search_protein

A fast protein sequence search tool using embedding-based similarity search. The tool searches for homologs of a query protein through an embedding-based search engine, and as a last steps aligns the results to the query using Mmseqs2 to provide the user sequence alignments.

Installation

Requirements

• Python 3.7+
• PyTorch (with CUDA support for GPU acceleration)
• FAISS
• usearch (the vector database), not the DNA/RNA search/clustering tool)
• MMseqs2
• Transformers
• NumPy
• scikit-learn

Setup

Create a conda environment with all dependencies:

# Create and activate environment
conda create -n search_fasta create  -c pytorch -c nvidia -c conda-forge -c bioconda   python=3.10 mmseqs2 "pytorch>=2.5" pytorch-cuda=12.1   "faiss-cpu>=1.8" numpy scikit-learn transformers einops
conda activate search_fasta
pip install usearch

Then download the repo

# Clone the repository
git clone https://github.com/RolandFaure/search_protein.git

And compile the C++ scripts

cd search_protein
make

Quick Start

Search Database

Search your pre-built database with the queries:

python search_database.py --help
  -h, --help            show this help message and exit
  --query_sequences QUERY_SEQUENCES
                        Fasta file of queries
  --database DATABASE   Path to the folder containing database files.
  --output OUTPUT, -o OUTPUT
                        Path to the output folder (created by embed_query.py if embedding step was run separately)
  --db-type {faiss,usearch}
                        Database type to use: faiss or usearch (default: usearch)
  --outfmt OUTFMT       Format of the mmseqs2 output [0], default is 0 which is a tabular format with header. See mmseqs2 documentation for details.
  -m MEMORY, --memory MEMORY
                        Maximum memory available in GB (mandatory)
  -t NUM_THREADS, --num_threads NUM_THREADS
                        Maximum number of threads available (mandatory)
  --force_cpu           Force the use of CPU even if GPUs are available (for embedding step).
  --version             show program's version number and exit

Output files:

results_folder/
├── PLM_aligned_proteins.tsv    # Set of proteins related to you queries according to the gLM2 protein language model (Logan50 proteins)
├── aligned_proteins.fasta  # Set of proteins aligning on your queries according to mmseqs2
├── aligned_proteins.mmseqs2  # Detail of the mmseqs2 alignments
├── all_proteins.fasta.gz  # All the proteins in Logan corresponding to the PLM aligned proteins in Logan50
└── intermediate_files/
    ├── query_embeddings.npy
    ├── query_embeddings.names.txt
    ├── query_results_intermediate.fasta
    ├── query_results.tsv
    ├── unique_centroids.fasta
    └── matches.top_hit

Main Output Files

• PLM_aligned_proteins.tsv: TSV files containing Logan50 proteins which gLM2 embeddings have a cosine distance of less 0.2 to the embeddings fo the query.

#query_name     result_name     result_sequences        cosine_distance
alpha      ERR11474596_7103_1      MLDWNTSSDIFVEKLLQRNYKSQSLHSQPRHRPQVDGIPYEFGYKGTIYPMNKSRNCIIILLLIPVLVHSTRNAAYFESLEMKIVEQVKLNRAQGKWQLVRELLGLKGTFLKPRWQHFAKTVSSRDFFGNWLPLMLEIERYLYSKKMYPDSYLSWDDHSSYRVRKKVYRRGYDDKGSRRPEHKWFPENAFSRELLDEVPVKRAVYKPFELYHGYSEKRRRRSSLSSLLDL* 0.015888094902038574

• all_proteins.fasta.gz: All Logan proteins found in the Logan50 clusters described in PLM_aligned_proteins.tsv.
• aligned_proteins.mmseqs2: MMseqs2 alignment of proteins aligned against the query (many of the proteins in all_proteins.fasta.gz actually do not align to the query using mmseqs)
• aligned_proteins.fasta: FASTA file containing proteins that align on the query

#target  query  identity        alignment_length        nb_mismatches   nb_gap_openings target_start     target_end       query_start    query_end      evalue  bitscore
SRR21885923_17279_1#87#695#-1   alpha   0.924   202     15      0       1       202     1       202     4.604E-129      393

Troubleshooting

Out of memory

The software automatically adjusts the number of threads used based on the memory available. This optimization is thought for single-sequence queries. This can typically lead to out-of-memory errors if querying several proteins.

In case of out-of-memory errors, decrease -m below the actual amount of available RAM, or decrease the number of available threads.

Problem loading the model

The script connects to the internet to load the gLM2 model the first time it runs. Make sure you have an internet connection.

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE file for details.