The science behind ProtSynq
Technical notes, dataset acknowledgments, and methodology descriptions for researchers who want to understand what's inside the scoring engine. If you're going to trust computational outputs for your next screen, you should know what produced them.
The ΔΔG formalism for point mutant stability
A protein's thermodynamic stability is characterized by its folding free energy ΔGfold — the free energy difference between the unfolded and folded states. When a point mutation is introduced at position i, the perturbation is captured as ΔΔG = ΔGfold(mutant) − ΔGfold(wild type). A negative ΔΔG means the mutation stabilizes the fold; a positive value means destabilization.
ΔΔG prediction is harder than it looks. The dominant contributors — van der Waals packing, electrostatics, solvation, and backbone strain — are all sensitive to the local structural environment of the mutated residue. Small errors in geometry propagate. And experimentally measured ΔΔG values reflect the thermodynamic average over a conformational ensemble, not a single structure.
ProtSynq approaches the problem with an ensemble scoring strategy: multiple physics-based and data-driven terms are computed independently, calibrated against experimental measurements, and combined using a learned weighting. This reduces the variance inherent in any single estimator and produces calibrated confidence intervals alongside each ΔΔG prediction.
Scanning the full single-substitution landscape
For a 350-residue protein, there are 350 × 19 = 6,650 possible single amino acid substitutions. ProtSynq evaluates all of them. The output is a per-residue mutation landscape: for each position, a heatmap row showing the predicted ΔΔG stability and ΔΔG binding for all 19 alternative amino acids.
The landscape view is not the same as a ranked list of top-5 variants — it exposes structure. Positions that are tolerant to substitution (broad flat rows in the heatmap) are informative for directed evolution. Positions with narrow fitness peaks (one or two acceptable substitutions) require more precision.
ProtSynq surfaces both the shortlist and the landscape, so researchers can calibrate their confidence: does position 123 have one good mutation or ten? Does the Pareto-optimal variant cluster in a specific secondary structure element?
What the learned models were trained on
The learned ΔΔG component of ProtSynq's scoring ensemble is a fine-tuned transformer model. Pre-training used a large corpus of protein sequence and structure data. Fine-tuning used experimentally measured ΔΔG values drawn from two curated databases:
ProThermDB — calorimetric measurements (DSC, ITC) and spectroscopic measurements (far-UV CD) of thermostability for thousands of single-substitution mutants across a range of protein families.
ProtaBank — curated experimental ΔΔG values from published mutagenesis studies, covering both thermostability and binding affinity measurements.
The training split was performed chronologically — data published before 2022 was used for training and validation; data from 2022 onward was held out for evaluation. This temporal split is more realistic than random split for assessing generalization to new experimental measurements.
Curation approach: we removed duplicate entries, excluded measurements with known quality flags in the source databases, and applied a physicality filter (rejecting ΔΔG values below −8 kcal/mol or above +8 kcal/mol as likely experimental artifacts). What remains is a smaller but cleaner training corpus.
How we evaluate performance
Performance is evaluated as Pearson correlation between predicted and experimentally measured ΔΔG on a held-out test set. The benchmark below reflects internal evaluation on our chronological holdout — not a published result, not a cherry-picked test set.
Read the methodology notes.
Our technical write-ups go deeper than this page — thermodynamic decomposition, calibration, handling of disordered inputs.