Scientific Foundation

Hybrid physics and machine learning, grounded in experimental data

Scala's stability model combines residue-level energy functions with sequence-fitness representations trained on deep mutational scanning datasets. Neither approach alone achieves the accuracy we need for practical wet-lab guidance.

Methodology

Why a hybrid model outperforms either component alone

Physics-based energy functions capture the thermodynamic intuition behind stability: packing, electrostatics, solvation, and backbone strain. But they struggle on remote homologs and multi-site combinations where context matters beyond the local interaction graph.

Sequence-fitness models trained on deep mutational scanning data capture evolutionary and functional context. But they can be overfit to the training protein family and miss non-conservative substitutions outside the training distribution.

Scala's ensemble combines residue-level energy deltas with learned fitness features, weighting each contribution by a per-family confidence estimate derived from sequence identity to known training data. This gives the model calibration: it knows when to lean on physics and when to defer to sequence context.

Scientific approach visualization showing hybrid model combining physics and machine learning

Benchmark Performance

Independent benchmark results

Evaluated on held-out test sets from ProThermDB and FireProtDB, using only sequence as input. Spearman rank correlation with experimental Tm shift or dG change.

Thermostability rank correlation (all families) 0.81

Spearman r on ProThermDB held-out set, 1,240 single-point variants

Industrial enzyme family accuracy 0.77

Lipases, cellulases, xylanases subset; same evaluation protocol

Multi-site combinatorial variant recall (top 10%) 68%

Fraction of actual top-10% stabilizing variants found in model's top-10% predictions

Zero-shot accuracy on novel families 0.63

Families with less than 20% sequence identity to any training protein

Benchmarks use publicly available datasets (ProThermDB v4, FireProtDB 2024). Full evaluation protocol available on request.

Publications

Foundation literature and Scala technical notes

Global analysis of protein folding using massively parallel design, synthesis, and testing

Rocklin GJ, Chidyausiku TM, Goreshnik I, et al.

Science, 357(6347):168-175, 2017. DOI: 10.1126/science.aan0693

Foundational thermostability dataset used in model training

Language models enable zero-shot prediction of the effects of mutations on protein function

Meier J, Rao R, Verkuil R, et al.

NeurIPS 2021 (Advances in Neural Information Processing Systems, vol. 34). bioRxiv: 2021.07.09.450648

Sequence-fitness representation approach informing Scala's learned features

Scala internal technical note: confidence calibration for out-of-distribution protein families

Scala Biodesign technical team, 2025.

Internal document. Available to research partners on request.

Scala internal technical note

Questions about our methods? Talk to the team.

We are happy to discuss benchmark methodology, dataset coverage, or how the model handles your specific protein family.