Dvir Aran
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Recognitions
Dvir Aran has a total of 302 co-authors. Here are the top co-authors by number of shared publications:
- Shahar Shelly
Ruth and Bruce Rappaport Faculty of Medicine, Technion — Israel Institute of Technology, Haifa, Israel
12 shared publications - Atul Janardhan Butte
Bakar Computational Health Sciences Institute, University of California, San Francisco, San Francisco, USA
11 shared publications - Asaf Hellman
Affiliation: Developmental Biology and Cancer Research, The Institute for Medical Research Israel-Canada (IMRIC), The Hebrew University-Hadassah Medical School, Jerusalem, Israel
8 shared publications - Marina Sirota
Bakar Computational Health Sciences Institute, University of California San Francisco, San Francisco, CA, USA
6 shared publications - Alon Gorenshtein
Beth Israel Deaconess Medical Center
4 shared publications - Amit Haim Shmilovitch
Department of Neurology, Rambam Medical Center, Haifa, Israel
4 shared publications
Latest Publications
- 22 October 2026
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 3 (11)
Comparative Cost of Care Recommended by Large Language Models versus Treating Physicians
- 22 October 2026
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 3 (11)
Comparative Cost of Care Recommended by Large Language Models versus Treating Physicians
- 1 September 2026
- Published by Elsevier in Cancer Cell
- Vol. 44 (9) , 1847-1861.e6
Tumor immune microenvironment remodeling predicts response to checkpoint inhibitor therapy
- 1 September 2026
- Published by Elsevier in Cancer Cell
- Vol. 44 (9) , 1847-1861.e6
Tumor immune microenvironment remodeling predicts response to checkpoint inhibitor therapy
- 19 August 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (10) , 3741-3747
Prospective evaluation of a large language model clinical decision support system in the emergency department
Prospective evidence for artificial intelligence (AI)-based clinical decision support in emergency departments remains limited. Here we conducted a DECIDE-AI stage 1 evaluation of SHAKED, a clinical decision support system built on multiple large language models, in a tertiary emergency department. Over 4 weeks, 1,138 patients were analyzed across two parallel units—one using SHAKED and one following routine rotations. Clinical adoption of SHAKED declined from 68% to 30%, owing to workload-sensitive disengagement (OR = 0.72 per shift hour, 95% CI 0.62 to 0.83). Physicians preferred the use of SHAKED for radiology consultations (OR = 2.98, 95% CI 1.58 to 5.63). No adverse events were detected, and expert review rated 99 of 100 sampled outputs as clinically appropriate. Emergency department length of stay did not differ between wings (4.9 h in both, P = 0.99). Intention-to-treat analysis showed a non-significant trend toward shorter consultation cycle time (−9.4 min, P = 0.077). These findings suggest that sustained clinician engagement, rather than algorithmic accuracy, may be the key barrier to effective clinical AI use in emergency departments. They inform randomized trial design but do not justify clinical deployment of AI clinical decision support at this stage. ClinicalTrials.gov identifier: NCT06902675.
- 19 August 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (10) , 3741-3747
Prospective evaluation of a large language model clinical decision support system in the emergency department
Prospective evidence for artificial intelligence (AI)-based clinical decision support in emergency departments remains limited. Here we conducted a DECIDE-AI stage 1 evaluation of SHAKED, a clinical decision support system built on multiple large language models, in a tertiary emergency department. Over 4 weeks, 1,138 patients were analyzed across two parallel units—one using SHAKED and one following routine rotations. Clinical adoption of SHAKED declined from 68% to 30%, owing to workload-sensitive disengagement (OR = 0.72 per shift hour, 95% CI 0.62 to 0.83). Physicians preferred the use of SHAKED for radiology consultations (OR = 2.98, 95% CI 1.58 to 5.63). No adverse events were detected, and expert review rated 99 of 100 sampled outputs as clinically appropriate. Emergency department length of stay did not differ between wings (4.9 h in both, P = 0.99). Intention-to-treat analysis showed a non-significant trend toward shorter consultation cycle time (−9.4 min, P = 0.077). These findings suggest that sustained clinician engagement, rather than algorithmic accuracy, may be the key barrier to effective clinical AI use in emergency departments. They inform randomized trial design but do not justify clinical deployment of AI clinical decision support at this stage. ClinicalTrials.gov identifier: NCT06902675.
- Open Access
- 8 August 2026
- Published by Springer Nature in Communications Medicine
Clinical liability cases reveal a coupling between legal defensibility and medical procedure escalation in large language models
- 8 August 2026
- Published by Springer Nature in Communications Medicine
Open AccessClinical liability cases reveal a coupling between legal defensibility and medical procedure escalation in large language models
- Open Access
Predicting Immunotherapy Response Through Chemotherapy-Induced Tumor Microenvironment Remodeling
Background: Chemotherapy before immunotherapy improves outcomes in multiple cancers, but the benefit is heterogeneous: only some tumors undergo the immune remodeling that makes subsequent immunotherapy effective, and no biomarker identifies these patients prospectively, leading to uniform use of chemo-immunotherapy and avoidable toxicity. Methods: A random forest model of tumor-microenvironment (TME) favorability, trained on 1,936 immunotherapy-treated patients, was applied to 334 paired pre-/post-chemotherapy biopsies across 19 studies to identify TME “converters,” from which a 50-gene baseline signature was derived. The signature was tested in IMvigor210 (the only cohort containing both treatment arms) and in five independent chemo-immunotherapy validation cohorts, and for specificity against 22 immunotherapy-only cohorts. Results: Of 209 paired-biopsy patients with unfavorable baseline TME, 61 (29%) converted to a favorable state after chemotherapy; converters were indistinguishable from non-converters at baseline by cell composition, motivating a gene-level signature. In IMvigor210, the signature discriminated responders among chemotherapy-primed patients with unfavorable baseline TME (AUC = 0.80, 95% CI: 0.67–0.94) but showed no signal in the trial’s immunotherapy-only arm (interaction OR = 4.42, p = 0.006). Across five independent chemo-immunotherapy cohorts, responders consistently scored higher than non-responders, yielding a significant pooled effect by meta-analysis (pooled g = 0.59, 95% CI: 0.11–1.08, p = 0.017) and individual significance in two (LUD2015-005, p = 0.009; GSE165252, p = 0.02). A meta-analysis of 22 immunotherapy-only cohorts constrained any effect there to a small magnitude (Hedges’ g = 0.08, 95% CI: −0.08 to 0.23), and five established immune signatures did not reproduce this treatment-context specificity. Converter tumors harbored coordinated baseline programs of proliferative stress, innate and adaptive immune readiness, and functional vasculature. Conclusions: A baseline transcriptomic signature predicts immunotherapy response specifically in chemotherapy-primed patients, a treatment-context specificity not shown by established immune biomarkers. These findings support prospective validation of the signature to guide chemotherapy– immunotherapy sequencing. Significance: Chemotherapy is now added to immunotherapy across a growing list of cancers, yet every eligible patient receives it even though only some tumors need it, and current biomarkers (PD-L1, tumor mutational burden, microsatellite instability) cannot identify which. We report the first biomarker to predict benefit from the chemotherapy component itself: a baseline gene-expression signature that flags likely responders in chemotherapy-primed settings while remaining silent under immunotherapy alone. This treatment-context specificity shifts the biomarker question from “who responds to immunotherapy” to “who needs the chemotherapy,” pointing toward a way to spare patients unnecessary toxicity and rationally sequence chemo-immunotherapy.
Open AccessPredicting Immunotherapy Response Through Chemotherapy-Induced Tumor Microenvironment Remodeling
Background: Chemotherapy before immunotherapy improves outcomes in multiple cancers, but the benefit is heterogeneous: only some tumors undergo the immune remodeling that makes subsequent immunotherapy effective, and no biomarker identifies these patients prospectively, leading to uniform use of chemo-immunotherapy and avoidable toxicity. Methods: A random forest model of tumor-microenvironment (TME) favorability, trained on 1,936 immunotherapy-treated patients, was applied to 334 paired pre-/post-chemotherapy biopsies across 19 studies to identify TME “converters,” from which a 50-gene baseline signature was derived. The signature was tested in IMvigor210 (the only cohort containing both treatment arms) and in five independent chemo-immunotherapy validation cohorts, and for specificity against 22 immunotherapy-only cohorts. Results: Of 209 paired-biopsy patients with unfavorable baseline TME, 61 (29%) converted to a favorable state after chemotherapy; converters were indistinguishable from non-converters at baseline by cell composition, motivating a gene-level signature. In IMvigor210, the signature discriminated responders among chemotherapy-primed patients with unfavorable baseline TME (AUC = 0.80, 95% CI: 0.67–0.94) but showed no signal in the trial’s immunotherapy-only arm (interaction OR = 4.42, p = 0.006). Across five independent chemo-immunotherapy cohorts, responders consistently scored higher than non-responders, yielding a significant pooled effect by meta-analysis (pooled g = 0.59, 95% CI: 0.11–1.08, p = 0.017) and individual significance in two (LUD2015-005, p = 0.009; GSE165252, p = 0.02). A meta-analysis of 22 immunotherapy-only cohorts constrained any effect there to a small magnitude (Hedges’ g = 0.08, 95% CI: −0.08 to 0.23), and five established immune signatures did not reproduce this treatment-context specificity. Converter tumors harbored coordinated baseline programs of proliferative stress, innate and adaptive immune readiness, and functional vasculature. Conclusions: A baseline transcriptomic signature predicts immunotherapy response specifically in chemotherapy-primed patients, a treatment-context specificity not shown by established immune biomarkers. These findings support prospective validation of the signature to guide chemotherapy– immunotherapy sequencing. Significance: Chemotherapy is now added to immunotherapy across a growing list of cancers, yet every eligible patient receives it even though only some tumors need it, and current biomarkers (PD-L1, tumor mutational burden, microsatellite instability) cannot identify which. We report the first biomarker to predict benefit from the chemotherapy component itself: a baseline gene-expression signature that flags likely responders in chemotherapy-primed settings while remaining silent under immunotherapy alone. This treatment-context specificity shifts the biomarker question from “who responds to immunotherapy” to “who needs the chemotherapy,” pointing toward a way to spare patients unnecessary toxicity and rationally sequence chemo-immunotherapy.
- Open Access
SOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps
Unsupervised representations are widely assumed to be neutral with respect to sensitive attributes when those attributes are withheld from training. We show that this assumption is false. Using SOMtime, a topology-preserving representation method based on high-capacity Self-Organizing Maps, we demonstrate that sensitive attributes such as age and income emerge as dominant latent axes in purely unsupervised embeddings, even when explicitly excluded from the input. On two large-scale real-world datasets (the World Values Survey across five countries and the Census-Income dataset), SOMtime recovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to 0.85, whereas PCA and UMAP typically remain below 0.23 (with a single exception reaching 0.31), and against t-SNE and autoencoders which achieve at most 0.34. Furthermore, unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters, demonstrating downstream fairness risks without any supervised task. These findings establish that \textit{fairness through unawareness} fails at the representation level for ordinal sensitive attributes and that fairness auditing must extend to unsupervised components of machine learning pipelines. We have made the code available at~ https://github.com/JosephBingham/SOMtime
Open AccessSOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps
Unsupervised representations are widely assumed to be neutral with respect to sensitive attributes when those attributes are withheld from training. We show that this assumption is false. Using SOMtime, a topology-preserving representation method based on high-capacity Self-Organizing Maps, we demonstrate that sensitive attributes such as age and income emerge as dominant latent axes in purely unsupervised embeddings, even when explicitly excluded from the input. On two large-scale real-world datasets (the World Values Survey across five countries and the Census-Income dataset), SOMtime recovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to 0.85, whereas PCA and UMAP typically remain below 0.23 (with a single exception reaching 0.31), and against t-SNE and autoencoders which achieve at most 0.34. Furthermore, unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters, demonstrating downstream fairness risks without any supervised task. These findings establish that \textit{fairness through unawareness} fails at the representation level for ordinal sensitive attributes and that fairness auditing must extend to unsupervised components of machine learning pipelines. We have made the code available at~ https://github.com/JosephBingham/SOMtime
- Open Access
- 12 January 2026
- Published by Oxford University Press (OUP) in Neuro-Oncology Advances
- Vol. 8 (1) , vdag134
Medulloblastoma immune microenvironment resembles early brain development and reveals novel basophil- B cell- associated signatures
Background: Medulloblastoma (MB) is the most common malignant pediatric brain tumor; yet, the mechanisms underlying immune evasion remain poorly understood. Since the developing brain’s immune environment naturally promotes tissue tolerance and remodeling, we hypothesized that MB may exploit developmental immune programs to escape surveillance. Here, we aimed to characterize the immune microenvironment of MB compared to healthy brain tissue across different developmental stages. Methods: We analyzed RNA-sequencing data from 124 MB tumors and 47 control brain samples spanning postconception weeks (PCWs) 8-37 through adulthood. Using xCell and DeconRNAseq, we quantified enrichment scores for 34 immune cell types and compared their profiles. Results: MB tumors exhibited a distinct immune profile most similar to PCW controls among all age groups tested (P < .001). Key findings included significantly elevated basophils, neutrophils, immature dendritic cells, memory B cells, Th2 cells, and NKT cells in MB samples. Conversely, macrophages (M1 and M2 subtypes), matured dendritic cells, regulatory T cells, and most B cell populations showed reduced enrichment. This pattern suggests a Treg-independent immunosuppressive mechanism. Notably, basophil enrichment in MB was comparable to PCW samples but significantly higher than older controls, potentially explaining the concurrent elevation of memory B cells. Conclusions: MB displays an immune microenvironment resembling early developmental stages, characterized by elevated basophils and memory B cells alongside reduced conventional immune surveillance mechanisms. These findings suggest that MB may co-opt developmental immune tolerance programs and highlight basophil - B cell interactions as potential therapeutic targets. The developmental similarity provides new insights into MB pathogenesis and immune evasion strategies.
- 12 January 2026
- Published by Oxford University Press (OUP) in Neuro-Oncology Advances
- Vol. 8 (1) , vdag134
Open AccessMedulloblastoma immune microenvironment resembles early brain development and reveals novel basophil- B cell- associated signatures
Background: Medulloblastoma (MB) is the most common malignant pediatric brain tumor; yet, the mechanisms underlying immune evasion remain poorly understood. Since the developing brain’s immune environment naturally promotes tissue tolerance and remodeling, we hypothesized that MB may exploit developmental immune programs to escape surveillance. Here, we aimed to characterize the immune microenvironment of MB compared to healthy brain tissue across different developmental stages. Methods: We analyzed RNA-sequencing data from 124 MB tumors and 47 control brain samples spanning postconception weeks (PCWs) 8-37 through adulthood. Using xCell and DeconRNAseq, we quantified enrichment scores for 34 immune cell types and compared their profiles. Results: MB tumors exhibited a distinct immune profile most similar to PCW controls among all age groups tested (P < .001). Key findings included significantly elevated basophils, neutrophils, immature dendritic cells, memory B cells, Th2 cells, and NKT cells in MB samples. Conversely, macrophages (M1 and M2 subtypes), matured dendritic cells, regulatory T cells, and most B cell populations showed reduced enrichment. This pattern suggests a Treg-independent immunosuppressive mechanism. Notably, basophil enrichment in MB was comparable to PCW samples but significantly higher than older controls, potentially explaining the concurrent elevation of memory B cells. Conclusions: MB displays an immune microenvironment resembling early developmental stages, characterized by elevated basophils and memory B cells alongside reduced conventional immune surveillance mechanisms. These findings suggest that MB may co-opt developmental immune tolerance programs and highlight basophil - B cell interactions as potential therapeutic targets. The developmental similarity provides new insights into MB pathogenesis and immune evasion strategies.
- Open Access
- 11 December 2025
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
CellMentor: cell-type aware dimensionality reduction for single-cell RNA-sequencing data
Single-cell RNA sequencing generates high-dimensional gene expression profiles for individual cells. A key step in its analysis is dimensionality reduction, which transforms these data into lower-dimensional representations for clustering and cell type identification. Current approaches often fail to balance technical noise removal with the preservation of biologically meaningful cell-type signals. Here, we present CellMentor, a fully supervised dimensionality reduction method based on non-negative matrix factorization that integrates cell type labels directly into its optimization objective. CellMentor minimizes variation within known populations while maximizing distinctions between types, producing low-dimensional embeddings optimized for cell type identification. Across diverse simulated and experimental datasets, CellMentor shows superior cell type separation, robust batch correction, and effective detection of rare cell populations, offering a valuable tool for integrative single-cell analyses across experiments.
- 11 December 2025
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessCellMentor: cell-type aware dimensionality reduction for single-cell RNA-sequencing data
Single-cell RNA sequencing generates high-dimensional gene expression profiles for individual cells. A key step in its analysis is dimensionality reduction, which transforms these data into lower-dimensional representations for clustering and cell type identification. Current approaches often fail to balance technical noise removal with the preservation of biologically meaningful cell-type signals. Here, we present CellMentor, a fully supervised dimensionality reduction method based on non-negative matrix factorization that integrates cell type labels directly into its optimization objective. CellMentor minimizes variation within known populations while maximizing distinctions between types, producing low-dimensional embeddings optimized for cell type identification. Across diverse simulated and experimental datasets, CellMentor shows superior cell type separation, robust batch correction, and effective detection of rare cell populations, offering a valuable tool for integrative single-cell analyses across experiments.
- Open Access
- 4 December 2025
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 4 (12) , e0001106
A multi-agent approach to neurological clinical reasoning
Large language models (LLMs) have demonstrated impressive capabilities in medical domains, yet their ability to handle the specialized reasoning patterns required in clinical neurology warrants systematic evaluation. Neurological assessment presents distinctive challenges that combine anatomical localization, temporal pattern recognition, and nuanced symptom interpretation—cognitive processes that are specifically tested in board certification examinations. We developed a comprehensive benchmark comprising 305 questions from Israeli Board Certification Exams in Neurology and classified each along three dimensions of complexity: factual knowledge depth, clinical concept integration, and reasoning complexity. We evaluated ten LLMs of varying architectures and specializations using this benchmark, testing base models, retrieval-augmented generation (RAG) enhancement, and a novel multi-agent system. Our analysis revealed significant performance variation across models and methodologies. The OpenAI-o1 model achieved the highest base performance (90.9% accuracy), while specialized medical models performed surprisingly poorly (52.9% for Meditron-70B). RAG enhancement provided variable benefits across models; substantial improvements for mid-tier models like GPT-4o (80.5% to 87.3%) and smaller models, but limited effectiveness on the highest complexity questions regardless of model size. In contrast, our multi-agent framework—which decomposes neurological reasoning into specialized cognitive functions including question analysis, knowledge retrieval, answer synthesis, and validation—achieved dramatic improvements, especially for mid-range models. The LLaMA 3.3-70B-based agentic system reached 89.2% accuracy compared to 69.5% for its base model, with particularly substantial gains on level 3 complexity questions across all dimensions. External validation on MedQA revealed dataset-specific RAG effects: while RAG improved board certification performance, it showed minimal benefit on MedQA questions (LLaMA 3.3-70B: + 1.4% vs + 3.9% on board exams), reflecting alignment between our specialized neurology textbook and board examination content rather than the broader medical knowledge required for MedQA. Most notably, the multi-agent approach transformed inconsistent subspecialty performance into remarkably uniform excellence, effectively addressing the neurological reasoning challenges that persisted even with RAG enhancement. We further validated our approach using an independent dataset comprising 155 neurological cases extracted from MedQA. The results confirm that structured multi-agent approaches designed to emulate specialized cognitive processes significantly enhance complex medical reasoning offering promising directions for AI assistance in challenging clinical contexts.
- 4 December 2025
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 4 (12) , e0001106
Open AccessA multi-agent approach to neurological clinical reasoning
Large language models (LLMs) have demonstrated impressive capabilities in medical domains, yet their ability to handle the specialized reasoning patterns required in clinical neurology warrants systematic evaluation. Neurological assessment presents distinctive challenges that combine anatomical localization, temporal pattern recognition, and nuanced symptom interpretation—cognitive processes that are specifically tested in board certification examinations. We developed a comprehensive benchmark comprising 305 questions from Israeli Board Certification Exams in Neurology and classified each along three dimensions of complexity: factual knowledge depth, clinical concept integration, and reasoning complexity. We evaluated ten LLMs of varying architectures and specializations using this benchmark, testing base models, retrieval-augmented generation (RAG) enhancement, and a novel multi-agent system. Our analysis revealed significant performance variation across models and methodologies. The OpenAI-o1 model achieved the highest base performance (90.9% accuracy), while specialized medical models performed surprisingly poorly (52.9% for Meditron-70B). RAG enhancement provided variable benefits across models; substantial improvements for mid-tier models like GPT-4o (80.5% to 87.3%) and smaller models, but limited effectiveness on the highest complexity questions regardless of model size. In contrast, our multi-agent framework—which decomposes neurological reasoning into specialized cognitive functions including question analysis, knowledge retrieval, answer synthesis, and validation—achieved dramatic improvements, especially for mid-range models. The LLaMA 3.3-70B-based agentic system reached 89.2% accuracy compared to 69.5% for its base model, with particularly substantial gains on level 3 complexity questions across all dimensions. External validation on MedQA revealed dataset-specific RAG effects: while RAG improved board certification performance, it showed minimal benefit on MedQA questions (LLaMA 3.3-70B: + 1.4% vs + 3.9% on board exams), reflecting alignment between our specialized neurology textbook and board examination content rather than the broader medical knowledge required for MedQA. Most notably, the multi-agent approach transformed inconsistent subspecialty performance into remarkably uniform excellence, effectively addressing the neurological reasoning challenges that persisted even with RAG enhancement. We further validated our approach using an independent dataset comprising 155 neurological cases extracted from MedQA. The results confirm that structured multi-agent approaches designed to emulate specialized cognitive processes significantly enhance complex medical reasoning offering promising directions for AI assistance in challenging clinical contexts.
- Open Access
- 3 October 2025
- Published by Springer Nature in Genome Biology
- Vol. 26 (1) , 1-25
xCell 2.0: robust algorithm for cell type proportion estimation predicts response to immune checkpoint blockade
Accurate estimation of cell type proportions from bulk gene expression data is essential for understanding the cellular heterogeneity underlying complex tissues and diseases. Here, we introduce xCell 2.0, an advanced version of the xCell algorithm, featuring a training function that permits the utilization of any reference dataset. xCell 2.0 generates cell type gene signatures using an improved methodology, including automated handling of cell type dependencies and more robust signature generation. We benchmark xCell 2.0 against eleven popular deconvolution methods using nine human and mouse reference sets and 26 validation datasets, encompassing 1711 samples and 67 cell types. Additionally, we validate xCell 2.0 using the independent Deconvolution DREAM Challenge dataset. xCell 2.0 outperforms all other tested methods across distinct reference datasets, demonstrating superior accuracy and consistency across diverse biological contexts. xCell 2.0 also shows the best performance in minimizing spillover effects between related cell types. In a test example of pan-cancer immune cell checkpoint blockage response prediction, xCell 2.0-derived TME features significantly improve prediction accuracy compared to models using only cancer type and treatment information, and outperformed other deconvolution methods and established prediction scores. xCell 2.0 is a versatile and robust tool for cell type deconvolution that maintains high performance across various reference types and biological contexts. It is available both via a locally hosted web application and as a Bioconductor-compatible package, equipped with a large collection of pre-trained cell type signatures for human and mouse research.
- 3 October 2025
- Published by Springer Nature in Genome Biology
- Vol. 26 (1) , 1-25
Open AccessxCell 2.0: robust algorithm for cell type proportion estimation predicts response to immune checkpoint blockade
Accurate estimation of cell type proportions from bulk gene expression data is essential for understanding the cellular heterogeneity underlying complex tissues and diseases. Here, we introduce xCell 2.0, an advanced version of the xCell algorithm, featuring a training function that permits the utilization of any reference dataset. xCell 2.0 generates cell type gene signatures using an improved methodology, including automated handling of cell type dependencies and more robust signature generation. We benchmark xCell 2.0 against eleven popular deconvolution methods using nine human and mouse reference sets and 26 validation datasets, encompassing 1711 samples and 67 cell types. Additionally, we validate xCell 2.0 using the independent Deconvolution DREAM Challenge dataset. xCell 2.0 outperforms all other tested methods across distinct reference datasets, demonstrating superior accuracy and consistency across diverse biological contexts. xCell 2.0 also shows the best performance in minimizing spillover effects between related cell types. In a test example of pan-cancer immune cell checkpoint blockage response prediction, xCell 2.0-derived TME features significantly improve prediction accuracy compared to models using only cancer type and treatment information, and outperformed other deconvolution methods and established prediction scores. xCell 2.0 is a versatile and robust tool for cell type deconvolution that maintains high performance across various reference types and biological contexts. It is available both via a locally hosted web application and as a Bioconductor-compatible package, equipped with a large collection of pre-trained cell type signatures for human and mouse research.