TransSurv: compatibility-aware transfer learning for target-cohort survival prediction
Abstract
Transcriptomic prognostic models are most useful for narrowly specified, event-limited target populations such as clinical-trial cohorts or stage-defined subtypes. Candidate auxiliary cohorts may carry related prognostic information but differ in disease context, assay platform, and population structure, so naive pooling—and transfer learning without compatibility screening—can erode target-cohort discrimination. We present TransSurv, a target-centered transfer-learning framework for high-dimensional Cox modeling that screens candidate auxiliary cohorts for empirical compatibility before borrowing, comes with non-asymptotic guarantees, and reports which auxiliary cohorts contributed. In the JAVELIN Renal 101 metastatic renal cell carcinoma trial—to our knowledge, the first transfer-learning survival evaluation with a randomized trial as the target—TransSurv raised held-out progression-free survival discrimination. It also achieved the highest mean discrimination in all five clinically defined target cohorts (kidney, ovarian, cervical, lung, breast) from The Cancer Genome Atlas (TCGA) and improved recurrence prediction in a cross-platform, cross-population Fudan University Shanghai Cancer Center triple-negative breast cancer analysis using TCGA and METABRIC auxiliaries.
Related
- BRIDGE-KO: auditable transfer learning with bridged knockoffs improves cross-cohort microbiome prediction of immunotherapy response
- Multi-omics survival prediction under modality incompleteness via dynamic modality graphs and product-of-experts fusion.
- Automated phytoplankton single‐cell counting from digital images with transfer learning for harmful algal bloom monitoring in the English Channel and North Sea
- Representation learning for low-data protein fitness prediction using sequence–structure embeddings and graph neural networks
- A comparative analysis of graph neural networks and classical machine learning for carbapenem-resistance prediction and genomic interpretation in Acinetobacter baumannii
BibTeX
@article{qu2026transsurv,
title = {{TransSurv: compatibility-aware transfer learning for target-cohort survival prediction}},
author = {Zheng Qu and Hongzhe Zhang and Katherine L Nathanson and Yi Fang and Hongzhe Li and Jing Wang},
journal = {Briefings in Bioinformatics},
year = {2026},
month = oct,
doi = {10.1093/bib/bbag532},
url = {https://doi.org/10.1093/bib/bbag532},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - TransSurv: compatibility-aware transfer learning for target-cohort survival prediction
AU - Zheng Qu
AU - Hongzhe Zhang
AU - Katherine L Nathanson
AU - Yi Fang
AU - Hongzhe Li
AU - Jing Wang
PY - 2026
DA - 2026/10/27
JO - Briefings in Bioinformatics
DO - 10.1093/bib/bbag532
UR - https://doi.org/10.1093/bib/bbag532
AB - Transcriptomic prognostic models are most useful for narrowly specified, event-limited target populations such as clinical-trial cohorts or stage-defined subtypes. Candidate auxiliary cohorts may carry related prognostic information but differ in disease context, assay platform, and population structure, so naive pooling—and transfer learning without compatibility screening—can erode target-cohort discrimination. We present TransSurv, a target-centered transfer-learning framework for high-dimensional Cox modeling that screens candidate auxiliary cohorts for empirical compatibility before borrowing, comes with non-asymptotic guarantees, and reports which auxiliary cohorts contributed. In the JAVELIN Renal 101 metastatic renal cell carcinoma trial—to our knowledge, the first transfer-learning survival evaluation with a randomized trial as the target—TransSurv raised held-out progression-free survival discrimination. It also achieved the highest mean discrimination in all five clinically defined target cohorts (kidney, ovarian, cervical, lung, breast) from The Cancer Genome Atlas (TCGA) and improved recurrence prediction in a cross-platform, cross-population Fudan University Shanghai Cancer Center triple-negative breast cancer analysis using TCGA and METABRIC auxiliaries.
SN - 1467-5463
SN - 1477-4054
ER -