Liquid biopsy records point to a growing focus on clonal hematopoiesis deconvolution, artifact suppression, fragmentomics, and methylation-based classification. The technical fight is shifting from detection to discrimination: separating real tumor signal from biological and analytical noise when burden is low and clinical stakes are high.
AI records increasingly describe workflow and orchestration problems — not just single-task prediction. And spatial biology continues to expand as a discovery tool while the translational question shifts from higher plex to smaller, deployable clinical signatures.
The majority of diagnostic-affiliated records (643 of 734) are poster sessions. But the higher-impact slots tell a different story: 25 late-breaking posters, 17 minisymposia, 8 Clinical Trials Minisymposia, and 2 Clinical Trials Plenary Sessions — the highest-visibility designation at AACR. A note on the source data's "oral" category: it includes all 35 Exhibitor Spotlight Presentations (vendor-organized, not competitively selected), which inflates the count if read uncritically.
One finding worth highlighting: what investigators use is not always what they choose to advertise. Sequencing appears in ~340 full texts but only 69 titles — it's infrastructure, rarely the headline. Liquid biopsy (122 title, 164 full-text), spatial biology (83 title, 117 full-text), and AI/ML (50 title, 128 full-text) all show tighter title-to-full-text ratios. These are technologies researchers are choosing to headline, not just employ. In a conference corpus, that ratio may be a rough proxy for how novel versus routine a technology still feels to its users.
One theme conspicuously absent: "companion diagnostics" as a named concept. Zero title mentions, 12 in full text. The term itself may be aging out of the scientific vocabulary even as the underlying practice of biomarker-guided treatment selection expands. PD-L1, the canonical CDx biomarker, appears in only 7 titles — while HER2 (15) and MSI/MMR (10) carry more weight.
The energy in diagnostics at AACR 2026 is in discovery and surveillance, not in the established CDx framework.
This brief is drawn from the ~9,600 abstracts accepted for AACR 2026. All counts are approximate and reflect keyword-based classification. Where the analysis moves from observation to interpretation, that transition is flagged.
Record counts reflect affiliation strings in the source data. A single record may be counted for multiple companies. Names not normalized.
Tempus shows the widest technology spread of any single company in the corpus, with records spanning AI-driven clinical data curation, predictive modeling, and deep learning pathology. Labcorp's portfolio covers methylation, CH classification, drug sensitivity testing, and comprehensive genomic profiling. Bruker concentrates nearly all of its 31 records in a single technology lane: spatial biology. These reflect different bets about where a diagnostics company adds value — and the corpus surfaces all three approaches at scale.
Further down the table, the technology identities sharpen. Guardant Health is closely associated with liquid biopsy. Natera spans MRD, screening, HRD scoring, and genomic characterization — one of the broader cross-functional footprints in the mid-table range. BostonGene is tumor microenvironment profiling. Lunit is AI pathology. The record counts are smaller, but the thematic focus in each portfolio is more defined.
Below the top 15, companies like Foundation Medicine (7), Biodesix (6), and Exact Sciences (5) have smaller footprints but disproportionate impact in specific niches — FMI in fragmentomics and non-bespoke MRD, Exact Sciences in MCED, GRAIL in health economics and implementation evidence.
About 25 records reference MRD in their titles, distributed across Biodesix, Natera, Mission Bio, AccuraGen, Illumina, Personalis, Myriad Genetics, Predicine, and others. MRD remains one of the most scientifically active and technically diverse areas in the corpus, with multiple architectures represented — including both personalized and standardized approaches. No single company accounts for more than 3 title-level MRD records, reflecting the breadth of activity across the field.
Methylation-based approaches appear in ~48 titles and ~81 full texts. Labcorp's enzymatic methyl-seq classification achieved AUC 0.98 via leave-one-out cross-validation in a 99-sample pan-cancer cohort across >31,000 model configurations. Separately, Foundation Medicine described a methylation-based approach to tumor fraction estimation that does not depend on detectable somatic mutations — a potentially valuable alternative when mutation-based approaches are limited by low tumor fraction. These methylation results sit alongside the fragmentomics and CH work as parallel efforts to extract trustworthy signal from plasma without relying solely on somatic variant detection.
If the signal-quality problem is solved — or even meaningfully improved — the clinical utility of liquid biopsy in early-stage disease, MRD monitoring, and population screening could expand substantially.
The MCED landscape at AACR 2026 doesn't look like a technology race converging on a single design. It looks like several companies making fundamentally different assumptions about what early detection should optimize for. Exact Sciences and Freenome are betting on analytical complexity — multiple analyte types that capture more biology. SeekIn and Quest are betting on simplicity and access — fewer analytes, lower cost, wider reach. Natera is betting on a different endpoint entirely — precancerous lesions rather than cancer. And GRAIL has pivoted to a different kind of evidence — not analytical performance, but health-systems data that might move payers.
The performance data reinforces the divergence. SeekIn's OncoSeek 2.0 reports 83.5% sensitivity at 90.1% specificity at ~$30/test. Exact Sciences' MP V2 reports 97.4% specificity but 41.4% overall sensitivity — a very different tradeoff, optimized for low false-positive rates. These aren't competing versions of the same product. They're competing visions of what population screening should look like.
Bruker is highly visible in spatially oriented records — its two name strings (Bruker Spatial Biology and Bruker Spatial Genomics) combine for 24 records, substantially more than any single competitor. Conference presence alone does not establish platform adoption or commercial leadership, but the concentration is notable.
The largest single-session concentration in the corpus is "Spatial Proteomics and Transcriptomics 3," which drew 21 diagnostic-affiliated records — suggesting spatial biology has moved past niche interest into a mainstream session category at AACR.
The central question for the field is how quickly research-grade spatial readouts can be translated into clinically deployable signatures. Bruker, Vizgen, Standard BioTools, Lunaphore, and others are expanding platform capability — more plex, more resolution, more throughput. But clinical laboratories typically need fewer markers that work reliably at scale. PredxBio's compression work — 51+ markers down to 6–8 — points directly at this translational gap.
Counts reflect diagnostic-affiliated abstracts only. Full AACR AI/ML corpus is substantially larger.
The more interesting question is not how often AI appears, but what kind of AI. The vocabulary in these abstracts has shifted in ways that carry architectural meaning. "Agentic" implies multi-step workflow orchestration. "Foundation model" implies pretraining at scale with fine-tuning for specific tasks. These are terms imported from the broader AI field, and their appearance in diagnostic company abstracts — not just academic ones — suggests the commercial sector is absorbing these paradigms faster than previous conference cycles would have predicted.
Natera's presence in this section is worth noting: beyond its 2 ML-titled abstracts (artifact mitigation, receptor subtype classification), the company has several large-scale RWD and genomics abstracts — including real-world genomic analysis of pancreatic cancer, ancestry-associated survival determinants in gynecologic cancers, and prognostic ctDNA analysis in NSCLC. These span the data-intensive, computationally heavy end of the spectrum that increasingly overlaps with AI methods.
External validation will be the next critical step. The abstracts document a vocabulary shift toward workflow-oriented and foundation-style systems; whether those architectures deliver durable performance improvements over simpler models remains to be tested in clinical settings.
The field appears to be moving from models to systems — from tools that answer a question to tools that manage a process.
The dual-affiliation pattern is worth reading carefully. These are not just academic collaborations where a diagnostic company provided reagents. Several of the clinical trial records feature diagnostic companies as named partners in registrational-grade clinical programs. That positioning — diagnostics embedded in the trial design, not layered on as a correlative afterthought — suggests the pharma-diagnostics relationship may be shifting from vendor-customer to co-development partner.
Pharma names appearing most frequently across dual-affiliation records: Genentech (6), AstraZeneca (4), Pfizer (3). On the diagnostics side: Roche (6), Creatv MicroTech (4), with Tempus, Bruker, Lunit, Biognosys, SeekIn, and Veracyte each at 2. The word "interception" in the MERIDIAN trial title deserves particular attention — it implies MRD is not being used to monitor recurrence passively but to trigger a therapeutic intervention, a fundamentally different clinical role.
Natera's 20 abstracts don't carry formal pharma co-affiliations, but their data is embedded in pharma treatment contexts — referencing pembrolizumab, nivolumab, and ipilimumab outcomes across MRD monitoring, ICI response prediction, and ctDNA clearance dynamics. The partnership signal is in the clinical data, not the author list.
Melanoma (7 records) — the poster child for immunotherapy — has a surprisingly thin diagnostic-affiliated footprint. This could suggest the biomarker landscape in melanoma is considered mature enough that new diagnostic work has lower marginal value. Pancreatic cancer (12) and glioma (8) are also relatively thin — both notoriously difficult diagnostic settings where the field may need more, not less, innovation.
Lung leads the table, consistent with the established companion diagnostic landscape in NSCLC and active MCED programs. Breast and colorectal are well represented — both areas with active MRD clinical development. The ~26 pan-cancer records reflect MCED and multi-biomarker approaches. The indication distribution roughly tracks where diagnostic innovation has the most commercial pull.
More than 700 unique institution name strings appear on diagnostic-affiliated records. MD Anderson leads with 25, followed by Broad Institute (20), Johns Hopkins (15), Mayo Clinic (13), NCI (13), and Dana-Farber (13). The presence of Yonsei University and KAIST among the top 15 may reflect growing Asia-Pacific engagement in spatial biology and computational genomics. For diagnostic companies evaluating collaboration strategies, this institutional distribution provides a useful starting map — though it captures conference activity, not the full scope of each institution's diagnostic research portfolio.
None of these emerging signals represent high-volume themes in the corpus. Platelet RNA has a handful of abstracts; urinary EV-miRNA is represented by a single company (Craif) across two studies. What makes them worth tracking is not frequency but the problems they point at: alternatives to mutation-based detection (fragmentomics), alternatives to blood-based sampling (urine, platelets), and the expansion of single-cell approaches from basic research into potential clinical monitoring. Each represents a bet that the current dominant paradigm — mutation calling from plasma cfDNA — leaves clinical value on the table.
The fragmentomics signal is particularly notable because it appeared across multiple companies (Foundation Medicine, DELFI Diagnostics, biomodal) and multiple applications — suggesting it is emerging as a standalone modality, not just a supplementary feature within existing assays.
1. The bottleneck is trust, not detection. CH deconvolution, artifact mitigation, tumor fraction estimation, and negative prediction all address the same underlying challenge: making liquid biopsy trustworthy when signal is weak.
2. Diagnostics is edging closer to action. Trial-linked records, pharmacodynamic biomarker use, and "interception" language suggest diagnostics may be moving from readout to trigger in some clinical contexts.
3. Interpretation is the innovation layer. Sequencing looks like infrastructure. AI, spatial biology, methylation, and fragmentomics are where the visible differentiation is happening.
4. MCED contains several product philosophies. From multi-analyte screening to protein-led risk models to pre-cancer interception, the programs reflect distinct views on what early detection should look like.
5. Spatial biology's next challenge is clinical compression. The field is mature as a discovery tool. The translational question is whether high-plex readouts can become smaller, deployable clinical signatures.
6. The sample matrix is widening beyond blood. Urinary EV-miRNA, urine proteomics, and platelet RNA represent efforts to expand liquid biopsy where blood draws are barriers to access and repeat testing.
The energy in diagnostics at AACR 2026 is in discovery and surveillance, not in the established CDx framework. The field is no longer asking whether diagnostics belongs at the center of treatment decisions. It's arguing about how fast to let it drive.