Loading…
Liquid-biopsy deep learning model that infers transcriptome-wide tumor gene expression from standard-depth cell-free DNA whole-genome sequencing.
No providers recorded yet. Browse all providers
A gene that is being actively transcribed leaves a physical trace in plasma. The nucleosome occupying its transcription start site is displaced, so the cell-free DNA fragments that survive circulation are depleted over that window, more variable in length and differently phased than the fragments protecting a silent gene. Reading that trace back into a per-gene expression value has previously required a specialized chromatin assay or an ultra-deep targeted panel, both of which cap how much of the transcriptome is visible at once.
Proteus infers gene-level tumor expression directly from ordinary whole-genome sequencing of plasma cfDNA at roughly 30–120x. It was developed in Gavin Ha's laboratory at Fred Hutchinson Cancer Center with the University of Washington. Its input features come from Triton, a companion cfDNA fragmentation and phased-nucleosome profiling pipeline from the same laboratory, unrelated to the GPU kernel language and inference server of that name. The model name is likewise shared with an unrelated protein backbone generation model from Westlake University; this entry covers the cfDNA model.
Where LEAF-1 represents individual cfDNA fragments and cfDecon deconvolves methylation reads into cell-type fractions, Proteus targets the tumor's transcriptional state, gene by gene, in units a standard RNA-seq pipeline already accepts.
Proteus is implemented in PyTorch as BioDVN, a biologically informed disentangling variational network with 4,512,344 trainable parameters. Triton emits per-gene TSS ±1,000 bp signal (2,000 bp × 9 channels), one-hot promoter sequence, and 11 TSS, 11 gene-body and 3 reference scalar features over MANE Select v1.3 transcripts; convolutions reduce the promoter map to 64 dimensions before parallel encoders (width 256, latent dimension 64) form the Gaussian tumor and background latents. Training used 40 patient-derived xenograft ctDNA models, 10 healthy-donor cfDNA samples, 80 in silico admixtures and two patient cohorts with matched tumor RNA-seq, optimized under a continuous ranked probability score objective with AdamW for 100 epochs and an exponential moving average of weights. A single production checkpoint serves every downstream cohort. On held-out cohorts it reached sample-level and gene-variance-weighted Spearman correlations of 0.86 and 0.26 (CRPC-WCDT) and 0.83 and 0.21 (BLCA-RA, an unseen cancer type), against 0.63 and 0.10 for EPIC-Seq in WGS mode, with AUROC 0.92–0.94 for on/off calls across the Cancer Surfaceome and Druggable Genome Tier 1 gene sets. Inference over roughly 18,000 genes takes under a minute per sample on a single GPU.
The immediate use is non-invasive assessment of therapeutic target expression in advanced cancer, where tissue biopsies are costly, spatially biased and often unavailable. Predicted expression tracked tumor RNA-seq for STEAP1 (r = 0.68, rising to 0.76 with abstention), NECTIN4, DLL3 and SEZ6 — targets of antibody-drug conjugates and cellular therapies in prostate, bladder and neuroendocrine disease. In a prostate cohort lacking matched RNA-seq, cfDNA-derived neuroendocrine scores separated pure NEPC from AR-driven disease (AUROC 0.90), and across 115 patients on 177Lu-PSMA-617 radioligand therapy, higher G2M-checkpoint scores tracked shorter survival (HR 1.40).
Proteus extends cfDNA analysis past mutation calling and tumor-fraction estimation toward serial readouts of tumor regulatory state, from sequencing many oncology programs already generate. Its more consequential argument is about evaluation: because expression rank is largely conserved across samples from one tissue, a naive predictor assigning each gene a random other sample's value still scores a sample-level correlation of 0.91, so the field's usual metric can look excellent while recovering no patient-specific signal. The gene-variance-weighted correlation adopted here is a stricter standard other cfDNA expression methods can be held to. Accuracy remains strongly tumor-fraction dependent, the model cannot resolve individual lesions or subclones within mixed cfDNA, and the work is a preprint awaiting peer review. Triton carries a Fred Hutchinson academic non-commercial license; Proteus source code and checkpoints are unreleased, with the authors stating they will be published on GitHub prior to publication.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.