EnergyFingerprint: mRNA Thermodynamic Profiling Combined with Protein Language Models Enables Zero-Shot Cross-Gene Missense Variant Classification

Accurate classification of missense variants remains a critical challenge in clinical genomics, particularly for genes with limited clinical data. I present EnergyFingerprint, a lightweight 1D convolutional neural network (~228K parameters) that classifies pathogenic and benign missense variants by integrating two orthogonal information axes: mRNA thermodynamic stacking profiles (nearest-neighbor free energy parameters from SantaLucia and Turner) and protein evolutionary conservation scores from ESM-1v, a 650M-parameter protein language model. Each variant is encoded as an 11-channel tensor over a 128-nucleotide window centered on the mutation site. I validate the model across eight genes spanning four protein families: cancer-associated (BRCA1, TP53, PTEN, PALB2), ABC transporter (CFTR), globin (HBB), voltage-gated ion channel (SCN1A), and intrinsically disordered protein (MECP2). Intra-gene classification achieves AUC 0.87–0.99 across all structured proteins. Crucially, a model trained solely on BRCA1 achieves zero-shot cross-gene transfer with AUC 0.85–0.97 for cancer genes, 0.78 for an ABC transporter (CFTR), 0.71 for a globin (HBB), and 0.64 for a transmembrane channel (SCN1A), without any retraining. However, transfer fails for MECP2 (AUC 0.43), a protein with ~50% intrinsically disordered regions, despite strong intra-gene performance (AUC 0.91). These results define a transferability gradient governed by structural similarity and reveal a sharp boundary between structured and disordered protein domains. The orthogonality of the two axes (r ≈ 0) explains their synergy: combined AUC reaches 0.94 versus ~0.86 for either axis alone. EnergyFingerprint demonstrates that fundamental biophysical and evolutionary signals, encoded compactly and processed by a small CNN, can match or complement large-scale variant effect predictors while offering interpretability and domain-specific modularity