# From coevolution to protein language models

> Residues in contact mutate together, so an alignment holds the fold (direct-coupling analysis). Language models trained only on sequences learn the same: ESM-2 scaled to 15 billion parameters predicts structures directly, and ESMFold built an atlas of more than 617 million metagenomic proteins.

- Page: https://www.labcanvas.com/templates/drug-discovery/coevolution-protein-language-models
- Open in the LabCanvas editor (LabCanvas account): https://www.labcanvas.com/editor.html?template=protein-language-models-coevolution
- Updated: 2026-10-14
- Category: Drug discovery & proteins
- Method: Protein language models
- Starting materials: Protein sequences (and their alignments)
- Product: Contacts, embeddings and structures
- Key result: Alignment and contact map are synthetic examples
- Figure: 1920 × 1080 px, 41 editable layers (5 generated objects with settings)

## Steps

1. **Coevolution**: Direct-coupling analysis of alignments predicts residue contacts across many protein families (Morcos 2011)
2. **Language model**: Trained on 86 billion amino acids in 250 million sequences, by filling in masked residues (Rives 2021)
3. **Structure**: ESM-2 scaled to 15 billion parameters; ESMFold predicts structure from a single sequence (Lin 2023)
4. **Atlas**: More than 617 million metagenomic structures, over 225 million with high confidence (Lin 2023)
5. **Generation**: ESM3 generated a fluorescent protein 58% identical to the nearest known one (Hayes 2025)

## Strengths

- No alignment needed at prediction time
- Fast enough for whole metagenomes
- Embeddings reused for many tasks

## Limitations

- Less accurate than alignment-based models on some proteins
- Large models need large compute to train

## Uses

- Structure prediction at scale
- Variant effect prediction
- Protein design

## References

1. Morcos et al., “Direct-coupling analysis of residue coevolution captures native contacts across many protein families”, PNAS 2011. https://doi.org/10.1073/pnas.1111471108 (primary)
2. Rives et al., “Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences”, PNAS 2021. https://doi.org/10.1073/pnas.2016239118 (primary)
3. Lin et al., “Evolutionary-scale prediction of atomic-level protein structure with a language model”, Science 2023. https://doi.org/10.1126/science.ade2574 (primary)
4. Hayes et al., “Simulating 500 million years of evolution with a language model”, Science 2025. https://doi.org/10.1126/science.ads0018 (primary)

Credit: After Morcos et al., PNAS 2011 (doi:10.1073/pnas.1111471108), Rives et al., PNAS 2021 (doi:10.1073/pnas.2016239118), Lin et al., Science 2023 (doi:10.1126/science.ade2574) and Hayes et al., Science 2025 (doi:10.1126/science.ads0018). Alignment and contacts synthetic; numbers from the abstracts.

## Questions

**What does the “From coevolution to protein language models” template show?** Residues in contact mutate together, so an alignment holds the fold (direct-coupling analysis). Language models trained only on sequences learn the same: ESM-2 scaled to 15 billion parameters predicts structures directly, and ESMFold built an atlas of more than 617 million metagenomic proteins.

**What are the steps shown in “From coevolution to protein language models”?** 1. Coevolution: Direct-coupling analysis of alignments predicts residue contacts across many protein families (Morcos 2011). 2. Language model: Trained on 86 billion amino acids in 250 million sequences, by filling in masked residues (Rives 2021). 3. Structure: ESM-2 scaled to 15 billion parameters; ESMFold predicts structure from a single sequence (Lin 2023). 4. Atlas: More than 617 million metagenomic structures, over 225 million with high confidence (Lin 2023). 5. Generation: ESM3 generated a fluorescent protein 58% identical to the nearest known one (Hayes 2025).

**What are the starting materials and the product?** Starting materials: Protein sequences (and their alignments). Product: Contacts, embeddings and structures.

**Which papers is “From coevolution to protein language models” based on?** Morcos et al., “Direct-coupling analysis of residue coevolution captures native contacts across many protein families”, PNAS 2011 (doi:10.1073/pnas.1111471108); Rives et al., “Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences”, PNAS 2021 (doi:10.1073/pnas.2016239118); Lin et al., “Evolutionary-scale prediction of atomic-level protein structure with a language model”, Science 2023 (doi:10.1126/science.ade2574); Hayes et al., “Simulating 500 million years of evolution with a language model”, Science 2025 (doi:10.1126/science.ads0018).

**Can I edit the “From coevolution to protein language models” template?** Yes. Open it in the LabCanvas editor: all 41 layers are editable, including 5 generated objects (structures, plots, equipment) with their own settings. Change the text, colours and layout, then export PNG, TIFF, SVG or PDF.

**Do I need an account to use this template?** Not to view it here. To open and edit it, sign in to LabCanvas or create a free account (the free plan keeps up to 3 designs). Downloading the figure is part of LabCanvas Pro (US$19 a month or US$199 a year).