1. Abstract
Personalized mRNA neoantigen vaccines fail or succeed on a short list. Rojas and colleagues showed that an individualized RNA vaccine can prime T cells after resection of pancreatic adenocarcinoma, with about half of adjuvant patients mounting de novo CD8 responses when up to 20 predicted neoantigens were encoded [1,2]. The remaining design problem is not whether to vaccinate. It is which peptides to put on the cassette, and how to show that those peptides came from the patient's own tumor file rather than from a curated cheat sheet.
On 1 September 2026, GenoBank.io closed that loop on Cancer Digital Twin case TN25-336147. A Caris somatic VCF owned by the patient wallet was read on the node, missense substitutions were mapped onto UniProt protein sequence, and native 8 to 11-mer flanks were emitted as WT and MUT pairs. The same compact table (gene, change, peptide, WT, HLA, IC50, VAF, TPM, oncogenicity) went to three independent agents named Fable, Agy, and Grok. A merger kept mutations with 2 or more votes, drivers first. Six epitopes entered the stitched ORF. Two real VCF substitutions were held. The vote is now a board on the gated vaccine visualizer, and the cassette FASTA is a patient-owned BioCID.
2. The Short List Is the Product
Neoantigen vaccination is a ranking problem dressed as immunology. Sahin and Ott established that personal mutanome RNA or peptide vaccines can expand T cells in melanoma [3,4]. Wells and the TESLA consortium then showed why most predicted binders never become epitopes: expression, clonality, and MHC presentation have to agree, and single-tool ranking is noisy [5]. Łuksza's fitness model made the same point from the tumor's side of the table. A clone that presents a true neoantigen is the clone the immune system can see [6]. MHCflurry 2.0 improved pan-allele MHC-I presentation scoring, which is necessary and not sufficient [7].
Autogene cevumeran (BNT122) is the clinical object in this patient's trial arm. The scientific object on the Cancer Digital Twin is narrower. Encode 4 to 8 MHC-I epitopes from a real somatic VCF, stitch them with a non-neoantigen CD4 helper on one ORF, and leave an audit trail that a third party can replay from BioCIDs. A long ranked list of 20 guesses is a TESLA-era artifact. A cassette of 6 peptides that three agents can defend is the design we ran.
Three agents see one table. Two votes put a peptide on the RNA. One vote leaves it off. A single ranker can hide its priors in prose. Named agents that must emit GENE:CHANGE identifiers cannot.
3. What Closed on 1 September 2026
Two failures had been sitting in the fabricate path. The VCF parser expected VEP ANN or CSQ fields. The Caris file for TN25-336147 speaks GI (gene), PC (protein change), FC (consequence), CI (clinical impact), and VF (variant fraction). KRAS G12D arrives as PC=G12D and VF=0.337, not as an HGVS CSQ block. Without those keys, every novel substitution was recorded as a protein change with no peptide, and the cassette fell back to a TN25-only curated table. That is a prior, not a discovery.
The second failure was social, not syntactic. A single ranker, even a good MHC predictor, is a single story. Fable, Agy (Gemini), and Grok were asked to vote independently on the same table and to merge only where they agree. Keep mutations with at least 2 votes. Sort drivers first. Do not invent genes. Do not put a 1-vote peptide on the RNA because it is pretty.
Both pieces now run inside biofs-node cassette fabrication on the production node, against bytes that never left Google Cloud Storage. The laptop did not download the VCF. The output is four BioCIDs owned by 0x5f5a60EaEf242c0D51A21c703f520347b96Ed19a: neoantigen candidates JSON, mRNA construct FASTA, windows JSON, and windows BED.
GI PC VF"] --> Flank["UniProt CDS flanks
8 to 11-mers"] TPM["RNA geneTPM"] --> Table["Compact table
8 mutations"] Flank --> Table Table --> Fable["Fable"] Table --> Agy["Agy"] Table --> Grok["Grok"] Fable --> Merge["Merge
2 or more votes
drivers first"] Agy --> Merge Grok --> Merge Merge --> ORF["Stitched ORF
6 epitopes"] ORF --> BioCID["Patient-owned BioCIDs"] BioCID --> Board["Consortium board"]
In words. The Caris VCF and RNA TPM join into a compact peptide table. Fable, Agy, and Grok vote independently. The merger keeps mutations with 2 or more votes, drivers first. The six-epitope ORF is registered as patient-owned BioCIDs and drawn on the board.
4. Methods
The somatic source is biocid://caris/0x5f5a60EaEf242c0D51A21c703f520347b96Ed19a/vcf/DNA_TN25-336147.vcf. Expression is the Caris RNA geneTPM file for the same case. HLA is the arcasHLA genotype already on the map: A*02:01, A*29:02, B*38:01, B*51:08, C*12:03, C*16:02. Fabrication refuses to attach the TN25 curated peptide table to any other wallet.
For each passing missense, the executor looks up the reviewed UniProt sequence whose primary gene name matches, confirms the WT residue at the HGVS.p position, substitutes the mutant amino acid, and emits every 8, 9, 10, and 11-mer that contains that residue. KRAS G12D on P01116 independently produces VGADGVGKSAL / VGAGGVGKSAL. ATR F2321L is taken from Q13535 (2,644 amino acids), not from a noisy gene_exact:ATR search that also returns SERPINA2 and ANTXR1. Frameshifts and in-frame indels are out of scope for this run.
K-mers collapse to one peptide per gene plus change before the vote. A peptide with an IC50 is preferred to a flank-only k-mer. Length 9 is the tiebreak. MHCflurry 2.0.6 is installed on the node and the presentation models were fetched. Predict failed on Keras 3. NetMHCpan is not on PATH. IC50 values on the six cassette inserts are therefore the map binder priors, not a fresh MHCflurry pass. That limit is stated on the board.
Each agent receives the identical compact table and returns a cassette of 4 to 8 mutation identifiers. Fable ranks by oncogenicity then clonality and keeps pathogenic and driver mutations. Agy ranks by expression times presentation and spreads HLA coverage, which is why it dropped KRAS G12D once ATR F2321L had already claimed C*16:02. Grok ranks drivers first with cassette-grade IC50 (at most 150 nM) and VAF at least 0.18, and holds proline-rich or weak-prior peptides. A vendor LLM overlay is armed for Claude (Fable) and Gemini (Agy). It did not fire on this run. The recorded votes are the independent agent writeups under that protocol. The merger counts GENE:CHANGE identifiers.
The construct is a fabricated DNA-alphabet surrogate for alignment: 5' UTR, MHC-I signal, PADRE helper AKFVAAWTLKAAA, GPGPG linkers, the six epitopes, stop, 3' UTR, polyA. It is not a GMP lot and it is not the BioNTech design FASTA. A synthetic HiFi FASTQ aligned to cassette_full with minimap2 2.24 is the coverage track on the visualizer (380 reads, mean depth about 372, all 6 epitopes present).
5. Results
The Caris VCF contains 3,253 records. Missense with FILTER equal to "." or PASS numbers 13. After CDS flanks and TPM join, 303 k-mers were kept and collapsed to 8 mutations. Six of those 8 were already in the TN25 map prior. Two were not. PTCH1 R1391Q (PPPVPGPGQ on UniProt Q13635, VAF 0.466, TPM 1,186, VUS) and KDM5A Q1515H (KRKLEKVEH on P29375, VAF 0.721, TPM 319) are VCF-derived peptides with native flanks. That is the discovery claim. The curated table is no longer the only source of amino acid sequence.
The three agents disagreed, which is the point of a consortium. Fable returned 4 epitopes, all drivers. Agy returned 7, dropped KRAS, and added both VCF-only flanks. Grok returned the 6 cassette-grade inserts and held PTCH1 and KDM5A. The merger kept every mutation with 2 or more votes, drivers first.
drivers first"] Agy --> M Grok --> M M --> K["Kept: ASXL1 ATR TP53 KRAS CSF1R ERC1"] M --> H["Held: PTCH1 KDM5A"]
In words. Fable is conservative (4 drivers). Agy spreads HLA and adds VCF-only flanks (7). Grok keeps cassette-grade IC50 (6). The merge keeps six and holds two.
| Mutation | Peptide | HLA | IC50 nM | VAF | TPM | Fable | Agy | Grok | Votes | Merge |
|---|---|---|---|---|---|---|---|---|---|---|
| ASXL1 V515M driver | IPSLPQETM | B*51:08 | 75 | 0.478 | 853 | yes | yes | yes | 3 | kept |
| ATR F2321L driver | GSDGKLYIM | C*16:02 | 77 | 0.363 | 313 | yes | yes | yes | 3 | kept |
| TP53 G266R Pathogenic | LLRRNSFEV | A*02:01 | 109 | 0.368 | 2527 | yes | yes | yes | 3 | kept |
| KRAS G12D Pathogenic | VGADGVGKSAL | C*16:02 | 95 | 0.337 | 257 | yes | yes | 2 | kept | |
| CSF1R A245T VUS | QHNNTKLTI | B*38:01 | 40 | 0.494 | 1260 | yes | yes | 2 | kept | |
| ERC1 S995Y | YEDDHFKSY | A*29:02 | 72 | 0.664 | 173 | yes | yes | 2 | kept | |
| PTCH1 R1391Q VCF flank | PPPVPGPGQ | 0.466 | 1186 | yes | 1 | held | ||||
| KDM5A Q1515H VCF flank | KRKLEKVEH | 0.721 | 319 | yes | 1 | held |
KRAS G12D is the test of the merge. Agy dropped it for HLA diversity. Fable and Grok kept it as a pathogenic driver. Two votes put VGADGVGKSAL on the RNA. A single ranker that optimizes HLA spread would have left the canonical PDAC neoantigen off a pancreatic cassette. The rule does not let one aesthetic win.
VGADGVGKSAL"] --> F["Fable: keep
Pathogenic driver"] K --> A["Agy: drop
C*16:02 already used"] K --> G["Grok: keep
driver plus IC50 95 nM"] F --> V["Votes = 2"] A --> V G --> V V --> R["Merge: kept"]
In words. Agy's HLA-spread rule would omit KRAS G12D. Two other agents keep it. The merge ships the peptide.
ASXL1 V515M IPSLPQETM B*51:08; ATR F2321L GSDGKLYIM C*16:02; TP53 G266R LLRRNSFEV A*02:01; KRAS G12D VGADGVGKSAL C*16:02; CSF1R A245T QHNNTKLTI B*38:01; ERC1 S995Y YEDDHFKSY A*29:02. Construct 692 nt. All 6 epitopes present on the aligned synthetic FASTQ. Owner wallet is the patient, not the operator.
6. The Board
A vote that lives in a log file is a vote that did not happen. The gated Cancer Digital Twin vaccine visualizer now carries a Neoantigen consortium board under the cassette schematic. Each epitope card shows Fable, Agy, and Grok as on or off. The matrix lists all 8 mutations, including the 2 that were held. The lede states the source: the patient's own Caris VCF, CDS flanks, then at least 2 votes, drivers first. Embed mode hides the page chrome and keeps the board, so the map iframe is the same document.
The live page is vaccine.html on the patient map, behind the existing email gate. Unauthenticated requests still see the lock page. Nothing about this board is a public leak of the VCF. It is a rendering of peptides the map already treated as the working neoantigen layer, now with a vote that can be audited.
7. What This Is Not
This is one case, one panel VCF, one HLA genotype. It is not a TESLA-scale benchmark and it is not a claim that the six peptides will prime T cells. Rojas's 0.50 blood-response rate was measured on an IV RNA-lipoplex product with up to 20 neoantigens [1]. This apparatus is a fabricated ORF for a patient-owned twin. Delivery, uridine chemistry, and GMP sequence are out of scope.
MHCflurry did not score the 303 k-mers on this run. NetMHCpan was absent. Fresh pan-allele IC50s would likely reshuffle PTCH1 and KDM5A, and might reshuffle 9-mer choice inside a mutation that currently rides a map prior. The consortium is built so that a later binder pass changes the table, not the vote rule.
The Natera somatic VCF was not in this vault's registry at fabricate time, so Caris is the somatic source. Germline subtraction is still a documented intent, not a second VCF join on this job. The vendor LLM overlay for Claude and Gemini is armed and did not fire. Independent agent writeups still disagreed, which is the behavior the merge is designed to harvest.
None of this teaches an unpublished apparatus. Issued patents US 11,915,808 and US 11,984,203 cover BioNFT ownership of biosamples and biodata. The screwdriver in this paper is the vote. The lock remains the BioNFT.
8. Why a Board, Not a Bigger Model
Frontier models are useful on neoantigen design when they are forced to disagree in public. A single large ranker can hide its priors in prose. Three named agents that must emit GENE:CHANGE identifiers cannot. Fable's conservatism, Agy's HLA spread, and Grok's driver-plus-IC50 rule are different enough that KRAS G12D and CSF1R A245T split 2 to 1 in opposite directions. The merge keeps both. That is the clinical intuition: do not drop a pathogenic driver because another epitope already uses the allele, and do not drop a strong, clonal, expressed binder because it is not in COSMIC.
The TESLA lesson still holds. Parameters of immunogenicity are empirical [5]. A consortium board does not replace MHC prediction, RNA abundance, or clonality. It stops a pipeline from pretending that one ranked list is a cassette. On a patient-owned BioFS node, that stop is also an ownership stop. The peptides are derived from a VCF the patient holds. The vote is derived from a table anyone with the BioCID can hash. The RNA, when it exists as a lot, should point at both.
The next honest step is a live MHCflurry or NetMHCpan pass on the 303 k-mers, a Natera somatic join if that BioCID is registered, and a vendor LLM overlay that actually returns JSON. None of those steps changes the rule that went live on 1 September 2026. Two votes. Drivers first. Show the board.
9. References
- Rojas LA, Sethna Z, Soares KC, Olcese C, Pang N, Patterson E, et al. Personalized RNA neoantigen vaccines stimulate T cells in pancreatic cancer. Nature. 2023;618:144-150. doi:10.1038/s41586-023-06063-y
- Sethna Z, Guasp P, Reiche C, Milighetti M, Ceglia N, Patterson E, et al. RNA neoantigen vaccines prime long-lived CD8+ T cells in pancreatic cancer. Nature. 2025;639:1042-1051. doi:10.1038/s41586-024-08508-4
- Sahin U, Derhovanessian E, Miller M, Kloke BP, Simon P, Löwer M, et al. Personalized RNA mutanome vaccines mobilize poly-specific therapeutic immunity against cancer. Nature. 2017;547:222-226. doi:10.1038/nature23003
- Ott PA, Hu Z, Keskin DB, Shukla SA, Sun J, Bozym DJ, et al. An immunogenic personal neoantigen vaccine for patients with melanoma. Nature. 2017;547:217-221. doi:10.1038/nature22991
- Wells DK, van Buuren MM, Dang KK, Hubbard-Lucey VM, Sheehan KCF, Campbell KM, et al. Key parameters of tumor epitope immunogenicity revealed through a consortium approach improve neoantigen prediction. Cell. 2020;183:818-834.e13. doi:10.1016/j.cell.2020.09.015
- Łuksza M, Riaz N, Makarov V, Balachandran VP, Hellmann MD, Solovyov A, et al. A neoantigen fitness model predicts tumour response to checkpoint blockade immunotherapy. Nature. 2017;551:517-520. doi:10.1038/nature24473
- O'Donnell TJ, Rubinsteyn A, Laserson U. MHCflurry 2.0: improved pan-allele prediction of MHC class I-presented peptides by incorporating antigen processing. Cell Syst. 2020;11:42-48.e7. doi:10.1016/j.cels.2020.06.010