnology and human health. Although genomic language models have advanced DNA representation learning, the extensive diversity of microorganisms and highly imbalanced taxonomic composition of existing pretraining corpora pose challenges for effective microbial genomic sequence modeling. Here we present MicroGlot, a taxonomy-informed microbial DNA foundation model pretrained on 3.70 million sequences comprising 378.3 billion nucleotides from 99,700 species. MicroGlot represents hierarchical relationships among taxa using hyperbolic embeddings and integrates this knowledge into a sparse mixture-of-experts architecture. Probing frozen embeddings across layers showed that MicroGlot captures both microbial phenotypic traits and taxonomic identity. Compared with a taxonomy-ablated variant pretrained under the same scheme, incorporating taxonomic knowledge consistently improved representation quality across model depth. Combining sparse computation with efficient architectural and training techniques, MicroGlot achieved competitive probing and fine-tuning performance with low computational requirements. Species-level expert-routing fingerprints aligned more closely with taxonomic groups than tetranucleotide-frequency profiles, indicating that expert usage recapitulates taxonomic structure. These results show that combining sparse architecture with taxonomic knowledge enables efficient and biologically informative genomic language modeling across diverse microbial taxa. MicroGlot is publicly available at https://huggingface.co/athanzli/MicroGlot .
Open AccessA Taxonomy-Informed Sparse DNA Foundation Model for Microbial Genomics
: Microorganisms are crucial to Earth’s ecosystems, with genomes encoding functions important to agriculture, biotechnology and human health. Although genomic language models have advanced DNA representation learning, the extensive diversity of microorganisms and highly imbalanced taxonomic composition of existing pretraining corpora pose challenges for effective microbial genomic sequence modeling. Here we present MicroGlot, a taxonomy-informed microbial DNA foundation model pretrained on 3.70 million sequences comprising 378.3 billion nucleotides from 99,700 species. MicroGlot represents hierarchical relationships among taxa using hyperbolic embeddings and integrates this knowledge into a sparse mixture-of-experts architecture. Probing frozen embeddings across layers showed that MicroGlot captures both microbial phenotypic traits and taxonomic identity. Compared with a taxonomy-ablated variant pretrained under the same scheme, incorporating taxonomic knowledge consistently improved representation quality across model depth. Combining sparse computation with efficient architectural and training techniques, MicroGlot achieved competitive probing and fine-tuning performance with low computational requirements. Species-level expert-routing fingerprints aligned more closely with taxonomic groups than tetranucleotide-frequency profiles, indicating that expert usage recapitulates taxonomic structure. These results show that combining sparse architecture with taxonomic knowledge enables efficient and biologically informative genomic language modeling across diverse microbial taxa. MicroGlot is publicly available at https://huggingface.co/athanzli/MicroGlot .
- Open Access
- 23 September 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (10) , 3682-3694
AI-based characterization of Alzheimer’s disease phenotypes from population-scale single-cell data
The complexity of Alzheimer’s disease (AD) manifests in diverse clinical phenotypes, including cognitive impairment and neuropsychiatric symptoms. However, the etiology of these phenotypes remains elusive. To address this, the PsychAD project generated a population-level single-nucleus RNA sequencing dataset comprising over 6 million nuclei from the prefrontal cortex of >1,000 individual brains, covering a variety of disease phenotypes. Here, leveraging this dataset, we developed a computational framework, called Phenotype Associated Single Cell encoder (PASCode), to score single-cell phenotype associations, and identified ∼1.5 million phenotype-associated cells (PACs) from 584 donors with AD-related phenotypes. PASCode ensembles multiple statistical methods into a graph neural model for robust scoring. Comparing PACs within 27 brain cell subclasses, we prioritized cell subpopulations and their expressed genes for various AD phenotypes. For instance, we identified microglia subpopulations implicated in AD pathology; reactive astrocyte subtypes with altered neuroprotective and neurotoxic gene expression that likely confer cognitive resilience; and enhanced excitatory/inhibitory imbalance and mitochondrial dysfunction in cognitively impaired AD donors. We also identified many PACs for multiple phenotypes, including the astrocytes between AD and depression showing specific gene expression patterns such as inflammation and endoplasmic reticulum stress pathways. These prioritized subpopulations, genes and pathways potentially offer valuable insights for precision diagnostic and therapeutic development. We also validated our findings in external population-scale datasets including AD and major depressive disorder, compiled an AD-phenotypic single-cell atlas and delivered the framework as an open-source tool with pre-trained models and a web application for community use.
- 23 September 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (10) , 3682-3694
Open AccessAI-based characterization of Alzheimer’s disease phenotypes from population-scale single-cell data
The complexity of Alzheimer’s disease (AD) manifests in diverse clinical phenotypes, including cognitive impairment and neuropsychiatric symptoms. However, the etiology of these phenotypes remains elusive. To address this, the PsychAD project generated a population-level single-nucleus RNA sequencing dataset comprising over 6 million nuclei from the prefrontal cortex of >1,000 individual brains, covering a variety of disease phenotypes. Here, leveraging this dataset, we developed a computational framework, called Phenotype Associated Single Cell encoder (PASCode), to score single-cell phenotype associations, and identified ∼1.5 million phenotype-associated cells (PACs) from 584 donors with AD-related phenotypes. PASCode ensembles multiple statistical methods into a graph neural model for robust scoring. Comparing PACs within 27 brain cell subclasses, we prioritized cell subpopulations and their expressed genes for various AD phenotypes. For instance, we identified microglia subpopulations implicated in AD pathology; reactive astrocyte subtypes with altered neuroprotective and neurotoxic gene expression that likely confer cognitive resilience; and enhanced excitatory/inhibitory imbalance and mitochondrial dysfunction in cognitively impaired AD donors. We also identified many PACs for multiple phenotypes, including the astrocytes between AD and depression showing specific gene expression patterns such as inflammation and endoplasmic reticulum stress pathways. These prioritized subpopulations, genes and pathways potentially offer valuable insights for precision diagnostic and therapeutic development. We also validated our findings in external population-scale datasets including AD and major depressive disorder, compiled an AD-phenotypic single-cell atlas and delivered the framework as an open-source tool with pre-trained models and a web application for community use.
- Open Access
- 23 September 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 9596
Personalized single-cell transcriptomics reveals molecular diversity in Alzheimer’s disease
Alzheimer’s disease (AD) is highly heterogeneous and driven by diverse molecular and cellular mechanisms. Functional genomics investigates these mechanisms from genetic variants to gene expression and regulation. We performed personalized functional genomics analysis on population-scale single-nucleus RNA-seq data, with cross-cohort validation across multiple cohorts comprising over 1900 individual brains, capturing donor-level cell type interactions and gene regulatory networks. Using a knowledge-guided graph neural network, we learned latent representations of each donor’s functional genomics that accurately classified AD phenotypes, identified molecularly defined subpopulations, and traced disease progression trajectories. Our importance scores, derived from graph attentions, identified significant inter-donor differences and prioritized personalized cell type genes and regulatory networks. Finally, we identified gene regulatory QTLs (grQTLs) linking genetic variants to donor-level regulatory changes, providing insights into gene regulatory relationships beyond traditional eQTLs. All results are summarized into a personalized functional genomics atlas for AD, including an open-source framework, iBrainMap, for general use.
- 23 September 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 9596
Open AccessPersonalized single-cell transcriptomics reveals molecular diversity in Alzheimer’s disease
Alzheimer’s disease (AD) is highly heterogeneous and driven by diverse molecular and cellular mechanisms. Functional genomics investigates these mechanisms from genetic variants to gene expression and regulation. We performed personalized functional genomics analysis on population-scale single-nucleus RNA-seq data, with cross-cohort validation across multiple cohorts comprising over 1900 individual brains, capturing donor-level cell type interactions and gene regulatory networks. Using a knowledge-guided graph neural network, we learned latent representations of each donor’s functional genomics that accurately classified AD phenotypes, identified molecularly defined subpopulations, and traced disease progression trajectories. Our importance scores, derived from graph attentions, identified significant inter-donor differences and prioritized personalized cell type genes and regulatory networks. Finally, we identified gene regulatory QTLs (grQTLs) linking genetic variants to donor-level regulatory changes, providing insights into gene regulatory relationships beyond traditional eQTLs. All results are summarized into a personalized functional genomics atlas for AD, including an open-source framework, iBrainMap, for general use.
- Open Access
- 4 May 2026
- Published by Oxford University Press (OUP) in Briefings in Bioinformatics
- Vol. 27 (3)
Benchmarking computational methods for multi-omics biomarker discovery in cancer
Multi-omics profiling characterizes cancer biology and supports biomarker discovery for prognosis and therapy selection. Although numerous computational multi-omics biomarker identification methods have been proposed, their ability to identify clinically relevant biomarkers has not been systematically evaluated, leaving it unclear whether the resulting biomarker nominations are reliable for downstream validation. Here, we systematically benchmark 20 representative statistical, machine learning and deep learning methods using curated gold-standard prognostic and therapeutic biomarkers across five real-world datasets. We evaluate performance in terms of both biomarker identification accuracy and stability. Overall, DeePathNet and DeepKEGG achieve the best performance. Across methods, effective biomarker recovery is associated with the integration of biological knowledge, global feature interactions, multivariate feature attribution, and effective regularization. Analysis of omics type contributions reveals method- and modality-specific biases, highlighting the importance of broader omics integration. We further evaluate methods on simulated datasets to probe sensitivity with controlled signal and noise. By aggregating results from top-performing methods, we construct consensus biomarker panels that nominate candidates for potential investigations. Finally, we provide user-friendly interfaces to allow researchers to benchmark new methods against the 20 baselines or apply selected methods for biomarker identification on custom multi-omics datasets. Our benchmark is publicly available at https://github.com/athanzli/CancerMOBI-Bench.
- 4 May 2026
- Published by Oxford University Press (OUP) in Briefings in Bioinformatics
- Vol. 27 (3)
Open AccessBenchmarking computational methods for multi-omics biomarker discovery in cancer
Multi-omics profiling characterizes cancer biology and supports biomarker discovery for prognosis and therapy selection. Although numerous computational multi-omics biomarker identification methods have been proposed, their ability to identify clinically relevant biomarkers has not been systematically evaluated, leaving it unclear whether the resulting biomarker nominations are reliable for downstream validation. Here, we systematically benchmark 20 representative statistical, machine learning and deep learning methods using curated gold-standard prognostic and therapeutic biomarkers across five real-world datasets. We evaluate performance in terms of both biomarker identification accuracy and stability. Overall, DeePathNet and DeepKEGG achieve the best performance. Across methods, effective biomarker recovery is associated with the integration of biological knowledge, global feature interactions, multivariate feature attribution, and effective regularization. Analysis of omics type contributions reveals method- and modality-specific biases, highlighting the importance of broader omics integration. We further evaluate methods on simulated datasets to probe sensitivity with controlled signal and noise. By aggregating results from top-performing methods, we construct consensus biomarker panels that nominate candidates for potential investigations. Finally, we provide user-friendly interfaces to allow researchers to benchmark new methods against the 20 baselines or apply selected methods for biomarker identification on custom multi-omics datasets. Our benchmark is publicly available at https://github.com/athanzli/CancerMOBI-Bench.
- Open Access
Benchmarking computational methods for multi-omics biomarker discovery in cancer
Multi-omics profiling characterizes cancer biology and supports biomarker discovery for prognosis and therapy selection. Although numerous computational multi-omics biomarker identification methods have been proposed, their ability to identify clinically relevant biomarkers has not been systematically evaluated, leaving it unclear whether the resulting biomarker nominations are reliable for downstream validation. Here we systematically benchmark 20 representative statistical, machine learning and deep learning methods using curated gold-standard prognostic and therapeutic biomarkers across five real-world datasets. We evaluate performance in terms of both biomarker identification accuracy and stability. Overall, DeePathNet and Deep-KEGG achieve the best performance. Across methods, effective biomarker recovery is associated with the integration of biological knowledge, global feature interactions, multivariate feature attribution, and effective regularization. Analysis of omics type contributions reveals method- and modality-specific biases, highlighting the importance of broader omics integration. We further evaluate methods on simulated datasets to probe sensitivity with controlled signal and noise. By aggregating results from top-performing methods, we construct consensus biomarker panels that nominate candidates for potential investigations. Finally, we provide user-friendly interfaces to allow researchers to benchmark new methods against the 20 baselines or apply selected methods for biomarker identification on custom multi-omics datasets. Our benchmark is publicly available at https://github.com/athanzli/CancerMOBI-Bench .
Open AccessBenchmarking computational methods for multi-omics biomarker discovery in cancer
Multi-omics profiling characterizes cancer biology and supports biomarker discovery for prognosis and therapy selection. Although numerous computational multi-omics biomarker identification methods have been proposed, their ability to identify clinically relevant biomarkers has not been systematically evaluated, leaving it unclear whether the resulting biomarker nominations are reliable for downstream validation. Here we systematically benchmark 20 representative statistical, machine learning and deep learning methods using curated gold-standard prognostic and therapeutic biomarkers across five real-world datasets. We evaluate performance in terms of both biomarker identification accuracy and stability. Overall, DeePathNet and Deep-KEGG achieve the best performance. Across methods, effective biomarker recovery is associated with the integration of biological knowledge, global feature interactions, multivariate feature attribution, and effective regularization. Analysis of omics type contributions reveals method- and modality-specific biases, highlighting the importance of broader omics integration. We further evaluate methods on simulated datasets to probe sensitivity with controlled signal and noise. By aggregating results from top-performing methods, we construct consensus biomarker panels that nominate candidates for potential investigations. Finally, we provide user-friendly interfaces to allow researchers to benchmark new methods against the 20 baselines or apply selected methods for biomarker identification on custom multi-omics datasets. Our benchmark is publicly available at https://github.com/athanzli/CancerMOBI-Bench .