Summary
A medRxiv preprint reports higher C-index estimates for several survival outcomes when Evo2-derived tumor-sequence features were added to existing risk models. The study analysed 102 patients with soft-tissue sarcoma and a separate 129-patient TCGA cohort; clinical use requires further validation.
A medRxiv preprint reports that adding tumor-sequence features generated by the genomic foundation model Evo2 raised the reported C-index for several survival outcomes in soft-tissue sarcoma. The computational study analysed 102 patients across three sarcoma types and assessed the models in a separate external cohort of 129 patients. The authors describe the comparisons as descriptive point estimates and say further confirmation is needed before clinical application.
How Evo2 was added to existing risk models
The researchers analysed formalin-fixed, paraffin-embedded tumour material and clinical data from patients with dedifferentiated liposarcoma (25), high-grade leiomyosarcoma (25) and undifferentiated pleomorphic sarcoma (52). They also used the TCGA-SARC cohort, comprising 129 patients with comparable histologies, for external validation.
Evo2 is a genomic foundation model trained to learn patterns in DNA and RNA sequences. The researchers converted reconstructable tumour alterations—including short variants, fusions and rearrangements—into sequence inputs, generated Evo2 representations and aggregated them into patient-level genomic features. They then used Cox proportional hazards models to assess those features alone and in combination with SARCULATOR, a clinicopathologic risk model, or CINSARC, a transcriptomic model.
The outcomes included overall survival (OS), progression-free survival (PFS), disease-free survival (DFS), relapse-free survival (RFS) and metastasis-free survival (MFS). Model discrimination was assessed with Harrell’s concordance index, or C-index, which measures how well a model ranks patients according to the timing of observed outcomes. It measures ranking performance rather than whether predicted absolute risks are well calibrated.
Reported results and clinical significance
In the 102-patient study cohort, adding Evo2 features to SARCULATOR raised the reported C-index from 0.620 to 0.770 for OS, 0.625 to 0.755 for PFS, 0.612 to 0.714 for DFS, and 0.633 to 0.771 for RFS.
In the external TCGA-SARC cohort, the reported estimates also rose when Evo2 was added to existing models:
| Model and outcome | Existing model | With Evo2 |
|---|---|---|
| SARCULATOR, OS | 0.595 | 0.716 |
| SARCULATOR, MFS | 0.548 | 0.667 |
| CINSARC, OS | 0.476 | 0.653 |
| CINSARC, MFS | 0.553 | 0.650 |
The authors also report that Evo2 alone outperformed SARCULATOR alone and CINSARC alone in TCGA-SARC. Taken together, the findings suggest that representations of altered tumour sequence context may add prognostic information alongside clinicopathologic and transcriptomic models. The reported comparisons are descriptive point estimates; the results do not establish how accurately individual patients’ absolute risks would be estimated in clinical practice.
The work is a medRxiv preprint. Its authors say that transportability, calibration and clinical utility need confirmation before these models are used in clinical care.