samusram/profun

Library of models for Protein Function prediction

ProFun

Library of models for Protein Function prediction

Installation

The majority of dependencies will be installed automatically via the command pip install git+https://github.com/SamusRam/ProFun.git.

If you want to use the BLAST-based model, please run these commands:

wget https://ftp.ncbi.nlm.nih.gov/blast/executables/LATEST/ncbi-blast-2.14.0+-x64-linux.tar.gz
tar zxvpf ncbi-blast-2.14.0+-x64-linux.tar.gz
# add ncbi-blast-2.14.0+/bin to PATH

If you want to use profile Hidden Markov models, please run the following commands:

conda install -c bioconda mafft -y
conda install -c bioconda hmmer -y

If you want to use Foldseek-based model, please run the following command:

conda install -c conda-forge -c bioconda foldseek -y

Basic usage

BLAST

Please see this notebook as a usage demo.

from profun.models import BlastMatching, BlastConfig
from profun.utils.project_info import ExperimentInfo

experiment_info = ExperimentInfo(validation_schema='public_lb', 
                                 model_type='blast', model_version='1nn')

config = BlastConfig(experiment_info=experiment_info, 
                      id_col_name='EntryID', 
                      target_col_name='term', 
                      seq_col_name='Seq', 
                      class_names=list(train_df_long['term'].unique()), 
                      optimize_hyperparams=False, 
                      n_calls_hyperparams_opt=None,
                      hyperparam_dimensions=None,
                      per_class_optimization=None,
                      class_weights=None,
                      n_neighbours=5,
                      e_threshold=0.0001,
                      n_jobs=100,
                      pred_batch_size=10
                    )

blast_model = BlastMatching(config)

# fit
blast_model.fit(train_df_long)

# predict
test_pred_df = blast_model.predict_proba(test_seqs_df.sample(42).drop_duplicates('EntryID'), return_long_df=True)

Profile Hidden Markov model

from profun.models import ProfileHMM, HmmConfig
from profun.utils.project_info import ExperimentInfo

experiment_info = ExperimentInfo(validation_schema='public_lb', 
                                 model_type='profileHMM', model_version='24additional')

config = HmmConfig(experiment_info=experiment_info, 
                     id_col_name='EntryID', 
                     target_col_name='term', 
                     seq_col_name='Seq', 
                     class_names=list(additional_classes), 
                     optimize_hyperparams=False, 
                     n_calls_hyperparams_opt=None,
                     hyperparam_dimensions=None,
                     per_class_optimization=None,
                     class_weights=None,
                     search_e_threshold=0.000001,
                     zero_conf_level=0.00001,
                     group_column_name='taxonomyID',
                     n_jobs=56,
                     pred_batch_size=20000)

hmm_model = ProfileHMM(config)
hmm_model.fit(train_df_long)
test_pred_df = hmm_model.predict_proba(test_seqs_df.drop_duplicates('EntryID'), return_long_df=True)

Foldseek-based classifier

Please see this notebook as a usage demo.

from profun.models import FoldseekMatching, FoldseekConfig
from profun.utils.project_info import ExperimentInfo

experiment_info = ExperimentInfo(validation_schema='public_lb', 
                                 model_type='foldseek', model_version='5nn')

config = FoldseekConfig(experiment_info=experiment_info, 
                        id_col_name='EntryID', 
                        target_col_name='term',
                        seq_col_name='Seq',
                        class_names=list(train_df_long_sample['term'].unique()), 
                        optimize_hyperparams=False, 
                        n_calls_hyperparams_opt=None,
                        hyperparam_dimensions=None,
                        per_class_optimization=None,
                        class_weights=None,
                        n_neighbours=5,
                        e_threshold=0.0001,
                        n_jobs=56,
                        pred_batch_size=10,
                        local_pdb_storage_path=None #then it stores structures into the working dir
                    )

model = FoldseekMatching(config)
model.fit(train_df_long)
test_pred_df = model.predict_proba(test_seqs_df.drop_duplicates('EntryID'), return_long_df=True)