tV triVax AI
TESLA public benchmark · CC BY 4.0
AI-DRIVEN THERAPEUTIC VACCINE DESIGN

Turn real experimental data
into vaccine decisions.

A working decision platform built on the public TESLA neoantigen benchmark, showing how peptide–HLA candidates are ranked and tested against patient-matched T-cell evidence.

Public research data · Prototype models · Not for clinical diagnosis or efficacy prediction
LEAD CONSTRUCT ATYKGVPYEVK
COMPOSITE SCORE 99.3
STATUS T-cell positive
THE PRODUCT THESIS

Not another prediction tool. A candidate decision layer.

714

Real pMHC records

Public peptides, matched HLA alleles, and experimental immunogenicity labels.

33

Experimentally positive

Candidates recognized in patient-matched T-cell assays.

5×

Cross-validated scoring

Every record is scored by a fold that never trained on it.

OPEN-SOURCE PLATFORM MAP

Eight modules. One end-to-end vaccine decision system.

01
SELECTED MODULE

Antigen Discovery

BSD-3-Clause
INPUTTumor VCF + RNA expression
→
PROCESSVariant annotation & mutant peptide generation
→
OUTPUTCandidate tumor antigens

Generate mutation-derived peptides that can enter patient-specific HLA prediction.

Execution status · Architecture mapped
LAST EXECUTED RESULT Loading reproducible run…
PLATFORM OUTPUT AI-ranked personalized vaccine candidate
PUBLIC PATIENT-LEVEL MAIN CHAIN

From a real cancer VCF to a manufacturable candidate.

PUBLIC BREAST CANCER SAMPLE

HCC1395 + matched normal HCC1395BL

SELECTED CONSTRUCT EPITOPES
REPRODUCIBLE ARTIFACT Loading…

VEP-annotated VCF · tumor/normal coverage · RNA expression · pVACtools · MHCflurry · UniProt · DNA Chisel

THE AI DECISION ENGINE

ML belongs at the decision layer—not as eight disconnected models.

Rank · Calibrate · Explain · Learn
MULTIMODAL CANDIDATE REPRESENTATION LIVE FEATURE FUSION

One candidate vector connects biology, safety, and manufacturability.

Peptide sequencePatient HLAPresentation RNA expressionDNA / RNA VAFClonality Human-proteome matchHydrophobicityConstruct risk
Current HCC1395 transparent score 62% presentation + 18% expression + 12% tumor VAF + 8% RNA VAF
CURRENT ML BENCHMARK

Grouped five-fold out-of-fold validation

0.735ROC AUC
0.116Average precision
8 / 33Top-50 positive recall

TESLA labels test whether the model prioritizes experimentally recognized pMHC candidates without scoring a record using a model trained on that record.

01

Learning to rank

Optimize the order of candidates rather than treating every peptide as an isolated binary classification.

Current: OOF baseline · Next: calibrated gradient boosting / neural ranker
02

Uncertainty and OOD

Separate a high score from high confidence and flag unfamiliar alleles, motifs, or assay domains.

Calibration · Confidence intervals · Out-of-distribution alerts
03

Explainable selection

Show which biological signals increased or reduced rank so scientists can review every decision.

Feature attribution · Counterfactuals · Audit trail
04

Active learning loop

Choose the next assay that delivers the most information, then use results to update future ranking.

Candidate → assay → label → retraining → better candidate
LIVE WORKFLOW

TESLA · Public neoantigen benchmark

Melanoma + NSCLC · Published cohort
AI analysis pipeline Ready
01
Dataset IngestionLoad public peptide–HLA pairs and assay labels
714 rows
02
Data QualityValidate sequence, HLA, and label consistency
714 valid
03
Feature EngineeringEncode sequence composition, anchor positions, and HLA
16 alleles
04
Out-of-fold ScoringGrouped five-fold validation prevents direct leakage
714 scored
05
Benchmark EvaluationCompare rankings with experimental T-cell labels
AUC 0.735
DECISION OUTPUT

Model rankings tested against real experimental outcomes.

RankPeptideHLAOOF score
SELECTED CANDIDATE

TESLA-006

94.2
Epitope FLTSVINRV
✓
Experimentally positive

Patient-matched T-cell testing detected recognition of this peptide–HLA pair.

ROC AUC0.735
Average precision0.116
Top-50 positive recall8 / 33
Evaluation designGrouped 5-fold OOF
FROM CANDIDATE TO CONSTRUCT

From ranked candidates to an optimized mRNA construct.

5′ Cap
5′ UTR
Signal
ATYKGVPYEVK
Linker
KIYTGEKPYK
3′ UTR
Poly(A)
GC content—
Coding length—
3D fast preview—
Day 60 memory—
WHY THIS CAN COMPOUND

Every experiment makes the next candidate decision smarter.

The platform compounds through a shared data layer, decision layer, and experimental feedback layer. Candidate selection becomes a learning R&D system—not a one-off prediction run.

triVax
Decision AI
Patient data
Model scores
Lab validation
Learning loop
Platform summary generated