Five signals from this year’s posters, and what they mean for anyone trying to take a biomarker from discovery to the clinic.

Walk the poster hall at ASCO and you get a different read on the field than you do from the plenary sessions. The plenaries tell you which drugs won. The posters tell you how the people running those trials are thinking: what they’re measuring, what they’re betting on, and where the science is quietly moving before it becomes consensus.

This year, one thing was unmistakable. The conversation has shifted from which therapy to which patient, which test, and how do we actually deploy it. Biomarkers weren’t a sidebar at ASCO 2026; they were the main event. Reading across the floor, five signals stood out. Each one is a window into where diagnostics is heading, and each one is a reminder that the hard part is rarely the discovery.

 

1. Liquid biopsy grew up, and went multi-modal

For years, “liquid biopsy” was shorthand for ctDNA. That framing is now obsolete. The most interesting blood-based work at ASCO 2026 wasn’t about a single analyte. It was about combining several.

Groups presented cfDNA fragmentomics layered with serum proteins to differentiate ovarian tumors; methylation signals from plasma cell-free DNA used to predict tumor-of-origin in cancers of unknown primary, validated across real-world cohorts in Asia and the Middle East; and extracellular-vesicle proteomics applied to lung-nodule triage and ovarian detection from uterine lavage. There were transformer-based “language models” reading cfDNA fragments to detect lymphoma from ultra-low-pass whole-genome sequencing, and personalized whole-genome ctDNA assays catching melanoma recurrence from a single early on-treatment blood draw.

The throughline: the blood holds far more signal than any one readout captures, and the winners will be the teams that integrate across genomic, epigenomic, fragmentomic, and proteomic layers into a single decision. That’s a much harder analytical and operational problem than running one assay well. It rewards platforms over point solutions.

 

2. AI was everywhere, but the bottleneck has moved to validation and deployment

It was hard to find a poster without a machine-learning model. Transcriptomic risk scores for metastatic prostate cancer. Foundation models reading whole bone-marrow slides for myelodysplastic syndromes. RNA-based homologous-recombination-deficiency algorithms predicting platinum response in pancreatic cancer. Deep-learning counterfactual survival models identifying which gastric-cancer patients can safely de-escalate adjuvant therapy. Quantified tumor-microenvironment features pulled from routine H&E slides to refine colon-cancer risk.

But the most honest poster on the floor may have been one that framed an AI paradox in precision oncology, using a blinded, prospective comparison of large language models against a molecular tumor board. The message underneath all the performance metrics was consistent: building a model is no longer the differentiator. The real challenge is proving it holds up outside the lab it was built in, across different cohorts, patient populations, and the messiness of real-world data. Several groups named the same culprits: single-institution training, modest performance on external validation, and reproducibility.

This is the gap between an impressive figure and a deployable test. It’s a validation problem, a data-diversity problem, and a regulatory-grade-evidence problem, and it’s exactly where most promising biomarkers stall.

 

3. The signal is multi-omic now, and the genome alone isn’t enough

A clear cohort of posters moved deliberately beyond DNA. Targeted phosphoproteomic assays predicting response to intensive chemotherapy in AML and durable response to CAR-T in B-cell malignancies. Plasma protein signatures, built with DIA mass spectrometry, predicting benefit from anti-angiogenic regimens in colorectal cancer. Lipidomics combined with proteins for early-stage ovarian detection. Microbiome and host-immune profiling explaining immunotherapy response in anal and biliary cancers. Multi-omic methylation-plus-expression work distinguishing early-onset colorectal cancer.

Biology is being read in layers, because the layers carry complementary information. For anyone developing biomarkers, this raises the bar: you increasingly need infrastructure that can run, and harmonize, proteomic, epigenomic, and transcriptomic measurements alongside the genomics, without each modality becoming its own bespoke project.

 

4. The point of a biomarker is to change a decision

The strongest translational posters weren’t content to detect or prognosticate. They aimed to move treatment.

De-escalation was a recurring theme: identifying the patient subset that gains little from an added agent, so it can be safely dropped. So was complementary biomarker thinking: tumor-infiltrating lymphocytes adding signal independent of PD-L1; CLDN18.2 emerging as a target and a stratifier in gastro-esophageal cancers; four-gene and HRD signatures refining who actually responds to PARP inhibition in BRCA-mutated ovarian cancer. Even the long-term KEYNOTE-811 follow-up in HER2-positive gastric cancer sat against a backdrop of sharper biomarker-defined subgroups.

The implication for partners is strategic, not just scientific. The biomarkers that matter are the ones wired to a clinical action: select, deselect, escalate, de-escalate. That’s what makes a companion or complementary diagnostic, and it’s what payers and regulators ultimately reward.

 

5. The new frontier is access, not just accuracy

Perhaps the quietest but most consequential signal: a wave of work aimed at getting diagnostics out of the academic center.

There were early-detection models built entirely from routine laboratory data, flagging multiple myeloma and MGUS progression without any new assay. Consumer wearables used to identify oncology patients at risk of delayed care. And implementation blueprints like PRIME-ROSE, mapping how to actually run precision-medicine programs across national health systems by harmonizing data, governance, and reimbursement so a test can scale beyond a single hospital.

This is the part the science rarely gets credit for. A biomarker that only works in a tertiary referral center with on-site sequencing isn’t a product; it’s a publication. The frontier now is decentralization: sample collection that meets the patient where they are, logistics that hold quality across geographies, and the operational backbone to turn a validated signature into a test real people can actually get.

 

The real gap isn’t discovery

Read across all five signals and a single pattern emerges. The science on display at ASCO 2026 was extraordinary, but almost none of the open problems were discovery problems. They were translation problems: validating a model across diverse, real-world cohorts; integrating multiple analytes into one robust readout; running regulated, multi-omic assays at consistent quality; and deploying all of it into decentralized settings at scale.

That’s the unglamorous middle of the diagnostics journey, the stretch between a beautiful poster and a test in routine use, and it’s where the field’s energy is clearly heading. It’s also, not coincidentally, the part we spend our days on. The discoveries will keep coming. The question for the next few years is who can carry them the rest of the way.

This is our read of the ASCO 2026 poster floor. If you’re working on a biomarker and thinking about the path from validation to deployment, multi-omic, decentralized, and regulatory-grade, we’d enjoy comparing notes.