henhaobio
Sign in

bioRxivBioinformaticsPreprint

FrustrAI-Seq: Scaling Local Energetic Frustration to the Protein Sequence Space

Jan-Philipp LeuschiD, Miriam Poley-GiliD, Miguel Fernandez-MartiniD, Franco L. SimonettiiD, Julius SchlensokiD, Nicola BordiniD, Burkhard RostiD, R. Gonzalo ParraiD, Michael HeinzingeriD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Proteins fold into their native three-dimensional (3D) structures by navigating complex energy landscapes shaped by the biophysical and biochemical properties of their sequence. Once folded, some sequence positions (dubbed residues) remain locally frustrated, reflecting functional constraints incompatible with optimal packing. This local energetic frustration provides important insights into protein function and dynamics, but its analysis typically relies on structure-based energy calculations and remains energetically costly at scale. Here, we introduce an ultra-fast sequence-based prediction of local energetic frustration directly from protein sequences using embeddings from protein language models (pLMs). Our method, coined FrustrAI-Seq, enables proteome-wide frustration profiling in minutes (17 minutes for the entire human proteome on a single Nvidia H100 GPU) while retaining biologically relevant performance as shown for the alpha-globin and beta-lactamase family. By eliminating the need for explicit structural or evolutionary information, this approach expands frustration analysis to protein regions and classes that were previously inaccessible, including intrinsically disordered regions and high-throughput de novo designed protein datasets. To support reproducibility and large-scale applications, we provide the largest freely available resource of precomputed local frustration scores to date (10^6 proteins), along with model weights and complete training and inference code at: github.com/leuschjanphilipp/FrustrAI-Seq.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (12 Aug 2026) · CC BY 4.0

A fine-tuned ProtT5 protein language model predicts per-residue energetic frustration from sequence (Pearson r 0.86 on unseen CATH topologies) and profiles the human proteome in ~17 minutes on one GPU.

Local energetic frustration flags residues tied to binding, allostery, catalysis and conformational switching, but computing it with FrustratometeR needs an experimental or predicted structure and becomes costly across large datasets. This work predicts it from sequence alone with a fine-tuned protein language model, extending frustration analysis to whole proteomes, intrinsically disordered regions and de novo designs, and releases precomputed scores for about one million proteins.

Key results

  1. ShownFrom sequence alone, a LoRA-tuned ProtT5 reproduces FrustratometeR residue frustration on CATH topologies unseen in training, beating an amino-acid-mean baseline and, for highly frustrated residues, the structure-based FrustraMPNN, though extreme values are under-predicted.

    Hold-out test set: Pearson r 0.86, mean absolute error (MAE) 0.33 vs r 0.76, MAE 0.45 (amino-acid mean baseline) and r 0, MAE 1.04 (shuffled labels). Highly frustrated class: F1 0.54 (precision 0.76, recall 0.42) vs 0.45 for FrustraMPNN (precision 0.50 at the same recall; FrustraMPNN macro F1 0.68). Minimally frustrated F1 0.85. Residues at positions ≥ 512: r 0.85, macro F1 0.72. · 2.2. Local Energetic Frustration Can Be…; 5.5. FrustraMPNN evaluation on the Funstration…; Tables

  2. ShownProteome-scale profiling is feasible: the human proteome took about 17 minutes on one H100 GPU, and about 136 million proteins from 22,105 UniProt proteomes were annotated; high-frustration proteins were enriched for membrane-transport Gene Ontology terms, mostly neutral ones for nuclear and gene-regulatory terms.

    Human proteome: 17 min (single NVIDIA H100) vs 58 h for FrustratometeR on AlphaFoldDB models using one Intel Xeon Platinum 8480+ core, structure prediction excluded; described as roughly three orders of magnitude faster across proteome sizes. GO analysis used the subset with protein- or transcript-level evidence (n = 255,794); top highly frustrated proteins were mainly transmembrane. · 2.9. Local frustration analysis at unprecedented…; 5.10. Proteome-wide inference; Discussion

  3. ShownPredictions carry over to designed proteins that lack evolutionary history: de novo designs and ProteinMPNN redesigns of natural folds.

    225 Protein Design Archive designs (86 enzymes, 23 fibres, 102 helical bundles, 14 nanoparticles): F1 0.86 / 0.78 / 0.52 for minimally / neutral / highly frustrated; minimally frustrated share 45% (FrustratometeR) vs 41% predicted; highly frustrated recall 0.40. ProteinMPNN sequences for 21 α-globin and 31 β-lactamase structures: Pearson r 0.91 and 0.88 vs FrustratometeR on ESMFold models. · 2.4. FrustrAI-Seq generalizes to de novo designs; 5.7. De novo dataset; 5.8. ProteinMPNN designs; Tables

  4. ShownWith no structural input, the single per-sequence profile resembles ensemble-averaged frustration more than any extreme conformer, and agreement with FrustratometeR declines as intrinsic disorder rises.

    414 ensembles (9,298 conformers): correlation is higher with ensemble mean or median profiles than with the most highly frustrated conformer; the best-matching conformer stays above about 0.75 up to 50-60% disordered residues and not below 0.6 beyond. Embeddings modestly improved a disorder predictor over base ProtT5 on all six disorder definitions, including CAID3. · 2.6. FrustrAI-Seq recovers conserved frustration…; 2.7. FrustrAI-Seq captures ensemble-averaged…; 5.9. Conformer dataset

  5. SuggestedPredicted frustration changes from single amino acid substitutions at functional sites agree with FrustratometeR, though an amino-acid-mean baseline comes close.

    Saturation scans at 12 α-globin protein-protein interaction sites (240 cases) and 6 β-lactamase catalytic sites (120 cases): Pearson r for Δ frustration vs FrustratometeR 0.76 and 0.88, compared with 0.74 and 0.87 for the amino-acid-mean baseline; per-site r ranged from -0.03 to 0.98 in α-globin and 0.70 to 0.97 in β-lactamase. · 2.8. Sensitivity of FrustrAI-Seq to mutations; 5.6. Use cases inference

Methods, data and reproducibility

ProtT5-XL-U50 tuned with low-rank adapters (LoRA, rank 4) plus a small CNN (shared first layer) feeding a regression head (squared-error loss on FrustratometeR index) and a class-weighted cross-entropy head; Funstration's 983,212 CATH FunFam sequences split by CATH topology into 887k train / 49k validation / 48k test; compared with FrustraMPNN, other pLM backbones, amino-acid-mean and shuffled baselines.

  • Funstration: CATH FunFam domains with TED/PDB structures (pLDDT ≥ 90, clustered at 80% identity) mapped to full-length UniProt sequences, FrustratometeR-labeled, split by CATH topology · n = 983,212 sequences; 186,477,919 residues; 8,259 FunFams
  • De novo designed proteins from the Protein Design Archive (enzymes, fibres, helical bundles, nanoparticles), split into single chains and deduplicated · n = 225 designs
  • ProteinMPNN sequences designed in prior work for α-globin and β-lactamase wild-type structures, compared against FrustratometeR on ESMFold models · n = 21 α-globin and 31 β-lactamase structures
  • α-globin family (FrustraEvo multiple sequence frustration alignment), experimental conformers of human α-globin, and saturation mutagenesis at protein-protein interaction sites · human (conformers, 2DN1-A); homologs from other species, e.g. guinea pig · n = 21 conformers; 12 sites, 240 residue-level mutation cases · PDB 2DN1 chain A; PDB 3A0G chain A
  • β-lactamase catalytic sites, all 19 substitutions at each · n = 6 sites; 120 residue-level cases · PDB 4BLM chain A
  • Conformational ensembles from PDB, PED and Jeon et al., clustered at 80% identity and binned by SETH-predicted disorder content · n = 414 ensembles; 9,298 conformers
  • UniProt reference proteomes for large-scale inference (sequence filters applied after retrieval); subset with protein- or transcript-level evidence used for GO enrichment · multiple species · n = 22,105 proteomes; 136,775,502 proteins retrieved; 255,794 high-evidence
  • Reference proteomes chosen for taxonomic coverage, used for runtime benchmarking · multiple species, including human (UP000005640) · n = Twenty-six proteomes · UP000000554, UP000001037, UP000005640, UP000006671, UP000189677, UP000444721 (only those processed individually)
Methods & tools
ProtT5-XL-U50, PEFT (LoRA), FrustratometeR, FrustraEvo, FrustraMPNN, ProteinMPNN, ESMFold, ESM2 650M, ProstT5, CD-HIT, SETH, CATH FunFams
Software & environment
Data versions stated: UniProt 2025-03, AlphaFoldDB v4; LoRA via the PEFT framework. No software versions, requirement files or containers given in the text.
Assay conditions & compute
Training: 4 NVIDIA H100 80 GB GPUs, distributed data parallel, early stopping (wall-clock time not stated). Inference: single H100 (fp16 for proteomes), about 17 min for the human proteome. FrustratometeR runtime baseline: one core of an Intel Xeon Platinum 8480+.
Reproducibility gaps
No release location stated for the 136-million-protein proteome predictions; training time and software versions not reported; de novo design IDs only in Supplementary Table 3.
Limitations (6)
  • One prediction per sequence, so frustration changes tied to conformational transitions, ligand binding, allostery or molecular dynamics trajectories cannot be resolved; structure-aware tools such as FrustratometeR remain the right choice there. · stated by the authors
  • Prediction quality may depend on how well a sequence is represented in the pLM embedding space (available evolutionary information); this has not been systematically evaluated for sparsely sampled sequence regions. · stated by the authors
  • The FrustraMPNN comparison carries caveats: FrustraMPNN was trained on labels from the legacy SeqDist 12 setting rather than 3, and may have seen the test proteins during training. · stated by the authors
  • Labels inherit FrustratometeR's coarse-grained, aqueous AWSEM Hamiltonian, which is blind to side chains, post-translational modifications, non-standard amino acids and lipid bilayers, so transmembrane regions tend to be labeled highly frustrated. · stated by the authors
  • Mutation analysis covers only 12 α-globin and 6 β-lactamase sites, beats the amino-acid-mean baseline by small margins (r 0.76 vs 0.74; 0.88 vs 0.87), and is benchmarked against FrustratometeR rather than experimental variant data. · noted in this summary
  • The speed comparison pits one H100 GPU against a single CPU core running FrustratometeR, so the reported three-orders-of-magnitude speed-up depends on that hardware pairing; parallel CPU runs are not reported. · noted in this summary

Read this if you need per-residue frustration estimates for large sequence sets, designed proteins or disordered regions without running structure-based energy calculations, or want a large precomputed frustration training set.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDraizen '26Beining '26Rausch '21Parra '24Freiberger '23Freiberger '22
77% semantic matchBioinformatics
Eli J. Draizen, Hana Jaafari, Carlos Bueno+8 more·bioRxiv·Nov 9, 2026