Tal Korem
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Tal Korem has a total of 365 co-authors. Here are the top co-authors by number of shared publications:
- Heekuk Park
Columbia University
47 shared publications - George I. Austin
Department of Biomedical Informatics, Columbia University Irving Medical Center, New York, NY, USA
46 shared publications - Federico Baldini
Program for Mathematical Genomics, Department of Systems Biology, Columbia University Irving Medical Center, New York, New York
41 shared publications - Julian A. Abrams
Department of Medicine, Columbia University Irving Medical Center, New York, New York, USA
41 shared publications - Quinn S. Solfisburg
Department of Medicine, Section of Hematology and Medical Oncology at Boston Medical Center, Boston, MA
40 shared publications - Brittany Baldwin-Hunter
Department of Medicine, Columbia University Irving Medical Center, New York, New York
39 shared publications
Latest Publications
- 16 April 2026
- Published by Elsevier in American Journal of Obstetrics & Gynecology MFM
- Vol. 8 (4) , 101900
Reply to Du: Differences in experimental setting make validation harder
- 16 April 2026
- Published by Elsevier in American Journal of Obstetrics & Gynecology MFM
- Vol. 8 (4) , 101900
Reply to Du: Differences in experimental setting make validation harder
- Open Access
End-to-end evaluation of pipelines for metagenome-assembled genomes reveals hidden performance gaps
The generation of Metagenome Assembled Genomes (MAGs) has become a standard and basic step in the analysis of metagenomic data. This multi-step process, which includes assembly, binning, refinement, and quality control, has many alternative approaches, algorithms, and parameters. Determining the ideal approach for a given ecosystem and study, or highlighting algorithmic gaps in need of additional research and development, requires rigorous benchmarking. We present MAG-E ( MAG pipeline E valuator), a generalizable and expandable framework for end-to-end evaluation of entire MAG pipelines: from assembly, through binning, to quality control and filtering. MAG-E relies on simulations that are built to match an ecosystem of interest and provide a ground truth for accurate evaluation. To demonstrate the capabilities of MAG-E, we benchmark two assemblers, six binning algorithms, three binning modes, and three quality control and refinement methods in the context of the human gut microbiome. Our findings offer multiple insights into optimal MAG generation in this context. We find that metaSPAdes consistently outperforms MEGAHIT in terms of recall (completeness), and that COMEBin overall outperforms alternative binning algorithms, but has lower precision than SemiBin2. While multi-sample binning results in higher precision, as previously shown, single-sample binning has higher recall and leads to better overall performance with modern binners. Binning refinement, which combines bins from multiple different algorithms, leads to reduced performance. We further show that CheckM2 systematically overestimates completeness and underestimates contamination, and that this is partially ameliorated when using GUNC. Finally, we analyze performance at the contig level, and demonstrate that binning algorithms systematically underperform for prophages and fail to bin contigs that are shared between genomes. Overall, MAG-E offers deep insights into successes and gaps in this important analytic process.
Open AccessEnd-to-end evaluation of pipelines for metagenome-assembled genomes reveals hidden performance gaps
The generation of Metagenome Assembled Genomes (MAGs) has become a standard and basic step in the analysis of metagenomic data. This multi-step process, which includes assembly, binning, refinement, and quality control, has many alternative approaches, algorithms, and parameters. Determining the ideal approach for a given ecosystem and study, or highlighting algorithmic gaps in need of additional research and development, requires rigorous benchmarking. We present MAG-E ( MAG pipeline E valuator), a generalizable and expandable framework for end-to-end evaluation of entire MAG pipelines: from assembly, through binning, to quality control and filtering. MAG-E relies on simulations that are built to match an ecosystem of interest and provide a ground truth for accurate evaluation. To demonstrate the capabilities of MAG-E, we benchmark two assemblers, six binning algorithms, three binning modes, and three quality control and refinement methods in the context of the human gut microbiome. Our findings offer multiple insights into optimal MAG generation in this context. We find that metaSPAdes consistently outperforms MEGAHIT in terms of recall (completeness), and that COMEBin overall outperforms alternative binning algorithms, but has lower precision than SemiBin2. While multi-sample binning results in higher precision, as previously shown, single-sample binning has higher recall and leads to better overall performance with modern binners. Binning refinement, which combines bins from multiple different algorithms, leads to reduced performance. We further show that CheckM2 systematically overestimates completeness and underestimates contamination, and that this is partially ameliorated when using GUNC. Finally, we analyze performance at the contig level, and demonstrate that binning algorithms systematically underperform for prophages and fail to bin contigs that are shared between genomes. Overall, MAG-E offers deep insights into successes and gaps in this important analytic process.
- Open Access
- 28 February 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
A generalizable cross-continent prediction of esophageal squamous cell carcinoma using the oral microbiome
Esophageal squamous cell carcinoma (ESCC) has a poor prognosis and limited tools for early detection. Saliva is easily accessible and its microbiome composition can serve as a marker for upper gastrointestinal tract disease. This study aims to evaluate the potential of an oral microbiome signature for classifying ESCC. In a cross-sectional study of 48 ESCC patients and 110 controls from South Africa, a region with high ESCC incidence, we studied the potential utility of an oral microbiome signature for the disease. We built models using nested cross-validation to evaluate whether this signature is generalizable to held-out samples and further evaluated generalizability in studies from China, a distinct geographic region. We find significant alterations in the oral microbiome in patients with ESCC including significantly reduced α diversity and increased abundance of Fusobacterium nucleatum. We also find that logistic regression models based on microbiome data can better classify ESCC in held-out samples (auROC=0.96) compared to clinical and demographic data (auROC = 0.69; DeLong p < 1 x 10-8). Lastly, we find that microbiome-based models trained across multiple studies can generalize well to geographically distinct studies. Our results show that the oral microbiome in individuals with ESCC is distinct from controls and that this signal can generalize across unseen samples, suggesting the potential of saliva to serve as a non-invasive screening tool for ESCC.
- 28 February 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessA generalizable cross-continent prediction of esophageal squamous cell carcinoma using the oral microbiome
Esophageal squamous cell carcinoma (ESCC) has a poor prognosis and limited tools for early detection. Saliva is easily accessible and its microbiome composition can serve as a marker for upper gastrointestinal tract disease. This study aims to evaluate the potential of an oral microbiome signature for classifying ESCC. In a cross-sectional study of 48 ESCC patients and 110 controls from South Africa, a region with high ESCC incidence, we studied the potential utility of an oral microbiome signature for the disease. We built models using nested cross-validation to evaluate whether this signature is generalizable to held-out samples and further evaluated generalizability in studies from China, a distinct geographic region. We find significant alterations in the oral microbiome in patients with ESCC including significantly reduced α diversity and increased abundance of Fusobacterium nucleatum. We also find that logistic regression models based on microbiome data can better classify ESCC in held-out samples (auROC=0.96) compared to clinical and demographic data (auROC = 0.69; DeLong p < 1 x 10-8). Lastly, we find that microbiome-based models trained across multiple studies can generalize well to geographically distinct studies. Our results show that the oral microbiome in individuals with ESCC is distinct from controls and that this signal can generalize across unseen samples, suggesting the potential of saliva to serve as a non-invasive screening tool for ESCC.
- Top Cited PapersOpen Access
- 28 November 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (48) , eadx6976
Distributional bias compromises leave-one-out cross-validation
Cross-validation is a common method for evaluating machine learning models. “Leave-one-out cross-validation,” in which each data instance is used to test a model trained on all other instances, is often used in data-scarce regimes. As common metrics such as the R 2 score cannot be calculated for a single prediction, predictions are commonly aggregated across folds for performance evaluation. Here, we prove that this creates “distributional bias”: a negative correlation between the average label of each training fold and the label of its corresponding test instance. As machine learning models tend to regress to the mean of their training data, this bias tends to negatively affect performance evaluation and hyperparameter optimization. We demonstrate that distributional bias exists across diverse tasks, models, and evaluation approaches, and can bias against stronger regularization. To address it, we developed a generalizable rebalanced cross-validation that is robust to distributional bias in both classification and regression, and demonstrates improved performance in simulations, machine learning benchmarks, and several published analyses.
- 28 November 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (48) , eadx6976
Top Cited PapersOpen AccessDistributional bias compromises leave-one-out cross-validation
Cross-validation is a common method for evaluating machine learning models. “Leave-one-out cross-validation,” in which each data instance is used to test a model trained on all other instances, is often used in data-scarce regimes. As common metrics such as the R 2 score cannot be calculated for a single prediction, predictions are commonly aggregated across folds for performance evaluation. Here, we prove that this creates “distributional bias”: a negative correlation between the average label of each training fold and the label of its corresponding test instance. As machine learning models tend to regress to the mean of their training data, this bias tends to negatively affect performance evaluation and hyperparameter optimization. We demonstrate that distributional bias exists across diverse tasks, models, and evaluation approaches, and can bias against stronger regularization. To address it, we developed a generalizable rebalanced cross-validation that is robust to distributional bias in both classification and regression, and demonstrates improved performance in simulations, machine learning benchmarks, and several published analyses.
- Open Access
Comparative metagenomics using pan-metagenomic graphs
Identifying microbial genomic factors underlying human phenotypes is a key goal of microbiome research. Sequence graphs are a highly effective tool for genome comparisons because they enable high-resolution de novo analyses that capture and contextualize complex genomic variation. However, applying sequence graphs to complex microbial communities remains challenging due to the scale and complexity of metagenomic data. Existing multi-sample sequence graphs used in these settings are highly complex, computationally expensive, less accurate than single-sample alternatives, and often involve arbitrary coarse-graining. Here, we present copangraph, a multi-sample sequence-graph-based analysis framework for comprehensive comparisons of genomic variation across metagenomes. Copangraph uses a novel homology-based graph, which provides both non-arbitrary, evolutionary-motivated grouping of sequences into the same node as well as flexibility in the scale of variation represented by the graph. Its construction relies on hybrid coassembly, a new coassembly approach in which single-sample graphs are first constructed separately and are then merged to create a multi-sample graph. We also present an algorithm that uses paired-end reads to improve detection of contiguous genomic regions, increasing accuracy. Our results demonstrate that copangraph captures sequence and variant information more accurately than alternative methods, provides graphs that are more suitable for comparative analysis than de Bruijn graphs, and is computationally tractable. We show that copangraph reflects meaningful metagenomic variation across diverse scenarios. Importantly, it enables significantly better performance than other metagenomic representations when predicting the gut colonization trajectories of Vancomycin-resistant Enterococcus. Our results underscore the value of our multi-sample, graph-based framework for comparative metagenomic analyses.
Open AccessComparative metagenomics using pan-metagenomic graphs
Identifying microbial genomic factors underlying human phenotypes is a key goal of microbiome research. Sequence graphs are a highly effective tool for genome comparisons because they enable high-resolution de novo analyses that capture and contextualize complex genomic variation. However, applying sequence graphs to complex microbial communities remains challenging due to the scale and complexity of metagenomic data. Existing multi-sample sequence graphs used in these settings are highly complex, computationally expensive, less accurate than single-sample alternatives, and often involve arbitrary coarse-graining. Here, we present copangraph, a multi-sample sequence-graph-based analysis framework for comprehensive comparisons of genomic variation across metagenomes. Copangraph uses a novel homology-based graph, which provides both non-arbitrary, evolutionary-motivated grouping of sequences into the same node as well as flexibility in the scale of variation represented by the graph. Its construction relies on hybrid coassembly, a new coassembly approach in which single-sample graphs are first constructed separately and are then merged to create a multi-sample graph. We also present an algorithm that uses paired-end reads to improve detection of contiguous genomic regions, increasing accuracy. Our results demonstrate that copangraph captures sequence and variant information more accurately than alternative methods, provides graphs that are more suitable for comparative analysis than de Bruijn graphs, and is computationally tractable. We show that copangraph reflects meaningful metagenomic variation across diverse scenarios. Importantly, it enables significantly better performance than other metagenomic representations when predicting the gut colonization trajectories of Vancomycin-resistant Enterococcus. Our results underscore the value of our multi-sample, graph-based framework for comparative metagenomic analyses.
- Open Access
A generalizable cross-continent prediction of esophageal squamous cell carcinoma using the oral microbiome
Esophageal squamous cell carcinoma (ESCC) is a disease with limited tools for early screening and a poor prognosis. Symptoms typically appear late, and early cancer is hard to detect without endoscopic screening, which is inaccessible in most high-risk areas. Saliva is easily accessible, and its microbiome composition can serve as a marker for upper gastrointestinal tract disease. We studied the potential utility of an oral microbiome signature for ESCC in South Africa, a region with a high incidence of the disease. In a cohort of 48 ESCC patients and 110 controls, we found marked alterations in the oral microbiome in patients with ESCC, including significantly reduced alpha diversity and increased Fusobacterium nucleatum . We devised machine learning models that classify ESCC using microbiome data, finding good performance on held-out samples (area under receiver operating characteristic curve of 0.96), and demonstrated generalization to data across independent studies conducted in different geographic regions (0.64-0.81). Overall, our results demonstrate the potential of the oral microbiome to serve as a non-invasive screening tool for ESCC.
Open AccessA generalizable cross-continent prediction of esophageal squamous cell carcinoma using the oral microbiome
Esophageal squamous cell carcinoma (ESCC) is a disease with limited tools for early screening and a poor prognosis. Symptoms typically appear late, and early cancer is hard to detect without endoscopic screening, which is inaccessible in most high-risk areas. Saliva is easily accessible, and its microbiome composition can serve as a marker for upper gastrointestinal tract disease. We studied the potential utility of an oral microbiome signature for ESCC in South Africa, a region with a high incidence of the disease. In a cohort of 48 ESCC patients and 110 controls, we found marked alterations in the oral microbiome in patients with ESCC, including significantly reduced alpha diversity and increased Fusobacterium nucleatum . We devised machine learning models that classify ESCC using microbiome data, finding good performance on held-out samples (area under receiver operating characteristic curve of 0.96), and demonstrated generalization to data across independent studies conducted in different geographic regions (0.64-0.81). Overall, our results demonstrate the potential of the oral microbiome to serve as a non-invasive screening tool for ESCC.
- Open Access
Spatially Mapping the Mechanical and Structural Properties of the Seedling Uterine Fibroid-Myometrium Interface
Uterine fibroids (leiomyomas) are highly prevalent, noncancerous tumors canonically described as stiff and collagen-dense. Still, the mechanical heterogeneity within uterine fibroids and the alterations that occur at the fibroid–myometrium interface remain poorly understood. This study quantitatively maps the local mechanical, structural, and compositional properties of seedling uterine fibroids ( < 1 cm) at the interface region. Relative to patient-matched myometrial tissues, uterine fibroids exhibited increased stiffness, decreased permeability, increased diffusivity, decreased hydration, greater collagen content, and distinct collagen organization. At the fibroid–myometrium interface, a band of aligned myometrial fibers immediately adjacent to the fibroid was observed, and heterogeneous spatial patterns in elastic modulus and permeability were identified. Ultimately, this study establishes foundational knowledge on the mechanics and structure of seedling uterine fibroids, facilitating future developments of clinically translatable detection tools.
Open AccessSpatially Mapping the Mechanical and Structural Properties of the Seedling Uterine Fibroid-Myometrium Interface
Uterine fibroids (leiomyomas) are highly prevalent, noncancerous tumors canonically described as stiff and collagen-dense. Still, the mechanical heterogeneity within uterine fibroids and the alterations that occur at the fibroid–myometrium interface remain poorly understood. This study quantitatively maps the local mechanical, structural, and compositional properties of seedling uterine fibroids ( < 1 cm) at the interface region. Relative to patient-matched myometrial tissues, uterine fibroids exhibited increased stiffness, decreased permeability, increased diffusivity, decreased hydration, greater collagen content, and distinct collagen organization. At the fibroid–myometrium interface, a band of aligned myometrial fibers immediately adjacent to the fibroid was observed, and heterogeneous spatial patterns in elastic modulus and permeability were identified. Ultimately, this study establishes foundational knowledge on the mechanics and structure of seedling uterine fibroids, facilitating future developments of clinically translatable detection tools.
- Open Access
Identification of Sample Processing Errors in Microbiome Studies Using Host Genetic Profiles
In microbiome studies, sample processing errors are frequent and difficult to detect, especially in large studies involving multiple sites, personnel, and sample types. We present two complementary approaches to identify such errors using host DNA profiled via metagenomic sequencing of microbiome samples. The first approach compares host SNPs inferred from metagenomics to independently obtained genotypes (e.g., microarray genotypes) to match samples to their donors, while the second method compares metagenomics-inferred SNPs between samples to identify samples supplied by the same donor. Furthermore, we demonstrate that combining these methods with experimental metadata provides greater confidence in the identification of errors. Analyzing a longitudinal vaginal microbiome dataset, we demonstrate the ability of our approach to identify mislabeled samples. Using subsampling, we further show that our methods are robust to low sequencing coverage. Overall, our analysis highlights the frequency of processing errors in microbiome studies. We therefore recommend applying error-detection methods in all studies with suitable data.
Open AccessIdentification of Sample Processing Errors in Microbiome Studies Using Host Genetic Profiles
In microbiome studies, sample processing errors are frequent and difficult to detect, especially in large studies involving multiple sites, personnel, and sample types. We present two complementary approaches to identify such errors using host DNA profiled via metagenomic sequencing of microbiome samples. The first approach compares host SNPs inferred from metagenomics to independently obtained genotypes (e.g., microarray genotypes) to match samples to their donors, while the second method compares metagenomics-inferred SNPs between samples to identify samples supplied by the same donor. Furthermore, we demonstrate that combining these methods with experimental metadata provides greater confidence in the identification of errors. Analyzing a longitudinal vaginal microbiome dataset, we demonstrate the ability of our approach to identify mislabeled samples. Using subsampling, we further show that our methods are robust to low sequencing coverage. Overall, our analysis highlights the frequency of processing errors in microbiome studies. We therefore recommend applying error-detection methods in all studies with suitable data.
- 27 August 2025
- Published by Elsevier in American Journal of Obstetrics & Gynecology MFM
- Vol. 7 (11) , 101773
A Simple data-driven dietary pattern associated with lower risk for adverse pregnancy outcomes
- 27 August 2025
- Published by Elsevier in American Journal of Obstetrics & Gynecology MFM
- Vol. 7 (11) , 101773
A Simple data-driven dietary pattern associated with lower risk for adverse pregnancy outcomes
- Top Cited PapersOpen Access
- 20 June 2025
- Published by Springer Nature in Nature Microbiology
- Vol. 10 (7) , 1570-1580
Guidelines for preventing and reporting contamination in low-biomass microbiome studies
Numerous important environments harbour low levels of microbial biomass, including certain human tissues, the atmosphere, plant seeds, treated drinking water, hyper-arid soils and the deep subsurface, with some environments lacking resident microbes altogether. These low microbial biomass environments pose unique challenges for standard DNA-based sequencing approaches, as the inevitability of contamination from external sources becomes a critical concern when working near the limits of detection. Likewise, lower-biomass samples can be disproportionately impacted by cross-contamination and practices suitable for handling higher-biomass samples may produce misleading results when applied to lower microbial biomass samples. This Consensus Statement outlines strategies to reduce contamination and cross-contamination, focusing on marker gene and metagenomic analyses. We also provide minimal standards for reporting contamination information and removal workflows. Considerations must be made at every study stage, from sample collection and handling through data analysis and reporting to reduce and identify contaminants. We urge researchers to adopt these recommendations when designing, implementing and reporting microbiome studies, especially those conducted in low-biomass systems.
- 20 June 2025
- Published by Springer Nature in Nature Microbiology
- Vol. 10 (7) , 1570-1580
Top Cited PapersOpen AccessGuidelines for preventing and reporting contamination in low-biomass microbiome studies
Numerous important environments harbour low levels of microbial biomass, including certain human tissues, the atmosphere, plant seeds, treated drinking water, hyper-arid soils and the deep subsurface, with some environments lacking resident microbes altogether. These low microbial biomass environments pose unique challenges for standard DNA-based sequencing approaches, as the inevitability of contamination from external sources becomes a critical concern when working near the limits of detection. Likewise, lower-biomass samples can be disproportionately impacted by cross-contamination and practices suitable for handling higher-biomass samples may produce misleading results when applied to lower microbial biomass samples. This Consensus Statement outlines strategies to reduce contamination and cross-contamination, focusing on marker gene and metagenomic analyses. We also provide minimal standards for reporting contamination information and removal workflows. Considerations must be made at every study stage, from sample collection and handling through data analysis and reporting to reduce and identify contaminants. We urge researchers to adopt these recommendations when designing, implementing and reporting microbiome studies, especially those conducted in low-biomass systems.