Kevin K. Tsang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kevin K. Tsang has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Aleck Cervantes
Center for Precision Medicine, City of Hope National Medical Center, Duarte, CA
3 shared publications - Alison Schwartz Levine
Division of Cancer Genetics and Prevention, Dana-Farber Cancer Institute, Boston, MA 02215, USA
3 shared publications - Benjamin L. Ebert
Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, MA
3 shared publications - Bita Nehoray
City of Hope Cancer Center, Duarte, California
3 shared publications - Christopher Ian Amos
Institute for Clinical and Translational Research, Baylor College of Medicine, Houston, TX, USA
3 shared publications - Danielle Castillo
City of Hope, Duarte, CA
3 shared publications
Latest Publications
- Open Access
- 26 November 2025
- Published by American Association for Cancer Research (AACR)
Data from Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
Background:Though germline TP53 pathogenic/likely pathogenic variants (PV) are associated with Li–Fraumeni syndrome, many detected by multigene panels represent aberrant clonal expansion (ACE), most due to clonal hematopoiesis (CH). Discerning ACE/CH from germline variants and postzygotic mosaicism (PZM) is critically needed for risk assessment and management.Methods:Participants in the Li-Fraumeni & TP53 Understanding & Progress (LiFT UP) study with a TP53 PV were eligible. Demographics, personal/family cancer history, and clinical laboratory test reports were obtained. DNA from multiple tissues was analyzed using a custom QIAseq assay (ACE panel) that included TP53 and other CH-associated genes; the ACE panel and eyebrow follicles were assessed in a workflow to discern TP53 PV clinical categories.Results:Among 134 participants there was a significant difference for the age at diagnosis (P < 0.001), component cancers (P = 0.007), and clinical testing criteria (P < 0.001), comparing germline with PZM or ACE. ACE panel analysis of DNA from 55 sets of eyebrow follicles (mean 1.4 ug) and 36 formalin-fixed, paraffin imbedded tissues demonstrated low variance (SE, 3%; P = 0.993) for TP53 variant allele fraction, with no significant difference (P = 0.965) between tissue types, and detected CH gene PVs. Of 55 multi-tissue cases, germline status was confirmed for 20, PZM in seven, ACE for 25, and three were indeterminate. Additional CH variants were detected in six ACE and two germline cases.Conclusions:We demonstrated an effective approach and tools for discerning germline TP53 status.Impact:Discernment of PZM and TP53-driven CH increases diagnostic accuracy and enables risk-appropriate care.
- 26 November 2025
- Published by American Association for Cancer Research (AACR)
Open AccessData from Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
Background:Though germline TP53 pathogenic/likely pathogenic variants (PV) are associated with Li–Fraumeni syndrome, many detected by multigene panels represent aberrant clonal expansion (ACE), most due to clonal hematopoiesis (CH). Discerning ACE/CH from germline variants and postzygotic mosaicism (PZM) is critically needed for risk assessment and management.Methods:Participants in the Li-Fraumeni & TP53 Understanding & Progress (LiFT UP) study with a TP53 PV were eligible. Demographics, personal/family cancer history, and clinical laboratory test reports were obtained. DNA from multiple tissues was analyzed using a custom QIAseq assay (ACE panel) that included TP53 and other CH-associated genes; the ACE panel and eyebrow follicles were assessed in a workflow to discern TP53 PV clinical categories.Results:Among 134 participants there was a significant difference for the age at diagnosis (P < 0.001), component cancers (P = 0.007), and clinical testing criteria (P < 0.001), comparing germline with PZM or ACE. ACE panel analysis of DNA from 55 sets of eyebrow follicles (mean 1.4 ug) and 36 formalin-fixed, paraffin imbedded tissues demonstrated low variance (SE, 3%; P = 0.993) for TP53 variant allele fraction, with no significant difference (P = 0.965) between tissue types, and detected CH gene PVs. Of 55 multi-tissue cases, germline status was confirmed for 20, PZM in seven, ACE for 25, and three were indeterminate. Additional CH variants were detected in six ACE and two germline cases.Conclusions:We demonstrated an effective approach and tools for discerning germline TP53 status.Impact:Discernment of PZM and TP53-driven CH increases diagnostic accuracy and enables risk-appropriate care.
- Open Access
- 26 November 2025
- Published by American Association for Cancer Research (AACR)
Supplementary Figure from Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
Supplementary Figure from Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
- 26 November 2025
- Published by American Association for Cancer Research (AACR)
Open AccessSupplementary Figure from Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
Supplementary Figure from Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
- Open Access
Probing Hidden States for Calibrated, Alignment-Resistant Predictions in LLMs
Scientific applications of large language models (LLMs) demand reliable, well-calibrated predictions, but standard generative approaches often fail to fully access relevant knowledge contained in their internal representations. As a result, models appear less capable than they are, with useful information remaining latent. We present PING (Probing INternal states of Generative models), an open-source framework that trains lightweight probes on frozen, HuggingFace-compatible transformers to deliver structured predictions with minimal compute overhead. Across diverse models and benchmarks including MMLU for broad coverage and MedMCQA for clinical focus, PING matches or exceeds generative accuracy while reducing Expected Calibration Error by up to 96%. Strikingly, on an LLM that has been explicitly safety-tuned to withhold medical information, PING recovered 87% of lost MedMCQA performance while generative accuracy is zero, showing this information still exists in the model’s latent space. The accompanying pingkit package makes these methods easy to deploy and is available through PyPI.
Open AccessProbing Hidden States for Calibrated, Alignment-Resistant Predictions in LLMs
Scientific applications of large language models (LLMs) demand reliable, well-calibrated predictions, but standard generative approaches often fail to fully access relevant knowledge contained in their internal representations. As a result, models appear less capable than they are, with useful information remaining latent. We present PING (Probing INternal states of Generative models), an open-source framework that trains lightweight probes on frozen, HuggingFace-compatible transformers to deliver structured predictions with minimal compute overhead. Across diverse models and benchmarks including MMLU for broad coverage and MedMCQA for clinical focus, PING matches or exceeds generative accuracy while reducing Expected Calibration Error by up to 96%. Strikingly, on an LLM that has been explicitly safety-tuned to withhold medical information, PING recovered 87% of lost MedMCQA performance while generative accuracy is zero, showing this information still exists in the model’s latent space. The accompanying pingkit package makes these methods easy to deploy and is available through PyPI.
- Open Access
- 27 March 2025
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 13 (1) , e67591
Biases in Race and Ethnicity Introduced by Filtering Electronic Health Records for “Complete Data”: Observational Clinical Data Analysis
Background: Integrated clinical databases from national biobanks have advanced the capacity for disease research. Data quality and completeness filters are used when building clinical cohorts to address limitations of data missingness. However, these filters may unintentionally introduce systemic biases when they are correlated with race and ethnicity. Objective: In this study, we examined the race and ethnicity biases introduced by applying common filters to 4 clinical records databases. Specifically, we evaluated whether these filters introduce biases that disproportionately exclude minoritized groups. Methods: We applied 19 commonly used data filters to electronic health record datasets from 4 geographically varied locations comprising close to 12 million patients to understand how using these filters introduces sample bias along racial and ethnic groupings. These filters covered a range of information, including demographics, medication records, visit details, and observation periods. We observed the variation in sample drop-off between self-reported ethnic and racial groups for each site as we applied each filter individually. Results: Applying the observation period filter substantially reduced data availability across all races and ethnicities in all 4 datasets. However, among those examined, the availability of data in the white group remained consistently higher compared to other racial groups after applying each filter. Conversely, the Black or African American group was the most impacted by each filter on these 3 datasets: Cedars-Sinai dataset, UK Biobank, and Columbia University dataset. Among the 4 distinct datasets, only applying the filters to the All of Us dataset resulted in minimal deviation from the baseline, with most racial and ethnic groups following a similar pattern. Conclusions: Our findings underscore the importance of using only necessary filters, as they might disproportionally affect data availability of minoritized racial and ethnic populations. Researchers must consider these unintentional biases when performing data-driven research and explore techniques to minimize the impact of these filters, such as probabilistic methods or adjusted cohort selection methods. Additionally, we recommend disclosing sample sizes for racial and ethnic groups both before and after data filters are applied to aid the reader in understanding the generalizability of the results. Future work should focus on exploring the effects of filters on downstream analyses.
- 27 March 2025
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 13 (1) , e67591
Open AccessBiases in Race and Ethnicity Introduced by Filtering Electronic Health Records for “Complete Data”: Observational Clinical Data Analysis
Background: Integrated clinical databases from national biobanks have advanced the capacity for disease research. Data quality and completeness filters are used when building clinical cohorts to address limitations of data missingness. However, these filters may unintentionally introduce systemic biases when they are correlated with race and ethnicity. Objective: In this study, we examined the race and ethnicity biases introduced by applying common filters to 4 clinical records databases. Specifically, we evaluated whether these filters introduce biases that disproportionately exclude minoritized groups. Methods: We applied 19 commonly used data filters to electronic health record datasets from 4 geographically varied locations comprising close to 12 million patients to understand how using these filters introduces sample bias along racial and ethnic groupings. These filters covered a range of information, including demographics, medication records, visit details, and observation periods. We observed the variation in sample drop-off between self-reported ethnic and racial groups for each site as we applied each filter individually. Results: Applying the observation period filter substantially reduced data availability across all races and ethnicities in all 4 datasets. However, among those examined, the availability of data in the white group remained consistently higher compared to other racial groups after applying each filter. Conversely, the Black or African American group was the most impacted by each filter on these 3 datasets: Cedars-Sinai dataset, UK Biobank, and Columbia University dataset. Among the 4 distinct datasets, only applying the filters to the All of Us dataset resulted in minimal deviation from the baseline, with most racial and ethnic groups following a similar pattern. Conclusions: Our findings underscore the importance of using only necessary filters, as they might disproportionally affect data availability of minoritized racial and ethnic populations. Researchers must consider these unintentional biases when performing data-driven research and explore techniques to minimize the impact of these filters, such as probabilistic methods or adjusted cohort selection methods. Additionally, we recommend disclosing sample sizes for racial and ethnic groups both before and after data filters are applied to aid the reader in understanding the generalizability of the results. Future work should focus on exploring the effects of filters on downstream analyses.
- 2 June 2022
- Published by American Association for Cancer Research (AACR) in Cancer Epidemiology, Biomarkers & Prevention
- Vol. 31 (8) , 1621-1629
Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
Background: Though germline TP53 pathogenic/likely pathogenic variants (PV) are associated with Li–Fraumeni syndrome, many detected by multigene panels represent aberrant clonal expansion (ACE), most due to clonal hematopoiesis (CH). Discerning ACE/CH from germline variants and postzygotic mosaicism (PZM) is critically needed for risk assessment and management. Methods: Participants in the Li-Fraumeni & TP53 Understanding & Progress (LiFT UP) study with a TP53 PV were eligible. Demographics, personal/family cancer history, and clinical laboratory test reports were obtained. DNA from multiple tissues was analyzed using a custom QIAseq assay (ACE panel) that included TP53 and other CH-associated genes; the ACE panel and eyebrow follicles were assessed in a workflow to discern TP53 PV clinical categories. Results: Among 134 participants there was a significant difference for the age at diagnosis (P < 0.001), component cancers (P = 0.007), and clinical testing criteria (P < 0.001), comparing germline with PZM or ACE. ACE panel analysis of DNA from 55 sets of eyebrow follicles (mean 1.4 ug) and 36 formalin-fixed, paraffin imbedded tissues demonstrated low variance (SE, 3%; P = 0.993) for TP53 variant allele fraction, with no significant difference (P = 0.965) between tissue types, and detected CH gene PVs. Of 55 multi-tissue cases, germline status was confirmed for 20, PZM in seven, ACE for 25, and three were indeterminate. Additional CH variants were detected in six ACE and two germline cases. Conclusions: We demonstrated an effective approach and tools for discerning germline TP53 status. Impact: Discernment of PZM and TP53-driven CH increases diagnostic accuracy and enables risk-appropriate care.
- 2 June 2022
- Published by American Association for Cancer Research (AACR) in Cancer Epidemiology, Biomarkers & Prevention
- Vol. 31 (8) , 1621-1629
Clonal Hematopoiesis and Mosaicism Revealed by a Multi-Tissue Analysis of Constitutional TP53 Status
Background: Though germline TP53 pathogenic/likely pathogenic variants (PV) are associated with Li–Fraumeni syndrome, many detected by multigene panels represent aberrant clonal expansion (ACE), most due to clonal hematopoiesis (CH). Discerning ACE/CH from germline variants and postzygotic mosaicism (PZM) is critically needed for risk assessment and management. Methods: Participants in the Li-Fraumeni & TP53 Understanding & Progress (LiFT UP) study with a TP53 PV were eligible. Demographics, personal/family cancer history, and clinical laboratory test reports were obtained. DNA from multiple tissues was analyzed using a custom QIAseq assay (ACE panel) that included TP53 and other CH-associated genes; the ACE panel and eyebrow follicles were assessed in a workflow to discern TP53 PV clinical categories. Results: Among 134 participants there was a significant difference for the age at diagnosis (P < 0.001), component cancers (P = 0.007), and clinical testing criteria (P < 0.001), comparing germline with PZM or ACE. ACE panel analysis of DNA from 55 sets of eyebrow follicles (mean 1.4 ug) and 36 formalin-fixed, paraffin imbedded tissues demonstrated low variance (SE, 3%; P = 0.993) for TP53 variant allele fraction, with no significant difference (P = 0.965) between tissue types, and detected CH gene PVs. Of 55 multi-tissue cases, germline status was confirmed for 20, PZM in seven, ACE for 25, and three were indeterminate. Additional CH variants were detected in six ACE and two germline cases. Conclusions: We demonstrated an effective approach and tools for discerning germline TP53 status. Impact: Discernment of PZM and TP53-driven CH increases diagnostic accuracy and enables risk-appropriate care.