PyTorch Frustratometer: An Ultra-fast, Parallelized Tool for Proteome-Scale Local Energetic Frustration Analysis
- 1Evolutionary Systems Biophysics Group, Life Sciences Department, Barcelona Supercomputing Center, Barcelona, Spain
Abstract
Most natural protein sequences fold into well-defined three-dimensional structural ensembles, satisfying interactions that have been shaped over evolutionary timescales. The principle of minimal frustration states that most native interactions are energetically more favorable than alternative non-native ones, resulting in a funnel-shaped folding energy landscape. The FrustratometeR tool measures how the energetics of a residue or contact compares to decoys with different amino acid identities or configurations. Contacts are classified as minimally frustrated when these are more stable than random alternatives, highly frustrated when less stable, or neutral when energetically indistinguishable from them. Clusters of highly frustrated interactions are often evolutionarily conserved and are typically associated with functional regions. Therefore, identifying and quantifying local frustration can provide insights into both protein stability and function. Here we present a PyTorch reimplementation of FrustratometeR that accelerates conventional frustration calculations by 70-13,000x on a single GPU, depending on chosen frustration mode and protein size, with substantially larger gains for exhaustive mutational scanning, enabling proteome-scale local frustration analysis. We benchmark against the original algorithm on the CATH v4.4 S40 dataset and show near-identical results across all three frustration modes and mutational scanning. Source code is freely available at https://github.com/ParraLab-ESBG/pytorch-frustratometer.
Related
- Convolutional LSTM and Self-Attention Networks in PyTorch for Decade-Ahead Forecasting of Land-Cover Time Series
- Cortical Spectral Dynamics in Flow and Frustration
- Frustration reduces interpersonal competition through dynamic interpersonal neural synchronization in dyads
- Detection of Frustration-related Operant Behavior in Rats via Machine Learning Methods
- A spin glass representation of hierarchical gene regulation and the partial resolution of frustration
BibTeX
@misc{draizen2026pytorch,
title = {{PyTorch Frustratometer: An Ultra-fast, Parallelized Tool for Proteome-Scale Local Energetic Frustration Analysis}},
author = {E. J. Draizen and H. Jaafari and C. Bueno and A. Llaberia and F. Clark and E. A. Galpern and F. L. Simonetti and M. I. Freiberger and P. G. Wolynes and D. U. Ferreiro and R. G. Parra},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.06.764892},
url = {https://www.biorxiv.org/content/10.64898/2026.11.06.764892v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - PyTorch Frustratometer: An Ultra-fast, Parallelized Tool for Proteome-Scale Local Energetic Frustration Analysis
AU - E. J. Draizen
AU - H. Jaafari
AU - C. Bueno
AU - A. Llaberia
AU - F. Clark
AU - E. A. Galpern
AU - F. L. Simonetti
AU - M. I. Freiberger
AU - P. G. Wolynes
AU - D. U. Ferreiro
AU - R. G. Parra
PY - 2026
DA - 2026/11/06
PB - bioRxiv
DO - 10.64898/2026.11.06.764892
UR - https://www.biorxiv.org/content/10.64898/2026.11.06.764892v1
AB - Most natural protein sequences fold into well-defined three-dimensional structural ensembles, satisfying interactions that have been shaped over evolutionary timescales. The principle of minimal frustration states that most native interactions are energetically more favorable than alternative non-native ones, resulting in a funnel-shaped folding energy landscape. The FrustratometeR tool measures how the energetics of a residue or contact compares to decoys with different amino acid identities or configurations. Contacts are classified as minimally frustrated when these are more stable than random alternatives, highly frustrated when less stable, or neutral when energetically indistinguishable from them. Clusters of highly frustrated interactions are often evolutionarily conserved and are typically associated with functional regions. Therefore, identifying and quantifying local frustration can provide insights into both protein stability and function. Here we present a PyTorch reimplementation of FrustratometeR that accelerates conventional frustration calculations by 70-13,000x on a single GPU, depending on chosen frustration mode and protein size, with substantially larger gains for exhaustive mutational scanning, enabling proteome-scale local frustration analysis. We benchmark against the original algorithm on the CATH v4.4 S40 dataset and show near-identical results across all three frustration modes and mutational scanning. Source code is freely available at https://github.com/ParraLab-ESBG/pytorch-frustratometer.
ER -