ecPath

Extrachromosomal DNA — circular, acentric fragments carrying amplified oncogenes — segregates unevenly at mitosis, inflates copy number, and marks some of the most aggressive tumors. Knowing whether a tumor carries it would change how a patient is stratified, but establishing it today means whole-genome sequencing with amplicon reconstruction, metaphase karyotyping, or FISH. None of those is part of a routine diagnostic workup, so ecDNA status goes unmeasured for most patients — while the one slide nearly every solid-tumor patient does get, stained with hematoxylin and eosin, sits unused for the question.

ecPath closes that gap by refusing to make the jump in a single step. A classifier trained to read ecDNA straight off tile features has to learn a genomic property from morphology using only a few hundred labeled slides, and the paper's own benchmark shows it largely fails (AUC 0.63 with ResNet-50 features, 0.65 with UNI features). ecPath instead inserts the tumor transcriptome as an intermediate. A multilayer perceptron over frozen tile features predicts bulk expression for 18,000 genes; a logistic regression over the reliably predicted subset of that inferred expression then calls ecDNA status. The authors call this transcriptomics-guided learning: matched expression labels are plentiful in TCGA and ecDNA labels are scarce, so the plentiful signal shapes the representation and the scarce one only has to fit a linear model on top.

The model was built at Sanford Burnham Prebys with collaborators at Rady Children's, UC San Diego, the National Cancer Institute and Leibniz University Hannover, and posted as a preprint in November 2024.

#Key Features

• Expression as a supervision bridge: The intermediate transcriptome step lifts mean ecDNA AUC from 0.65 to 0.77 over predicting ecDNA directly from the same slide features.
• Frozen pathology-foundation features: Tiles are encoded by UNI, a vision transformer whose features supported predictions for 7,074 genes on average across 16 tumor types, against 4,084 for a ResNet-50 encoder. The released pipeline defaults to ResNet-50 and switches to UNI when the user supplies an approved access token.
• Ready-to-run per-cancer checkpoints: Inference ships with trained weights for low-grade glioma, glioblastoma and stomach adenocarcinoma, each with a decision threshold pre-calibrated by Youden's J (0.268, 0.226 and 0.567), so no re-fitting is required on a new cohort.
• Discriminates ecDNA from other amplifications: Against samples carrying breakage-fusion-bridge, complex non-cyclic or linear amplifications rather than ecDNA, AUCs are 0.76, 0.75 and 0.79.
• Morphology that a pathologist can read: Tiles weighted most heavily in the predictions were reviewed and found enriched for denser, larger and more pleomorphic nuclei.

#Technical Details

Pretraining used 6,262 FFPE whole-slide images from 6,189 patients across 16 TCGA tumor types. Slides are tissue-masked with Sobel edge detection, Macenko color-normalized and cut into 256 × 256-pixel tiles at 20× following the DeepPT preprocessing protocol; UNI returns 1,024 features per tile. Genes are retained when predicted and measured expression correlate at Pearson r > 0.4 with FDR-corrected q < 0.05 — 7,074 per tumor type on average, and 12,510 for low-grade glioma. ecDNA labels came from AmpliconRepository calls made with AmpliconArchitect on PCAWG whole genomes, covering 797 slides from 766 patients, 207 of them ecDNA-positive; nine tumor types had at least ten samples in each class. For each type, 150 ecDNA-associated genes feed a logistic regression trained in repeated nested five-fold cross-validation, 200 iterations for 1,000 models in all. From predicted expression, ecDNA AUC is 0.82 in glioblastoma, 0.78 in low-grade glioma and 0.73 in stomach adenocarcinoma; measured expression sets the ceiling at 0.83, 0.92 and 0.81. A model restricted to ecDNA amplifications harboring EGFR reached AUC 0.73.

#Applications

The practical use is triage: scoring archival or newly cut H&E slides to nominate which tumors are worth confirmatory sequencing, and to assign ecDNA status retrospectively in cohorts that were never sequenced. A brain-specific model trained on TCGA gliomas scored 985 primary tumors from the Children's Brain Tumor Network — a separate institution and an entirely pediatric population — at AUC 0.72. Applied to 802 TCGA patients with no whole-genome data, it called 265 tumors ecDNA-positive, and predicted status recovered the same poor survival association in low-grade glioma that sequencing-derived labels show.

#Impact

ecPath makes a genomic feature legible from the cheapest and most widely collected image in oncology, and its benchmark against direct image-to-ecDNA classifiers argues concretely for routing scarce labels through an abundant intermediate — the same design Path2Omics applies to expression and methylation. The limits are stated plainly: accuracy suits screening rather than diagnosis, the released inference path covers only three tumor types and .svs slides, the external cohort is pediatric while training was adult, and the work remains a preprint awaiting peer review. The code carries no license file, and the authors have filed a provisional patent on detecting ecDNA from histopathology images.