In the largest clinical validation of a blood-based cancer test ever run — roughly 50,000 participants — adding a protein layer for advanced adenoma moved sensitivity only from about 15% to 16%. Pierre Arsène, founder and CEO of Mursla Bio, cites that JAMA-published figure as the central caution for multiomics programs in 2026: more layers add dimensions but erode coherence. "No cohort in the world is large enough to train that coherence back computationally," he says.

That argument frames a round of expert interviews from five companies — Bruker, Thermo Fisher Scientific, 10x Genomics, Mission Bio, and Mursla Bio — on where multiomics projects will hit friction in the coming year. Their answers converge on four bottlenecks: sample preparation, data integration, AI implementation, and cost.

Where do multiomics workflows fail first?

At the bench, sample preparation. "I sound like a broken record because I can't stop talking about the importance of sample prep in multiomics work," says Michael Easterling, PhD, vice president of MALDI/Imaging at Bruker Daltonics. He cites section-to-section variability, internal standards, FFPE samples, and long runs that complicate downstream algorithms — and notes these issues are less well defined for multiomics imaging than for mature bulk methods.

Poor freezing during biopsies offers a concrete failure mode: disrupted cell structures and ice crystal artifacts. Easterling's prescription is a solid tissue-harvest workflow with an unbroken chain of custody to the analyzer, plus engagement with the spatial imaging community to learn established practices.

Scott Jelinsky, PhD, senior director of R&D at Thermo Fisher Scientific, frames the decision as front-loaded portfolio planning: "Different molecular classes have different requirements for sample collection, storage, and preparation. Therefore, careful experimental design and standardization are essential." Decisions on aliquoting and storage made at study start determine what can later be measured and how confidently datasets can be interpreted together.

Why is integration not plug-and-play?

Each omics layer carries its own data structure, distributions, and variability. "Combining them isn't plug-and-play," says Vanee Pho-Conners, PhD, vice president of global marketing at Mission Bio. Many labs also lack the cross-disciplinary mix of biostatistics, machine learning, and biology the analysis demands.

Najiba Mammadova, PhD, associate director of product management at 10x Genomics, highlights a more fundamental question: whether researchers measuring different omics from separate preparations are even looking at the same cells and biological features. Her answer is measuring multiple modalities from the same cell or tissue — 10x's MultiPro Human Discovery Panel, for example, integrates into existing Flex scRNA-seq workflows and reads 326 protein targets, which the company describes as the largest antibody-based single-cell protein panel. This reduces the need to reconstruct biology after the fact, she says.

Timing adds another layer of difficulty. Jelinsky notes that proteins, RNA, and metabolites respond on different timescales, so simultaneous measurements may not align. Published criticisms of multiomics back these concerns, flagging data-integration weaknesses and a tendency toward false positives.

What can AI fix — and what can't it?

Mammadova sees the rate-limiting question shifting from "Can we measure this?" to "Can we interpret it?" — where AI can identify patterns, accelerate interpretation, and build models predicting disease progression or treatment response. Concrete examples exist: a Medical University of South Carolina group developed an AI algorithm that helps lasers target individual cells rather than clusters in single-cell MALDI workflows, and Mission Bio applies machine learning to jointly cluster DNA, RNA, and protein — triomics — from the same cell to establish disease resistance.

Easterling draws the hard line. "If a model's trained on an inconsistently prepared section, it'll give you a confident biological story that's completely wrong. And the reader of this data may have no way to tell." Jelinsky agrees that AI is a hypothesis generator, not a replacement for experimental design or biological validation.

What about extracellular vesicles?

EV analysis carries a distinct problem: each omics layer is measured from an unknown blend of cells, with no way to confirm a protein and an RNA signal share a tissue of origin. Mursla addresses this with tissue-specific EVs captured from two milliliters of plasma via antibody-coated magnetic beads targeting hepatocyte-derived vesicles. Its EvoLiver liver-cancer surveillance test for cirrhotic patients — three proteins and five microRNAs, FDA Breakthrough Device Designation in hand — relies on prospective collection under a common protocol because blood-handling standards for EVs lag standard liquid biopsy.

What happens next?

Cost becomes the binding constraint as studies scale, Mammadova argues; making the technologies scalable and cost-effective will determine how far they spread beyond well-funded programs. Prajan Divakar, PhD, of Bruker Spatial Biology predicts spatial multiomic layers will grow increasingly critical for clinical translation, citing the CosMx imager's measurement of the whole transcriptome plus over 64 proteins in FFPE tissue. Jelinsky's closing judgment sets the planning standard for R&D managers: the future lies not in maximizing measurements but in selecting the combination that answers a specific biological question.