Training on in silico host-tissue mixtures reduces host attraction in RNA-seq-based tissue-of-origin classification of metastatic biopsies
Background Transcriptome classifiers that infer the primary site of metastases are trained mostly on primary tumors, but metastatic biopsies also contain normal host-organ cells. Such classifiers can name the biopsied organ instead of the primary, an error we call host attraction. A previous method reduced errors in liver metastases by training a microRNA classifier on in silico tumor–liver mixtures. Using internally preregistered analyses, we tested whether mixtures with several host tissues reduce host attraction in a whole-transcriptome classifier. Results We trained logistic regression classifiers for 32 tumor types with or without in silico tumor–normal mixtures from ten host tissues (HostMix-TOO). In POG570, HostMix-TOO reduced host attraction from 22.8% to 2.4% of at-risk biopsies (n = 378; −20.4 percentage points, 95% CI − 24.3 to − 16.4) and increased top-1 accuracy from 68.2% to 75.6% (n = 512; 7.4 points, 95% CI 3.9 to 10.7). Non-inferiority of accuracy for biopsies from the organ of origin was not shown (86.8% versus 80.2%, n = 91; one-sided 95% lower bound − 12.1 points, margin − 10). In six auxiliary RNA-seq cohorts, all three primary hypotheses were confirmed: host attraction fell from 20.4% to 5.6% (n = 427), top-1 accuracy rose from 52.7% to 69.8% (n = 729), mostly in one bladder cancer cohort, and accuracy for biopsies from the organ of origin was non-inferior (n = 315). Published classifiers SCOPE and CUP-AI-Dx showed host attraction in 12.2–30.4% of at-risk biopsies (HostMix-TOO: 2.4–5.6%), although CUP-AI-Dx was more accurate in auxiliary cohorts. In five microarray cohorts, HostMix-TOO reduced host attraction but lowered top-1 accuracy from 72.4% to 45.5% (n = 536). A post hoc ablation attributed the host-attraction reduction to the mixtures and the microarray loss to feature standardization: without standardization, mixtures reduced microarray host attraction from 25.6% to 9.2% (n = 207) with 77.6% top-1 accuracy. Conclusions For RNA-seq biopsies, training on in silico host-tissue mixtures reduced host attraction and improved accuracy in two preregistered evaluations. Post hoc analyses traced the microarray loss to feature standardization, which requires confirmation. Benchmarks of tissue-of-origin classifiers on metastatic biopsies should report host attraction, with its denominator, alongside top-1 accuracy.
Authors
- Juwon Kang (ORCID: https://orcid.org/0009-0008-1561-2778)
- Junjeong Choi
Institutions
- Yonsei University (KR)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11293314/v1
- Primary Topic
- Gene expression and cancer classification
- Type
- preprint