Thermodynamic Profiling for Synonymous Variant Classification: When the Message Changes Without Changing the Protein

Abstract

In synonymous variants, the signal of pathogenicity is detectable as a local, signed thermodynamic perturbation of nearest-neighbor nucleotide stacking — orthogonal to protein-level predictors — that a two-feature interpretable model reproduces at the level of a dedicated deep classifier, and whose clarity increases with the quality of clinical annotation - arguing for genuine signal over ascertainment artifact. A dedicated follow-up control, however, qualifies the clinical scope: on ClinVar the confident pathogenic synonymous set is dominated by splice-region variants (~65% within ≤3 nt of an internal exon–intron border) that SpliceAI classifies at AUC 0.97, with the stacking signal adding nothing once SpliceAI is included; the result therefore stands as an interpretable, orthogonal correlate, not an independent clinical classifier. Synonymous genetic variants — nucleotide changes that do not alter the encoded protein — are increasingly recognized as contributors to human disease through mechanisms including mRNA destabilization, aberrant splicing, altered translation kinetics, and disruption of cis-regulatory elements. However, current state-of-the-art variant effect predictors (REVEL, AlphaMissense, CADD) operate primarily at the protein level and are therefore blind or insensitive to synonymous pathogenicity. Here I present an extension of the EnergyFingerprint framework to synonymous variant classification. Using nearest-neighbor stacking free energy (Delta G) profiles — the same thermodynamic engine validated for missense classification (with an AUC of 0.72 to 0.99 across 8 genes) — I demonstrate that the biophysical channel alone (9 channels, no evolutionary signal) discriminates pathogenic from benign synonymous variants with an AUC of 0.683 plus or minus 0.018 in 5-fold cross-validation across 641 pathogenic and 130,143 benign variants from 505 ClinVar genes. Leave-one-gene-out validation (with an AUC of 0.625, where 72 percent of 54 genes score above chance) confirms cross-gene generalization of learned thermodynamic disruption patterns. The anti-correlation with missense transfer models (AUC of 0.386) demonstrates that synonymous pathogenicity operates through fundamentally different thermodynamic geometries than protein-level damage. Mechanistic analysis reveals G-to-A transition enrichment (of 2.60 times) consistent with CpG-mediated constraint, and a dependence on CDS length suggesting the 128-nucleotide context window captures more relevant signal in compact genes. In this work, I introduce a qualitative conceptual framework (represented as Lambda equals S coupled with Phi) to hypothesize that biological transcript viability (Lambda) emerges from the non-linear coupling between discrete sequence information (S) and the local thermodynamic form factor (Phi). Under this framework, I show that a synonymous variant leaves S unchanged (with Delta S equal to zero) but perturbs Phi, offering a principled mechanistic lens to interpret synonymous pathogenicity through mRNA physics without relying on amino acid changes.