J-SNACKKB / autoeval

Module to auto evaluate FLIP datasets via bio-trainer

AutoEval

This repository contains AutoEval, a module for a fast and easy evaluation of FLIP benchmarking tasks. It uses biotrainer to train the task-specific models and bio-embeddings or custom embedders to embed proteins.

Its way of working is as simple as

python run-autoeval.py scl_mixed_soft residues_to_class ./results --embedder Rostlab/prot_t5_xl_uniref50

where

• scl_mixed_soft indicates the task and the split to be evaluated,
• residues_to_class the protocol used for the tasks,
• ./results the output directory,
• and --embedder Rostlab/prot_t5_xl_uniref50 the embedder from bio-embeddings to be used

The different options are summarized below.

Installation and running

1. Make sure you have poetry installed:

curl -sSL https://install.python-poetry.org/ | python3 - --version 1.4.2

2. Install dependencies and biotrainer via poetry:

# In the base directory:
poetry install
# Optional: Add bio-embeddings to compute embeddings
poetry install --extras "bio-embeddings"
# You can also install all extras at once
poetry install --all-extras

To run AutoEval:

• with Poetry:

# Option 1:
poetry run autoeval DATASET_SPLIT PROTOCOL WORKING_DIR [...]

# Option 2:
autoeval DATASET_SPLIT PROTOCOL WORKING_DIR [...]

The provieded run-autoeval.py can also be used.

• with Docker:

# Build
docker build -t autoeval .
# Run
docker run --rm \
    -v "$(pwd)/examples/docker":/mnt \
    -v bio_embeddings_weights_cache:/root/.cache/bio_embeddings \
    -u $(id -u ${USER}):$(id -g ${USER}) \
    biotrainer:latest /mnt/config.yml

Options

Parameter	Usage
split	Name of the split, e.g. aav_des_mut. The different options are listed at the end of this file.
protocol	Task-specific training protocol to use from the available ones in biotrainer: residue_to_class, residues_to_class, sequence_to_class and sequence_to_value.
working_dir	Path to the working directory.
-e / --embedder	Embedder to use if different from the one in the default configuration. It can be from the ones available in bio-embeddings, e.g. esm1b; or a custom embedder (see details here).
-f / --embeddingsfile	Path to the file containing precomputed embeddings if available.
-m / --model	Model to use if different fro them one in the default configuration. It should be one from the ones available in biotrainer, e.g. FNN or CNN.
-c / --config	Config file different from the provided one in configsbank for the indicated split.
-mins / --minsize	Only use proteins the given minimum length.
-maxs / --maxsize	Only use proteins the given maximum length.
-mask / --mask	If set, use the masks in the file mask.fasta from the split to filter the residues. It also accepts a path to a different masks file.

Default configurations

For every task, the original configuration is the one used by default (defined in the configsbank folder). A different configuration can be used by changing the input arguments of AutoEval or by copying and changing the given one. The default can be overwritten using --config NEW_CONFIG.yml.

Dataset	Type of task	Recommended pLM Embeddings	Recommended model	Reference	Available in Configsbank
AAV	sequence_to_value	-	FNN	[Dallago 2021]	⚠️
GB1	sequence_to_value	-	FNN	[Dallago 2021]	⚠️
Meltome	sequence_to_value	-	FNN	[Dallago 2021]	⚠️
SCL	residues_to_class	ProtT5 (ProtT5-XL-UniRef50)	LightAttention	[Stärk 2021]	✅
Bind	residue_to_class	ProtT5 (ProtT5-XL-UniRef50)	CNN	[Littmann 2021]	✅
SAV	sequence_to_class	ProtT5 (ProtT5-XL-U50)	FNN	[Marquet 2021]	⚠️
Secondary Structure	residue_to_class	ProtT5 (ProtT5-XL-U50)	CNN	-	✅
Conservation	residue_to_class	ProtT5 (ProtT5-XL-U50)	CNN	[Marquet 2021]	✅

Availability semaphore:

• ✅: Available in configsbank is the closest possible way to the best configuration in the reference.
• ⚠️: The best configuration is not possible due to, e.g., a lack of features in biotrainer. The best possible alternative is the one available.
• ❌: Not available in configsbank. Somecases can be used anyhow under user's responsability.

Available splits

In order to reference the split to be evaluated the pattern dataset_split must be followed. For example, the split seven_vs_many from the dataset aav must be referenced as aav_seven_vs_many.

Dataset	Splits
AAV (aav_*)	des_mut, mut_des, one_vs_many, two_vs_many, seven_vs_many, low_vs_high, sampled
Meltome (meltome_*)	mixed_split, human, human_cell
GB1 (gb1_*)	one_vs_rest, two_vs_rest, three_vs_rest, low_vs_high, sampled
SCL (scl_*)	mixed_soft, mixed_hard, human_soft, human_hard, balanced, mixed_vs_human_2
Bind (bind_*)	one_vs_many, two_vs_many, from_publication, one_vs_sm, one_vs_mn, one_vs_sn
SAV (sav_*)	mixed, human, only_savs
Secondary Structure (secondary_structure_*)	sampled
Conservation (conservation_*)	sampled