Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset
PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML · Peer-reviewed
What this source contributed
Records in this release that cite this source. A record can cite several sources.
Results
Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.
Claims
Statements this source supports, each with where in the source it was checked.
- ELASPIC-2: description: Description Stability predictor that operates on both an ESM sequence embedding and a graph-neural-network structural embedding. (Introduction)
- ESM therm: description: Description Protein language model fine-tuned on the folding stability of 528,000 natural and de novo sequences from 461 protein domains, which also accommodates deletions, insertions and multiple-point mutations. (Abstract; Introduction)
- Evidence summary: protein variant stability prediction: Summary Three published comparisons add evidence beyond the AMFR assay. Each uses its own dataset, split and metric, so values cannot be pooled. Pancotti et al. 2022 ran 21 predictors on S669, 669 experimental ddG variants in proteins dissimilar to the main training sets, plus their reverse variants; several of the tools, including ACDC-NN, DDGun, INPS and I-Mutant, were developed by the paper's own authors. The highest Pearson correlation on direct variants was 0.5 (MAESTRO), but MAESTRO fell to 0.2 on reverse variants with a bias of -0.57, while tools built to be antisymmetric, such as INPS-Seq, kept 0.43 in both directions with a bias of 0. On all 1,338 variants the highest correlations were 0.62 (PremPS) and 0.61 (ACDC-NN and INPS-Seq). Dieckhaus et al. 2024, the ThermoMPNN developers, held out proteins clustered at 25% identity in the mega-scale proteolysis set: ThermoMPNN reached Pearson 0.75 and RMSE 0.71 kcal/mol, retrained RaSP 0.71 and Rosetta 0.53 with RMSE 5.18. On a homologue-free FireProtDB split ThermoMPNN reached 0.65 against 0.59 for retrained PROSTATA and 0.47 for retrained RaSP. Chu et al. 2024, the ESM therm developers, show the limit of transfer: on held-out mega-scale domains of 40 to 72 residues RaSP and ELASPIC-2 reached Spearman 0.64 and Rosetta 0.61, but on six single proteins of 177 to 501 residues their fine-tuned model scored between -0.11 and 0.06, while ELASPIC-2 scored 0.47 to 0.58 on five of them and Rosetta 0.33 to 0.49. Before choosing a method, a user still needs to validate transfer to proteins larger than the small domains these models were trained on, and to check antisymmetry on reverse variants, which only one of these comparisons tests across many tools. None of the comparisons covers the AMFR construct, and none establishes whole-protein function, fitness or pathogenicity. (protein-stability-20261009-source-pancotti2022: Section 2, Table 1; protein-stability-20261009-source-dieckhaus2024: Table 2; protein-stability-20261009-source-chu2024: Tables 1 and 2)
- RaSP: description: Description Convolutional-neural-network stability predictor trained on Rosetta-computed stability changes, operating directly on the input protein structure. (Introduction)
- Relevance of Acetyltransferase dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by Acetyltransferase dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'Acetyltransferase dataset'; Table 2; Results)
- Relevance of Bgl3 dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by Bgl3 dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'Bgl3 dataset'; Table 2; Results)
- Relevance of BglB dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by BglB dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'BglB dataset'; Table 2; Results)
- Relevance of Lipase EstA dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by Lipase EstA dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'Lipase EstA dataset'; Table 2; Results)
- Relevance of Mega-scale dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by Mega-scale dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'Mega-scale dataset'; Table 2; Results)
- Relevance of Methyltransferase dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by Methyltransferase dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'Methyltransferase dataset'; Table 2; Results)
- Relevance of PTEN dataset stability ranking (Chu et al. 2024) to "Assess methods for protein stability experiments": Assessed by PTEN dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1) (protein-stability-20261009-source-chu2024: Table 1 row 'PTEN dataset'; Table 2; Results)
Models, benchmarks and other records
Methods (5)
Configurations (7)
Protocols (7)
- Acetyltransferase dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
- Bgl3 dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
- BglB dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
- Lipase EstA dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
- Mega-scale dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
- Methyltransferase dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
- PTEN dataset stability ranking, Spearman correlation (Chu et al. 2024 Table 1)
Datasets (7)
- Acetyltransferase dataset: Gentamicin 3-N-acetyltransferase, chemical stability
- Bgl3 dataset: Beta-glucosidase (Bgl3), catalytic susceptibility to heat shock
- BglB dataset: Beta-glucosidase (BglB), melting temperature (Tm)
- Lipase EstA dataset: Lipase EstA, melting temperature (T50)
- Mega-scale dataset: multiple, cDNA display proteolysis
- Methyltransferase dataset: Thiopurine S-methyltransferase, protein abundance
- PTEN dataset: PTEN, protein abundance
Evaluations (49)
- ELASPIC-2 on Acetyltransferase dataset (Chu et al. 2024)
- ELASPIC-2 on Bgl3 dataset (Chu et al. 2024)
- ELASPIC-2 on BglB dataset (Chu et al. 2024)
- ELASPIC-2 on Lipase EstA dataset (Chu et al. 2024)
- ELASPIC-2 on Mega-scale dataset (Chu et al. 2024)
- ELASPIC-2 on Methyltransferase dataset (Chu et al. 2024)
- ELASPIC-2 on PTEN dataset (Chu et al. 2024)
- ESM therm on Acetyltransferase dataset (Chu et al. 2024)
- ESM therm on Bgl3 dataset (Chu et al. 2024)
- ESM therm on BglB dataset (Chu et al. 2024)
- ESM therm on Lipase EstA dataset (Chu et al. 2024)
- ESM therm on Mega-scale dataset (Chu et al. 2024)
- ESM therm on Methyltransferase dataset (Chu et al. 2024)
- ESM therm on PTEN dataset (Chu et al. 2024)
- ESM-2 (3B) on Acetyltransferase dataset (Chu et al. 2024)
- ESM-2 (3B) on Bgl3 dataset (Chu et al. 2024)
- ESM-2 (3B) on BglB dataset (Chu et al. 2024)
- ESM-2 (3B) on Lipase EstA dataset (Chu et al. 2024)
- ESM-2 (3B) on Mega-scale dataset (Chu et al. 2024)
- ESM-2 (3B) on Methyltransferase dataset (Chu et al. 2024)
- ESM-2 (3B) on PTEN dataset (Chu et al. 2024)
- ESM-2 (35M) on Acetyltransferase dataset (Chu et al. 2024)
- ESM-2 (35M) on Bgl3 dataset (Chu et al. 2024)
- ESM-2 (35M) on BglB dataset (Chu et al. 2024)
- ESM-2 (35M) on Lipase EstA dataset (Chu et al. 2024)
Use cases (1)
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
17 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Artifact SHA-256 bb9830cbd6bab1b3e0ddbb52edb2a9afd4b06e427dbfeb19e6a07f843ebd914f Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Artifact url https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11293664/fullTextXML Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Doi 10.1371/journal.pcbi.1012248 Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Licence CC-BY-4.0 Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Media type application/xml Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Publication status peer_reviewed Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Retrieved at 2026-10-09T21:01:08Z Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Url https://doi.org/10.1371/journal.pcbi.1012248 Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Version PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Description Primary source retrieved and hashed for the protein stability use-case pass. Source metadata | Protein stability prediction by fine-tuning a protein language model on a mega-scale dataset No field-specific location recorded Version: PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Sources and history
Release 2026-11-21-adc15377fab2 · Record review: source checked
Primary source retrieved and hashed for the protein stability use-case pass.
Technical metadata and extraction receipts
Stable ID: protein-stability-20261009-source-chu2024
- areas
- proteins-complexes
- contexts
- research
- url
- https://doi.org/10.1371/journal.pcbi.1012248
- artifact url
- https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11293664/fullTextXML
- version
- PLOS Computational Biology 20(7):e1012248, published 2024-07-22; PMC11293664 full-text XML
- retrieved at
- 2026-10-09T21:01:08Z
- artifact sha256
- bb9830cbd6bab1b3e0ddbb52edb2a9afd4b06e427dbfeb19e6a07f843ebd914f
- doi
- 10.1371/journal.pcbi.1012248
- publication status
- peer_reviewed
- licence
- CC-BY-4.0
- media type
- application/xml