plm-entropy 0.1

The pLM entropy Python package

pLM entropy

Code and data for: Lytras, S., Strange, A., Ito, J., and Sato K. (2025). Inferring context-specific site variation with evotuned protein language models. BioRxiv

pLM entropy is protein language model (pLM)-based metric to assess protein site conservation and variability. We test this metric using versions of two popular pLMs (ESM-2 and protT5) fine-tuned on the diversity of different Influenza A virus serotype hemagglutinin (HA) proteins.

The plm_entropy Python package

Installation

The python package for embedding sequences though ESM-2 and protT5 based models and calculating pLM entropy, along with a number of helper functions is available to install through pip. We further provide a conda environment yml file for easily installing all compatibilities.

To install the conda environment with the python package:

git clone

cd plm_entropy

conda env create -f pLM_entropy.yml

pip install update plm_entropy

Alternatively, the plm_entropy package can also be installed via pip from PyPI.

Usage example

Find a detailed usage example in the following Jupyter notebook.

Functions

embed_entropy_esm2(modnam, inputf, outf)

options:

• modnam: path to the pLM weights (ESM-2 or fine-tuned version only)
• inputf:
  • .csv file with at least one node column containing sequence identifiers and one seq column containing the amino acid sequences, or
  • fasta file with amino acid sequences (accepted extensions: .fa, .fas, .fasta)
• outf: output files name (without file extension)
• (optional) save_pickle: (default = False)
• (optional) save_logit: (default = False)
• (optional) torch_device: (default = cuda:0)

embed_entropy_protT5(modnam, inputf, outf)

options:

• modnam: path to the pLM weights (protT5MLM or fine-tuned version only)
• inputf:
  • .csv file with at least one node column containing sequence identifiers and one seq column containing the amino acid sequences, or
  • fasta file with amino acid sequences (accepted extensions: .fa, .fas, .fasta)
• outf: output files name (without file extension)
• (optional) save_pickle: (default = False)
• (optional) save_logit: (default = False)
• (optional) torch_device: (default = cuda:0)

aln_plm_entropy(entrdf_f, asralf)

Transforms the csv file with pLM entropy values outputted from the embed_entropy_esm2() or embed_entropy_protT5() functions so that values correspond to the alignment positions of the aligned input amino acid sequences.

options:

• entrdf_f: -site_entropy.csv output file containing per site pLM entropy values, calculated with the embed_entropy_esm2() or embed_entropy_protT5() functions
• asralf: protein sequence alignment of the sequences for which pLM entropy values were calculated in fasta format

calc_aln_entropy(alfile)

options:

• alfile:
• (optional) exclude_internal_nodes:

al_mod_entr_correl(alentropy, modentr_f)

options:

• alfile:
• modentr_f:
• (optional) exclude_internal_nodes:

pLM entropy applied to Influenza A virus HA evotuned pLMs

See Lytras, S., Strange, A., Ito, J., and Sato K. (2025). Inferring context-specific site variation with evotuned protein language models. BioRxiv for details.

Evotuned pLMs

The weights for the ESM-2 and protT5 -based evotuned pLMs presented in the manuscript above can be downloaded from the zenodo repository below:

Data & code available in this repository

• Evotuning

  • IAV HA sequence datasets for evotuning

  • Jupyter notebook for evotuning ESM-2

  • Python script for evotuning protT5 (and adding the MLM head)

• Inference

  • Jupyter notebook for inferring pLM entropy values using the evotuned models with the plm_entropy Python package.

  • Example data for the inference.

• Testing data

  • IAV HA sequences and phylogenies used in the manuscript.
  • Note that protein sequences retrieved from the GISAID database have been omitted from the repository.

• Additional tests

  • Jupyter notebook for predicting which protein sites will mutate in a given sequence context using pLM entropy.

  • Code for testing influence of pretrained ESM-2 by resetting weights before evotuning

  • Data and code for H3N2 tree backbone aBSREL branch-specific selection analysis