Genomics' blind spot
A synonymous variant is a DNA mutation that does not change the encoded amino acid. The codon changes, but the resulting protein is identical. For decades, synonymous variants were considered "silent" — with no functional effect.
We know that is false. Synonymous variants cause cystic fibrosis, cancer, and dozens of diseases. But two of the most widely used genomics predictors — REVEL and AlphaMissense — are blind to them: they operate at the protein level, so if the protein doesn't change they have no signal to measure. CADD does score synonymous variants, since it integrates nucleotide-level annotations, and SpliceAI scores them better than our thermodynamic signal when they fall near an exon border.
This is where QMetrika's two research lines converge.
If the protein doesn't change but the patient becomes ill, the pathology is in the mRNA, not in the protein. And the mRNA has its own physics that a synonymous mutation can alter.
The framework: Λ = S ⊗ Φ
S — genetic code (protein sequence) · semantic identity
Φ — biophysical constraints (mRNA thermodynamic form factor)
⊗ — systemic dependency between both axes
Every mRNA transcript has two layers of information. S is the digital message: the amino acid sequence it encodes. Φ is the analog channel: the mRNA thermodynamics that determine how efficiently that message is translated — local stability, translation speed, cotranslational folding.
In a missense variant, both axes change: ΔS ≠ 0, ΔΦ ≠ 0. In a synonymous variant, the argument is elegant: ΔS = 0, ΔΦ ≠ 0. The protein doesn't change, but the mRNA form factor does. If that causes disease, the pathology is purely thermodynamic.
This framework is presented in Paper 1 as a qualitative conceptual framework — not as formal algebra. Paper 2 makes it concrete in the opposite, stronger direction: the signal decomposes into two interpretable physical observables (signed stacking ΔΔG + G>A) that reproduce the deep classifier.
Pathogenic mechanisms
How can a variant that doesn't change the protein cause disease? Four mechanisms that the ΔG profile can capture:
mRNA destabilization
Alteration of local stacking that reduces thermodynamic stability → accelerated mRNA degradation → less protein produced.
Aberrant splicing
Local secondary structure change that alters splice sites or disrupts exonic enhancers/silencers (ESE/ESS).
Translation speed
Rare codons or stacking changes that slow the ribosome → cotranslational misfolding of the protein.
Cis regulation
Disruption of regulatory elements that depend on the nucleotide sequence, not the protein sequence.
Results
Dedicated CNN: AUC 0.683 ± 0.018 using only mRNA thermodynamic signal (ΔG channels). No evolutionary information, no conservation, no 3D structure. Pure nucleotide physics. It does not exceed SpliceAI.
A note on the AUCs: they are computed on ClinVar, where the value may reflect gene identity or variant position rather than the effect of the variant itself.
LOGO (Leave-One-Gene-Out): AUC 0.625 — the signal generalizes across genes; it is not an artifact of a particular gene.
Mechanistic finding — G>A enrichment: Pathogenic synonymous variants are 2.60× enriched in G>A transitions. This has a direct thermodynamic explanation: G>A reduces stacking (guanine stacks more strongly than adenine), and this effect is captured by the ΔG profile.
CDS length effect: Genes with longer CDS have stronger thermodynamic signal. This is consistent with the model: longer mRNAs have more opportunities for a local perturbation to propagate.
What this implies
An AUC of 0.683 does not compete with the best missense predictors (which reach 0.99), nor with SpliceAI on synonymous variants close to an exon border. But it doesn't have to. The question is not "Is it better than AlphaMissense?" but "Does a thermodynamic signal exist in synonymous variants?" And the answer is yes.
This points to a gap that protein-level predictors cannot reach: REVEL and AlphaMissense are blind to synonymous variants. EnergyFingerprint operates at the nucleotide level, and its biophysical channel works for synonymous variants without modification. The gap is not empty: CADD scores them, and SpliceAI does it better than we do when the mechanism is splicing.
EnergyFingerprint is not competing with AlphaMissense on its turf. It looks at a different axis: the physics of the mRNA.
The trilogy
This paper is the first of three that progressively unfold the Λ = S ⊗ Φ framework:
Three papers, one escalating question
"The signal exists" (this paper) — A dedicated CNN demonstrates that the thermodynamic signal discriminates pathogenic synonymous variants from benign ones. The Λ = S ⊗ Φ framework is presented as a qualitative conceptual framework.
"The signal is interpretable" (horizon 2027) — The dedicated CNN reduces to two physical observables: the signed nearest-neighbor stacking ΔΔG (Turner parameters) plus the G>A mutational bias. A two-feature model matches the deep network and generalizes across genes — full interpretability of the σ operator, a step toward the traceability the clinic demands.
"The predicted physics occurs in the cell" (horizon 2027-2028) — Direct experimental validation. DMS-MaPseq measures in vivo accessibility. Reporter assays measure expression. Does predicted ΔΦ = actual structural change?
L1 × L2 convergence
This work is the natural convergence of QMetrika's two research lines. L1 (EnergyFingerprint) provides the CNN and the transferability gradient. L2 (mRNA Design) demonstrates that the ΔG profile is a useful and orthogonal dimension.
Together they predict exactly this result: if ΔG is informative for missense and orthogonal to MFE, it should also be informative for synonymous variants — where it is the only available signal, because the protein doesn't change.
The natural extension is the evolutionary channel. ESM-1v operates at the protein level — for synonymous variants it does not directly apply. But a masked marginal at the DNA level (not protein) could capture evolutionary pressure on the nucleotide sequence. That is the subject of Paper 2.
Update (Sep 2026): a later control qualifies this result. Among ClinVar's confident pathogenic synonymous variants, ~65% fall within ≤3 nt of an exon border, and SpliceAI —a free tool— classifies them better than the thermodynamic signal (AUC 0.97 vs σ+G>A 0.59 in the deep-exonic stratum). In practice, the ClinVar positive set is a set of splicing variants. This does not refute that thermodynamics matters for expression/translation — it says ClinVar cannot prove it, and that functional data is needed (MPRA / reporter / DMS-MaPseq). The result stands as interpretability (σ+G>A ≈ CNN), not as an independent clinical classifier.