Case · SR-OST-001 · published
A 9-year-old Rottweiler presented for lameness with a lytic lesion on the distal radius. Biopsy confirmed osteosarcoma. We took the same tumor through DarkScan V2 + Sci-JEPA to show what a personalized canine mRNA vaccine pipeline looks like end-to-end — patient summary, ranking, provenance, and cost.
Patient (de-identified)
The pipeline, in plain language
The tumor and a matched-normal saliva sample arrived from a partner vet school as FFPE blocks with a redacted clinical history. We ran whole-exome and bulk-RNA sequencing through our pipeline at ~$8 of compute on RunPod A100 80GB inference with orchestration on a $7/month Railway dyno.
DarkScan V2 acted as an intelligent funnel: from roughly a million raw variants to a top ten neoantigen candidates. Each stage applies a cheap filter first (expression, MHC presentability against canine DLA-88) before the expensive Sci-JEPA and immunogenicity scoring runs on what survives.
The output is the same shape a human-oncology tumor board would expect — a ranked peptide list with rationale, allele binding, self-identity, and cross-species literature support — ready for the next conversation.
DarkScan V2 funnel — measured on this case
Stage 1
Raw variants
WES + RNA tumor vs. matched-normal calls
1,047,392
Stage 2
Expressed in tumor
RNA-seq read-supported (≥3 reads)
11,204
Stage 3
MHC-presentable
DLA-88 + DLA-12 binding, IC50 ≤ 500 nM
1,386
Stage 4
Non-self
Host-proteome subtraction (self-identity < 80%)
187
Stage 5
Top ranked
Sci-JEPA cross-species + immunogenicity scoring
10
The speedup buys turnaround time. The accuracy comes from keeping the expensive Sci-JEPA + immunogenicity inference for the ~10² candidates that survive the upstream filters, rather than running it on the full ~10⁶.
Top candidate
Peptide
VSPSFSSTL
Source protein
MAGEC1
MHC allele (canine analog)
DLA-88*HLA-B*27:05
Why it ranks first
Cancer-testis antigen with very low normal-tissue expression. Sci-JEPA flags a strong pediatric-OS corollary; in-silico immunogenicity high.
Note on alleles: canine MHC is reported as DLA-88; we show the matched human-MHC corollary (HLA-B*27:05) so readers comparing to pediatric-osteosarcoma literature can cross-reference. Both alleles are used in the full clinical report.
Full ranking (top 10)
| # | Peptide | Source | IC50 (nM) | Self-identity | Rationale |
|---|---|---|---|---|---|
| 1 | VSPSFSSTL | MAGEC1 | 26 | 62% | CT-antigen; strong pediatric-OS corollary in Sci-JEPA |
| 2 | YLLDDLLAR | RUNX2 | 142 | 71% | Osteoblast TF; canine-expressed; moderate-affinity DLA-88 |
| 3 | SLLPAIVEL | TP53 (R175H) | 88 | 68% | Hotspot analog; cross-species literature support |
| 4 | TLDNVISGV | EZH2 | 112 | 65% | Polycomb component; recurrent in human OS as well |
| 5 | AMFQDPQER | MAGEA4 | 204 | 70% | CT-antigen family; lower affinity but high specificity |
| 6 | QLAEKVLEK | AURKA | 168 | 67% | Mitotic kinase; targetable by alisertib class |
| 7 | RVASRTLLL | MET (exon 14 skip) | 94 | 63% | Splice analog of human MET ex14 skip — interesting CDx fit |
| 8 | GVYDGREHTV | PBK | 276 | 72% | Mitotic checkpoint; recurrent across canine OS cases |
| 9 | FLDEFMEGV | MYC | 189 | 69% | Amplification-driven; high baseline expression in tumor |
| 10 | ALLEPSDTV | BIRC5 (survivin) | 221 | 74% | Apoptosis inhibitor; pan-cancer relevance |
Cost & provenance
Total compute
$7.94
RunPod A100 80GB · pipeline-run cost
Orchestration
$7/mo
Railway dyno · amortized
Wall-clock TAT
6h 12m
From sample receipt to report
For investigators & collaborators
Includes the complete pipeline trace, all 187 non-self candidates with binding + self-identity scores, vaccine construct draft, suggested IND-enabling next steps, and citations. We email it after a one-click confirmation.
What this means for you
When your vet brings up sequencing or a clinical trial, you have a real, end-to-end example of what the pipeline actually does. No marketing — just the trace.
This is the same shape of output we'd return on a case you submit. Familiar columns, familiar units, plus the provenance metadata your reviewers will ask for.
Replicable run IDs, dataset versions, model versions, ledger hashes. Send us a comparable canine osteosarcoma sample and we'll re-run with the same configuration.