AI, biocomputation and data science applied to mRNA: variant classification and therapeutic design from its thermodynamic form factor — the physical axis no predictor uses. Sensitive data never leaves your infrastructure.
No signup. Each research line distills an interpretable physical observable into a tool that runs entirely on your machine — no GPU, no data leaving it.
The thermodynamic signal σ + G>A that matches the deep network. Available as a web tool, CLI and VEP plugin.
pip install ef-synonymous
from raw data to an actionable result
algorithms that speak molecular biology
cnn · llm · governed multi-agent
the mRNA form factor others don't see
The lab's first vertical. A biophysical platform that operates where REVEL and AlphaMissense are blind: the thermodynamic profile of the mRNA and, in particular, synonymous variants — a niche with no competitor. A shared engine, nearest-neighbor ΔG, feeds three fronts: variant classification, mRNA design and repeat-expansion disorders.
The core finding: the signal reduces to an interpretable 2-variable physical model that matches the deep network — auditable as ACMG evidence.
Reviewable scientific evidence: three preprints and a technical benchmark report. A trilogy in progress, with peer review underway.
Multichannel 1D CNN, 42 validated genes, AUC 0.87–0.99. Zero-shot transferability gradient.
ΔG as an axis orthogonal to MFE (R²=0.144). A dual optimizer that outperforms commercial vaccines.
AUC 0.683. Λ = S ⊗ Φ framework. The CNN reduces to 2 physical observables (signed ΔΔG + G>A bias), auditable as ACMG evidence (PP3/BP4).
Formal comparison across 7 genes. EF beats CADD in SCN1A (0.913 vs 0.832). Additive value demonstrated.
Preprints with DOI, open-source code and two PyPI packages. Glass-box, no signup.