34 Neoantigens, One Patient, One Batch What Moderna and Merck's Phase 3 Win Demands of mRNA Manufacturing

A seven-year question was answered on 19 August 2026 — and it was not the last question

On 19 August 2026, Merck and Moderna announced that INTerpath-001 had met its endpoints.

The Phase 3 trial enrolled 1,137 patients with completely resected stage IIB–IV cutaneous melanoma, randomised 2:1 to intismeran autogene — an individualized neoantigen therapy, previously known as mRNA-4157 or V940 — plus pembrolizumab, or pembrolizumab alone. At a pre-specified interim analysis, the combination significantly improved recurrence-free survival, the primary endpoint, and distant metastasis-free survival, a key secondary endpoint. No new safety signals were reported. Overall survival data remain immature and the study continues.

Two firsts are contained in that paragraph. It is the first positive Phase 3 result for an individualized neoantigen therapy of any kind. And it is the first positive Phase 3 result for any mRNA-based cancer treatment.

The supporting Phase 2b evidence had been building for years. KEYNOTE-942 randomised 157 patients from July 2019, and by the five-year analysis presented at ASCO in June 2026 the combination had shown a 49% reduction in the risk of recurrence or death (HR 0.51; 95% CI 0.294–0.887) and a 59% reduction in the risk of distant metastasis or death (HR 0.411; 95% CI 0.200–0.843). The Phase 3 hazard ratios have not yet been disclosed; full data will be presented at a future medical meeting.

Figure: Every published estimate from the Phase 2b programme sits well below 1.0, and the confidence intervals tightened as follow-up lengthened. What the Phase 3 read-out adds is not a new direction but the statistical weight of 1,137 patients.

Most of the commentary this week has been about oncology, and rightly so. This article is about something else — the manufacturing problem that a positive Phase 3 creates, and why that problem is, at its root, an enzyme problem.

Individualized therapy inverts every economic assumption built into mRNA manufacturing since 2020

Consider what has to happen for one patient to receive this therapy.

The tumour is resected. Tissue undergoes whole-exome and RNA sequencing. Machine learning algorithms compare tumour to germline, identify the somatic mutations that generate novel peptide sequences, and rank them by predicted immunogenicity and HLA binding. Up to 34 of those neoantigens are selected and concatenated into a single synthetic mRNA construct — a sequence that has never existed before and will never be made again. That construct is transcribed, capped, tailed, purified, formulated into lipid nanoparticles, filled, tested, released, and shipped.

For one person.

Then the process begins again, with a completely different sequence, for the next patient.

This is not how mRNA has been manufactured until now. A pandemic vaccine campaign is one sequence, one process, one set of specifications, executed at enormous scale. The economics reward volume; quality burden concentrates into a small number of very large batches; and every fixed cost — analytical development, process validation, facility depreciation — is amortised across billions of doses.

Individualized neoantigen therapy inverts every term in that equation. The total mass of mRNA involved across an entire commercial programme might be measured in kilograms. The number of independent GMP manufacturing events is measured in tens of thousands.

This is scale-out, not scale-up. And the industry already has a hard-won precedent for what scale-out costs.

Figure: The same total mass of product, redistributed across three orders of magnitude more manufacturing events. Batch size falls; batch count rises. Every fixed per-batch cost is multiplied rather than amortised.

Autologous CAR-T already demonstrated that per-batch cost does not fall with commercial scale

Autologous CAR-T is the closest analogue we have, and it is a sobering one.

Published estimates put total production cost for an autologous CAR-T product somewhere between $373,000 and $475,000, with narrower manufacturing cost-of-goods analyses landing near $95,780 per dose. Materials and labour together account for roughly 50–70% of autologous cell therapy cost. Vein-to-vein time runs two to six weeks. Manufacturing failure rates in the published literature range from roughly 1% to 13%, improving with commercial maturity but never disappearing.

The structural lesson is the one that matters here. As one industry analysis put it: to treat a thousand patients, you must successfully, repeatably and compliantly run your single-patient process a thousand separate times. Per-batch cost-of-goods is high and flat. There is almost no per-dose cost reduction at commercial scale, because there is no scale in the conventional sense — only repetition.

Every unit operation you add to a personalized process is therefore not a one-time cost. It is a cost multiplied by your patient count, forever.

Figure: In a campaign model, cost per dose falls as fixed costs spread across volume. In a per-patient model the curve is flat: each batch carries its own QC, release testing and changeover burden regardless of how many patients are treated.

Schematic representation of the cost structures described in the autologous cell therapy literature.

That single observation determines everything that follows.

Chromatography that pays for itself at campaign scale becomes unaffordable when every patient is a batch

Which brings us to the impurity that has shaped mRNA process design for a decade: double-stranded RNA.

dsRNA is an unavoidable byproduct of in-vitro transcription. It is also a potent pathogen-associated molecular pattern, recognised by RIG-I and MDA5 in the cytosol, TLR3 in the endosome, and PKR and the OAS/RNase L system. The downstream consequences are well characterised — type I interferon induction, PKR-mediated eIF2α phosphorylation and translational shutdown, RNase L activation and RNA degradation. The net effect is less protein expressed per unit of mRNA delivered, plus reactogenicity.

The magnitude is not marginal. Karikó and colleagues demonstrated in 2011 that removing dsRNA contaminants from nucleoside-modified mRNA by HPLC eliminated interferon and cytokine induction and increased protein translation by 10- to 1,000-fold in primary cells. That finding is the reason mRNA purification trains look the way they do.

The problem is that those purification trains were designed for large campaigns.

Reversed-phase ion-pair HPLC delivers excellent dsRNA clearance but is the least scalable option in the toolbox, and its acetonitrile eluent brings solvent handling and environmental burdens that Karikó's own group described as unaffordable for most laboratories. Oligo-dT affinity chromatography, optimised and run continuously, can achieve above 90% yield with high purity — but resin capacity sits around 5 mg mRNA per mL, column dead volume and hold-up losses become punishing at milligram scale, and cleaning validation between patient batches is a per-batch burden rather than an amortised one. Cellulose-based dsRNA removal clears at least 90% of dsRNA contaminant but recovers only around 65–80% of the mRNA. Tangential flow filtration recovers well, above 95%, but does not solve dsRNA.

Figure: Labour and facility are not where mRNA cost lives. Raw materials dominate the drug substance, and the drug substance dominates the product — which is why a yield loss in purification is not a process inefficiency but a direct multiplier on the cost of treating a patient.

Chain these together and the arithmetic is unforgiving. A realistic purification train — affinity capture at roughly 90%, a dsRNA-removal step at roughly 70%, polishing and TFF at roughly 95% — compounds to about 60% overall recovery. Published estimates for full mRNA purification trains cluster at 50–70% cumulative recovery, meaning 30–50% of transcribed product is lost.

Figure: Step recoveries compound multiplicatively. Individually defensible unit operations combine into a train that discards roughly two fifths of what was transcribed — a loss absorbed once per campaign, or ten thousand times per year.

In a campaign model, that loss is a line item you engineer against over time. In a per-patient model, it is a line item you pay ten thousand times a year, on a batch that may be only milligrams to begin with, where the analytics for release testing consume a meaningful fraction of what survives, and where there is no second batch to fall back on because the sequence is unique to a patient who is waiting.

There is also a risk dimension that pure cost accounting misses. Every chromatographic step is a changeover, and every changeover between two patient batches is a cross-contamination control point that must be validated and documented. Multiply that by your patient count.

The conclusion follows directly. At per-patient scale, you cannot purify your way to quality. The economics will not carry it and the operational risk compounds. Quality has to be built into the reaction instead.

Which makes it an enzyme problem.

dsRNA is generated by the polymerase itself, which makes specific activity an inadequate specification

To control dsRNA upstream, you have to understand its origins — and they sit almost entirely with T7 RNA polymerase and the template it is given.

Three mechanisms dominate.

Self-templated 3′ extension. The nascent transcript's 3′ end folds back on itself, and the polymerase extends it. Gholamalipour and colleagues showed in 2018 that most 3′ additions arise through this cis fold-back route, with T7 rebinding the resulting hairpin and extending it — a phenomenon they demonstrated even on chemically synthesised, run-off RNA. This is the polymerase's terminal transferase behaviour acting where it is not wanted.

Promoter-independent antisense transcription. T7 initiates from the free end of the linear DNA template, transcribing the complementary strand and generating RNA that anneals to the intended product. Triana-Alonso and colleagues documented this in 1995, finding that across 64 linear templates only around 15% yielded solely the expected run-off product; the remainder generated longer aberrant species.

DNA-terminus-initiated transcription. More recent work published in Nucleic Acids Research in 2024 identified transcription initiating directly at the DNA terminus as the primary source of full-length dsRNA, showed that guanosine and cytosine residues at the template end enhance it, and — crucially for enzyme engineering — demonstrated that aromatic residues at position 47 in the polymerase's C-helix suppress it.

Figure: All three routes are properties of the polymerase and the template geometry, not of the downstream process. Two of the three initiate at the free end of the linear template, which is why linearization chemistry is a dsRNA control parameter.

Schematic. Compiled from Triana-Alonso et al. 1995, Gholamalipour et al. 2018, and Nucleic Acids Research 2024;52(14):8443.

Two things follow from this mechanistic picture, and both are commercially significant.

First, dsRNA formation is a property of the polymerase, not merely of the process. T7's terminal transferase and RNA-dependent RNA polymerase activities are mechanistically linked to dsRNA generation. Different T7 preparations, at identical nominal specific activity, will produce different dsRNA burdens on the same template. Specific activity is a necessary specification. It is a grossly insufficient one.

Second, template geometry matters as much as the enzyme. Because antisense transcription initiates at the template's free end, linearization chemistry is a dsRNA control parameter: 3′ overhangs promote byproduct formation, and blunt or 5′-overhang linearization suppresses it. Incomplete linearization is worse still, producing run-through and plasmid-length transcripts. Work published in Frontiers in Molecular Biosciences in 2023 confirmed that purifying the linearized template before transcription measurably reduces dsRNA formation.

The upstream levers are therefore real, and they are specific.

Two levers control dsRNA before chromatography: engineer the enzyme, or digest the impurity

An engineered polymerase raised 3′ end homogeneity from 6–12% to above 90%

This is no longer theoretical. Dousis and colleagues at Moderna reported in Nature Biotechnology in 2023 a rationally and computationally engineered double-mutant T7 RNA polymerase producing substantially less immunostimulatory RNA than wild type. The headline number is the one worth internalising: 3′ end homogeneity rose from 6–12% for wild-type T7 to above 90% for the engineered variant, with potency maintained and — the point of the exercise — a simplified process and faster manufacturing.

Read that in the context of everything above. A polymerase that does not generate the impurity is worth more than any column that removes it, because the column cost recurs with every patient while the enzyme selection is made once.

Figure: Rational engineering of the polymerase moved 3′ end homogeneity by roughly an order of magnitude. The impurity that purification trains were built to remove is substantially not formed in the first place.

Other engineering routes have followed. Thermostable variants permit elevated-temperature transcription that disfavours duplex formation, though at some risk of mRNA degradation. Attenuating the RNA-dependent RNA polymerase activity by weakening RNA rebinding suppresses the fold-back mechanism directly. Combinatorial screening approaches have produced multi-site mutants with reduced terminal transferase and RDRP activities. Commercial low-dsRNA T7 variants are now available from several suppliers, and their performance differences are real and measurable.

Reaction design compounds the effect: controlled nucleotide and magnesium concentrations, temperature and reaction-time optimisation, template purity, and linearization geometry. N1-methylpseudouridine, meanwhile, reduces immune recognition of dsRNA by lowering binding affinity to sensors including PKR, though its effect on the quantity of dsRNA formed is modest — a distinction that matters when writing specifications.

RNase III trades a column for a vessel, but the trade is substrate-dependent

RNase III cleaves double-stranded RNA, recognising duplexes of around 22 nucleotides. It offers something chromatography cannot at per-patient scale: an aqueous, single-vessel, potentially closed and automatable operation with no organic solvent, no column, no changeover cleaning, and no resin lifetime to validate. For a process that must run ten thousand times in parallel small batches, that architectural difference is worth more than a marginal purity advantage.

It is not a free lunch, and the honest version of this argument requires saying so. RNase III is substrate-dependent. Application data from NEB show that with optimised magnesium conditions, luciferase, eGFP and EPO constructs retain roughly 95% or better single-stranded RNA integrity after dsRNA reduction — but a highly structured construct lost integrity under the same conditions. A 2026 report went further, showing that one commercial ShortCut RNase III preparation could digest single-stranded Cas9 mRNA into fragments as short as nine nucleotides. Structured mRNA is at risk of over-digestion, and neoantigen concatemers are not simple sequences.

The practical implication is that RNase III conditions require per-construct optimisation — which, in a personalized therapy where every construct is new, is a genuine process-development challenge rather than a settled protocol. This is precisely the kind of problem that belongs in a platform qualification exercise rather than in a per-patient workflow.

Other non-chromatographic options are maturing alongside it. Cellulose-based removal, low-pH denaturation coupled to affinity capture, and ligand-based dsRNA capture all trade differently across clearance, recovery and complexity. None of them is a finished answer. All of them are better matched to small parallel batches than an HPLC skid.

Figure: Clearance, recovery and scale-down suitability pull against each other. RP-IP-HPLC clears dsRNA best and scales down worst; TFF recovers best and clears dsRNA least. The methods best suited to per-patient batches sit in the upper right, and none of them is yet a finished answer.

Residual DNA control is driven by regulation rather than economics, and it does not scale down either

dsRNA dominates the technical conversation, but there is a second impurity where the driver is regulatory rather than economic, and where the control point is again an enzyme.

Residual DNA template is a defined limit. WHO recommends 10 ng per dose for products derived from continuous cell lines, with fragment sizes below that of a functional gene — conventionally around 200 base pairs. Authorised mRNA vaccines were held to that 10 ng/dose limit, and independent regulatory testing has confirmed compliance across batches.

The control point is DNase I digestion following transcription, and it involves a genuine process tension. Digestion runs above 35°C under agitation — conditions that also risk mRNA loss. The process window must clear DNA below limits while preserving yield, and the DNase protein itself then becomes a residual-protein control requiring downstream removal. Enzyme quality, animal-origin-free sourcing and full documentation are not peripheral concerns here; they are filing content.

For a personalized therapy, this compliance burden does not scale down. Every patient batch requires its own residual DNA result, its own batch record, and traceability to the raw material lots used. Documentation scales with batch count, not with volume — which is a sentence worth reading twice if you supply raw materials to this industry.

When batch count replaces batch size, five raw material requirements become disproportionate

Step back and the pattern across all of the above is consistent.

When manufacturing shifts from volume to frequency, the definition of a good raw material changes with it:

• Lot-to-lot consistency becomes the dominant risk. In a campaign model, an enzyme lot is qualified once and consumed over months. In a high-frequency model, one lot is drawn across hundreds of independent patient batches — so any drift in specific activity, dsRNA propensity, residual nuclease or residual host cell DNA propagates across all of them simultaneously, and is discovered late.

• Small-format GMP supply becomes operationally necessary. Personalized manufacturing consumes enzymes in small repeated aliquots across many dispensing events, not in kilograms. Suppliers geared exclusively to campaign-scale formats generate waste, open-vial risk and reconciliation burden.

• Documentation must scale with batch count. Ten thousand batch records require ten thousand traceable raw material entries. Drug Master File support, animal-origin-free declarations, complete traceability and disciplined change notification are baseline, not differentiators.

• Lead time becomes a clinical parameter. The clock starts at resection. A raw material stock-out does not land on a procurement dashboard; it lands on a patient timeline. Against typical GMP enzyme lead times of eight to sixteen weeks, and minimum order quantities designed for pandemic-scale campaigns, the mismatch with per-patient cadence is structural.

• Qualification decisions made now are effectively permanent. Once a process locks through Phase 3, changing a critical raw material carries change-control costs that dwarf any unit price difference. The selections being made in Phase 1 programmes today are decisions about the next decade.

Requirement	Campaign model (2020–2025)	Per-patient model
Lot-to-lot consistency	One lot qualified once, consumed over months. Drift is caught within a single campaign.	One lot drawn across hundreds of patient batches. Drift propagates simultaneously and is discovered late.
Supply format	Bulk formats. Fewer, larger dispensing events.	Small repeated aliquots across many dispensing events. Bulk-only supply creates waste and open-vial risk.
Documentation burden	Scales with volume. Amortised across billions of doses.	Scales with batch count. Ten thousand batch records require ten thousand traceable raw material entries.
Lead time	A procurement variable, managed with forecasting and safety stock.	A clinical variable. The clock starts at resection; a stock-out lands on a patient timeline.
Change control	Costly but contemplated. Second sources can be qualified in parallel.	Effectively immovable after Phase 3 lock. Phase 1 selections are decade-long commitments.

Table 1. The same five requirements, read against each manufacturing model. Nothing in the right-hand column is new as a category — what changes is which requirements dominate.

Case study: one supplier's answer to the per-patient specification

The following section reflects data and product information supplied by Hzymes Biotech, a manufacturing partner, and has not been independently verified by this publication.

An ultra-high-throughput screening platform for enzyme engineering

Hzymes Biotech, spun out of Professor Guangyu Yang’s laboratory at Shanghai Jiao Tong University, has developed an aptamer-based fluorescence-activated droplet sorting platform (Ab-FADS) for directed evolution of T7 RNA polymerase. The company reports screening throughput of up to 10⁸ variants per day at 99.4% sorting accuracy, including under non-standard conditions such as elevated temperature and modified nucleotides — the same class of multi-parameter screening that produced Moderna’s published double-mutant discussed above [25].

The M30 Mutant

Four rounds of directed evolution using this platform are reported to have produced a lead mutant, M30, with the following performance data:

Metric	Result
Catalytic efficiency	~10× increase vs. wild-type
dsRNA byproduct at 37°C	Reduced to 10% of wild-type level
dsRNA byproduct at 50°C	Reduced to 0.1% of wild-type level
3′-end consistency	Increased from 5.6% to 60%
Transcription fidelity	Comparable to wild-type enzyme

In LNP-delivered mouse studies, M30-derived mRNA is reported to produce IL-6 and IFN-α levels close to baseline — consistent with the reduced-dsRNA mechanism discussed earlier in this article. The company attributes this to a “strong DNA binding, weak RNA rebinding” selectivity profile in the mutant, which would suppress the self-templated extension and antisense fold-back mechanisms described above [25].

Thermostable variants for circRNA synthesis

The company also reports thermostable T7 RNA polymerase variants with an approximately 270-fold longer half-life and 58-fold higher catalytic activity at 52°C, stable across a 37–55°C range — positioned for one-step circRNA synthesis, where transcription and circularization occur in a single elevated-temperature reaction rather than requiring a separate high-temperature circularization step [25].

GMP-scale, small-batch manufacturing

Against the small-format GMP supply and documentation requirements in Table 1, Hzymes reports GMP-scale manufacturing of its low-dsRNA mutants under a documented quality control system, supplying raw material to active mRNA, saRNA and circRNA clinical development pipelines [25]. Full specifications are available on the manufacturer’s product page.

Whether engineered mutants like this become the industry default, or one of several viable routes toward the same goal, the direction is the same: dsRNA control is migrating from the purification suite to the reaction vessel.

And these pressures are not evenly distributed geographically. GMP-grade enzymes, nucleotides and capping reagents remain concentrated among a small number of suppliers. Developers building personalized mRNA capability outside the United States and European Union — and there are now many, from the WHO technology transfer hub network to national programmes in India, Brazil, Korea and Russia, where a personalized mRNA melanoma therapy received clinical-use authorisation in November 2025 and dosed its first patient in April 2026 — face those lead times and order quantities with fewer qualified alternatives and thinner regulatory support.

Figure: Melanoma is the proof, not the market. The manufacturing volume, and with it the raw material demand, arrives if the lung, kidney and bladder programmes read out the same way.

For those programmes, raw material access is not a procurement question. It is a question of whether personalized therapy is achievable at all.

Frequently Asked Questions

Why can’t purification alone solve the dsRNA problem in personalized mRNA manufacturing?

Chromatography-based purification trains recover roughly 50–70% of transcribed product, meaning 30–50% is lost at every run. In a campaign model that loss is engineered against once, across billions of doses. In a per-patient model the same loss, plus the associated changeover and cross-contamination validation burden, recurs on every single batch — which is why the economics that work at pandemic scale break down at per-patient scale.

What raw-material properties matter most when a process runs thousands of one-patient batches instead of a few large campaigns?

Five properties become disproportionately important: lot-to-lot consistency (because one lot now spans hundreds of independent batches), small-format GMP supply, documentation that scales with batch count rather than volume, lead times short enough to fit a clinical timeline rather than a procurement forecast, and qualification decisions made early, since changing a raw material after Phase 3 lock is effectively immovable.

How can dsRNA be controlled before chromatography?

Two upstream levers exist. One is engineering the T7 RNA polymerase itself — Moderna’s published double-mutant raised 3′ end homogeneity from 6–12% to above 90%, and the case study above describes one supplier extending that approach with its M30 mutant. The other is enzymatic digestion of dsRNA post-transcription with RNase III, which trades a column for a single vessel but requires per-construct optimisation since it is substrate-dependent.

References

Clinical evidence

1. Weber JS, Carlino MS, Khattak A, et al. Individualised neoantigen therapy mRNA-4157 (V940) plus pembrolizumab versus pembrolizumab monotherapy in resected melanoma (KEYNOTE-942): a randomised, phase 2b study. Lancet. 2024;403(10427):632–644. DOI: 10.1016/S0140-6736(23)02268-7

2. Khattak A, Meniawy T, Carlino MS, et al. Five-year results from the randomised phase 2b KEYNOTE-942 / mRNA-4157-P201 study. J Clin Oncol. 2026;44(suppl 16):9500. Presented at ASCO Annual Meeting, June 2026.

3. Merck & Co., Moderna Inc. INTerpath-001 Phase 3 topline results announcement. Press release, 19 August 2026.

4. Rojas LA, Sethna Z, Soares KC, et al. Personalized RNA neoantigen vaccines stimulate T cells in pancreatic cancer. Nature. 2023;618:144–150. DOI: 10.1038/s41586-023-06063-y

dsRNA formation and T7 RNA polymerase mechanism

5. Triana-Alonso FJ, Dabrowski M, Wadzack J, Nierhaus KH. Self-coded 3′-extension of run-off transcripts produces aberrant products during in vitro transcription with T7 RNA polymerase. J Biol Chem. 1995;270(11):6298–6307. DOI: 10.1074/jbc.270.11.6298

6. Gholamalipour Y, Karunanayake Mudiyanselage A, Martin CT. 3′ end additions by T7 RNA polymerase are RNA self-templated, distributive and diverse in character — RNA-Seq analyses. Nucleic Acids Res. 2018;46(18):9253–9263. DOI: 10.1093/nar/gky796

7. Nacheva GA, Berzal-Herranz A. Preventing nondesired RNA-primed RNA extension catalyzed by T7 RNA polymerase. Eur J Biochem. 2003;270(7):1458–1465. DOI: 10.1046/j.1432-1033.2003.03510.x

8. DNA-terminus-dependent transcription by T7 RNA polymerase and its C-helix mutants. Nucleic Acids Res. 2024;52(14):8443–8458.

9. Engineered T7 RNA polymerase reduces dsRNA formation by lowering terminal transferase and RNA-dependent RNA polymerase activities. FEBS J. 2025.

Impurity control and purification

10. Karikó K, Muramatsu H, Ludwig J, Weissman D. Generating the optimal mRNA for therapy: HPLC purification eliminates immune activation and improves translation of nucleoside-modified, protein-encoding mRNA. Nucleic Acids Res. 2011;39(21):e142. DOI: 10.1093/nar/gkr695

11. Dousis A, Ravichandran K, Hobert EM, Moore MJ, Rabideau AE. An engineered T7 RNA polymerase that produces mRNA free of immunostimulatory byproducts. Nat Biotechnol. 2023;41:560–568. DOI: 10.1038/s41587-022-01525-6

12. Baiersdörfer M, Boros G, Muramatsu H, et al. A facile method for the removal of dsRNA contaminant from in vitro-transcribed mRNA. Mol Ther Nucleic Acids. 2019;15:26–35. DOI: 10.1016/j.omtn.2019.02.018

13. Purification of linearized template plasmid DNA decreases double-stranded RNA formation during IVT reaction. Front Mol Biosci. 2023;10:1248511. DOI: 10.3389/fmolb.2023.1248511

14. ShortCut RNase III is capable of digesting single mRNA transcripts into <10 nucleotide products. 2026.

15. New England Biolabs. RNase III protocol modification can enable safe dsRNA removal from IVT samples. Application note.

16. Rosa SS, Prazeres DMF, Azevedo AM, Marques MPC. mRNA vaccines manufacturing: challenges and bottlenecks. Vaccine. 2021;39(16):2190–2200. DOI: 10.1016/j.vaccine.2021.03.038

Regulatory and quality standards

17. World Health Organization. Evaluation of the quality, safety and efficacy of messenger RNA vaccines for the prevention of infectious diseases: regulatory considerations. WHO Technical Report Series No. 1039, Annex 3. Geneva: WHO; 2021.

18. United States Pharmacopeia. Analytical procedures for quality of mRNA vaccines and therapeutics (draft guidelines, 3rd edition). Rockville, MD: USP; August 2024.

19. Residual DNA quantification in mRNA COVID-19 vaccine batches using orthogonal analytical methods. npj Vaccines. 2025.

Manufacturing economics

20. Kis Z, Kontoravdi C, Shattock R, Shah N. Resources, production scales and time required for producing RNA vaccines for the global pandemic demand. Vaccines. 2021;9(1):3. DOI: 10.3390/vaccines9010003

21. Comparative cost analysis of autologous CAR-T cell manufacturing. Front Oncol. 2024;14:1433432. DOI: 10.3389/fonc.2024.1433432

22. Manufacturing GMP-grade mRNA vaccine-associated enzymes: a techno-economic assessment. Stellenbosch University; 2024.

23. Advancing RNA. COGS reduction strategies for mRNA manufacturing (industry cost model, panel session). 2025.

24. Merck & Co., Moderna Inc. Moderna and Merck present 5-year data for intismeran autogene in combination with KEYTRUDA in high-risk stage III/IV melanoma following complete resection. Press release, January 2026.

Manufacturing partner data

25. Yang G, et al. Directed evolution of T7 RNA polymerase minimizes dsRNA byproduct and enables high-fidelity mRNA synthesis for demanding therapeutic applications. Research. Shanghai Jiao Tong University. Manufacturer-supplied data, cited with disclosure; not independently verified by this publication.