FusOn-pLM / fuson_plm /benchmarking /caid /process_fusion_structures.py
svincoff's picture
caid benchmark
bae913a
# Process fusion structures and the structures of their head and tail proteins for pLDDTs
import requests
import json
import pandas as pd
import numpy as np
import requests
import re
import os
import shutil
from Bio.PDB import MMCIFParser
import Bio.PDB as PDB
from Bio import pairwise2
from Bio.pairwise2 import format_alignment
from bs4 import BeautifulSoup
import pdb
from fuson_plm.utils.logging import log_update, open_logfile
#@markdown Define AlphaFoldStructure class
class AlphaFoldStructure:
'''
This class processes an mmCIF file, either uploaded or downloaded from the AlphaFold2 database, to provide comprehensive information.
'''
def __init__(self, fold_path=None, uniprot_to_download=None, uniprot_output_dir= None, secondary_structure_types=None):
# If the user provided a PDB path, convert their file to mmcif. Isolate the suffix
if fold_path is not None:
fold_fname = fold_path.split('/')[-1]
prefix, suffix = fold_fname.split('.')
if suffix == 'pdb': # convert to cif
# make a directory for converted cif files
conversion_path = 'mmcif_converted_files'
if not(os.path.exists(conversion_path)):
os.makedirs(conversion_path)
fold_path = self.__convert_pdb_to_mmcif__(fold_path, f'{conversion_path}/{prefix}.cif')
self.file_path = fold_path
# If user provided a uniprot ID to download, download it and save it as the file path so it can be processed
if uniprot_to_download is not None:
if fold_path is not None:
log_update("WARNING: both a fold_path and a uniprot_to_download were provided. Running default: downloading the CIF file for provided UniProt ID.")
self.file_path = self.__download_mmCIF(uniprot_to_download, output_path=uniprot_output_dir)
# Either they provide acceptable secondary structure types, or query the internet for them
if secondary_structure_types is None:
self.secondary_structure_types = self.__pull_secondary_structure_types()
else:
self.secondary_structure_types = secondary_structure_types
# If there's a CIF file, initialize the object
if self.file_path:
self.cif_lines = self.__parse_cif()
self.secondary_structures = self.__extract_secondary_structures()
self.structure_dict = self.__calc_pLDDTs()
self.sequence = self.structure_dict['seq']
self.plddts = self.structure_dict['res_pLDDTs']
self.avg_pLDDT = self.structure_dict['avg_pLDDT']
self.residues_df = self.__create_residues_summary_dataframe()
self.secondary_structures_df = self.__create_secondary_structures_summary_dataframe()
# Otherwise, print an error.
else:
log_update("ERROR: structure could not be created. No CIF file found.")
def __convert_pdb_to_mmcif__(self, pdb_filename, mmcif_filename):
parser = PDB.PDBParser()
structure = parser.get_structure('structure', pdb_filename)
io = PDB.MMCIFIO()
io.set_structure(structure)
io.save(mmcif_filename)
return mmcif_filename
def __download_mmCIF(self, uniprot_id, output_path=None):
'''
Download mmCIF file with provided uniprot_id and optional output_path for the downloaded file.
Return: path to downloaded file if successful, None otherwise
'''
full_file_name = f"AF-{uniprot_id}-F1-model_v4.cif" # define file name that will be found on the AlphaFold2 database.
# if output path not provided, just save locally under full_file_name
if output_path is None:
output_path = full_file_name
else:
output_path = f"{output_path}/{full_file_name}"
# request the URL for the file
url = f"https://alphafold.ebi.ac.uk/files/{full_file_name}"
response = requests.get(url)
if response.status_code == 200:
with open(output_path, 'wb') as file:
file.write(response.content)
#log_update(f"File downloaded successfully and saved as {output_path}")
else:
log_update(f"Failed to download file. Status code: {response.status_code}")
return None
return output_path
def __pull_secondary_structure_types(self):
'''
Pull a dictionary of secondary structure types and their descriptions from the PDB mmCIF website (necessary for annotating the CIF file)
Only called if the user does not provide such a dictionary themselves.
'''
# request the .html tree from the website with all secondary structure terms
url = "https://mmcif.wwpdb.org/dictionaries/mmcif_pdbx_v50.dic/Items/_struct_conf_type.id.html"
response = requests.get(url)
if response.status_code != 200:
raise Exception("Failed to retrieve mmCIF dictionary")
# Parse the response content
soup = BeautifulSoup(response.content, 'html.parser')
# Debug: Print the soup to understand the structure
# log_update(soup.prettify())
# write the prettified soup to a txt file
#with open('mmcif_dictionary.txt', 'w') as f:
# f.write(soup.prettify())
# Find the h4 header with the class "panel-title" and text "Controlled Vocabulary"
header = soup.find('h4', class_='panel-title')
if header is None or 'Controlled Vocabulary' not in header.text:
raise Exception("Could not find the 'Controlled Vocabulary' header")
# Debug: Print the found header
#log_update(f"Found header: {header}")
# The table should be the next sibling of the header
table = header.find_next('table')
if table is None:
raise Exception("Could not find the table following the 'Controlled Vocabulary' header")
# Debug: Print the found table (only the opening <table> tag)
#log_update(f"Found table (showing header line): {str(table).split('<thead')[0]}")
# Iterate through rows in the table and process each entry
secondary_structure_types = {}
rows = table.find_all('tr')
for row in rows[1:]: # Skip the header row
cols = row.find_all('td')
if len(cols) > 1:
type_id = cols[0].text.strip()
description = cols[1].text.replace('\t', ' ').strip()
# Replace multiple spaces with a single space
description = re.sub(' +', ' ', description)
# If this is a protein secondary structure (the table also contains nucleic acid structures), add it to teh dictionary
if '(protein)' in description:
secondary_structure_types[type_id] = description
return secondary_structure_types
def get_secondary_structure_types(self):
'''
Display secondary structure types
'''
log_update("Secondary Structure Types in mmCIF files:")
for ss_type, description in self.secondary_structure_types.items():
log_update(f"{ss_type}: {description}")
return self.secondary_structure_types
def __parse_cif(self):
'''
Read cif file lines from self.file_path
'''
with open(self.file_path, 'r') as file:
lines = file.readlines()
return lines
def __extract_secondary_structures(self):
'''
Iterate through the lines of the cif files to find each secondary structure.
Returns a tuple for each amino acid that has a secondary structure annotation. Tuple contains:
1. Structure Type (e.g. STRN)
2. Structure ID (e.g. STRN1)
3. Description (e.g. beta strand)
4. Position (e.g. 3)
'''
secondary_structures = []
parsing_secondary_structure = False
# iterate throhugh cif lines
for line in self.cif_lines:
# hone in on the right section of the cif file
if line.startswith("_struct_conf.conf_type_id"):
parsing_secondary_structure = True
continue
# if we're in the right section...
if parsing_secondary_structure:
if line.startswith("#"):
parsing_secondary_structure = False # no longer in the right section
continue
# still in the right section
columns = line.split()
# iterate through columns to find each piece of info we need
if len(columns) >= 7:
sec_struc_type = columns[6]
sec_struc_id = columns[13]
start_res = int(columns[2])
end_res = int(columns[9])
sec_struc_name = self.secondary_structure_types.get(sec_struc_type, 'Unknown')
# make tuple for this position in the sequence
for pos in range(start_res, end_res + 1):
secondary_structures.append((sec_struc_type, sec_struc_id, sec_struc_name, pos))
return secondary_structures
def __calc_pLDDTs(self):
'''
This method iterates through the cif file to return a dictionary with a few key pieces of info:
1. Sequence
2. pLDDTs for each residue
3. Average pLDDT
'''
# define dictionary needed to translate into single-letter AA code
aa_dict = {
"ALA": "A", "CYS": "C", "ASP": "D", "GLU": "E", "PHE": "F",
"GLY": "G", "HIS": "H", "ILE": "I", "LYS": "K", "LEU": "L",
"MET": "M", "ASN": "N", "PRO": "P", "GLN": "Q", "ARG": "R",
"SER": "S", "THR": "T", "VAL": "V", "TRP": "W", "TYR": "Y"
}
parser = MMCIFParser(QUIET=True) # create a parser
data = parser.get_structure("structure", self.file_path) # parse structure
# count models and chains (should be 1 model and 1 chain; don't use this class to parse a complex)
model = data.get_models()
models = list(model)
chains = list(models[0].get_chains())
# iterate through the chains and get amino acid letters and pLDDTs
all_pLDDTs = []
for n in range(len(chains)):
chainname = chr(n + 65) # turn chain number into letter (e.g. 1 --> "A" so we have Chain A instead of Chain 1)
residues = list(chains[n].get_residues()) # extract all residues
seq = ''
pLDDTs = [0] * len(residues) # initialize empty pLDDT array for this chain
# iterate through all residues in this chain
for i in range(len(residues)):
r = residues[i]
# which amino acid is here?
try:
seq += aa_dict[r.get_resname()]
# error if it's not a real amino acid
except KeyError:
log_update('residue name invalid')
break
# look at each atom. Get its pLDDT (bfactor). make sure bfactor for all atoms within one residue are equal.
atoms = list(r.get_atoms())
bfactor = atoms[0].get_bfactor()
for a in range(len(atoms)):
# if not all atoms within an AA have the same pLDDT, error.
if atoms[a].get_bfactor() != bfactor:
break
pLDDTs[i] = bfactor # add pLDDT for this residue to the list.
all_pLDDTs.extend(pLDDTs) # add pLDDTs for this chain to list of all pLDDTs
avg_pLDDT = np.mean(all_pLDDTs) # average pLDDTs across all chains
return_dict = {
'avg_pLDDT': round(avg_pLDDT, 2),
'res_pLDDTs': all_pLDDTs,
'seq': seq
}
return return_dict
def __create_residues_summary_dataframe(self):
'''
Create a dataframe that summarizes the secondary structure information for each residue.
Columns:
1. Position: amino acid position (e.g. 3)
2. Residue: amino acid 1-letter code (e.g. A)
3. pLDDT: alphafold2's pLDDT score for this residue to 2 decimal places (e.g. 77.54)
4. Structure Type: type of secondary structure (e.g. STRN)
5. Structure ID: ID of this secondary structure (e.g. STRN1)
5. Description: description of this secondary structure (e.g. beta strand)
6. Disordered: is this residue disordered or not? A residue is not disordered if it's in a HELX or STRN. (True/False)
'''
# Convert the secondary structures to a dataframe
df_secondary_structures = pd.DataFrame(self.secondary_structures, columns=['Structure Type', 'Structure ID', 'Description', 'Position'])
# Add Residue and pLDDT columns to the dataframe
df_temp = pd.DataFrame(
data={
'Position': list(range(1, len(self.sequence) + 1)),
'Residue': list(self.sequence),
'pLDDT': self.plddts
})
df_secondary_structures = pd.merge(df_secondary_structures, df_temp, on='Position', how='right')
# Determine if each residue is disordered or not based on what Structure Type it's in. If helix or strand, it's ordered. If anything else or NaN, it's disordered.
df_secondary_structures['Disordered'] = df_secondary_structures['Structure Type'].apply(
lambda x: False if (type(x)==str and (('HELX' in x) or ('STRN' in x))) else True
)
return df_secondary_structures
def __create_secondary_structures_summary_dataframe(self):
'''
Create a dataframe grouped by each Structure ID, providing a summary of each secondary structure in the chain.
Columns:
1. Structure ID: ID of this secondary structure (e.g. STRN1)
2. Start: start position of this secondary structure (e.g. 3)
3. End: end position of this secondary structure (e.g. 12)
4. Start Residue: amino acid 1-letter code of the start position (e.g. A)
5. End Residue: amino acid 1-letter code of the end position (e.g. L)
6. Disordered: is this residue disordered or not? A residue is not disordered if it's in a HELX or STRN. (True/False)
7. Description: description of this secondary structure (e.g. beta strand)
8. Structure Type: type of secondary structure (e.g. STRN)
9. avg_pLDDT: average pLDDT for this secondary structure (e.g. 77.54)
'''
# Apply groupby on self.residues_df to reorganize it by Structure ID
secondary_structures_df = self.residues_df.groupby('Structure ID').agg({
'Position': ['first', 'last'],
'Residue': ['first','last'],
'Disordered': 'first',
'Description': 'first',
'Structure Type': 'first',
'pLDDT': 'mean'
}).reset_index()
# Flatten the multi-level columns
secondary_structures_df.columns = ['Structure ID', 'Start', 'End', 'Start Residue', 'End Residue', 'Disordered', 'Description', 'Structure Type', 'avg_pLDDT']
secondary_structures_df['avg_pLDDT'] = secondary_structures_df['avg_pLDDT'].round(2)
# Display the summarized DataFrame
return secondary_structures_df
def get_residues_df(self):
return self.residues_df
def get_secondary_structures_df(self):
return self.secondary_structures_df
def get_full_sequence(self):
return ''.join([res for res in self.residues_df['Residue']])
def get_average_plddt(self):
plddt_values = [plddt for plddt in self.residues_df['pLDDT'] if plddt is not None]
return sum(plddt_values) / len(plddt_values) if plddt_values else None
def pull_secondary_structure_types():
url = "https://mmcif.wwpdb.org/dictionaries/mmcif_pdbx_v50.dic/Items/_struct_conf_type.id.html"
response = requests.get(url)
if response.status_code != 200:
raise Exception("Failed to retrieve mmCIF dictionary")
soup = BeautifulSoup(response.content, 'html.parser')
# Debug: Print the soup to understand the structure
# log_update(soup.prettify())
# write the prettified soup to a txt file
with open('mmcif_dictionary.txt', 'w') as f:
f.write(soup.prettify())
# Find the h4 header with the class "panel-title" and text "Controlled Vocabulary"
header = soup.find('h4', class_='panel-title')
if header is None or 'Controlled Vocabulary' not in header.text:
raise Exception("Could not find the 'Controlled Vocabulary' header")
# Debug: Print the found header
#log_update(f"Found header: {header}")
# The table should be the next sibling of the header
table = header.find_next('table')
if table is None:
raise Exception("Could not find the table following the 'Controlled Vocabulary' header")
# Debug: Print the found table (only the opening <table> tag)
#log_update(f"Found table (showing header line): {str(table).split('<thead')[0]}")
secondary_structure_types = {}
rows = table.find_all('tr')
for row in rows[1:]: # Skip the header row
cols = row.find_all('td')
if len(cols) > 1:
type_id = cols[0].text.strip()
description = cols[1].text.replace('\t', ' ').strip()
# Replace multiple spaces with a single space
description = re.sub(' +', ' ', description)
if '(protein)' in description:
secondary_structure_types[type_id] = description
return secondary_structure_types
# Process structures downloaded from FusionPDB
def process_fusionpdb_fusion_files(files, level_2_3_structure_info, folder, save_path=None):
# get secondary structure types so we can process PDBs
secondary_structure_types = pull_secondary_structure_types()
# Initialize 3 columns to store structural info - the AA seq in the fold (should match), the Avg pLDDT, and the per-residue pLDDTs (comma-separated, 2 decimal pts.)
level_2_3_structure_info['Fold AA seq'] = ['']*len(level_2_3_structure_info)
level_2_3_structure_info['Avg pLDDT'] = [0]*len(level_2_3_structure_info)
level_2_3_structure_info['pLDDTs'] = ['']*len(level_2_3_structure_info)
# pre-loop processed
pre_loop_processed = []
if os.path.exists(save_path):
pre_loop_processed = pd.read_csv(save_path)
pre_loop_processed = pre_loop_processed['Structure Link'].tolist()
pre_loop_processed = [x.split('/')[-1] for x in pre_loop_processed]
log_update(f"Total structures already processed: {len(pre_loop_processed)}")
log_update("\nProcessing fusion structures...")
# only process structures we haven't processed yet
for i, structure in enumerate(files):
log_update(f'\tProcessing #{i+1}: {structure}')
# make sure we haven't already processed it and aren't wasting time
if structure in pre_loop_processed:
log_update(f"\t\tAlready processed. Continuing...")
continue
# create AlphaFoldStructure object
obj = AlphaFoldStructure(fold_path=f'{folder}/{structure}', secondary_structure_types=secondary_structure_types)
aa_seq = obj.get_full_sequence()
avg_plddt = obj.get_average_plddt()
residues_df = obj.get_residues_df()
all_plddts = ",".join(residues_df['pLDDT'].astype(str).tolist())
log_update(f"\t\tAvg pLDDT: {round(avg_plddt,2)}\tFold AA seq: {aa_seq}\tFirst 5 pLDDTs: {','.join(all_plddts.split(',')[0:5])}")
level_2_3_structure_info.loc[level_2_3_structure_info['Structure Link'].str.contains(f"/{structure}"), 'Fold AA seq'] = aa_seq
level_2_3_structure_info.loc[level_2_3_structure_info['Structure Link'].str.contains(f"/{structure}"), 'Avg pLDDT'] = avg_plddt
level_2_3_structure_info.loc[level_2_3_structure_info['Structure Link'].str.contains(f"/{structure}"), 'pLDDTs'] = all_plddts
# write level_2_3_structure_info to csv
cur_df = level_2_3_structure_info.loc[level_2_3_structure_info['Structure Link'].str.contains(f"/{structure}")].reset_index(drop=True)
if os.path.exists(save_path):
cur_df.to_csv(save_path,mode='a',header=False,index=False)
else:
cur_df.to_csv(save_path,index=False)
# now reload the completed dataframe
level_2_3_structure_info = pd.read_csv(save_path)
return level_2_3_structure_info
def process_fusionpdb_head_tail_files(ht, save_path='heads_and_tails_structures_processed.csv'):
# ht is a list of head and tail proteins we have to process.
log_update("\nProcessing head and tail structures...")
# get secondary structure types so we can process PDBs
secondary_structure_types = pull_secondary_structure_types()
# make directory to save alphafold DB structures of heads and tails
os.makedirs('raw_data/fusionpdb/head_tail_af2db_structures',exist_ok=True)
# pre-loop processed
pre_loop_processed = []
if os.path.exists(save_path):
pre_loop_processed = pd.read_csv(save_path)
pre_loop_processed = pre_loop_processed['UniProtID'].tolist()
log_update(f"Heads and tails already processed: {len(pre_loop_processed)}")
ht_structures_df = pd.DataFrame(
data = {
'UniProtID': ['']*len(ht),
'Avg pLDDT': ['']*len(ht),
'All pLDDTs': ['']*len(ht),
'Seq': ['']*len(ht)
}
)
for i, uniprotid in enumerate(ht):
log_update(f'\tProcessing #{i+1}: {uniprotid}')
aa_seq, avg_plddt, all_plddts = None, None, None
# make sure we haven't processed it yet!
if uniprotid in pre_loop_processed:
log_update(f"\t\tAlready processed. Continuing")
continue
try:
obj = AlphaFoldStructure(uniprot_to_download=uniprotid, secondary_structure_types=secondary_structure_types,
uniprot_output_dir='raw_data/fusionpdb/head_tail_af2db_structures')
aa_seq = obj.get_full_sequence()
avg_plddt = obj.get_average_plddt()
residues_df = obj.get_residues_df()
all_plddts = ",".join(residues_df['pLDDT'].astype(str).tolist())
log_update(f"\t\tAvg pLDDT: {round(avg_plddt,2)}\tFold AA seq: {aa_seq}\tFirst 5 pLDDTs: {','.join(all_plddts.split(',')[0:5])}")
except:
log_update(f"\t\tAvg pLDDT: {None}\tFold AA seq: {None}\tFirst 5 pLDDTs: {None}")
# Fill in info for combined ht df
ht_structures_df.loc[i, 'UniProtID'] = uniprotid
ht_structures_df.loc[i, 'Avg pLDDT'] = avg_plddt
ht_structures_df.loc[i, 'All pLDDTs'] = all_plddts
ht_structures_df.loc[i, 'Seq'] = aa_seq
# write level_2_3_structure_info to csv
cur_df = pd.DataFrame(ht_structures_df.iloc[i,:]).T.reset_index(drop=True)
if os.path.exists(save_path):
cur_df.to_csv(save_path,mode='a',header=False,index=False)
else:
cur_df.to_csv(save_path,index=False)
# ensure we got everything
ht_structures_df = pd.read_csv(save_path)
level_2_3 = pd.read_csv(f'processed_data/fusionpdb/intermediates/giant_level2-3_fusion_protein_head_tail_info.csv')
level_2_3['FusionGene'] = level_2_3['FusionGene'].str.replace('-','::')
heads = level_2_3['HGUniProtAcc'].tolist()
tails = level_2_3['TGUniProtAcc'].tolist()
ht = heads + tails
ht = set([x for x in ht if type(x)==str])
ht = set(','.join(ht).split(','))
log_update(f"total heads and tails: {len(ht)}")
log_update(f"total processed: {len(ht_structures_df)}\t{len(ht_structures_df['UniProtID'].unique())}")
# which ones are missing?
missing = set(ht) - set(ht_structures_df['UniProtID'].unique())
log_update(f"missing: {len(missing)}")
log_update(missing)
# Some heads and tails are not in the alphxwafold database. I folded these myself, externally.
ht_structures_df = ht_structures_df.replace('',np.nan)
need_to_fold = ht_structures_df[ht_structures_df['Avg pLDDT'].isna()]['UniProtID'].tolist()
with open('processed_data/fusionpdb/intermediates/uniprotids_not_in_afdb.txt','w') as f:
for uniprotid in need_to_fold:
f.write(f'{uniprotid}\n')
idmap = pd.read_csv(f'raw_data/fusionpdb/not_in_afdb_idmap.txt',sep='\t')
idmap = idmap[idmap['Entry'].isin(need_to_fold)].reset_index(drop=True)
idmap = idmap[['Entry','Sequence']].rename(columns={
'Entry': 'ID'})
idmap['Length'] = idmap['Sequence'].apply(len)
log_update("Investigating heads and tails that were not in the AF2 database:")
log_update(f"\tMin length: {min(idmap['Length'])}")
log_update(f"\tMax length: {max(idmap['Length'])}")
idmap = idmap.sort_values(by='Length',ascending=True).reset_index(drop=True)
# Q9NNW7
id='Q9NNW7'
if id in idmap['ID'].tolist():
ht_structures_df.loc[
ht_structures_df['UniProtID']=='Q9NNW7', 'Avg pLDDT'
] = 91.68
ht_structures_df.loc[
ht_structures_df['UniProtID']=='Q9NNW7', 'Seq'
] = idmap.loc[
idmap['ID']=='Q9NNW7', 'Sequence'
].item()
## Q16881
id='Q16881'
if id in idmap['ID'].tolist():
ht_structures_df.loc[
ht_structures_df['UniProtID']==id, 'Avg pLDDT'
] = 89.55
ht_structures_df.loc[
ht_structures_df['UniProtID']==id, 'Seq'
] = idmap.loc[
idmap['ID']==id, 'Sequence'
].item()
# Q86V15
id='Q86V15'
if id in idmap['ID'].tolist():
ht_structures_df.loc[
ht_structures_df['UniProtID']==id, 'Avg pLDDT'
] = 48.14
ht_structures_df.loc[
ht_structures_df['UniProtID']==id, 'Seq'
] = idmap.loc[
idmap['ID']==id, 'Sequence'
].item()
return ht_structures_df
def process_fusions_and_hts():
# Process the structures of fusion proteins downloaded from FusionPDB
level_2_3_structure_info_og = pd.read_csv('processed_data/fusionpdb/intermediates/giant_level2-3_fusion_protein_structure_links.csv')
# figure out which ones we have
folder = 'raw_data/fusionpdb/structures'
# get all the structure files in folder
files = os.listdir(folder)
log_update(f"total pdbs: {len(files)}")
log_update(f"examples: {files[:5]}")
os.makedirs('processed_data/fusionpdb', exist_ok=True)
# process the full fusion pdbs
level_2_3_structure_info = process_fusionpdb_fusion_files(files, level_2_3_structure_info_og, folder, save_path='processed_data/fusionpdb/intermediates/giant_level2-3_fusion_protein_structures_processed.csv')
# process the head and tail pdbs
level_2_3 = pd.read_csv(f'processed_data/fusionpdb/intermediates/giant_level2-3_fusion_protein_head_tail_info.csv')
level_2_3['FusionGene'] = level_2_3['FusionGene'].str.replace('-','::')
# Get the heads and tails, see how many unique proteins we're working with
heads = level_2_3['HGUniProtAcc'].tolist()
tails = level_2_3['TGUniProtAcc'].tolist()
ht = heads + tails
ht = set([x for x in ht if type(x)==str])
ht = set(','.join(ht).split(','))
log_update(f"Unique heads/tails: {len(ht)}")
heads_tails_analyzed = process_fusionpdb_head_tail_files(list(ht), save_path='processed_data/fusionpdb/heads_tails_structural_data.csv')
# In the level_2_3 database, we only have the fusions with documented heads and tails.
level_2 = pd.read_csv(f'raw_data/fusionpdb/FusionPDB_level2_curated_09_05_2024.csv')
level_3 = pd.read_csv(f'raw_data/fusionpdb/FusionPDB_level3_curated_09_05_2024.csv')
joined_23 = pd.concat([level_2,level_3]).reset_index(drop=True)
joined_23['FusionGene'] = joined_23['FusionGene'].str.replace('-','::') # use new notation with head::tail
log_update(f"\nnumber of duplicated fusion gene rows: {len(joined_23[joined_23['FusionGene'].duplicated()])}")
# make the dictionary
fo_gid_dict = dict(zip(joined_23['FusionGene'],joined_23['FusionGID']))
log_update(len(fo_gid_dict))
# let's clean giant level 2 and level 3
# first, drop anyting where Fold AA seq is nan. there is no fold.
level_2_3_structure_info_clean = level_2_3_structure_info.replace('',np.nan) # make sure there are nans where there should be
level_2_3_structure_info_clean = level_2_3_structure_info_clean.dropna(subset=['Fold AA seq']).reset_index(drop=True)
log_update(f"length of processed structure file: {len(level_2_3_structure_info_clean)}")
level_2_3_structure_info_clean['pLDDT'] = level_2_3_structure_info_clean['Avg pLDDT'].round(2)
level_2_3_structure_info_clean = level_2_3_structure_info_clean.drop(columns=['Avg pLDDT'])
level_2_3_structure_info_clean['FusionGene'] = level_2_3_structure_info_clean['FusionGene'].str.replace('-','::')
level_2_3_structure_info_clean['FusionGID'] = level_2_3_structure_info_clean['FusionGene'].apply(lambda x: fo_gid_dict[x])
# now let's use the FusionPDB database we processed as ground truth for sequence, rather than the webpage
log_update("Using FusionPDB as ground truth for sequences...")
raw_download = pd.read_csv('../../data/raw_data/FusionPDB.txt',sep='\t',header=None)
raw_download['FusionGene'] = raw_download[7]+ '::' + raw_download[11]
raw_download = raw_download.rename(columns={18:'Raw Download AA Seq'})
log_update(f"FusionPDB raw download size: {len(raw_download)}")
level_2_3_structure_info_clean_ids = set(level_2_3_structure_info_clean['FusionGene'].tolist())
level_2_3_structure_info_clean_seqs = set(level_2_3_structure_info_clean['Fold AA seq'].tolist())
raw_download_ids = set(raw_download['FusionGene'].tolist())
raw_download_seqs = set(raw_download['Raw Download AA Seq'].tolist())
log_update(f"Number of overlapping gene IDs: {len(level_2_3_structure_info_clean_ids.intersection(raw_download_ids))}")
log_update(f"Number of overlapping sequences: {len(level_2_3_structure_info_clean_seqs.intersection(raw_download_seqs))}")
# attempt a merge on Raw Download AA Seq with both. ofthe ogs
# Merging with the AlphaFold sequence
test_merge_1 = pd.merge(
level_2_3_structure_info_clean.rename(columns={'Fold AA seq': 'Raw Download AA Seq'}),
raw_download,
on=['FusionGene','Raw Download AA Seq'],
how='inner'
)
test_merge_1 = test_merge_1.drop(columns=['AA seq'])
test_merge_1['Seq Source'] = ['AlphaFold,Raw Download']*len(test_merge_1)
log_update(f"Merge on AlphaFold AA Seq and raw Download AA Seq. len={len(test_merge_1)}")
# Merging with the webpage sequence
test_merge_2 = pd.merge(
level_2_3_structure_info_clean.rename(columns={'AA seq': 'Raw Download AA Seq'}),
raw_download,
on=['FusionGene','Raw Download AA Seq'],
how='inner'
)
test_merge_2 = test_merge_2.drop(columns=['Fold AA seq'])
test_merge_2['Seq Source'] = ['Webpage,Raw Download']*len(test_merge_2)
log_update(f"Merge on Webpage AA Seq and Raw Download AA Seq. len={len(test_merge_2)}")
test_merge = pd.concat([test_merge_1,test_merge_2])
test_merge['Len(AA seq)'] = test_merge['Raw Download AA Seq'].apply(lambda x: len(x))
# drop duplicates
test_merge = test_merge.drop_duplicates().reset_index(drop=True)
# for anything that has a CIF, keep the CIF
log_update(f"len test_merge before keeping CIFs over identical PDBs: {len(test_merge)}")
test_merge = test_merge.sort_values(by='Structure Type',ascending=True).reset_index(drop=True).groupby(['Hgene', 'Hchr', 'Hbp', 'Hstrand', 'Tgene', 'Tchr',
'Tbp', 'Tstrand', 'Len(AA seq)', 'FusionGene',
'Level', 'Raw Download AA Seq', 'pLDDT', 'pLDDTs','FusionGID', 'Seq Source']).agg(
{
'Structure Link': 'first',
'Structure Type': 'first'
}
).reset_index()
log_update(f"len after: {len(test_merge)}")
# for anything with multiple seq sources, concatenate them
log_update(f"len test_merge before combining seq sources: {len(test_merge)}")
test_merge = test_merge.groupby(['Structure Link','Hgene', 'Hchr', 'Hbp', 'Hstrand', 'Tgene', 'Tchr',
'Tbp', 'Tstrand', 'Len(AA seq)', 'FusionGene','Structure Type',
'Level', 'Raw Download AA Seq', 'pLDDT', 'pLDDTs', 'FusionGID', ]).agg(
{
'Seq Source': lambda x: ','.join(x)
}
).reset_index()
test_merge['Seq Source'] = test_merge['Seq Source'].apply(lambda x: ','.join(set(x.split(','))))
log_update(f"len after: {len(test_merge)}")
# are there cases of multiple folds for the same sequence? miraculously, yes!
dup_seqs = test_merge[test_merge['Raw Download AA Seq'].duplicated()]['Raw Download AA Seq'].unique().tolist()
# for anything with multiple folds / seq, randomly choose the first one.
log_update(f"len test_merge before randomly choosing first fold when one seq has multiple folds: {len(test_merge)}")
test_merge = test_merge.groupby(['Hgene', 'Hchr', 'Hbp', 'Hstrand', 'Tgene', 'Tchr',
'Tbp', 'Tstrand', 'Len(AA seq)', 'FusionGene',
'Level', 'Raw Download AA Seq', 'FusionGID', ]).agg(
{
'Structure Link': 'first',
'Structure Type': 'first',
'Seq Source': 'first',
'pLDDT': 'first',
'pLDDTs': 'first'
}
).reset_index()
log_update(f"len after: {len(test_merge)}")
# how many columns DO NOT have the right AlphaFold sequences?
source_str = test_merge['Seq Source'].value_counts().reset_index().rename(columns={'index': 'Seq Source','Seq Source': 'count'}).to_string(index=False)
source_str = "\t\t" + source_str.replace("\n","\n\t\t")
log_update(f"Distribution of sequence sources:\n{source_str}")
# dropping anything where AF sequence is wrong. Don't want to use these.
test_merge = test_merge.loc[test_merge['Seq Source'].str.contains('AlphaFold')].reset_index(drop=True)
log_update(f"Dropped rows where AlphaFold sequence was incorrect. New DataFrame length: {len(test_merge)}")
# make sure there's only one FusionGID number for each sequence for each GID
assert len(test_merge[test_merge.duplicated(['FusionGID','Raw Download AA Seq'])])==0
# round pLDDTs
test_merge['pLDDT'] = test_merge['pLDDT'].round(2)
# Finally, select only the columns we want
test_merge_v2 = test_merge[
['FusionGID', 'FusionGene', 'Raw Download AA Seq','Len(AA seq)', 'Hgene', 'Hchr', 'Hbp', 'Hstrand', 'Tgene', 'Tchr', 'Tbp', 'Tstrand',
'Level','Structure Link', 'Structure Type', 'pLDDT', 'pLDDTs', 'Seq Source']
].rename(
columns={
'Raw Download AA Seq': 'Fusion_Seq',
'Seq Source': 'Fusion_Seq_Source',
'Structure Link': 'Fusion_Structure_Link',
'Structure Type': 'Fusion_Structure_Type',
'pLDDT': 'Fusion_pLDDT',
'pLDDTs': 'Fusion_AA_pLDDTs',
'Len(AA seq)': 'Fusion_Length'
}
)
log_update(f"Unique FusionGIDs: {len(test_merge_v2['FusionGID'].unique())}")
log_update(f"Number of structures: {len(test_merge_v2)}")
# Note that test_merge_v2 will still have duplicates where it's same seq, different ID
log_update("\nChecking for duplicate sequences..")
log_update(f"\tThe structure-based fusion database of length {len(test_merge_v2)} has {len(test_merge_v2['Fusion_Seq'].unique())} unique fusion sequences.")
dup_seqs = test_merge_v2[test_merge_v2['Fusion_Seq'].duplicated()]['Fusion_Seq'].tolist()
dup_seqs_df = test_merge_v2.loc[test_merge_v2['Fusion_Seq'].isin(dup_seqs)].reset_index(drop=True)
dup_seqs_df['FusionGID'] = dup_seqs_df['FusionGID'].astype(str)
dup_seqs_df = dup_seqs_df.groupby('Fusion_Seq').agg({
'FusionGID': lambda x: ','.join(x),
'FusionGene': lambda x: ','.join(x)
})
dup_seqs_df_str = dup_seqs_df.to_string(index=False)
dup_seqs_df_str = "\t"+dup_seqs_df_str.replace("\n","\n\t")
log_update(f"\tShowing FUsionGIDs and FusionGenes for duplicated sequences below:\n{dup_seqs_df_str}")
# round pLDDT column
heads_tails_analyzed['Avg pLDDT'] = heads_tails_analyzed['Avg pLDDT'].round(2)
# merge treating data as head data
level_2_3_v2 = pd.merge(
level_2_3,
heads_tails_analyzed.rename(columns={'UniProtID': 'HGUniProtAcc', 'Avg pLDDT': 'HG_pLDDT', 'All pLDDTs': 'HG_AA_pLDDTs', 'Seq': 'HG_Seq'}),
on='HGUniProtAcc',
how='left'
)
# merge treating data as tail data
level_2_3_v2 = pd.merge(
level_2_3_v2,
heads_tails_analyzed.rename(columns={'UniProtID': 'TGUniProtAcc', 'Avg pLDDT': 'TG_pLDDT', 'All pLDDTs': 'TG_AA_pLDDTs', 'Seq': 'TG_Seq'}),
on='TGUniProtAcc',
how='left'
)
# giant_level2_3 with valid structures only, no duplicate sequences, and head and tail proteins' uniprot IDs, pLDDTs, and sequences
test_merge_v2.to_csv(f'processed_data/fusionpdb/FusionPDB_level2-3_cleaned_structure_info.csv',index=False)
log_update("Saved file with all fusion structure pLDDTs to: processed_data/fusionpdb/FusionPDB_level2-3_cleaned_structure_info.csv")
# level_2_3 with the head and tail proteins' uniprot IDs, pLDDTs, and sequences
level_2_3_v2.to_csv(f'processed_data/fusionpdb/FusionPDB_level2-3_cleaned_FusionGID_info.csv',index=False)
log_update("Saved file with all fusion protein heads and tails, and their structure pLDDTs to: processed_data/fusionpdb/FusionPDB_level2-3_cleaned_FusionGID_info.csv")
def main():
with open_logfile("process_fusion_structures_log.txt"):
process_fusions_and_hts()
if __name__ == "__main__":
main()