Michelle M. Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Michelle M. Li has a total of 168 co-authors. Here are the top co-authors by number of shared publications:
- Marinka Zitnik
Harvard University, USA
14 shared publications - Emily Alsentzer
Brigham and Women’s Hospital, Harvard University , Boston, MA 02115
4 shared publications Yepeng Huang
Department of Biomedical Informatics, Harvard Medical School
4 shared publications- Ayush Noori
Department of Engineering Science, University of Oxford, Oxford, United Kingdom
3 shared publications - Daniel Marbach
Roche Pharma Research and Early Development, Roche Innovation Center Basel, F. Hoffmann-La Roche Ltd, Basel, Switzerland
3 shared publications - Wanxiang Shen
College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, China
3 shared publications
Latest Publications
- Open Access
Multimodal protein representations across human brain cell types
The central nervous system (CNS) comprises diverse cell types whose molecular context shapes the roles of the proteins they express. Computational models offer a scalable way to study the roles of proteins across human brain cell types, with protein language models and interactome- or single-cell-based models capturing complementary information encoded in protein sequences and cellular network contexts. Human brain tissue is hard to access, so brain cell types make up a small fraction of the atlases single-cell models train on and of the benchmarks used for their evaluation. To capture these complementary sources of information in the CNS, we introduce SEMPER, a multimodal representation learning model that aligns protein language model representations with protein-protein interaction networks from 442 cell types of the adult human brain, giving each protein a sequence-informed representation in every cell type in which its gene is expressed. We show that SEMPER representations predict cell type-specific protein complexes and predict genes essential to the survival of dopaminergic neurons, measured in a genome-wide CRISPR screen. In zero-shot retrieval, SEMPER recovers clinically pursued drug targets for Parkinson's disease, Alzheimer's disease, multiple sclerosis, Huntington's disease and amyotrophic lateral sclerosis at higher rates than cell type-specific models. To further improve target recovery, we combine SEMPER with a large language model that reasons over protein functional annotations. The combination recovers more targets, with SEMPER providing experimentally derived cellular context that complements literature-based reasoning of the large language model.
Open AccessMultimodal protein representations across human brain cell types
The central nervous system (CNS) comprises diverse cell types whose molecular context shapes the roles of the proteins they express. Computational models offer a scalable way to study the roles of proteins across human brain cell types, with protein language models and interactome- or single-cell-based models capturing complementary information encoded in protein sequences and cellular network contexts. Human brain tissue is hard to access, so brain cell types make up a small fraction of the atlases single-cell models train on and of the benchmarks used for their evaluation. To capture these complementary sources of information in the CNS, we introduce SEMPER, a multimodal representation learning model that aligns protein language model representations with protein-protein interaction networks from 442 cell types of the adult human brain, giving each protein a sequence-informed representation in every cell type in which its gene is expressed. We show that SEMPER representations predict cell type-specific protein complexes and predict genes essential to the survival of dopaminergic neurons, measured in a genome-wide CRISPR screen. In zero-shot retrieval, SEMPER recovers clinically pursued drug targets for Parkinson's disease, Alzheimer's disease, multiple sclerosis, Huntington's disease and amyotrophic lateral sclerosis at higher rates than cell type-specific models. To further improve target recovery, we combine SEMPER with a large language model that reasons over protein functional annotations. The combination recovers more targets, with SEMPER providing experimentally derived cellular context that complements literature-based reasoning of the large language model.
- 22 October 2026
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 3 (11)
Clinical AI in 2025–2026 — Six Themes from the SAIL Year in Review
- 22 October 2026
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 3 (11)
Clinical AI in 2025–2026 — Six Themes from the SAIL Year in Review
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 12856-12867
Causal Representation Learning from Network Data
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 12856-12867
Open AccessCausal Representation Learning from Network Data
- Open Access
- 3 July 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (8) , 1-13
Generalizable AI predicts immunotherapy outcomes across cancers and treatments
Immune checkpoint inhibitors (ICIs) are a standard treatment across cancers, yet most patients do not respond, and existing biomarkers generalize poorly across tumor types and therapies. Here we present COMPASS, a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept bottleneck transformer. COMPASS encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor−microenvironment interaction and signaling pathways. Trained on 10,184 tumors across 33 cancer types, COMPASS achieves better average performance than 22 methods across 16 clinical cohorts spanning seven cancers and six ICIs, improving accuracy by 8.5% and area under the precision-recall curve by 15.7% on average across cohorts. COMPASS generalizes to cancer types and treatments not represented during fine-tuning and may inform indication selection and patient stratification. In survival analyses, patients classified by COMPASS as responders had longer overall survival (hazard ratio = 4.7, P < 0.0001). Personalized response maps connect gene expression to immune concepts, identifying programs associated with response and resistance; in immune-inflamed non-responders, COMPASS highlights programs including TGFβ signaling, endothelial exclusion, CD4+ T cell dysfunction and B cell deficiency. COMPASS predicts immunotherapy response and provides hypothesis-generating mechanistic insight for trial design and translational studies.
- 3 July 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (8) , 1-13
Open AccessGeneralizable AI predicts immunotherapy outcomes across cancers and treatments
Immune checkpoint inhibitors (ICIs) are a standard treatment across cancers, yet most patients do not respond, and existing biomarkers generalize poorly across tumor types and therapies. Here we present COMPASS, a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept bottleneck transformer. COMPASS encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor−microenvironment interaction and signaling pathways. Trained on 10,184 tumors across 33 cancer types, COMPASS achieves better average performance than 22 methods across 16 clinical cohorts spanning seven cancers and six ICIs, improving accuracy by 8.5% and area under the precision-recall curve by 15.7% on average across cohorts. COMPASS generalizes to cancer types and treatments not represented during fine-tuning and may inform indication selection and patient stratification. In survival analyses, patients classified by COMPASS as responders had longer overall survival (hazard ratio = 4.7, P < 0.0001). Personalized response maps connect gene expression to immune concepts, identifying programs associated with response and resistance; in immune-inflamed non-responders, COMPASS highlights programs including TGFβ signaling, endothelial exclusion, CD4+ T cell dysfunction and B cell deficiency. COMPASS predicts immunotherapy response and provides hypothesis-generating mechanistic insight for trial design and translational studies.
- Top Cited Papers
- 3 February 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (2) , 439-448
Scaling medical AI across clinical contexts
Medical artificial intelligence (AI) tools, including clinical language models, vision–language models and multimodal health record models, are used to summarize clinical notes, answer questions and support decisions. Their adaptation to new populations, specialties or care settings often relies on fine-tuning, prompting or retrieval from external knowledge bases. These strategies can scale poorly and risk contextual errors—outputs that appear plausible but miss critical patient or situational information. We envision context switching as an emergent solution. Context switching adjusts model reasoning at inference, without retraining. Generative models can tailor outputs to patient biology, care setting or disease. Multimodal models can switch between notes, laboratory results, imaging and genomics, even when some data are missing or delayed. Agent models can coordinate tools and roles based on task and user context. In each case, context switching enables medical AI to adapt across specialties, populations and geographies. This approach requires advances in data design, model architectures and evaluation frameworks, and establishes a foundation for medical AI that scales to an infinite number of contexts, while remaining reliable and suited to real-world care.
- 3 February 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (2) , 439-448
Top Cited PapersScaling medical AI across clinical contexts
Medical artificial intelligence (AI) tools, including clinical language models, vision–language models and multimodal health record models, are used to summarize clinical notes, answer questions and support decisions. Their adaptation to new populations, specialties or care settings often relies on fine-tuning, prompting or retrieval from external knowledge bases. These strategies can scale poorly and risk contextual errors—outputs that appear plausible but miss critical patient or situational information. We envision context switching as an emergent solution. Context switching adjusts model reasoning at inference, without retraining. Generative models can tailor outputs to patient biology, care setting or disease. Multimodal models can switch between notes, laboratory results, imaging and genomics, even when some data are missing or delayed. Agent models can coordinate tools and roles based on task and user context. In each case, context switching enables medical AI to adapt across specialties, populations and geographies. This approach requires advances in data design, model architectures and evaluation frameworks, and establishes a foundation for medical AI that scales to an infinite number of contexts, while remaining reliable and suited to real-world care.
- Top Cited PapersOpen Access
Medea: An omics AI agent for therapeutic discovery
AI agents promise to empower biomedical discovery, but realizing this promise requires the ability to complete transparent, long-horizon analyses using tools. Agents must make intermediate decisions explicit, and validate each decision and output against data and tool constraints as the analysis unfolds. We present M edea , an AI agent that takes an omics objective and executes a transparent multi-step analysis using tools. M edea comprises four modules: research planning with context and integrity verification, code execution with pre- and post-run checks, literature reasoning with evidence-strength assessment, and a consensus stage that reconciles evidence across datasets, tools, and literature. M edea uses 20 tools spanning single-cell and bulk transcriptomic datasets, cancer vulnerability maps, pathway knowledge bases, and machine learning models. We evaluate M edea across 5,679 analyses in three open-ended domains: target identification across five diseases and cell type contexts (2,400 analyses), synthetic lethality reasoning in seven cell lines (2,385 analyses), and immunotherapy response prediction in bladder cancer (894 patient analyses). In evaluations that vary large language models, tool sets, omics objectives, and agentic modules, M edea improves the performance of existing approaches by up to 46% for target identification, 22% for synthetic lethality, and 24% for immunotherapy response prediction, while maintaining low failure rates and calibrated abstention. M edea shows that verification-aware AI agents improve performance by producing transparent analyses, not simply more efficient workflows.
Top Cited PapersOpen AccessMedea: An omics AI agent for therapeutic discovery
AI agents promise to empower biomedical discovery, but realizing this promise requires the ability to complete transparent, long-horizon analyses using tools. Agents must make intermediate decisions explicit, and validate each decision and output against data and tool constraints as the analysis unfolds. We present M edea , an AI agent that takes an omics objective and executes a transparent multi-step analysis using tools. M edea comprises four modules: research planning with context and integrity verification, code execution with pre- and post-run checks, literature reasoning with evidence-strength assessment, and a consensus stage that reconciles evidence across datasets, tools, and literature. M edea uses 20 tools spanning single-cell and bulk transcriptomic datasets, cancer vulnerability maps, pathway knowledge bases, and machine learning models. We evaluate M edea across 5,679 analyses in three open-ended domains: target identification across five diseases and cell type contexts (2,400 analyses), synthetic lethality reasoning in seven cell lines (2,385 analyses), and immunotherapy response prediction in bladder cancer (894 patient analyses). In evaluations that vary large language models, tool sets, omics objectives, and agentic modules, M edea improves the performance of existing approaches by up to 46% for target identification, 22% for synthetic lethality, and 24% for immunotherapy response prediction, while maintaining low failure rates and calibrated abstention. M edea shows that verification-aware AI agents improve performance by producing transparent analyses, not simply more efficient workflows.
- Open Access
Graph AI generates neurological hypotheses validated in molecular, organoid, and clinical systems
Neurological diseases are the leading global cause of disability, yet most lack disease-modifying treatments. We present PROTON, a heterogeneous graph transformer that generates testable hypotheses across molecular, organoid, and clinical systems. To evaluate PROTON, we apply it to Parkinson's disease (PD), bipolar disorder (BD), and Alzheimer's disease (AD). In PD, PROTON linked genetic risk loci to genes essential for dopaminergic neuron survival and predicted pesticides toxic to patient-derived neurons, including the insecticide endosulfan, which ranked within the top 1.29% of predictions. In silico screens performed by PROTON reproduced six genome-wide $α$-synuclein experiments, including a split-ubiquitin yeast two-hybrid system (normalized enrichment score [NES] = 2.30, FDR-adjusted $p < 1 \times 10^{-4}$), an ascorbate peroxidase proximity labeling assay (NES = 2.16, FDR $< 1 \times 10^{-4}$), and a high-depth targeted exome sequencing study in 496 synucleinopathy patients (NES = 2.13, FDR $< 1 \times 10^{-4}$). In BD, PROTON predicted calcitriol as a candidate drug that reversed proteomic alterations observed in cortical organoids derived from BD patients. In AD, we evaluated PROTON predictions in health records from $n = 610,524$ patients at Mass General Brigham, confirming that five PROTON-predicted drugs were associated with reduced seven-year dementia risk (minimum hazard ratio = 0.63, 95% CI: 0.53-0.75, $p < 1 \times 10^{-7}$). PROTON generated neurological hypotheses that were evaluated across molecular, organoid, and clinical systems, defining a path for AI-driven discovery in neurological disease.Open AccessGraph AI generates neurological hypotheses validated in molecular, organoid, and clinical systems
Neurological diseases are the leading global cause of disability, yet most lack disease-modifying treatments. We present PROTON, a heterogeneous graph transformer that generates testable hypotheses across molecular, organoid, and clinical systems. To evaluate PROTON, we apply it to Parkinson's disease (PD), bipolar disorder (BD), and Alzheimer's disease (AD). In PD, PROTON linked genetic risk loci to genes essential for dopaminergic neuron survival and predicted pesticides toxic to patient-derived neurons, including the insecticide endosulfan, which ranked within the top 1.29% of predictions. In silico screens performed by PROTON reproduced six genome-wide $α$-synuclein experiments, including a split-ubiquitin yeast two-hybrid system (normalized enrichment score [NES] = 2.30, FDR-adjusted $p < 1 \times 10^{-4}$), an ascorbate peroxidase proximity labeling assay (NES = 2.16, FDR $< 1 \times 10^{-4}$), and a high-depth targeted exome sequencing study in 496 synucleinopathy patients (NES = 2.13, FDR $< 1 \times 10^{-4}$). In BD, PROTON predicted calcitriol as a candidate drug that reversed proteomic alterations observed in cortical organoids derived from BD patients. In AD, we evaluated PROTON predictions in health records from $n = 610,524$ patients at Mass General Brigham, confirming that five PROTON-predicted drugs were associated with reduced seven-year dementia risk (minimum hazard ratio = 0.63, 95% CI: 0.53-0.75, $p < 1 \times 10^{-7}$). PROTON generated neurological hypotheses that were evaluated across molecular, organoid, and clinical systems, defining a path for AI-driven discovery in neurological disease. - 23 October 2025
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 2 (11)
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
- 23 October 2025
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 2 (11)
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
- Open Access
- 20 August 2025
- Published by ScienceOpen
Contextual AI for Single-cell Therapeutic Target Discovery
- 20 August 2025
- Published by ScienceOpen
Open AccessContextual AI for Single-cell Therapeutic Target Discovery
- Open Access
Generalizable AI predicts immunotherapy outcomes across cancers and treatments
Immune checkpoint inhibitors are standard across cancers, yet most patients do not respond and existing biomarkers generalize poorly across tumor types, drugs and clinical settings. We present C ompass , a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept-bottleneck transformer. C ompass encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor-microenvironment interactions, and signaling pathways. Trained on 10,184 tumors across 33 cancer types, C ompass outperforms 22 baseline methods in 16 independent clinical cohorts spanning seven cancers and six immune checkpoint inhibitors, increasing accuracy by 8.5% and area under the precision-recall curve by 15.7%, with minimal additional training. The model generalizes to unseen cancer types and treatments, supporting indication selection and patient stratification in early-phase clinical trials. In survival analyses, C ompass -stratified responders have longer overall survival (hazard ratio = 4.7, p < 0.0001). Personalized response maps connect gene expression to immune concepts, revealing mechanisms of response and resistance; in immune-inflamed non-responders, C ompass highlights programs including TGF- β signaling, endothelial exclusion, CD4+ T cell dysfunction, and B cell deficiency. By combining interpretability with transfer learning, C ompass enables robust prediction and mechanistic insight to inform trial design and translational studies.
Open AccessGeneralizable AI predicts immunotherapy outcomes across cancers and treatments
Immune checkpoint inhibitors are standard across cancers, yet most patients do not respond and existing biomarkers generalize poorly across tumor types, drugs and clinical settings. We present C ompass , a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept-bottleneck transformer. C ompass encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor-microenvironment interactions, and signaling pathways. Trained on 10,184 tumors across 33 cancer types, C ompass outperforms 22 baseline methods in 16 independent clinical cohorts spanning seven cancers and six immune checkpoint inhibitors, increasing accuracy by 8.5% and area under the precision-recall curve by 15.7%, with minimal additional training. The model generalizes to unseen cancer types and treatments, supporting indication selection and patient stratification in early-phase clinical trials. In survival analyses, C ompass -stratified responders have longer overall survival (hazard ratio = 4.7, p < 0.0001). Personalized response maps connect gene expression to immune concepts, revealing mechanisms of response and resistance; in immune-inflamed non-responders, C ompass highlights programs including TGF- β signaling, endothelial exclusion, CD4+ T cell dysfunction, and B cell deficiency. By combining interpretability with transfer learning, C ompass enables robust prediction and mechanistic insight to inform trial design and translational studies.