Source code for tka.external_models.moshkov

import importlib
import os
from typing import List, Union

import chemprop
import numpy as np
import pandas as pd

from tka.utils import (is_valid_smiles, load_l1000_ordered_feature_columns,
                       load_mobc_ordered_feature_columns,
                       prepare_df_for_mobc_predictions,
                       transform_moshkov_outputs)


[docs]def load_assay_metadata() -> pd.DataFrame: """Loads assay metadata of the assays used by Moshkov et al.""" with importlib.resources.path("tka.data", "assay_metadata.csv") as file_path: return pd.read_csv(file_path)
[docs]def predict_from_smiles( smiles_list: List[str], checkpoint_dir: str, auc_modality_filter: dict = {} ) -> pd.DataFrame: """ Make predictions from a list of SMILES strings using a trained checkpoint. Args: smiles_list (List[str]): List of SMILES strings for which to make predictions. checkpoint_dir (str): Directory containing the trained checkpoint. auc_modality_filter (dict, optional): If supplied, assays whose prediction accuracies are lower than auc at a given modality, will be dropped. The dict has three keys - **'auc'** and **'modality'**, **'version'**. Allowed modalities are in the following list: ['late_fusion_cs_ge', 'late_fusion_cs_ge_mobc', 'late_fusion_cs_mobc', 'late_fusion_ge_mobc', 'cpcl_es_op', 'cp_es_op', 'ges_es_op', 'ge_cp_es_op', 'ge_es_op', 'ge_mobc_cp_es_op', 'ge_mobc_es_op', 'ge_mo_cp_es_op', 'ge_mo_es_op', 'mobc_cp_es_op', 'mobc_es_op', 'mo_cp_es_op', 'mo_es_op'] Allowed auc thresholds are any floating point values between 0.5 and 1.0. Allowed versions are 'v3' and 'v4'. Returns: pd.DataFrame: Predictions with SMILES as indices and assays as columns. Examples: >>> predict_from_smiles( ... smiles_list=["CCC", "CCCC", "CH4"], ... checkpoint_dir=".../Moshkov(etal)-single-models/2021-02-cp-es-op" ... ) smiles AmyloidFormation.Absorb.AB42_1_1 ... HoxA13DNABinding.FluorOligo.HoxDNA_93_259 CCC 0.000082 ... 0.442998 CCCC 0.000082 ... 0.442998 CH4 Invalid SMILES ... Invalid SMILES (3, 270) """ arguments = [ "--test_path", "/dev/null", "--preds_path", "/dev/null", "--checkpoint_dir", checkpoint_dir, "--no_features_scaling", ] args = chemprop.args.PredictArgs().parse_args(arguments) preds = chemprop.train.make_predictions(args=args, smiles=smiles_list) return transform_moshkov_outputs( identifier_col_vals=smiles_list, output=preds, use_full_assay_names=True, auc_modality_filter=auc_modality_filter, )
[docs]def predict_from_mobc( df_real: pd.DataFrame, checkpoint_dir: str, mobc_features_version: str, auc_modality_filter: dict = {}, ) -> pd.DataFrame: """ Make predictions from a dataframe of batch effect corrected morphology profiles from CellProfiler and a trained model checkpoint. Args: df_real (pd.DataFrame): a pd.DataFrame with the columns being CellProfiler features (1746 features) and the index column being the identification column checkpoint_dir (str): Directory containing the trained checkpoint. mobc_features_version (str): One of ["v4", "v3"] auc_modality_filter (dict, optional): If supplied, assays whose prediction accuracies are lower than auc at a given modality, will be dropped. The dict has three keys - **'auc'** and **'modality'**, **'version'**. Allowed modalities are in the following list: ['late_fusion_cs_ge', 'late_fusion_cs_ge_mobc', 'late_fusion_cs_mobc', 'late_fusion_ge_mobc', 'cpcl_es_op', 'cp_es_op', 'ges_es_op', 'ge_cp_es_op', 'ge_es_op', 'ge_mobc_cp_es_op', 'ge_mobc_es_op', 'ge_mo_cp_es_op', 'ge_mo_es_op', 'mobc_cp_es_op', 'mobc_es_op', 'mo_cp_es_op', 'mo_es_op'] Allowed auc thresholds are any floating point values between 0.5 and 1.0. Allowed versions are 'v3' and 'v4'. Returns: pd.DataFrame: Predictions with df_real's first column as indices and assays as columns. Examples: In the following code, identifier_col remains to the only data left besides CellProfiler features. Also, sphering normalization is used to modify df_real and this is why df_dmso is required. >>> import pandas as pd >>> from tka.utils import prepare_df_for_mobc_predictions >>> # Load dataset for prediction >>> df = pd.read_csv("path/to/dataset.csv") >>> df_dmso = df_filter.loc[df_filter["Metadata_broad_sample"] == "DMSO"] >>> df_real = df_filter.loc[df_filter["Metadata_broad_sample"] != "DMSO"] >>> out_df = prepare_df_for_mobc_predictions( ... df_dmso=df_dmso, df_real=df_real, identifier_col="Metadata_pert_id" ... ) >>> predict_from_mobc( ... df_real = out_df, ... checkpoint_dir = ".../2023_Moshkov_NatComm/models/2023-01-mobc-es-op", ... mobc_features_version = "v4", ... auc_modality_filter = { ... "auc": 0.9, ... "modality": "mobc_es_op", ... "version": "v4" ... } ... ) smiles AmyloidFormation.Absorb.AB42_1_1 ... HoxA13DNABinding.FluorOligo.HoxDNA_93_259 BRD-K18619710 0.000000e+00 ... 0.000000e+00 BRD-K20742498 3.456357e-10 ... 1.632998e-03 ... ... ... ... Shape: (X, 270) """ # Check if identifier column has valid SMILES values. if not all([is_valid_smiles(x) for x in df_real.index]): smiles_list = ["CCCC" for _ in range(len(df_real))] else: smiles_list = list(df_real.index) # The following lines of code are to adhere to ChemProp's parameter format rules with open("tmp_smiles.csv", "w") as file: for item in ["smiles"] + smiles_list: file.write(item + "\n") # Load the MOBC ordered features to generate .npz file mobc_features = load_mobc_ordered_feature_columns(version=mobc_features_version) # Save the pd.DataFrame features in numpy so that you can load it from a path np.savez("out.npz", features=df_real[mobc_features].to_numpy()) arguments = [ "--test_path", "tmp_smiles.csv", "--preds_path", "/dev/null", "--checkpoint_dir", checkpoint_dir, "--features_path", "out.npz", "--no_features_scaling", ] args = chemprop.args.PredictArgs().parse_args(arguments) preds = chemprop.train.make_predictions(args=args) # Remove temporary files os.remove("out.npz") os.remove("tmp_smiles.csv") return transform_moshkov_outputs( identifier_col_vals=list(df_real.index), output=preds, use_full_assay_names=True, auc_modality_filter=auc_modality_filter, )
[docs]def predict_from_ge( df: List[str], gene_id: str, checkpoint_dir: str, auc_modality_filter: dict = {} ) -> pd.DataFrame: """ Make predictions from a pd.DataFrame of standard scaled gene expressions and a trained model checkpoint. Args: df (pd.DataFrame): a pd.DataFrame with the columns being L1000 features (977 features) and the index column being the identification column gene_id (str): type of identifier present in the header row - one of "affyID", "entrezID" or "ensemblID" checkpoint_dir (str): Directory containing the trained checkpoint. auc_modality_filter (dict, optional): If supplied, assays whose prediction accuracies are lower than auc at a given modality, will be dropped. The dict has three keys - **'auc'** and **'modality'**, **'version'**. Allowed modalities are in the following list: ['late_fusion_cs_ge', 'late_fusion_cs_ge_mobc', 'late_fusion_cs_mobc', 'late_fusion_ge_mobc', 'cpcl_es_op', 'cp_es_op', 'ges_es_op', 'ge_cp_es_op', 'ge_es_op', 'ge_mobc_cp_es_op', 'ge_mobc_es_op', 'ge_mo_cp_es_op', 'ge_mo_es_op', 'mobc_cp_es_op', 'mobc_es_op', 'mo_cp_es_op', 'mo_es_op'] Allowed auc thresholds are any floating point values between 0.5 and 1.0. Allowed versions are 'v3' and 'v4'. Returns: pd.DataFrame: Predictions with df's first column as indices and assays as columns. Examples: >>> df ENSG00000132423 ENSG00000182158 ENSG00000122873 ENSG00000213585 ... 0 -0.559783 1.127299 0.767661 -0.103637 ... 1 1.055605 -0.131212 0.170593 0.485176 ... ... ... ... ... ... ... (10, 977) # Assuming df is a pd.Dataframe with shape (X, 977) # and the columns are either ensembl, entrez or affyIDs. >>> predict_from_ge( ... df=df, ... gene_id="ensemblID", ... checkpoint_dir=".../Moshkov(etal)-single-models/2021-02-mobc-es-op" ... ) smiles AmyloidFormation.Absorb.AB42_1_1 ... HoxA13DNABinding.FluorOligo.HoxDNA_93_259 ... 0 0.013138 ... 0.207173 ... 1 0.064487 ... 0.389113 ... ... ... ... ... ... (10, 270) """ # Generate and save a dummy smiles CSV file to comply with chemprop_predict # Serves no real purpose and does not affect the final predictions in any way dummy_smiles = ["CCCC" for _ in range(len(df))] with open("tmp_smiles.csv", "w") as file: for item in ["smiles"] + dummy_smiles: file.write(item + "\n") valid_gene_ids = ["affyID", "entrezID", "ensemblID"] if gene_id not in valid_gene_ids: raise ValueError( f"Invalid gene_id argument -> ({gene_id}). Should be one of {valid_gene_ids}." ) # Load the MOBC ordered features to generate .npz file l1000_features = load_l1000_ordered_feature_columns(gene_id) # Save the pd.DataFrame so that you can load it from a path np.savez("out.npz", features=df[l1000_features].to_numpy()) arguments = [ "--test_path", "tmp_smiles.csv", "--preds_path", "/dev/null", "--checkpoint_dir", checkpoint_dir, "--features_path", "out.npz", "--no_features_scaling", ] args = chemprop.args.PredictArgs().parse_args(arguments) preds = chemprop.train.make_predictions(args=args) # Remove temporary files os.remove("out.npz") os.remove("tmp_smiles.csv") return transform_moshkov_outputs( identifier_col_vals=list(df.index), output=preds, use_full_assay_names=True, auc_modality_filter=auc_modality_filter, )
if __name__ == "__main__": # predict_from_smiles( # smiles_list=["CCC"], # checkpoint_dir="/home/filip/Downloads/Moshkov(etal)-single-models/2021-02-cp-es-op" # ) common_path = "/home/filip/Documents/TKA/2023_Moshkov_NatComm/analysis/" df_real = pd.read_csv(common_path + "real.csv") df_real = df_real.iloc[:10, :] df_dmso = pd.read_csv(common_path + "dmso.csv") out_df = prepare_df_for_mobc_predictions( df_dmso=df_dmso, df_real=df_real, identifier_col="Metadata_pert_id" ) out = predict_from_mobc( df_real=out_df, checkpoint_dir="/home/filip/Downloads/2021-02-mobc-es-op-20231030T082254Z-001/2021-02-mobc-es-op", auc_modality_filter={"auc": 0.8, "modality": "mobc_es_op"}, ) print(out)