StabilityArc: Decoding Protein Sequence Embeddings into Generalizable Stability Landscapes

Aaron L. Feller; Andrew D. Ellington; Claus O. Wilke

q-bio.BMSep 30, 2026

StabilityArc: Decoding Protein Sequence Embeddings into Generalizable Stability Landscapes

Authors: Aaron L. Feller, Andrew D. Ellington, Claus O. Wilke

Organizations: The University of Texas at Austin Interdisciplinary Life Sciences · The University of Texas at Austin Integrative Biology

Abstract

Every protein has a unique stability landscape, but the physical consequences of mutation are governed by recurring biochemical constraints. We test whether a shared decoder, trained on measurements from diverse proteins, can interpret these constraints in an unseen target, enabling cross-protein transfer for initial experimental round prescreening. We present StabilityArc , which maps frozen ESMC-600M residue representations through a shared RoPE transformer to an Lx20 matrix of substitution effects; a symmetric, contact-aware residual aids in predicting epistasis in simultaneous substitutions. In 66 strict leave-one-protein-out evaluations covering 134,794 ProteinGym variants, StabilityArc achieves 0.7134 Spearman correlation, exceeding the strongest zero-shot baseline, ProSST-2048 (0.6526), by 0.0608. We further explore the utility of this method by providing the score as a prior for Kermut, achieving Spearman correlation of 0.8280 across three supervised split schemes, improving on Kermut's reported 0.8167.

Figures & tables

Figure 1: Cross-protein learning and deployment of StabilityArc . During training (top), alternate proteins ( $q$ ) are passed through a frozen ESMC-600M encoder to train the shared 4-layer RoPE transformer adapter against ground-truth experimental fitness landscapes. All labels from the target protein are strictly withheld. At deployment (bottom), the target wild-type sequence ( $p$ ) is encoded to provide structural representations and contact maps, which the learned adapter decodes into a complete $L\times 20$ single-substitution stability landscape. For simultaneous substitutions, a multi-mutant scoring module (right) combines these single effects with a learned epistasis interaction term using attention-derived contacts.

Method	Singles	Doubles	Overall
ESM3	0.6460	0.3262	0.6406
VenusREM	0.6370	0.3803	0.6503
ProSST-4096	0.6243	0.4100	0.6522
ProSST-2048	0.6385	0.3801	0.6526
ESM-C-600M (frozen prior)	0.4887	0.2355	0.5277
StabilityArc (ours)	0.6946	0.4119	0.7134

Table 1: Macro Spearman on 66 proteins assayed for stability. All comparator model values are from the ProteinGym score export; overall score concatenates singles and doubles within each assay.

Method	Random	Modulo	Contig.	Mean
Kermut (reported)	0.9046	0.7871	0.7583	0.8167
Kermut + StabilityArc	0.9094	0.7984	0.7763	0.8280
Gain	+0.0048	+0.0113	+0.0180	+0.0113

Table 2: Supervised stability benchmark. Kermut values are reported by Groth et al. (2024) .

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Spearman	AUROC	Recall	Precision	Accuracy
StabilityOracle (supplied)	0.718	0.852	0.376	0.938	0.669
StabilityArc (ours)	0.606	0.804	0.527	0.835	0.705
$\Delta$ ( StabilityArc $-$ StabilityOracle)	$-$ 0.112	$-$ 0.048	+0.151	$-$ 0.103	+0.037

Appendix

Table 3: Paired pooled T2837 comparison on 1,884 shared raw PDB-chain mutations from 19 targets with more than 30 shared mutations.

\Delta

is StabilityArc minus StabilityOracle; both use

\Delta\Delta G\geq 1

kcal/mol as the positive class.

Metric	Singles	Doubles	All variants
Macro Spearman	0.6946	0.4119	0.7134

Appendix

Table 4: Zero-shot StabilityArc performance by mutation order.

Explore similar work

Constraint-Aware Optimization for Robust Protein Stability Prediction
Jun 6, 2026A Shivram, Aneesh S. Chivukula, Manik Gupta +1Protein Structure Prediction Lattice Mutation
Department of Computer Science and Information Systems, Birla Institute of Technology and Science Pilani, Hyderabad Campus, Hyderabad, 500078, Telangana, India · Department of Biological Sciences, Birla Institute of Technology and Science Pilani, Hyderabad Campus, Hyderabad, 500078, Telangana, India
ProtSent: Protein Sentence Transformers
May 7, 2026Dan Ofer, Oriel Perets, Michal Linial +1Language Modeling Transformer Architectures
Department of Computer and Information Science Ben-Gurion University of the Negev
LC-SEPLM: long-range contact-supervised adaptation for sequence-only protein representation learning
Jul 24, 2026Chen Wang, Boming Kang, Qinghua CuiProtein Structure Prediction Sequence Modeling
School of Electrical and Electronic Engineering, Hubei University of Technology, Hubei, China · Department of Biomedical Informatics, State Key Laboratory of Vascular Homeostasis and Remodeling, School of Basic Medical Sciences, Peking University, 38 Xueyuan Rd, Beijing, 100191, China. · School of Sports Medicine, Wuhan Institute of Physical Education, No. 461 Luoyu Rd. Wuchang District, Wuhan 430079, Hubei Province, China

Jun 6, 2026, cs.LGJ/K move · Enter open · S save

Constraint-Aware Optimization for Robust Protein Stability Prediction

A Shivram, Aneesh S. Chivukula, Manik Gupta, Sourav Chowdhury

Department of Computer Science and Information Systems, Birla Institute of Technology and Science Pilani, Hyderabad Campus, Hyderabad, 500078, Telangana, India · Department of Biological Sciences, Birla Institute of Technology and Science Pilani, Hyderabad Campus, Hyderabad, 500078, Telangana, India

Protein Structure Prediction Lattice Mutation Feature Alignment Language Modeling

Open paper AskPDF arXiv

Multimodal

ΔΔG

predictors integrating protein language models with inverse-folding representations achieve strong in-distribution accuracy on the Megascale dataset but exhibit limited robustness on out-of-distribution (OOD) proteins, persistent forward-reverse bias on paired-mutation benchmarks, and under-representation of rare stabilizing mutations. Existing approaches address these limitations primarily through additional architectural components, leaving optimization-level intervention comparatively underexplored. We introduce a constraint-aware optimization framework combining Balanced Mean Squared Error, a Siamese anti-symmetric regularizer, and a novel OOD-margin consistency loss on the per-position feature representation, requiring no architectural changes to the SPURS backbone. Across eleven benchmarks and three random seeds, the framework improves Spearman correlation on S669 from 0.486 to 0.540 (

σ=0.002

across seeds), matching the published SPURS baseline (0.50) without architectural modification, and on S461 from 0.653 to 0.711, with consistent smaller gains on five additional OOD datasets. A controlled diagnostic on Ssym reveals that anti-symmetric training does not eliminate systematic forward-reverse bias, indicating that gains arise through implicit regularization rather than exact thermodynamic constraint enforcement.

Abstract

Explore similar work

Constraint-Aware Optimization for Robust Protein Stability Prediction

ProtSent: Protein Sentence Transformers

LC-SEPLM: long-range contact-supervised adaptation for sequence-only protein representation learning