Progen2 Tools#

Configuration File: remote_tools/progen2_tools.json Tool Type: Remote Tools Count: 2

This page contains all tools defined in the progen2_tools.json configuration file.

Available Tools#

run_progen2_generate (Type: RemoteTool)#

Generate new protein sequences with the ProGen2 protein language model by nucleus sampling, optio…

run_progen2_generate tool specification

Tool Information:

  • Name: run_progen2_generate

  • Type: RemoteTool

  • Description: Generate new protein sequences with the ProGen2 protein language model by nucleus sampling, optionally continuing a prompt. Use it to propose candidate sequences, extend a partial design, or sample a family’s neighbourhood. Give ‘prompt’ an N-terminal fragment to condition on, or leave it out for an unconditioned sample. Output is bounded on purpose: at most 16 sequences of at most 1024 tokens each, which is what the provider’s VRAM budget is declared at. Results are reproducible for a given ‘seed’.

Parameters:

  • prompt (string) (optional) An N-terminal fragment to continue, in single-letter amino-acid code, or a single-record FASTA. Omit it (or give ‘1’) to sample without conditioning. The prompt is NOT closed with a C-terminus marker, because generation continues it.

  • num_samples (integer) (optional) How many sequences to sample (1-16). Peak GPU memory is roughly linear in num_samples x max_length.

  • max_length (integer) (optional) Total token budget per sequence INCLUDING the prompt (ProGen2’s own –max-length). At most 1024. Sampling stops early at a terminus marker.

  • temperature (number) (optional) Sampling temperature (ProGen2’s –t; its default is 0.2, which samples conservatively). Higher is more diverse and less protein-like.

  • top_p (number) (optional) Nucleus sampling threshold (ProGen2’s –p, default 0.95).

  • seed (integer) (optional) Random seed (ProGen2’s –rng-seed, default 42). The same seed and the same arguments give the same sequences.

Example Usage:

query = {
    "name": "run_progen2_generate",
    "arguments": {
    }
}
result = tu.run(query)

run_progen2_score (Type: RemoteTool)#

Score one protein sequence with the ProGen2 protein language model: the log-likelihood it assigns…

run_progen2_score tool specification

Tool Information:

  • Name: run_progen2_score

  • Type: RemoteTool

  • Description: Score one protein sequence with the ProGen2 protein language model: the log-likelihood it assigns to the sequence, as a sum over residues (ll_sum), a per-residue mean (ll_mean) and the corresponding perplexity. Both are the average of the left-to-right and right-to-left readings, which is what ProGen2’s own likelihood.py reports. Use it to rank designed sequences, compare variants of one protein, or judge how protein-like a sequence is. Higher (less negative) ll_mean means the model finds the sequence more likely. Per-residue log-probabilities are opt-in and bounded, because a long sequence is a long array; the whole array can be written to a .npy file instead.

Parameters:

  • sequence (string) (required) One protein sequence in single-letter amino-acid code, or a single-record FASTA. ProGen2’s own terminal markers are accepted: ‘1<seq>2’ reads N to C and ‘2<seq>1’ reads C to N. Given amino acids alone the sequence is wrapped ‘1<seq>2’. At most 2046 residues (progen2-base’s context window is 2048 tokens).

  • per_token_log_prob (boolean) (optional) Return the log-probability of every scored token, left to right, one number per position. Refused, not clipped, above 2048 positions – ask for ‘per_token_file’ instead.

  • per_token_file (boolean) (optional) Write the per-token log-probabilities to a float32 .npy file on the provider host and return its path, shape and size instead of its contents.

  • decimals (integer) (optional) Decimal places for returned floats (default 6).

Example Usage:

query = {
    "name": "run_progen2_score",
    "arguments": {
        "sequence": "example_value"
    }
}
result = tu.run(query)