# Protein language model

## Embed, score and generate protein sequences

Embed protein sequences, score substitutions without assay data, generate new sequences, and fold the results with ESM-2, ESM-C, ProGen2, ESM-IF1, AbLang, and ESMFold.

[Run ESM-2](https://proteiniq.io/app/esm-2)[Open the ubiquitin example](https://proteiniq.io/app/esm-2?jobId=k098v2zryv285vv2uyguze14)

![ESM-2 on ProteinIQ with human ubiquitin, a heatmap of layer 33 embeddings with 76 residues by 1280 dimensions](https://proteiniq.io/_next/image?url=%2Fimages%2Fuse-cases%2Fscreens%2Fesm-2-ubiquitin-last-layer-light.webp&w=1920&q=75&dpl=dpl_7UnP1WeaWTToeLYR9CGTBoRfu8WH)![ESM-2 on ProteinIQ with human ubiquitin, a heatmap of layer 33 embeddings with 76 residues by 1280 dimensions](https://proteiniq.io/_next/image?url=%2Fimages%2Fuse-cases%2Fscreens%2Fesm-2-ubiquitin-last-layer-dark.webp&w=1920&q=75&dpl=dpl_7UnP1WeaWTToeLYR9CGTBoRfu8WH)

## Trusted by 16,500+ researchers and published in

- [Bioorganic Chemistry](https://proteiniq.io/customers/furan-morpholine-metal-complexes-ramachandran-plot)
- [Journal of Dairy Science](https://proteiniq.io/customers/mle-19-salmonella-dairy-boltz-2)
- [Antibiotics](https://proteiniq.io/customers/palm-sugar-bacteriocin-ramachandran-plot)
- [Frontiers in Cellular and Infection Microbiology](https://proteiniq.io/customers/tuberculosis-multi-epitope-vaccine-alphafold-2)
- [Microbial Biotechnology](https://proteiniq.io/customers/pet-lipase-accessibility-sasa)
- [Journal of Medicinal Chemistry](https://proteiniq.io/customers/ace-inhibitor-peptide-stability)
- [BMC Plant Biology](https://proteiniq.io/customers/chloroplast-dna-controls-random-dna)
- [Translational Neurodegeneration](https://proteiniq.io/customers/coa6-coq10-binding-protenix)
- [International Journal of Molecular Sciences](https://proteiniq.io/customers/halogenated-auxins-lipophilicity-molecular-descriptors)
- [Microorganisms](https://proteiniq.io/customers/amnoonvirus-pb2-validation-molprobity)

1. - [ESM-2](https://proteiniq.io/app/esm-2)
   - [ESM-C](https://proteiniq.io/app/esm-c)

   ## Embed the sequences

   Embeddings turn each sequence into numbers that other models can use. ESM-2 returns per-residue and mean-pooled vectors, and ESM-C comes in three sizes. Human ubiquitin gives a 76 by 1280 matrix from layer 33.
2. - [ESM-C Mutation Scoring](https://proteiniq.io/app/esm-c-mutation-scoring)
   - [EvoIF](https://proteiniq.io/app/evoif)
   - [ProteusAI](https://proteiniq.io/app/proteusai)

   ## Score variants

   ESM-C Mutation Scoring ranks every single substitution by log-likelihood ratio, and EvoIF adds a structure and a homolog alignment. ProteusAI fits a model to your own fitness data. These are rankings, not measured effects.
3. - [ProGen2](https://proteiniq.io/app/progen)
   - [EvoDiff](https://proteiniq.io/app/evodiff)
   - [ProFam](https://proteiniq.io/app/profam)
   - [ESM-IF1](https://proteiniq.io/app/esm-if1)

   ## Generate new sequences

   ProGen2, EvoDiff and ProFam generate sequences from context, motifs or a family alignment, and ESM-IF1 proposes sequences for a backbone. A generated sequence is a hypothesis until it is folded and tested.
4. - [AbLang-2](https://proteiniq.io/app/ablang-2)
   - [AbLang](https://proteiniq.io/app/ablang)

   ## Work with antibodies

   AbLang-2 scores paired or unpaired antibody sequences and returns embeddings, and AbLang restores masked residues. Both are antibody-specific, which a general model is not.
5. - [ESMFold](https://proteiniq.io/app/esmfold)

   ## Fold and check candidates

   ESMFold predicts a structure from sequence alone and reports pLDDT, pTM and optional PAE. A confident fold does not show that a sequence works.

## Run protein language model on ProteinIQ

[Compare plans](https://proteiniq.io/pricing)

## More protein engineering workflows

- [Protein sequence design Create or optimize amino-acid sequences against structural, functional, or developability goals.](https://proteiniq.io/use-cases/protein-sequence-design)
- [Inverse folding Design sequences for a fixed backbone and refold each candidate to check it.](https://proteiniq.io/use-cases/inverse-folding)
- [Antibody design Design antibody CDRs and variable domains, then triage the candidates.](https://proteiniq.io/use-cases/antibody-design)

[See all use cases](https://proteiniq.io/use-cases)

## Read the scores before you trust them

Background guides on the methods, scores, and numbers behind these workflows.

- [![Ubiquitin as a row of 76 residue tokens with leucine 43 masked, a fan of the 20 amino acids above it and arcs to the residues it touches in 3D.](https://proteiniq.io/_next/image?url=%2Fimages%2Fguides%2Fprotein-language-models%2Ffeatured.webp&w=1920&q=75&dpl=dpl_7UnP1WeaWTToeLYR9CGTBoRfu8WH) Proteins · 18 min ### What is a protein language model, and what is it used for A protein language model is a neural network trained to predict amino acids from the rest of a sequence. Learn how ESM-2, ESM C, ESM3, ProtT5 and ProGen2 differ, what their embeddings and likelihoods are used for, how they compare with alignment-based methods like AlphaFold2, and which licences allow commercial use.](https://proteiniq.io/guides/protein-language-models)
- [![A teal helix that continues into a loose chain, above a row of one bead per residue with vertical bars that are tall under the helix and short under the chain](https://proteiniq.io/_next/image?url=%2Fimages%2Fguides%2Fplddt%2Ffeatured.webp&w=1920&q=75&dpl=dpl_7UnP1WeaWTToeLYR9CGTBoRfu8WH) Structures · 17 min ### What is pLDDT, and what counts as a good score pLDDT is the per-residue confidence score that AlphaFold, ESMFold and Boltz attach to predicted structures. Learn what the 90, 70 and 50 cutoffs mean and why neither a high nor a low score settles whether a region is real.](https://proteiniq.io/guides/plddt)
- [![Two folded protein domains joined by a linker, with ghost outlines showing uncertain relative placement.](https://proteiniq.io/_next/image?url=%2Fimages%2Fguides%2Fpae%2Ffeatured.webp&w=1920&q=75&dpl=dpl_7UnP1WeaWTToeLYR9CGTBoRfu8WH) Structures · 16 min ### What is PAE, and how to read a PAE plot PAE (predicted aligned error) is the pairwise confidence matrix that AlphaFold, Boltz and Chai-1 write next to a predicted structure. Learn how to read its blocks and stripes and use it for domains and interfaces.](https://proteiniq.io/guides/pae)

[See all guides](https://proteiniq.io/guides)

## FAQs

AI answers from our docs. No sign-in needed.

## Focus on research, not infrastructure

Start free with 200 credits a month and no installs. Upgrade when your project needs more compute.

[Start for free](https://proteiniq.io/sign-up)[View pricing](https://proteiniq.io/pricing)
