Protein Design · Salesforce Research
ProGen2
Generates protein and antibody sequences conditionally using autoregressive language modelling.
What ProGen2 is
ProGen2 is an autoregressive protein language model developed by Salesforce Research. It generates plausible amino acid sequences, scaling up to 6.4 billion parameters, and allows generation to be conditioned on specific protein family or functional tags.
For immunological applications, the model family includes ProGen2-OAS, a specialised checkpoint trained on 554 million antibody sequences sourced from the Observed Antibody Space database.
How it works
ProGen2 operates as a causal autoregressive language model trained on protein sequence data. It takes an amino acid sequence prompt as a prefix and iteratively predicts subsequent residues based on learned sequence probabilities.
The model outputs generated completions along with their corresponding log-likelihood scores. In addition to general protein models, an antibody-specific variant was trained using 554 million sequences from the Observed Antibody Space.
Key facts
- Model architecture
- Autoregressive causal language model
- Maximum parameter count
- 6.4B
- Antibody training data
- 554 million sequences (Observed Antibody Space)
- GPU memory requirement
- 24 GB
What you can use it for
- Conditional Protein Generation
- Generating novel amino acid sequences guided by predefined protein family or functional tags.
- Antibody Sequence Design
- Designing and sampling immunoglobulin sequences using the ProGen2-OAS checkpoint.
- Sequence Extension and Scoring
- Extending partial sequence prompts and evaluating the log-likelihood of the resulting sampled proteins.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
ProGen2
Strengths and limits
Good at
- Scales up to 6.4 billion parameters for generating protein sequences.
- Supports conditioned generation using functional and protein family tags.
- Provides a specialised checkpoint trained on over half a billion antibody sequences.
- Calculates log-likelihood metrics alongside generated sequence samples.
Watch out for
- Requires 24 GB of GPU memory to run.
- Operates exclusively on sequence representations without direct 3D structural inputs.
Inputs and outputs
You provide
Sequence prefixoptional
Optional prefix to condition generation on.
Samplesoptional
You get back
Sampled sequences with log-likelihood.
Licence
Commercial use permitted · BSD-3-Clause
The BSD-3-Clause licence allows use in commercial research and products.
Questions
Can I use ProGen2 commercially?
Commercial use permitted. The BSD-3-Clause licence allows use in commercial research and products.
How do I run ProGen2 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 20s, and a run costs about 168 credits. Visitors without an account get 1,000 credits a month.
What inputs does the model require?
The model takes an initial amino acid sequence prefix as a prompt, alongside a parameter specifying the desired number of sequence samples to generate.
Which checkpoint should be used for antibody research?
The ProGen2-OAS checkpoint should be selected, as it was specifically trained on 554 million sequences from the Observed Antibody Space.
What information is included in the output?
The model returns a table containing the sampled amino acid sequences together with their respective log-likelihood values.
How to cite it
Nijkamp, et al. arXiv:2206.13517
Written from ProGen2's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.