This repository contains a Python package called plmutils that provides some tools for generating embeddings of protein sequences using pre-trained protein language models (PLMs).
It also includes code to predict whether putative open reading frames (ORFs) are coding or noncoding based on their embeddings. See the section below about ORF prediction for more details.
It accompanies the pub "Using protein language models to predict coding and noncoding transcripts".
First clone this repo and cd into it. Then create a conda environment from the envs/dev.yml file:
conda env create -n plmutils-env -f envs/dev.yml
conda activate plmutils-envInstall the plmutils package in editable mode:
pip install -e .Finally, check that pytorch can find the GPU:
python -c "import torch; print(torch.cuda.is_available())"Although a GPU is not required to generate embeddings, it provides a significant speedup (of 10x or more) compared to using the CPU.
The main package defines several generic (i.e., task-nonspecific) commands:
plmutils translate
This command uses orfipy to find putative ORFs in a fasta file of transcripts and translates them to protein sequences. The translated sequences are saved to a new fasta file. The --longest-only option can be used to retain only the longest ORF for each transcript.
plmutils embed
This command uses a pre-trained protein language model to embed a fasta file of protein sequences. Currently only ESM-2 models are supported. The resulting matrix of embeddings is saved as a numpy array in a .npy file. The order of the rows of this matrix corresponds to the order of the sequences in the input fasta file.
plmutils train
This command trains a generic binary classifier given two embedding matrices, one for the positive class and one for the negative class. The trained classifier is optionally saved to a user-specified directory.
plmutils predict
This command generates predictions given an embedding matrix and a pre-trained classifier generated by the train command above.
This repo is prospectively organized as a generic tool for working with PLM embeddings but it was motivated by, and also includes code specific to, the concrete task of predicting whether putative open reading frames (ORFs) are coding or noncoding. This code is confined to the plmutils.tasks.orf_prediction module. It include both the code to construct training and test datasets and the code to train and evaluate classification models. It relies on extant annotated transcriptomes (i.e., datasets of coding and noncoding transcripts) to obtain a set of putative ORFs that are likely to be "real" (from a coding transcript) or "not real" (from a noncoding transcript). The ESM embeddings of these ORFs are then used to train a classifier to predict whether a given putative ORF is likely to be "real" or not on the basis of its ESM embedding. Refer to this jupyter notebook for more details.
We use ruff for formatting and linting; use make format and make lint to run formatting and linting locally. The CLIs are written using click. There is a single conda environment that defines all direct dependencies of the plmutils package.
We use pytest for testing. All tests and test data are in the tests/ subdirectory. There are currently very limited automated tests. To run the tests, use make test from the root directory of the repo.
There is also a test dataset that can be used to manually test the ORF-prediction task. See here for more details.
This repo was initially developed as a tool to identify sORFs for the peptigate pipeline. This effort was motivated by the fact that existing tools for identifying sORFs are limited, poorly maintained, and generally operate at the transcript rather than ORF level (i.e., they predict whether a given transcript is coding or noncoding, rather than whether a given putative ORF represents a real protein/peptide or not). However, given the versatility and generality of embeddings produced by large PLMs like ESM-2, we decided to prospectively organize this repo in a more generic way to facilitate its use for future tasks that can be addressed with embeddings of protein sequences.