TLmutation: predicting the effects of mutations using transfer learning
TLmutation leverages deep mutational scanning datasets to predict the functional consequences of mutations in homologous proteins. Find our publication here: https://pubs.acs.org/doi/full/10.1021/acs.jpcb.0c00197
In this example, we will use TLmutation to transfer from chemokine receptors CXCR4 to CCR5.
TLmutation first requires a model to represent the protein sequence. The current iteration of TLmutation uses a Pott's model and is built using the EVcouplings package.
https://github.com/debbiemarkslab/EVcouplings
Below is an example script to execute the EVcoupling pipeline.
import os
os.environ['QT_QPA_PLATFORM']='offscreen'
from evcouplings.utils import read_config_file, write_config_file
from evcouplings.utils import read_config_file
from evcouplings.utils.pipeline import execute
source_protein_config = read_config_file("CXCR4-config.txt")
source_protein_outcfg = execute(**source_protein_config)
target_protein_config = read_config_file("CCR5-config.txt")
target_protein_outcfg = execute(**target_protein_config)
In the current iteration of TLmutation, the supervised transfer is from the evolutionary statistical energy to a functional assay. Here, deep mutational scanning (DMS) is the studied experimental assay.
The DMS dataset should be in a .csv format as shown.
mutant,linear
...
E2L,0.605017852
E2I,-2.179323351
E2V,0.148787296
E2A,-0.222549939
E2S,0.32657855
E2T,-1.829107102
E2N,1.136931418
E2Q,-0.18022787
E2D,-0.929420795
...
G3V,-1.085390056
G3A,-0.943412925
G3S,0.372614715
G3T,-0.321523109
G3N,-1.660989957
G3Q,1.538280976
G3D,-2.056820679
...
Now that we have a model, we will incorporate the experimental DMS dataset into the model of the source protein (CXCR4).
The following script is a part of transfer_mut.py.
from jmhe import jmhe
import pandas as pd
from jmhe import score
import pickle
Couplings_model_source = '/home/zshamsi2/projects/functionalCouplings/CXCR4-CCR5/jmhe/CXCR4/Model.model'
dataAdress_source = "/home/zshamsi2/projects/functionalCouplings/CXCR4-CCR5/jmhe/CXCR4/BiFC/exp.csv"
# Train the source model based on its own DMS experiments
data = pd.read_csv(dataAdress_source, sep=",", comment="#")
model_source = jmhe(Couplings_model=Couplings_model_source)
# Predict from initial model
dx2 = model_source.predict(X=dataAdress_source)
s1 = score(dx2, data_exp=dataAdress_source)
print('Score for CXCR4 before training: ', s1)
# Predict after fitting with experimental data
model_source.sp = s1[0]
if s1[1]>0:
model_source.positiveL = True
else:
model_source.positiveL = False
print(model_source.positiveL)
model_source.fit(data)
pickle.dump(model_source, open('model_cxcr4.pkl', 'wb'))
#model_source = pickle.load(open('model_cxcr4.pkl', 'rb'))
dx = model_source.predict(X=dataAdress_source)
# Calculate the Spearman's correlation score
s2 = score(dx, data_exp=dataAdress_source)
print('Score for CXCR4 after training: ', s2)
After training the model on the DMS data, we observed an increase in the Spearman's correlation, signifying the incorporation of experimental data greatly benifits the model.
Score for CXCR4 before training: ('sp', 0.19921471564043003)
Score for CXCR4 after training: ('sp', 0.51009788834287362)
First and as a comparison, we will evaluate how the model will fair without transfer learning.
# Transfer from source protein to target protein
Couplings_model_target='/home/zshamsi2/projects/functionalCouplings/CXCR4-CCR5/jmhe/CCR5/Model.model'
model_target = jmhe(Couplings_model=Couplings_model_target)
#DMS dataset for the target protein, for evaluating purposes
dataAdress_target = "/home/zshamsi2/projects/functionalCouplings/CXCR4-CCR5/jmhe/CCR5/BiFC/exp.csv"
data_target_target = pd.read_csv(dataAdress_target, sep=",", comment="#")
# To compare, let's see how predicting the target protein without transfer will do
dx2 = model_target.predict(X=dataAdress_target)
# Calculate the Spearman's correlation score without transfer to the actual DMS dataset
s1 = score(dx2, data_exp=dataAdress_target)
print('Score for CCR5 before transfer: ',s1)
Without transfer, we recieve the following Spearmans' coorelation.
Score for CCR5 before transfer: ('sp', 0.32721716144910645)
Let us now see how transfering from the source protein compare. We will now transfer the weights obtained from the source model to the target model (CCR5). This requires a sequence alignment file of the two proteins and be generated from any sequence alignment server or program.
CXCR4-CCR5.aln
MDYQVSSPIYDIN-----------YYTSEPCQKINVKQIAARLLPPLYSLVFIFGFVGNMLVILILINCKRLKSMTDIYLLNLAISDLFFLLTVPFWAHYAAAQWDFGNTMCQLLTGLYFIGFFSGIFFIILLTIDRYLAVVHAVFALKARTVTFGVVTSVITWVVAVFASLPGIIFTRSQKEGLHYTCSSHFPYSQYQFWKNFQTLKIVILGLVLPLLVMVICYSGILKTLLRCRNEKKRHRAVRLIFTIMIVYFLFWAPYNIVLLLNTFQEFF-GLNNCSSSNRLDQAMQVTETLGMTHCCINPIIYAFVGEKFRNYLLVFFQKHIAKRFCKCCSIFQQEAPERASSVYTRSTGEQEISVGL
---MEGISIYTSDNYTEEMGSGDYDSMKEPCFREENANFNKIFLPTIYSIIFLTGIVGNGLVILVMGYQKKLRSMTDKYRLHLSVADLLFVITLPFWAVDAVANWYFGNFLCKAVHVIYTVNLYSSVLILAFISLDRYLAIVHATNSQRPRKLLAEKVVYVGVWIPALLLTIPDFIFANVSEADDRYICDRFYPND---LWVVVFQFQHIMVGLILPGIVILSCYCIIISKLSHSKGHQKR-KALKTTVILILAFFACWLPYYIGISIDSFILLEIIKQGCEFENTVHKWISITEALAFFHCCLNPILYAFLGAKFKTSAQHALTSVSRGS---SLKILSKGKRGGHSSVSTESESSSFHSS--
# Sequence alignment file can be generated from any alignment server or program
algn_file = '/home/zshamsi2/projects/functionalCouplings/CXCR4-CCR5/case_PMEP_jmhe/CXCR4-CCR5.aln'
model_target_transfer = model_source.transfer(Couplings_model_target=Couplings_model_target, algn_file=algn_file)
#pickle.dump(model_target_transfer, open('model_ccr5_transfered.pkl', 'wb'))
dx2_transfer = model_target_transfer.predict(X=dataAdress_target)
# Calculate the Spearman's correlation score after transfer to the actual DMS dataset
s1_transfer = score(dx2_transfer, data_exp=dataAdress_target)
print('Score for CCR5 after transfer from CCR5', s1_transfer)
After transfer, we obtain the Spearman's coorelation of the target protein after transfer and is greater than without the transfer of knowledge from the source protein.
Score for CCR5 after transfer from CCR5 ('sp', 0.34294314055843561)