Calibrate multiomics with reference materials before feeding AI: a common ruler makes batches comparable

2026-09-03

An 18-institution team led by Fudan University proposes in Nature Biotechnology that standard reference materials be co-profiled with study samples, with results reported as sample-to-reference ratios (SRR), so multiomics data become reproducible across batches, labs and platforms and fit for AI training.

What problem this solves

AI is eating multiomics data. A single transcriptomic, proteomic or metabolomic measurement yields tens of thousands of features per sample, exactly the high-dimensional input models want. But the data has a metrology blind spot: the same biological sample produces different raw signals (intensity, I) on two instruments, in two labs or across two batches, in unpredictable ways. Training on such data teaches the model to learn instrument fingerprints as if they were biology.

This is not news. The MAQC (MicroArray Quality Control) consortium, founded at the US FDA in 2006, has repeatedly shown batch effects to be the leading root cause of irreproducibility in omics. The dominant fix has been computational correction after the fact (ComBat, limma, RUV), each of which assumes batch structure can be estimated from within the data itself. A a systematic comparison (Molania et al., Nat. Biotechnol. 41, 82–95, 2023) showed that without an external anchor these methods often remove real biological signal along with the batch. The "for artificial intelligence" in the title makes the stakes explicit: the stronger the downstream model, the more measurement error upstream gets amplified into systematic bias.

Method

The core mechanism in one sentence: give every batch a common ruler, report results as how many times the ruler the sample is, and the instrument sensitivity factor cancels in the division.

• Reference material (RM): a well-characterized, stable biological standard (for example a mixed plasma or cell-line reference) that every lab uses.
• Co-profiling: each 96-well plate reserves about four wells for the RM, which goes through the exact same extraction, instrument run and detection as study samples.
• SRR (sample-to-reference ratio): sample signal divided by the same-plate RM signal. The derivation in Fig. 1 is plain: instrument response I = f × C, where f is that instrument's sensitivity and C is the true concentration, so Isample/IRM = (f×Csample)/(f×CRM) = Csample/CRM. Two instruments with different sensitivities, each carrying the same RM, converge on the same concentration ratio: f cancels.
• Reporting change: databases store the SRR rather than raw intensity.

In analytical chemistry this is old discipline: numbers from an uncalibrated instrument do not count. In clinical testing, the ICH M10 guideline (the international harmonized standard for bioanalytical method validation) already mandates internal-standard ratio reporting. The paper's move is to bring that discipline into multiomics and argue why downstream AI needs it.

Results

This is a four-page comment (Brief Communication): a proposal plus a conceptual Fig. 1, no new experiments, no baseline table. What it offers instead is a chain of argument assembled from two decades of evidence:

The paper gives no head-to-head comparison of SRR against ComBat-style correction, and no quantified cross-lab SRR comparability numbers. Those are said to live in ref. 10 (Liu et al., bioRxiv 2026.08.09.740864); the preprint was blocked by Cloudflare and the comment sits behind a paywall, so the details could not be verified.

Why it matters

For anyone doing AI4Science, this is a data-supply-side comment. Three points land directly:

• It moves the omics-AI reproducibility problem upstream into measurement. Prior discussion concentrated on models or correction algorithms. This comment says the defect is present when the data leaves the instrument, and no downstream pipeline fully recovers it. If training data are incomparable across batches, the "features" a model learns include instrument fingerprints, and external validation collapses the moment the platform changes.
• The scheme is a transfer of existing industrial discipline, not a new invention. Internal-standard ratio reporting is standard analytical chemistry, and ICH M10 is the legal framework for it. Requiring omics data to pass the same gate before entering an AI pipeline is cheap to adopt, and closer to institutionalized reproducibility than re-running a correction pipeline before every training run. The transfer cost is about four wells per plate, roughly 4% of a 96-well plate's throughput.
• Multimodal fusion presupposes comparable modalities. Proteomics and metabolomics each produce raw signals in their own units; concatenating them into one model concatenates numbers measured by two different rulers. SRR anchors every modality to the same physical benchmark, giving multimodal fusion a unified metrological base.

Verdict: a concept-proposal comment, not a methods paper. The concept, calibrate at measurement rather than correct downstream, holds up, backed by twenty years of MAQC/Quartet data. For AI practitioners it reads as a reminder to check whether training data were calibrated, not as a drop-in tool.

Limitations

• Paywalled and proposal-heavy. Four pages; the method details, SRR computation workflow and cross-lab validation numbers live in the references. Ref. 10, the companion preprint, was blocked by Cloudflare and could not be read. Judging actual effect must wait for its empirical results.
• SRR rests on a linear approximation. I = f×C assumes linear instrument response and constant f across the dynamic range. Real mass-spec and sequencing responses are non-linear at both ends; whether ratio reporting stays unbiased in high-dynamic-range settings (common in metabolomics) is not discussed.
• Cost and throughput. Reserving wells for RM is a throughput loss, with no cost estimate given. Whether it scales to UK-Biobank-sized projects is left unexamined.
• Commentary, not regulation. Several authors are FDA officials, but this is personal academic opinion, not an FDA position or compliance requirement; the paper itself notes the views are the authors'.
• The AI relevance is argued, not measured. No experiment pits a real AI model trained on SRR data against one trained on raw intensities. "Fit for AI" is an inference; the "robust AI/ML model" in Fig. 1 is a schematic, not a result.

Terms

• Multiomics: measuring multiple molecular layers (transcriptome, proteome, metabolome) on the same sample for a fuller biological picture.
• Reference material (RM): a characterized, stable standard sample used as a common ruler in measurement.
• SRR (sample-to-reference ratio): sample signal divided by same-plate reference signal; cancels instrument sensitivity via a shared ruler.
• Batch effect: systematic measurement differences introduced by handling, instrument or timing — unrelated to biology.
• MAQC/Quartet: the omics quality-control consortium founded at the FDA in 2006 and its successors, and the source of two decades of batch-effect data.
• ICH M10: the international harmonized guideline for bioanalytical method validation, the existing legal framework for internal-standard ratio reporting.

Source

• Paper

What people are saying

• Nature Biotech paper: reference materials as common calibrator to make multiomics data AI-ready kshameer

All paper explainers