On August 19, 2026, Merck and Moderna reported that INTerpath-001 (NCT05933577) met its primary endpoint of recurrence-free survival and its key secondary endpoint of distant metastasis-free survival at a prespecified interim analysis. The trial randomized 1,137 patients with completely resected stage IIB–IV cutaneous melanoma 2:1 to intismeran autogene (1 mg every three weeks, up to nine doses) plus pembrolizumab versus pembrolizumab alone, over roughly 56 weeks. Each dose encodes up to 34 neoantigens selected from that patient's tumor. No new safety signals were reported; overall survival remains immature. This is the first Phase 3 win for an individualized neoantigen therapy and the first for an mRNA cancer therapy — and the first regimen to beat pembrolizumab monotherapy in adjuvant melanoma.
The predecessor Phase 2b, KEYNOTE-942, produced an RFS hazard ratio of 0.51 (95% CI 0.294–0.887) and a DMFS hazard ratio of 0.411 (95% CI 0.200–0.843). Merck has not yet released the Phase 3 effect sizes. Whether the confirmatory hazard ratios land near the Phase 2b estimates or regress substantially toward 1.0 is the single most consequential unreported number in the field, because it sets the commercial and clinical threshold every follow-on platform will be measured against.
Into that moment steps BreakBio, a small Miami company positioning itself as the third personalized-cancer-drug platform after Moderna and BioNTech. Its differentiator is antigen breadth: BreakVax, an IND-authorized personalized multi-epitope synthetic peptide vaccine whose BreakBioAI engine ranks "40+ mutated and non-mutated unique targets" per patient by integrating tumor DNA sequencing, RNA sequencing, and mass-spectrometry immunopeptidomics — deliberately including cancer-testis and embryonic antigens alongside neoantigens. Two things about this thesis deserve correction before it gets repeated. First, the contrast with BioNTech is inverted: BNT111, the FixVac candidate that met its Phase 2 ORR endpoint in anti-PD-(L)1-refractory melanoma, is not a neoantigen product at all — it encodes a fixed, off-the-shelf set of four non-mutated melanoma antigens (NY-ESO-1, MAGE-A3, tyrosinase, TPTE), expressed by over 90% of cutaneous melanomas. Non-mutated targeting is not the road not taken; it is a road with two decades of traffic and a lot of wreckage. Second, BreakBio's own materials name triple-negative breast cancer, not colorectal, as the lead indication for its 2026 Phase 1/2 — and TNBC is not a low-mutation tumor in the way MSS colorectal is. The genuinely novel claim is narrower and more interesting than the framing suggests: not that non-mutated antigens work, but that per-patient mass-spectrometric selection of them works where fixed selection did not.
The Antigen Source Question
The case against non-mutated targets is immunological, not technical. Cancer-testis and embryonic antigens are self-proteins with restricted expression; the high-avidity T cells that would recognize them are largely deleted by central tolerance, leaving a repertoire too weak to clear established tumors. The clinical record follows the biology. MAGRIT, the adjuvant MAGE-A3 Phase 3 in over 2,200 resected NSCLC patients, failed to separate from placebo on disease-free survival. A 2025 systematic review in Critical Reviews in Oncology/Hematology pooling 73 single-arm cancer-testis antigen vaccine trials (n = 1,435) found an objective response rate of 13.0% (95% CI 11.0–16.0) and a disease control rate of 39.0%, with melanoma (16.2%), intravenous delivery (22.4%), and dendritic-cell platforms (19.3%) faring better; across 12 randomized trials, overall survival improved modestly (HR 0.80, 95% CI 0.66–0.97) while disease-free survival did not move at all (HR 0.98, 95% CI 0.89–1.08). That DFS null is the exact endpoint class BreakBio would need to beat in an adjuvant setting.
BreakBio's implicit rebuttal is that those trials picked antigens the wrong way — off a population-level list, without per-patient evidence that the peptide is actually presented on that patient's tumor. Personalized immunopeptidomic selection changes the selection function, and that is a real change. It does not change the tolerance biology. A cancer-testis peptide confirmed by mass spectrometry on a specific patient's HLA is a better-verified target than the same peptide assumed from bulk expression data, but the T cells available to see it were pruned in the thymus either way. The mechanism that has to rescue this is precursor repertoire, not presentation confidence — and presentation confidence is what immunopeptidomics measures.
Meanwhile the mutation-derived side has strengthened outside melanoma. TNBC-MERIT, the individualized mRNA neoantigen trial reported in Nature and covered in early 2026, treated 14 adjuvant TNBC patients with up to 20 patient-specific neoantigens on two lipid-nanoparticle-delivered mRNAs, without concurrent chemotherapy or checkpoint blockade. All 14 mounted vaccine-induced or amplified T cell responses; some CD8+ clones expanded to 6–17% of circulating CD8+ T cells, magnitudes the authors note are rarely seen with conventional cancer vaccines. Eleven of 14 remained relapse-free at up to six years. It is uncontrolled and n = 14, so it proves immunogenicity, not efficacy — but it is uncomfortable evidence for the premise that neoantigen-only strategies run out of signal in breast cancer. Reported TNBC medians of roughly 7–8 mutations per megabase sit below the conventional ≥10/Mb "high TMB" cutoff yet well above MSS colorectal. BreakBio has chosen an indication where the neoantigen-only comparator already has a positive immunogenicity readout.
What Mass Spectrometry Actually Sees
The load-bearing assumption in BreakBio's pitch is that immunopeptidomics adds targets in low-mutation tumors. Recent method work suggests it does — but mostly non-mutated ones, for reasons that are physical rather than biological. NeoDiscMS, published in Nature Communications in 2025, pairs NGS-guided real-time spectral matching with personalized inclusion lists of roughly 1,500 target peptides. Across three melanoma cell lines it recovered four mutated neoantigens, three of them previously demonstrated immunogenic in patients. Across three uveal melanoma tissue samples with tumor mutational burdens of 2.6–3.6 mutations per megabase it recovered zero neoantigens — and nine tumor-associated antigen peptides instead. That is the low-mutation case in miniature: when you point a state-of-the-art targeted immunopeptidomics workflow at a low-TMB tumor, what comes back is shared antigens, because those are the abundant ligands MS can actually see.
Read charitably, this is evidence for BreakBio's operational thesis and against its clinical one. The pipeline will find targets; the targets will disproportionately be the tolerance-constrained class. BreakBio's "40+ targets per patient" figure should be interpreted with that composition question attached — the number that matters is the mutated/non-mutated split per patient, and in a TNBC cohort at 7–8 mut/Mb with MS sensitivity as the binding constraint, the split will likely skew non-mutated. No public disclosure specifies it.
Sample input is the other underappreciated constraint. NeoDiscMS operates on 5–10 million cells per measurement for cell lines, or roughly 10 mg of tissue per injection. Core-needle biopsies from a resected or metastatic TNBC lesion do not reliably yield that on demand, and immunopeptidomics consumes tissue that also has to feed DNA-seq, RNA-seq, and pathology. Sensitivity is improving — NeoDiscMS reports up to 20% more TAA-derived peptides than standard data-dependent acquisition, and a two-fold increase in reproducible target identifications at high dilution — but the field's own reviews still flag limited sensitivity for low-abundance peptides and technical variability as unresolved. A per-patient MS step is a new single point of failure in a manufacturing chain that already has several.
Where the Ranking Breaks Down
The strongest version of BreakBio's argument is not about antigen class at all — it is about ranking. Prediction-first pipelines impose a hard top-N cut, and that cut demonstrably discards real targets. The NeoDiscMS authors highlight a validated immunogenic neoantigen, YPAAVNTIVAI, that ranked 186th by prediction and would have been excluded from any vaccine taking the top 40 candidates; direct MS evidence justified its inclusion. Against a background where only about 1% of somatic mutations induce spontaneous or vaccine-induced T cell responses, and where current reviews note that pipelines still largely omit TCR recognition and similarity to known immunogenic epitopes, the ordering problem is real and unsolved. Multi-modal evidence that reorders candidates is a defensible product claim.
It is worth being precise about what the open-source state of the art already does, because that sets the bar for "proprietary AI." pVACtools v6 adds peptide presentation scoring, immunogenicity prediction, anchor residue analysis, percentile scoring, and pVACsplice for splice-derived neoantigens, alongside class I and class II support and synthetic-long-peptide design utilities. ImmunoNX, a WDL/Cromwell workflow that has supported over 185 patients across 11 clinical trials, wraps pVACtools in consensus variant calling, automated HLA typing, and a two-stage immunogenomics review using pVACview and manual IGV inspection; on the HCC1395 benchmark it narrows 322 initial predictions to 78 high-confidence candidates, and completes vaccine design within three months. A commercial platform claiming superiority has to beat that, and the comparison is currently unauditable in either direction.
One design choice deserves specific scrutiny. pVACtools v6 includes reference proteome similarity checking — a filter whose purpose is to remove candidates that look like self, on both safety and efficacy grounds. BreakBio's architecture deliberately re-admits a large class of exactly those peptides. That may be the right call, and it is coherent with per-patient MS confirmation of presentation. But it inverts a standard filter, and it means the Phase 1 has to carry an autoimmunity readout that a neoantigen-only trial does not: cancer-testis antigens have restricted normal expression, embryonic antigens less reliably so, and multi-epitope peptide vaccines dosed with adjuvant against 40+ self-derived targets are not a setting where prior neoantigen safety data transfers cleanly.
Forty-Five Days Is Aggressive
BreakBio advertises "biopsy to personalized treatment in 45 days." For calibration: TNBC-MERIT averaged 69 days from sample receipt to vaccine release for a 20-neoantigen mRNA product without an MS step; ImmunoNX describes three months to complete vaccine design; and industry CMC discussion frames the realistic window as two to three months covering biopsy, design, manufacture, control, and release. BreakBio is claiming to beat all of these while adding immunopeptidomics — the slowest, most tissue-hungry, least automated assay in the stack — and while synthesizing 40+ peptides per patient. Forty-five days is not impossible, but it is the operational claim most likely to slip first, and slippage matters clinically: in the adjuvant setting the window between resection and recurrence is finite, and in metastatic disease it is shorter.
The regulatory surface is genuinely unsettled, and more so for an AI-first design process than for the manufacturing itself. Standard potency assays are not feasible for these products — neither the vaccine nor the disease is comparable between subjects, no surrogate model exists, and a meaningful potency readout would require T cells from an already-vaccinated patient. Sterility testing is a timeline bottleneck that rapid-release alternatives may not fully solve, and conventional stability protocols are structurally incompatible with a one-lot-per-patient product. Most pointedly for BreakBio: there is currently no internationally approved regulatory framework for assessing the AI algorithms used in personalized vaccine design. A company whose core differentiation is an algorithm is differentiating in the one layer regulators have the least settled machinery to evaluate.
None of this is disqualifying — it is the same terrain Moderna and BioNTech crossed, and an IND authorization means FDA has already accepted a version of the answer. But it does reframe what a small company is taking on. BreakBio is a 1–50 person organization, MTEC-affiliated and Endless Frontier Labs-selected, running per-patient DNA-seq, RNA-seq, mass spectrometry, deep-learning ranking, and GMP peptide synthesis on a 45-day clock, in an indication where the comparator platform just posted 100% immunogenicity. It is pursuing Breakthrough Therapy Designation, which on current evidence would be premature — BTD generally wants preliminary clinical evidence of substantial improvement, and BreakVax has none yet.
What to Watch
The INTerpath-001 hazard ratios at the upcoming medical meeting. Whether Phase 3 RFS lands near the KEYNOTE-942 HR of 0.51 or regresses toward 0.7–0.8 determines the price of admission for every follow-on platform, and the DMFS curve separation timing tells you how long adjuvant vaccine benefit takes to appear. Watch also for the filing pathway Merck and Moderna choose and how CBER handles per-lot potency for an approvable individualized product — that precedent is worth more to BreakBio than its own Phase 1 design.
BreakBio's ClinicalTrials.gov registration. As of this writing no BreakVax record surfaces in public registry searches despite an announced IND and a stated 2026 dosing start. The registration will disclose what the marketing does not: the mutated/non-mutated target split, whether MS confirmation is required or advisory for target inclusion, adjuvant and route, checkpoint-inhibitor combination or monotherapy, sample size, and — critically — whether the primary endpoint is safety and immunogenicity or something efficacy-adjacent. A monotherapy design in adjuvant TNBC is the clean scientific test; a pembrolizumab-combination design in metastatic disease is the commercially safer one and will make attribution much harder.
Three specific readouts from the Phase 1, in order of information value. First, the per-patient target composition and the failure rate of the immunopeptidomics step — how many enrolled patients yield insufficient tissue or zero MS-confirmed targets. Second, the immunogenicity split by antigen class: if CD8 responses to non-mutated targets are materially weaker or lower-avidity than to the neoantigen fraction in the same patients, the tolerance objection is confirmed inside a single trial, and that is the fastest way this thesis resolves. Third, grade ≥3 immune-related adverse events with attention to tissues expressing the selected cancer-testis and embryonic antigens — the safety profile of 40+ self-derived epitopes is genuinely unknown.
Actual versus advertised turnaround. Publish-or-it-didn't-happen applies: median days from biopsy to first dose, reported with the distribution and the manufacturing failure rate, not the target. Also watch whether independent benchmarking of BreakBioAI against pVACtools v6 and ImmunoNX ever appears — a head-to-head on a public dataset such as HCC1395 would convert a marketing claim into a method claim. And watch the wider field for the counterfactual: if BNT122-01's ctDNA-positive adjuvant colorectal readout shows neoantigen vaccines working in an MSS setting, the low-mutation gap BreakBio is built to fill narrows considerably.
Sources
- Source article
- Merck and Moderna Announce Phase 3 INTerpath-001 Trial of Inti… — August 2026
- BreakBio — Personalized Cancer Immunotherapy (company site) — 2026
- BreakBio Corp company profile, MTEC — 2026
- BioNTech Announces Positive Topline Phase 2 Results for BNT111… — 2024
- Personalized mRNA Cancer Vaccine Gives TNBC Patients Strong Mu… — 2026
- Sensitive neoantigen discovery by real-time mutanome-guided im… — 2025
- ImmunoNX: a robust bioinformatics workflow to support personal… — 2025
- pVACtools v6: A comprehensive suite for neoantigen prediction,… — 2026
- Advances and challenges in neoantigen prediction for cancer im… — 2025
- Juggling CMC trade-offs with personalised cancer vaccines, pha…
- Epitope analysis for influenza vaccine design — 2004-08-23
- Quantifying Influenza Vaccine Efficacy and Antigenic Distance — 2005-03-19