Likelihood Fitness Bridging (LFB)

Installation

conda create --name lfb python=3.11 mmseqs2 -c conda-forge -c bioconda
conda activate lfb
pip install .

For running Evo 2 predictions, install according to instructions in the repo and ensure HAL Tools are installed - see the repo or Cactus for binaries.

Running LFB

For protein language models

./scripts/plm/predict_lfb.py

Input:

• --tsv_alignment_file: a tsv file (as produced by mmseqs covertalis), with no header, with at least the fields: query, target, qstart, qend, tstart, tend, cigar, qseq, tseq. These can be produced using ./scripts/plm/produce_alignment.sh. Note the query should be the sequence for scoring.

• --model: a huggingface path to an ESM style model, or a path to a progen model if the flag --progen is provided.

Output:

• --output: a csv file of model predictions across variants of the protein - all possible point mutations unless --variants_table is provided.

Example:

python ./scripts/plm/predict_lfb.py \
--tsv_alignment_file "./data/test_data/BLAT_ECOLX_Stiffler_2015.tsv" \
--output "./data/test_data/test_output.csv" \
--model "facebook/esm2_t6_8M_UR50D" \
--inference_mode "masked-marginals" \
--percentage_identity "0.3" \
--random_subsample_num "10"

Alignment Filtering and Random Subsampling:

• --coverage: sets a minimum coverage threshold, ensuring that only sequences covering a sufficient portion of the query are used.
• --percentage_identity: filters based on the minimum acceptable identity fraction, so only sequences similar enough to the query are considered.
• --random_subsample_num: allows you to randomly select a subset of the filtered alignment (including the query) - almost equivalent performance can be seen with values as low as 10 or 15.

For genomic languague models

See the scripts in ./scripts/glm

• 1_make_bed_from_vcf.py: Create a bed file, usable with hal liftover of the loci of variants in a vcf file.

• 2_liftover_to_zoonomia: Mapping over the bed file of variants to other species in hal alignment using hal liftover.

• 3_get_sequences.py: Extracting sequences centered around variants using lifted-over variants to a fasta file.

• 4_score_evo.py: Score the variants in given fasta file using evo2.

References

MMseqs

https://github.com/soedinglab/MMseqs2

Zoonomia

https://zoonomiaproject.org

Hal toolkit

https://github.com/ComparativeGenomicsToolkit/hal

Evo 2

https://github.com/ArcInstitute/evo2

ProGen2

https://github.com/enijkamp/progen2

ESM-2

https://github.com/facebookresearch/esm

and on huggingface,

https://huggingface.co/facebook/esm2_t6_8M_UR50D

Preprint

The preprint for LFB is available here.