Evo2 Tools#

Configuration File: remote_tools/evo2_tools.json Tool Type: Remote Tools Count: 2

This page contains all tools defined in the evo2_tools.json configuration file.

Available Tools#

run_evo2_embed (Type: RemoteTool)#

Embed one DNA sequence with the Evo 2 genomic foundation model and return the mean-pooled embeddi…

run_evo2_embed tool specification

Tool Information:

  • Name: run_evo2_embed

  • Type: RemoteTool

  • Description: Embed one DNA sequence with the Evo 2 genomic foundation model and return the mean-pooled embedding vector of one hidden layer, plus the sequence’s log-likelihood. Use it to compare, cluster or search genomic regions by learned representation. The full per-position matrix is NOT returned inline (4096 floats per base): ask for ‘position_range’ to get a bounded window of position vectors, or ‘per_position_file’ to get the whole matrix written as a .npy file whose path comes back.

Parameters:

  • sequence (string) (required) ONE DNA sequence in IUPAC nucleotide codes, or a single-record FASTA. At most 8192 bp.

  • layer_name (string) (optional) Which hidden layer to take the embedding from, as ‘blocks.<i>.mlp.l3’. Defaults to the last block’s. Only these are offered: the optimization kit folds the input-projection weights, so a hook on a projection layer would return that folded layout.

  • position_range (array) (optional) Inclusive 1-based [start, end] window of position vectors to return inline. Refused, not clipped, when (end - start + 1) * embedding_dim exceeds 20000 floats – at Evo 2’s 4096 dimensions that is 4 positions.

  • per_position_file (boolean) (optional) Write the whole (length x embedding_dim) float32 matrix to a .npy file on the provider host and return its path, shape and size instead of its contents.

  • per_position_log_prob (boolean) (optional) Also return the log-probability of every observed base, one number per position – L-1 of them, for 1-based positions 2..L, since the first base has no preceding context. Refused for sequences longer than 4000 bp.

  • decimals (integer) (optional) Decimal places for returned floats (default 5).

Example Usage:

query = {
    "name": "run_evo2_embed",
    "arguments": {
        "sequence": "example_value"
    }
}
result = tu.run(query)

run_evo2_score (Type: RemoteTool)#

Score DNA sequences with the Evo 2 genomic foundation model: the mean log-likelihood the model as…

run_evo2_score tool specification

Tool Information:

  • Name: run_evo2_score

  • Type: RemoteTool

  • Description: Score DNA sequences with the Evo 2 genomic foundation model: the mean log-likelihood the model assigns each sequence, and its perplexity. Use it to compare a reference allele against an alternate one (score both windows and take the difference), to rank designed or synthetic sequences by how genome-like they are, or to flag unusual regions. Returns one score per FASTA record, in input order.

Parameters:

  • sequence (string) (required) DNA in IUPAC nucleotide codes: one sequence, or a FASTA of up to 8 records. At most 8192 bp per record and 16384 bp in total.

  • reduce (string) (optional) ‘mean’ for the mean per-base log-likelihood (the default, and comparable between sequences of different length), or ‘sum’ for the pseudo-log-likelihood of the whole sequence (which grows with length).

  • average_reverse_complement (boolean) (optional) Score each sequence on both strands and average. Doubles the work; use it when the strand of the input is arbitrary.

  • decimals (integer) (optional) Decimal places for returned floats (default 5).

Example Usage:

query = {
    "name": "run_evo2_score",
    "arguments": {
        "sequence": "example_value"
    }
}
result = tu.run(query)