QMetrika Labs [ research ] [ preprints ] [ tools ] [ blog ] es/en
// thermodynamic bio-ai

Energy turned into knowledge

AI and biocomputation applied to mRNA. A shared biophysical engine — nearest-neighbor stacking ΔG — operates on a physical axis that MFE and protein-level predictors don't use. Each line distills an interpretable physical observable, not a black box: auditable models that run locally.

§ 01research lines
3 fronts
  1. L2flagship line

    Therapeutic mRNA optimization

    Sequence design from the stacking axis (ΔG), which LinearDesign and MFE don't capture. As a computational proof of concept, without experimental validation, the dual optimizer (ΔG + MFE) reaches more favourable ΔG and MFE values than BNT162b2 and mRNA-1273; 78 % of the stacking axis variance is neither local composition nor secondary structure. Code not public; open to collaboration and licensing.

    P-L2 →
  2. L1variants

    Variant classification

    Thermodynamic stacking profiles for missense and synonymous variants. For synonymous variants, the signal reduces to 2 interpretable physical observables (signed ΔΔG + G>A bias).

    P-L1 →P1 →
  3. —repeats

    Repeat-expansion disorders

    Physical observable of repeat-interruption architecture (type and position) in DM1, Huntington, SCA1 and Fragile X.

    energorna →
§ 02preprints and technical notes
not peer reviewed · DOI on Zenodo
  1. P-L2preprint

    mRNA Stacking Optimization: Codon design guided by nearest-neighbor thermodynamics

    Stacking ΔG as a dimension of its own next to MFE. Dual optimizer: a computational proof of concept on ΔG and MFE, without experimental validation.

    10.5281/zenodo.20228979
  2. P-L1preprint

    EnergyFingerprint: Thermodynamic stacking profiles for missense variant classification

    Multichannel 1D CNN. 8 genes in the preprint, intra-gene AUC 0.72–0.99†; 42-gene panel, results unpublished. Zero-shot transferability gradient.

    10.5281/zenodo.19831154
  3. TECH-2technical report

    Benchmark: EnergyFingerprint vs CADD, REVEL, AlphaMissense

    Comparison across 7 genes. Overall, EF sits below CADD (ΔAUC −0.024, p = 0.048), REVEL (−0.028) and AlphaMissense (−0.040)†; in SCN1A it beats CADD (0.913 vs 0.832, 5 folds). Whether EF adds value to these predictors is a hypothesis yet to be tested.

    10.5281/zenodo.21309529
  4. P1preprint

    Synonymous variants alter mRNA thermodynamic profiles: a CNN-based pathogenicity signal

    AUC 0.683 ± 0.018 on ClinVar†, not exceeding SpliceAI. The CNN reduces to 2 interpretable physical observables (signed ΔΔG + G>A bias).

    10.5281/zenodo.20275792

† AUC computed on ClinVar: it may reflect gene identity or variant position rather than the effect of the variant itself.

§ 03open tools
no sign-up · no GPU · local
  1. L1synonymous

    ef-synonymous

    Interpretable 2-variable physical model for synonymous variants. Web, CLI and VEP plugin. pip install ef-synonymous

    webpypigithub
  2. —repeats

    energorna

    Repeat-interruption architecture; glass-box and model-free. pip install energorna

    webpypigithub