China just put AI cancer vaccines on a production line
China has industrialised what the West is still piloting — the policy and competitive consequences land before the clinical ones do.
TL;DR
- Likang Life Sciences broke ground in Beijing on a production line for AI-designed personalised mRNA cancer vaccines (LK101). Facility completion: October 2026. Investment: ¥110m / US$16.1m.
- Target turnaround: 24 hours from tumour DNA sequencing to patient-specific neoantigen vaccine. Conventional bespoke vaccine workflows take 4–8 weeks.
- Regulatory status: NMPA trial approval (2023), FDA IND (2025). Already in Hainan medical pilot zone at ~US$22,000 per injection, ~US$140,000 per seven-injection course.
- Reported early data: 100% five-year survival in liver cancer (n=24), doubled progression-free survival in melanoma. Read these numbers with the caution n=24 demands.
- This is not a one-company story. Everest Medicines holds a parallel FDA IND with its EVER-NEO-1 neoantigen algorithm. KAIST in Korea unveiled a B-cell-targeting AI vaccine model in January. The Asia-Pacific neoantigen-AI cluster is now denser than the Western one.
What's actually happening
Personalised cancer vaccines work by sequencing a patient's tumour, identifying neoantigens — mutant protein fragments unique to that tumour — and manufacturing an mRNA vaccine that trains the patient's immune system to attack cells carrying them. The hard parts are: (a) computationally predicting which mutations will produce neoantigens the immune system can actually see, and (b) doing the manufacturing fast enough that the patient is still treatable when the vaccine arrives.
The Western leaders here — Moderna with Merck on mRNA-4157 / V940, BioNTech with Genentech on autogene cevumeran — have run convincing Phase 2 data in melanoma and pancreatic cancer. Their manufacturing cycles sit in the 4–8 week range. That timeline is the bottleneck. It is also where Likang is making its bet.
The 24-hour claim is not about replacing clinical trials. It is about replacing a process. AI-driven neoantigen prediction (Likang's stack and Everest's EVER-NEO-1 do this), automated mRNA sequence design, and an LNP delivery formulation that doesn't require re-optimisation per patient — pulled into one continuous production line — collapse the weeks-of-handoffs into a near-real-time pipeline. The hardware is being built now. Whether the 24-hour target actually holds at scale is a 2027 question, not a 2026 one.
The number you should pause on
100% five-year survival in liver cancer is the kind of figure that travels well on social media and badly through peer review. The sample is n=24. In hepatocellular carcinoma — where five-year survival under standard care varies wildly by stage, region, and access to transplant — a 24-patient cohort tells you something has happened but not yet what it means. The melanoma signal (doubled progression-free survival) is more clinically familiar territory because Moderna and BioNTech have already plotted similar shapes, but the same n-size caveat applies.
The honest read: the early signals are consistent with the broader mRNA-cancer-vaccine field looking real. They are not, on their own, evidence that Likang has solved cancer. Anyone telling you otherwise — and the next ten days of LinkedIn will not be short of them — is selling something.
Why this is genuinely new
Strip the hype and what's left is still significant. Three things are happening simultaneously, and the combination is what changes the picture:
- Industrialisation of personalisation. A purpose-built factory for n-of-1 therapeutics is a category shift. Western players have done this in research-grade GMP suites; a dedicated production line at clinical commercial scale is a different commitment, and a different cost base.
- AI as the load-bearing step, not the marketing step. The neoantigen prediction stack is the rate-limiting science. China now has two companies (Likang, Everest) with FDA INDs cleared on AI-derived candidates, and one Korean academic group (KAIST) with a B-cell-targeting prediction model heading toward 2027 trials. That is a credible regional pipeline.
- A pricing and access asymmetry. Hainan availability at ~US$140,000 per course is brutally expensive for a Chinese patient and remarkably cheap relative to a comparable Western personalised therapy estimate. If this manufacturing cycle holds and Chinese reimbursement absorbs even part of it, the access-pattern argument starts running in an unfamiliar direction.
Who actually benefits, who's exposed
| Group | What changes for them |
|---|---|
| Late-stage cancer patients with options exhausted | A new — and largely Asia-located — clinical-trial pathway worth tracking, particularly for melanoma, HCC, and select solid tumours. |
| Western mRNA-vaccine programmes (Moderna/Merck, BioNTech/Genentech) | Competitive pressure on cycle time, not on science. Their Phase 3 readouts in 2026–2027 are still the gating events for the category. |
| US FDA and EMA reviewers | A second IND from a Chinese sponsor (after Everest's EVM14) signals that AI-derived neoantigen designs from China are now a routine part of the review queue, not an outlier. |
| Chinese biotech investors | The category is no longer speculative. NMPA-approved trials, FDA INDs, and physical manufacturing put it past the vapour stage. |
| Western policy-makers on bio-AI export controls | The frame shifts. Export controls on training compute don't constrain a neoantigen prediction stack that already exists and is in clinical use. |
| Everyone else | This is genuine progress on a real disease, on a non-Western node, with a long way to run. That's the whole story. |
The Western mirror — and where the comparison breaks
The cleanest analogue to LK101 is Moderna/Merck's mRNA-4157 (V940), in Phase 3 for adjuvant melanoma. The trial design, the platform logic, and the neoantigen-prediction approach are recognisably the same family. Where the comparison breaks is timeline-to-clinic-at-scale. Moderna's manufacturing has been the bottleneck the company has spoken about openly; BioNTech has signalled similar constraints. Likang's bet is that the bottleneck is solvable with a purpose-built production line, sooner, in Beijing.
The bet may or may not pay off in 2027. The geopolitical signal is independent of that. China now has a visible industrial commitment to AI-driven personalised oncology that the West, despite stronger upstream science in some respects, has not yet built. The infrastructure asymmetry compounds.
What this means for you
If you're a patient or a patient's family member, this does not change anything you should do this week. Personalised mRNA cancer vaccines remain experimental, and the meaningful access pathways are clinical trials. The Likang and Everest INDs in the US are worth flagging to your oncologist if you're already in a tumour type these candidates target (melanoma, HCC, select solid tumours). Anyone offering you a US$22,000 injection in Hainan as a confirmed cure is not your friend.
If you're a clinician or researcher in immuno-oncology, the watch-list shortens. Track: (i) Likang's IND-enabling Phase 1 design when posted on ClinicalTrials.gov, (ii) Everest's EVM16 IIT read-outs, (iii) KAIST/Neogen Logic's 2027 IND target. The neoantigen prediction sub-field has moved from "interesting open problem" to "two production-grade vendor stacks", and the methods papers underneath them will matter more than the press releases.
If you're a policy reader or investor, the meta-story is the one that travels. The category went from research-grade to industrial-grade outside the US/EU before doing so inside. Whatever your prior was on the East–West bio-AI gap, this is a data point that updates it in one direction. The size of the update depends on whether Likang ships the October build and whether the 24-hour cycle survives contact with patients.
If you write about AI for a living, notice this: the most consequential AI application of the week is not a model release, a benchmark, or an agent demo. It is a factory. That is probably true more often than the discourse admits.
Uncertainty ledger
- 24-hour claim unverified at scale. Production-line targets and real-world cycle times routinely diverge. First independent clinical-cycle data likely 2027.
- Sample sizes are small. n=24 liver cancer, similarly small melanoma cohorts. Phase 3 evidence does not yet exist for any AI-designed personalised cancer vaccine, from any sponsor, globally.
- Manufacturing-line claim is under construction, not operational. Completion target October 2026. Slip risk material.
- Pricing trajectory unknown. Current Hainan pricing (~US$140k/course) is not a guide to scaled or reimbursed pricing.
- Regulatory geopolitics. A future US bio-tech export-control posture toward Chinese AI-derived therapeutics would change the access map fast.
Bottom Line
China didn't beat the West to the cancer vaccine — the science is converging across multiple regions, and Phase 3 data from Moderna and BioNTech remain the events that move the category. What China did this week was put a personalised AI therapeutic on an industrial production line before anyone else. That is a smaller claim than the headlines make it, and a bigger one than the n=24 critics will say it is. The factory is the news. Track the October build, the Phase 1 dosing schedule, and the second-IND tempo. Everything else is downstream.
Sources
- South China Morning Post, "China's first AI-powered cancer vaccine production line set to launch in Beijing", 29 June 2026 — Tier 1
- Ynetnews, "China unveils AI system to develop personalized cancer vaccines in 24 hours", 29 June 2026 — Tier 2
- Caixin Global, "Weekend Long Read: A Chinese Startup's Moonshot Cancer Cure", 25 April 2026 (background context on Likang trial history, regulatory status, pricing) — Tier 1
- Benzinga, "AI-powered Cancer Vaccine In The Works" (Everest Medicines / EVER-NEO-1 / EVM14 / EVM16 context), 25 March 2025 — Tier 3 (used for ecosystem context only, not load-bearing)
- Chosun, "KAIST Unveils B-Cell Targeting AI Cancer Vaccine", 2 January 2026 — Tier 2
- Nature, "Evaluating neoantigen-vaccine responses through mechanistic and model-based frameworks", 24 June 2026 — Tier 1 (methodology context)
