Methodology, data sources, and benchmark notes.

ProtSynq's ΔΔG predictions are produced by an ensemble of three distinct scoring terms, calibrated and combined using a learned weighting scheme. The ensemble approach reduces variance compared to any individual estimator.

Term 1 — Rosetta-based free energy decomposition. Local structural perturbations around the mutated position are minimized using a simplified energy function derived from the Rosetta force field. This term captures packing density changes, backbone strain, and short-range electrostatics. We use a stripped-down version of the calculation to maintain per-residue scan times below 30 seconds.

Term 2 — Sequence-based evolutionary conservation. A multiple sequence alignment of homologous proteins is used to compute a position-specific log-odds score for the proposed substitution. High-conservation positions penalize non-conservative substitutions regardless of local structure. This term prevents the physics component from suggesting mutations that evolutionary pressure has already rejected.

Term 3 — Learned ΔΔG residual. A transformer model fine-tuned on experimental ΔΔG data from ProThermDB and ProtaBank predicts a residual correction on top of the physics and conservation terms. The learned component captures effects that the two physics-adjacent terms systematically underestimate — particularly for hydrophobic core packing and surface salt bridge networks in larger proteins.

Weights for combining the three terms were learned by optimizing Pearson correlation on a validation set. The weighting is protein-class-specific: globular single-domain proteins, multi-domain proteins, and membrane-adjacent segments use slightly different weight profiles.

Binding ΔΔG predictions require the complex structure (e.g., antibody–antigen, enzyme–substrate). The scoring approach differs from stability ΔΔG in that it focuses on the change in interface energy rather than the change in overall folding stability.

ProtSynq identifies the binding interface by computing buried surface area for each residue in both the bound and unbound states. Residues contributing more than 10 Ų of buried surface area are classified as interface residues and scored with the interface-specific scoring function.

The interface scoring function weights three contributions: direct contact energy (van der Waals and electrostatics between the two chains), backbone conformational strain introduced by the mutation, and predicted loss of key hydrogen bonds and salt bridges across the interface. The current model does not account for entropic contributions from interface loop dynamics — this is a known limitation for highly flexible interfaces.

The learned components of ProtSynq were trained on experimentally measured ΔΔG values from two primary sources:

ProThermDB. A database of protein thermostability measurements curated from the literature. We used calorimetric (DSC, ITC) and spectroscopic (far-UV CD) measurements. Entries with flagged quality issues were excluded. The database is publicly available at prothermdb.ddbj.nig.ac.jp.

ProtaBank. A curated database of protein–protein interaction data including binding affinity measurements and mutational effects. Available at protabank.org.

Curation steps: (1) deduplicated entries using protein UniProt accession + position + mutation identity; (2) excluded measurements outside the physically plausible range (|ΔΔG| > 8 kcal/mol); (3) performed a chronological train/validation/test split with the test set containing only data published in 2022 or later.

We do not claim to have improved on the source databases themselves — our contribution is the curation pipeline and the integration into the ensemble scoring framework.

Being explicit about limitations is part of operating a scientific tool responsibly. These are the cases where ProtSynq's predictions are less reliable:

Disordered regions and flexible loops. Predictions for intrinsically disordered regions or long surface loops are unreliable. The physics-based term depends on a well-defined local structure; loop flexibility violates that assumption.

Combinatorial mutations. ProtSynq scores single substitutions. When two mutations are introduced simultaneously, epistatic effects can make the combined ΔΔG non-additive. We do not model combinatorial effects. For multi-mutation libraries, treat ProtSynq's shortlist as individual candidate positions, not as pre-validated combination sets.

Orphan proteins (low homology). The evolutionary conservation term degrades for proteins with few homologs in major sequence databases. ProtSynq flags jobs where the best multiple sequence alignment has fewer than 50 effective sequences.

Membrane proteins. The current scoring function was calibrated on soluble proteins. Predictions for integral membrane proteins or membrane-adjacent segments are not validated and should be treated with extra skepticism.

ΔΔG magnitude, not ranking. ProtSynq's Pearson correlation of 0.78 reflects ranking ability. The absolute ΔΔG magnitudes in kcal/mol are less reliable than the relative ranking. Use ProtSynq to prioritize variants, not to predict exact Tm shifts.