RaPiD-mAb-MS delivers a genuine 20–25x throughput advance for monoclonal antibody peptide mapping. The methodology is serious work. The framing around it is where INGA314 finds the calibration gaps
A case study in calibrated criticism. The paper is good. The headline is not the paper.
Click to access 2026.05.14.725248v1.full.pdf
A small detail before we begin
I went to read the paper a colleague had flagged on bioRxiv — DOI 10.64898/2026.05.14.725248. The archive index returned by the search engines said the paper at that DOI was titled “Trypsin exhibits exopeptidase-like activity toward N-terminal arginine that biases proteomic analyses.” That title set my expectations entirely: a foundational measurement-validity claim against the most-used enzyme in proteomics. Provocative. Worth a hard look.
The actual paper at that DOI is a methods paper on direct-infusion mass spectrometry for monoclonal antibodies, by the Coon lab at Wisconsin together with industry collaborators from AbbVie, Lilly, Johnson & Johnson, and Genentech. Different topic, different authors, different field within biochemistry. The archive index was simply mislabeled.
This is the kind of thing INGA314 exists to catch. The bioRxiv index is a piece of upstream metadata most analytical pipelines — human or AI — will trust by default. If you build a literature-monitoring system on top of it, you will sometimes get the wrong paper attached to the right DOI, and your downstream analysis will be confidently wrong about what the paper says. The lesson is older than mass spectrometry: verify the artifact before you analyze it.
With that out of the way, here is the real paper, and what INGA314 finds when run across it.
What the paper accomplishes
Salome, Morgenstern, Hebert et al. present RaPiD-mAb-MS, and it is a genuine methodological advance.
Peptide mapping has been the analytical backbone of monoclonal antibody characterization for two decades. Every antibody drug program runs it, repeatedly, at every stage from discovery through manufacturing. The conventional workflow is LC-MS based, with chromatographic gradients of 30–90 minutes per sample, and the throughput ceiling that creates has been an accepted constraint of biologic drug development — slowing candidate triage, limiting the number of formulation conditions that can be screened, and forcing pharma analytical groups to deep-characterize only the leads they have already pre-selected on cruder criteria.
The Coon lab and their industrial collaborators eliminate the chromatography. They directly infuse tryptic digests into a high-resolution Orbitrap mass spectrometer through an automated nano-electrospray chip, achieving ~60 seconds of MS time per sample and demonstrating that the resulting data quantify the modifications that matter — oxidation, deamidation, glycosylation, isomerization, sequence variants, antibody-drug conjugate attachment sites — with linear-regression correlations against LC-MS routinely above R² = 0.99. They validate across 28 unique antibodies and over 2,000 samples. They demonstrate a single 1,152-sample forced-degradation campaign completed in 35.5 hours of MS time. They show, in a separate workflow, that a chemical labeling step (PIMT enzymatic methylation) can extend the method to isoaspartate detection, which is otherwise invisible to mass spectrometry alone.
This is real work. The throughput claim is genuine. The downstream implications — full-factorial design of experiments on formulation conditions, broader candidate screening before lead selection, the eventual feasibility of machine-learning models trained on standardized peptide-mapping data at scale — are reasonable directions, and the paper takes the first concrete step toward them.
What INGA314 examines is something narrower. We are not asking is this paper right? We are asking where does the language outrun the evidence, and by how much? For a method this consequential, that calibration matters more, not less. The five gaps below are the kind of soft edges that compound when a method moves from preprint to adoption decision, and they are worth surfacing precisely because the underlying work deserves to be evaluated on its real merits rather than its rounded-up framing.
The INGA314 lens
The framework runs four detector categories across any document:
- Scope violations — claims that reach beyond what was actually tested
- Proxy elevation — treating a measurement as the thing itself
- Causal inflation — drawing causal conclusions from correlational structure
- Confidence inflation — overstating certainty, especially in Discussion sections
For methodology papers in regulated industries — pharma, diagnostics, biologics QC — these categories matter because the gap between “the method works” and “the method should replace existing practice” is where investment, regulatory, and procurement decisions get made.
Let’s walk it.
Pass 1: The speed claim
The abstract says “accelerating the method by up to 100-fold.”
Inside the paper, the numbers tell a different story:
- The MS acquisition is ~60s per sample
- The sample-to-sample cycle time is ~120s (acquisition plus changeover, stated in the paper)
- The 1,152-sample run took 35.5 hours of MS time → ~111s per sample real-world
- The 4-hour sample preparation is still required, on top of the MS time
- The LC-MS comparator gradient is 40 minutes (~2,400s), but modern fast LC-MS uses 15–20 minute gradients
The honest end-to-end comparison against a modern LC-MS workflow is closer to 20–25x faster, not 100x. The 100x number is true only in the most favorable framing — MS-time only, against the slowest plausible LC-MS comparator.
This is not a fabrication. It is scope-of-comparison inflation, factor ~4–5x. Common, almost universal in methodology abstracts. Worth noting precisely because it is so common that we no longer notice it.
Pass 2: The “all phases” claim that the paper itself contradicts
The abstract: “RaPiD-mAb-MS is positioned to accelerate all phases of antibody-based drug discovery & development.”
Page 3 of the same paper: “While conventional LC-MS remains key for final quality control due to the entrenchment of FDA-approved methods, RaPiD-mAb-MS breaks the analytical throughput bottleneck essential for accelerating discovery and pre-clinical development.”
These two sentences do not agree. The second one is the honest one — the method is positioned for discovery and pre-clinical, not for regulated QC and batch release. The first one is the abstract’s marketing version.
INGA314 calls this internal-contradiction confidence inflation. The author already knows the constraint; they wrote it down themselves. The abstract simply omits it. Anyone reading only the abstract — which is most readers, and most automated literature-monitoring systems — gets the inflated scope and not the qualification.
Pass 3: When “comparable to LC-MS” needs unpacking
The abstract claim of comparable results across modifications is mostly defensible, but it contains three quiet caveats that matter for anyone considering this method for critical-quality-attribute monitoring:
Hydrophobic CDR3 peptides. The NIST mAb’s heavy-chain CDR3 peptide (DMIFNFYFDVWGQGTIVTVSSASTK) was not detected by direct infusion without solvent modification. CDR3 is the binding-determining region of an antibody — the region whose modifications matter most for potency. The paper mentions this in one sentence, attributing it to solubility limits in the aqueous spray solvent. It does not quantify how often this failure occurs across the 28-antibody dataset. A method with a known failure mode on the region you most need to monitor is a real constraint, not a footnote.
Peptide-level, not residue-level. Figure 5’s caption notes that the reported occupancies are “peptide-level (summed) occupancies rather than residue-localized measurements.” For peptides containing multiple modifiable sites — multiple methionines, multiple asparagines — the method reports aggregate modification, not per-residue. For regulatory QC and developability decisions that require site localization, this is a meaningful gap. The abstract does not flag it.
IsoAsp requires a different workflow. The paper handles isoaspartate detection by adding an overnight enzymatic labeling step using PIMT (protein L-isoaspartyl O-methyltransferase). This generates new covalent species that can then be detected by direct infusion. It is a different measurement than direct detection of isoAsp by LC-MS, and it requires its own normalization to a “100% native” reference sample. Calling this “comparable to LC-MS” elides a meaningful methodological substitution.
None of these caveats invalidates the work. They constrain the scope. A reader who absorbs only the abstract will not know they exist.
Pass 4: The machine-learning recommendation
The paper closes with a linear machine-learning model fit to the 1,152-sample dataset. It returns an “ideal” formulation: histidine buffer at pH 6.5 with proline as excipient. The text frames this as evidence the method enables novel formulation discovery at scale.
Two issues:
Best-vs-worst framing. The 1.6 percentage-point reduction in overall modification is calculated between the best and worst buffer-pH combinations. Best-vs-typical would be the more policy-relevant number, and would be smaller.
Single-cohort generalization. No held-out validation. Four antibodies, NIST-mAb-weighted, under forced-degradation conditions. The recommendation may or may not generalize to a real-world formulation development decision on a different antibody.
Circular structural logic. The recommendation is offered as evidence that the method is needed — but the method is what produced the recommendation. “This suggests the critical need for testing conditions thoroughly with large-scale study designs, such as those enabled by RaPiD-mAb-MS.” The output of the method justifies the method. This is a soft form of causal inflation: presenting a methodological capability as a substantive discovery.
The result itself — histidine + proline — is broadly consistent with existing formulation literature, which means the ML model has partially recovered known good practice. That is reassuring for the method’s internal validity. It is not, by itself, evidence of novel discovery.
Pass 5: The conflict-of-interest architecture
This paper has one of the densest commercial-interest disclosures I have seen in a methodology preprint, and it deserves direct attention.
- The corresponding author and a co-author founded CeleramAb Inc. to commercialize this methodology.
- The corresponding author consults for Thermo Fisher, which manufactures the Orbitrap instruments and ionization platform the method requires.
- Six authors hold intellectual property on the work.
- Four authors are consultants of CeleramAb.
- The industrial collaborators at AbbVie, Lilly, J&J, and Genentech provided both the samples and the LC-MS comparator data against which the new method was benchmarked.
None of this is undisclosed. All of it is in the paper. But disclosure does not eliminate the structural concern: the entire comparison framework — what was tested, against what, by whom, with which sample-prep protocols — was designed and operated by parties with direct commercial interest in the new method outperforming. The deamidation discrepancy of 1–3 percentage points was attributed to AbbVie’s sample-prep inducing deamidation, rather than to RaPiD-mAb-MS under-counting it. Both are defensible interpretations of the same data. The paper picks the one favorable to the new method.
For a methodology paper destined for academic discussion, this is acceptable. For a methodology that wants to be adopted in pre-clinical development or eventually in regulated QC, independent replication by an uninvolved lab does not yet exist, and that is the load-bearing absence.
Composite Inflation Table
| Dimension | Inflation magnitude | Severity |
|---|---|---|
| Speed claim (100x in headline) | ~4–5x relative to honest end-to-end comparison | Moderate |
| “All phases” of development scope | Contradicted within the same paper | Moderate |
| Modification breadth (“comparable to LC-MS”) | Partial — true for some, indirect for isoAsp, peptide-level only for multi-site | Low-moderate |
| Modality generalization (bispecifics, fusions, vaccines, viruses in Discussion) | Full speculative extension, zero supporting data | Discussion-section only |
| ML formulation recommendation | Single-cohort, no held-out validation, best-vs-worst framing, circular justification | Moderate |
| Independence of validation | Dense commercial COI throughout; no truly independent replication | Structural |
Composite validity score: ~0.72 on the INGA314 scale. This is a strong methodology paper with normal-magnitude inflations concentrated in the headline framing and the Discussion section. Nothing crosses into fabrication or misrepresentation of data. The structural concern is not the science — it is the absence of independent replication for a method whose authors have founded a company to sell it.
Bottom line
RaPiD-mAb-MS is a real advance and deserves to be recognized as one. Direct infusion of tryptic mAb digests with high-resolution Orbitrap detection delivers a credible 20–25x throughput improvement over conventional LC-MS for pre-clinical peptide mapping, with strong correlation to the established methods on the modifications that matter most, demonstrated at industrial scale across 28 antibodies and over 2,000 samples. This is the kind of methodological step-change that meaningfully expands what is feasible in early biologic development. Pharma analytical groups working in discovery and developability should take it seriously.
The five gaps INGA314 surfaces do not change that verdict. They sharpen it.
The headline says 100x. The honest number is 20–25x — still transformative, but it should be claimed at its real magnitude.
The abstract says “all phases of development.” The paper itself correctly excludes regulated QC, and the method is appropriately positioned for the pre-clinical and developability stages where its speed compounds with iteration.
The Discussion projects to bispecifics, fusion proteins, vaccines, and viruses. None of those were tested. They are reasonable next directions, not present-tense capabilities.
The ML extension demonstrates capability, not discovery. The formulation recommendation it returns is broadly consistent with existing literature.
The commercial-interest density is high, the disclosure is complete, and independent replication by an uninvolved lab does not yet exist. For a method of this potential significance, that replication is the load-bearing next step.
For an investor evaluating CeleramAb: defensible Series A story for pre-clinical and developability throughput. The key diligence question is not whether the method works — it does — but how often the failure modes (hydrophobic CDR3 peptides, multi-site peptide aggregation, isoAsp requiring a parallel workflow) intersect with the customer’s most-monitored attributes.
For a pharma analytical group evaluating adoption: highly useful at the discovery and developability stages, where the throughput multiplier compounds with iteration. Pair with conventional LC-MS for any attribute that requires residue-level localization or that is destined for a regulatory filing. The method is ready for that role today.
For a reader of the abstract: read the paper. The abstract and the paper are not the same document, and in this case the paper is the more impressive of the two.
Why this case study matters for INGA314
This is not a paper with logical failures. It is a paper with the normal, low-magnitude inflations that accumulate in every methodology abstract in every field, often unnoticed by the authors themselves. Catching them requires reading slowly enough to notice when two sentences in the same paper disagree, and calibrated enough not to manufacture flaws that aren’t there.
The bioRxiv index that mislabeled the DOI is a separate matter — but it is the kind of upstream metadata error that any automated literature pipeline will silently propagate. Source verification, then content verification. Both passes matter. Both are routine in INGA314 workflows.
The headline of this post could have been “Coon Lab Paper Inflates 100x Speed Claim by 5x.” That would have been technically accurate and substantively misleading. The accurate version is: the paper is good, the abstract is rounded up, and the method is worth taking seriously for the use cases it actually supports.
That distinction — between catching real overreach and inventing fake controversy — is what calibration looks like in practice.
INGA314 detects and quantifies logical failures in high-stakes documents. They find papers. We find flaws.
Contact: daridor@inga314.ai · inga314.ai
System D 