ASCARIS / code /process_input.py
fatmacankara's picture
Update code/process_input.py
7d03af3
raw
history blame
2.55 kB
import pandas as pd
import streamlit as st
def clean_data(input_set):
data = pd.DataFrame()
try:
if ',' in input_set:
input_set = [i.strip() for i in input_set.split(',')]
initial_len = len(input_set)
for i in input_set:
data = data.append(pd.Series([j.strip() for j in i.split('-')]), ignore_index=True)
data.columns = ['uniprotID', 'wt', 'pos', 'mut']
data = data[((~data.uniprotID.isna()) & (~data.wt.isna()) & (~data.pos.isna()) & (~data.mut.isna()))]
if initial_len != len(data):
st.write(f'{initial_len- len(data)} of {initial_len} datapoints is omitted. Check your input.')
elif '\t' in input_set:
input_set = [i.strip() for i in input_set.split('\t')]
initial_len = len(input_set)
for i in input_set:
data = data.append(pd.Series([j.strip() for j in i.split('-')]), ignore_index=True)
data.columns = ['uniprotID', 'wt', 'pos', 'mut']
data = data[((~data.uniprotID.isna()) & (~data.wt.isna()) & (~data.pos.isna()) & (~data.mut.isna()))]
if initial_len != len(data):
st.write(f'{initial_len- len(data)} of {initial_len} datapoints is omitted. Check your input.')
elif '-' in input_set:
data = data.append(pd.Series([j.strip() for j in input_set.split('-')]), ignore_index=True)
data.columns = ['uniprotID', 'wt', 'pos', 'mut']
elif '.txt' in input_set:
data = pd.read_csv(input_set, sep='\t', names=['uniprotID', 'wt', 'pos', 'mut'])
else:
data = pd.DataFrame(columns = ['uniprotID', 'wt', 'pos', 'mut'])
data = data[['uniprotID', 'wt', 'pos', 'mut']]
# Exclude termination codons, synonymous mutations and any non-standard residues such as Sec, 4 or 6.
aa_list = ['A', 'R', 'N', 'D', 'C', 'E', 'Q', 'G', 'H', 'I', 'L', 'K', 'M', 'F', 'P', 'S', 'T', 'W', 'Y', 'V']
data.wt = data.wt.str.strip()
data.mut = data.mut.str.strip()
data = data[data.wt.isin(aa_list)]
data = data[data.mut.isin(aa_list)]
for i in data.index:
data.at[i, 'datapoint'] = data.at[i, 'uniprotID'] + data.at[i, 'wt'] + str(data.at[i, 'pos']) + data.at[i, 'mut']
data = data.astype(str)
return data
except ValueError:
st.write('Your input is in the wrong format. Please see the example.')
return pd.DataFrame()