ProGen2

What is ProGen2?

ProGen2 is a family of autoregressive protein language models from Salesforce Research. It learns the token patterns of natural and metagenomic protein sequences, then either samples new sequences from a context string or scores an existing sequence under the model.

ProGen2 works entirely in sequence space. It does not require a structure, multiple sequence alignment, or target receptor. That makes it useful for broad sequence exploration, but generated candidates still need structural, functional, and experimental validation.

ProteinIQ runs the Salesforce source at commit c27a419 and uses the published model checkpoint archives.

How to use ProGen2 online

Choose Generate to sample sequence continuations, or Score likelihood to evaluate one token string. Enter an uppercase context or sequence, select a checkpoint, and run the job on a cloud GPU.

Input

Input	Format	Notes
Context / sequence	One continuous token string or one FASTA record	Use uppercase A-Z except J, with optional control tokens 1 or 2. Wrapped lines are accepted. Lowercase letters, spaces within the sequence, and multiple FASTA records in one run are rejected.

The default generation context is 1. A partial sequence such as 1MKTLL asks ProGen2 to continue from that prefix. Score likelihood expects at least two tokens.

In a workflow, an incoming sequence artifact supplies the context input for that job. Multi-FASTA artifacts are materialized as individual ProGen2 jobs, and each sequence is validated against the selected model and maximum total length.

Modes

Mode	Behavior	Main result
Generate	Samples one or more complete context-plus-continuation strings with nucleus sampling.	Clean protein sequences and a FASTA file.
Score likelihood	Runs the published likelihood.py behavior and averages left-to-right and right-to-left scores.	ll_sum and ll_mean.

Shared settings

Setting	Default	Description
Model	progen2-large	Published ProGen2 checkpoint used for generation or scoring.
Random seed	42	Seeds Python and PyTorch with deterministic CUDA behavior.

Generation settings

Setting	Range	Default	Description
Number of sequences	1 to 20, with lower limits for larger models	1	Number of return sequences sampled in one call.
Top-p	0.01 to 1.0	0.95	Nucleus sampling threshold. Lower values restrict each step to a smaller high-probability token set.
Temperature	0.01 to 2.0	0.2	Scales the sampling distribution. Lower values are more conservative; higher values increase diversity.
Maximum total length	2 to 2048, with lower limits for larger models	256	Total token length of the submitted context plus generated continuation. It must be greater than the context length.

Temperature 0.2 and top-p 0.95 are defaults from the published command-line program. They are not universal paper defaults: the ProGen2 experiments evaluate several sampling temperatures and top-p values.

Choosing a model

Model	Parameters	Training focus	Service envelope
progen2-small	151M	UniRef90 and BFD30	Up to 20 sequences, 2048 total tokens
progen2-medium	764M	UniRef90 and BFD30	Up to 12 sequences, 2048 total tokens
progen2-base	764M	UniRef90 and BFD30	Up to 12 sequences, 2048 total tokens
progen2-oas	764M	Observed Antibody Space	Up to 12 sequences, 2048 total tokens
progen2-large	2.7B	UniRef90 and BFD30	Up to 4 sequences, 1024 total tokens
progen2-BFD90	2.7B	BFD90	Up to 4 sequences, 1024 total tokens
progen2-xlarge	6.4B	UniRef90 and BFD30	1 sequence, 512 total tokens

The compute envelopes keep the largest requests within the available GPU memory and job duration. For general generation, progen2-large is a practical starting point. For antibody-like sequences, progen2-oas is trained on immune repertoire data. The xlarge checkpoint is substantially slower and is best reserved for focused runs.

Context and control tokens

The training format uses 1 and 2 as terminal markers around a protein sequence. Each training sequence is provided in both forward and reversed order, as described in the ProGen2 paper. These tokens do not select a training dataset; the selected checkpoint determines the model used for the run. Both tokens can terminate a sampled sequence.

ProteinIQ preserves the source program's terminal truncation behavior, then removes a leading 1 or 2 and a trailing 1 or 2 from the user-facing sequence. The JSON detail file keeps both the source truncation and the raw decoded completion for provenance.

Results

Output	Format	Meaning
Results table	Spreadsheet	Generated sequences and lengths, or likelihood metrics, together with model and source commit.
Generated sequences	FASTA	One clean sequence per generated sample. Available only in Generate mode and reusable in workflows.
Generation details	JSON	Raw decoded completions, source-truncated completions, clean sequences, token counts, and provenance.
Likelihood scores	CSV and JSON	Averaged bidirectional ll_sum and ll_mean values.
Likelihood log	TXT	Console output produced by the pinned Salesforce likelihood.py entrypoint.

ll_sum scales with sequence length, while ll_mean is normalized across scored positions. These values are model scores, not probabilities of biological function, stability, expression, or experimental success. Scores are most useful for comparisons made with the same checkpoint and token conventions.

ProGen2 vs alternative sequence design tools

• RFdiffusion and EvoDiff support structural or motif conditioning. ProGen2 has no structural awareness.
• ProFam is designed for family-conditioned generation from homologous sequence context.
• PepMLM focuses on peptide binder design.
• ProteinMPNN designs sequences for a supplied backbone. ProGen2 does not use a backbone.
• RFantibody and IgDesign use antibody-specific structural context. progen2-oas is useful when antibody-like sequence generation without a structure is sufficient.

Generated candidates should be checked for sequence quality, structural plausibility, liabilities, and target-specific behavior before experimental prioritization.