机器学习之DeepSequence软件使用学习3-预测突变效应

import theano
import numpy as np
import sys
import pandas as pd
import scipy
#scipy 模块是 Python 中用于科学计算和数据分析的重要模块之一。它包含了许多高级的数学函数和工具，包括数值积分、优化、线性代数、统计等。
from scipy.stats import spearmanr
#scipy.stats模块是SciPy库中用于统计分析的模块，它提供了许多用于概率分布和统计测试的函数。%matplotlib inline
import matplotlib.pyplot as plt

我们将介绍加载模型和预测突变影响的基本函数。

下载预训练参数。

请首先使用 download_pretrained.sh 脚本下载预训练参数。

加载模型。

sys.path.insert(0, "../DeepSequence")import model
import helper
import train

突变影响预测。

突变影响预测辅助函数始终针对比对中的焦点序列。我们可以单独请求预测突变效应。
为了获得可靠的突变效应预测结果，我们建议从模型中取 Monte Carlo 500-2000 个样本（使用 N_pred_iterations 参数）。
我们可以预测单个、双重、三重突变等的影响。突变以元组列表的形式组织，其中元组为（Uniprot位置，野生型氨基酸，突变氨基酸）。

PABP

首先让我们加载一个模型。我们不需要在这里计算序列权重，因为我们不是在训练模型，而且在 CPU 上进行这项计算可能会很慢。
在 "Explore model parameters.ipynb" 笔记本中，helper.py 代码被修改以预先指定 DataHelper 类使用的数据集。然而，我们可以传入一个比对名称和一些额外参数，这样就不必修改 helper.py 文件。

data_params = {"alignment_file":"datasets/PABP_YEAST_hmmerbit_plmc_n5_m30_f50_t0.2_r115-210_id100_b48.a2m"}pabp_data_helper = helper.DataHelper(alignment_file=data_params["alignment_file"],working_dir=".",#指定当前程序的工作目录为当前所在的目录。calc_weights=False#不计算权重)model_params = {"batch_size"        :   100,"encode_dim_zero"   :   1500,"encode_dim_one"    :   1500,"decode_dim_zero"   :   100,"decode_dim_one"    :   500,"n_patterns"        :   4,"n_latent"          :   30,"logit_p"           :   0.001,"sparsity"          :   "logit","encode_nonlin"     :   "relu","decode_nonlin"     :   "relu","final_decode_nonlin":  "sigmoid","output_bias"       :   True,"final_pwm_scale"   :   True,"conv_pat"          :   True,"d_c_size"          :   40}pabp_vae_model   = model.VariationalAutoencoder(pabp_data_helper,batch_size              =   model_params["batch_size"],encoder_architecture    =   [model_params["encode_dim_zero"],model_params["encode_dim_one"]],decoder_architecture    =   [model_params["decode_dim_zero"],model_params["decode_dim_one"]],n_latent                =   model_params["n_latent"],n_patterns              =   model_params["n_patterns"],convolve_patterns       =   model_params["conv_pat"],conv_decoder_size       =   model_params["d_c_size"],logit_p                 =   model_params["logit_p"],sparsity                =   model_params["sparsity"],encode_nonlinearity_type       =   model_params["encode_nonlin"],decode_nonlinearity_type       =   model_params["decode_nonlin"],final_decode_nonlinearity      =   model_params["final_decode_nonlin"],output_bias             =   model_params["output_bias"],final_pwm_scale         =   model_params["final_pwm_scale"],working_dir             =   ".")print ("Model built")

我们查看一下datasets文件夹下的PABP_YEAST_hmmerbit_plmc_n5_m30_f50_t0.2_r115-210_id100_b48.a2m文件，显示有456123行，每三行为一组数据，表明有152041组数据
可以看到，比对文件是使用的类似于FASTA格式的氨基酸序列。

Encoding sequences
Neff = 151528.0
Data Shape = (151528, 82, 20)
Model built

加载预训练模型在 ‘params’ 文件夹中的参数。

file_prefix = "PABP_YEAST"
pabp_vae_model.load_parameters(file_prefix=file_prefix)
print ("Parameters loaded")

Parameters loaded

print (pabp_data_helper.delta_elbo(pabp_vae_model,[(126,"G","A")], N_pred_iterations=500))

-2.03463650668

print (pabp_data_helper.delta_elbo(pabp_vae_model,[(126,"G","A"), (137,"I","P")], N_pred_iterations=500))

-10.8308351474

注意，这里我单独计算了(137,“I”,“P”)的elbo，计算结果如下

>>> print (pabp_data_helper.delta_elbo(pabp_vae_model,[(137,"I","P")], N_pred_iterations=500))
-9.21612828741

可以看到，计算的结果不是每个突变的简单叠加。

print (pabp_data_helper.delta_elbo(pabp_vae_model,[(126,"G","A"), (137,"I","P"), (155,"S","A")], N_pred_iterations=500))

-16.058655309

这里我将同样的程序跑了三遍，但结果有略微出入。

>>> print (pabp_data_helper.delta_elbo(pabp_vae_model,[(126,"G","A"), (137,"I","P"), (155,"S","A")], N_pred_iterations=500))
-15.8832967199
>>> print (pabp_data_helper.delta_elbo(pabp_vae_model,[(126,"G","A"), (137,"I","P"), (155,"S","A")], N_pred_iterations=500))
-15.7348273612
>>> print (pabp_data_helper.delta_elbo(pabp_vae_model,[(126,"G","A"), (137,"I","P"), (155,"S","A")], N_pred_iterations=500))
-15.6141800809

我又计算了(155,“S”,“A”)的单独的突变影响，结果如下：

>>> print (pabp_data_helper.delta_elbo(pabp_vae_model,[(155,"S","A")], N_pred_iterations=500))                             
-4.43953605237

可见虽然不是每个位点突变的简单相加，但也近似于相加的结果。

我们可以预测所有单个突变的影响。优选使用此函数及以下函数，因为它们能够利用对突变数据进行小批量处理所带来的加速优势。

pabp_full_matr_mutant_name_list, pabp_full_matr_delta_elbos \= pabp_data_helper.single_mutant_matrix(pabp_vae_model, N_pred_iterations=500)

print (pabp_full_matr_mutant_name_list[0], pabp_full_matr_delta_elbos[0])

('K123A', 0.5887526915685584)

我们还可以以批处理模式从文件中预测突变的影响。

pabp_custom_matr_mutant_name_list, pabp_custom_matr_delta_elbos \= pabp_data_helper.custom_mutant_matrix("mutations/PABP_YEAST_Fields2013-singles.csv", \pabp_vae_model, N_pred_iterations=500)print (pabp_custom_matr_mutant_name_list[12], pabp_custom_matr_delta_elbos[12])

('N127D', -6.426795215037501)

我们也可以编写一个快速的函数来从一个突变文件计算 Spearman 系数（rho）。

def generate_spearmanr(mutant_name_list, delta_elbo_list, mutation_filename, phenotype_name):measurement_df = pd.read_csv(mutation_filename, sep=',')mutant_list = measurement_df.mutant.tolist()expr_values_ref_list = measurement_df[phenotype_name].tolist()mutant_name_to_pred = {mutant_name_list[i]:delta_elbo_list[i] for i in range(len(delta_elbo_list))}# If there are measurements wt_list = []preds_for_spearmanr = []measurements_for_spearmanr = []for i,mutant_name in enumerate(mutant_list):expr_val = expr_values_ref_list[i]# Make sure we have made a prediction for that mutantif mutant_name in mutant_name_to_pred:multi_mut_name_list = mutant_name.split(':')# If there is no measurement for that mutant, pass over itif np.isnan(expr_val):pass# If it was a codon change, add it to the wt vals to averageelif mutant_name[0] == mutant_name[-1] and len(multi_mut_name_list) == 1:wt_list.append(expr_values_ref_list[i])# If it is labeled as the wt sequence, add it to the average listelif mutant_name == 'wt' or mutant_name == 'WT':wt_list.append(expr_values_ref_list[i])else:measurements_for_spearmanr.append(expr_val)preds_for_spearmanr.append(mutant_name_to_pred[mutant_name])if wt_list != []:measurements_for_spearmanr.append(np.mean(average_wt_list))preds_for_spearmanr.append(0.0)num_data = len(measurements_for_spearmanr)spearman_r, spearman_pval = spearmanr(measurements_for_spearmanr, preds_for_spearmanr)print ("N: "+str(num_data)+", Spearmanr: "+str(spearman_r)+", p-val: "+str(spearman_pval))

generate_spearmanr(pabp_custom_matr_mutant_name_list, pabp_custom_matr_delta_elbos, \"mutations/PABP_YEAST_Fields2013-singles.csv", "log")

N: 1188, Spearmanr: 0.6509305755221257, p-val: 4.0800344026520655e-144

PDZ

data_params = {"alignment_file":"datasets/DLG4_RAT_hmmerbit_plmc_n5_m30_f50_t0.2_r300-400_id100_b50.a2m"}pdz_data_helper = helper.DataHelper(alignment_file=data_params["alignment_file"],working_dir=".",calc_weights=False)pdz_vae_model   = model.VariationalAutoencoder(pdz_data_helper,batch_size              =   model_params["batch_size"],encoder_architecture    =   [model_params["encode_dim_zero"],model_params["encode_dim_one"]],decoder_architecture    =   [model_params["decode_dim_zero"],model_params["decode_dim_one"]],n_latent                =   model_params["n_latent"],n_patterns              =   model_params["n_patterns"],convolve_patterns       =   model_params["conv_pat"],conv_decoder_size       =   model_params["d_c_size"],logit_p                 =   model_params["logit_p"],sparsity                =   model_params["sparsity"],encode_nonlinearity_type       =   model_params["encode_nonlin"],decode_nonlinearity_type       =   model_params["decode_nonlin"],final_decode_nonlinearity      =   model_params["final_decode_nonlin"],output_bias             =   model_params["output_bias"],final_pwm_scale         =   model_params["final_pwm_scale"],working_dir             =   ".")print ("Model built")file_prefix = "DLG4_RAT"
pdz_vae_model.load_parameters(file_prefix=file_prefix)print ("Parameters loaded\n\n")pdz_custom_matr_mutant_name_list, pdz_custom_matr_delta_elbos \= pdz_data_helper.custom_mutant_matrix("mutations/DLG4_RAT_Ranganathan2012.csv", \pdz_vae_model, N_pred_iterations=500)generate_spearmanr(pdz_custom_matr_mutant_name_list, pdz_custom_matr_delta_elbos, \"mutations/DLG4_RAT_Ranganathan2012.csv", "CRIPT")

Encoding sequences
Neff = 102246.0
Data Shape = (102246, 84, 20)
Model built
Parameters loadedN: 1577, Spearmanr: 0.6199244929585085, p-val: 4.31636475994128e-168

B-lactamase

对于包含更多待预测突变的较大蛋白质，运行时间可能会更长。针对这种情况，我们建议使用支持 GPU 的计算。

data_params = {"dataset":"BLAT_ECOLX"}blat_data_helper = helper.DataHelper(dataset=data_params["dataset"],working_dir=".",calc_weights=False)blat_vae_model   = model.VariationalAutoencoder(blat_data_helper,batch_size              =   model_params["batch_size"],encoder_architecture    =   [model_params["encode_dim_zero"],model_params["encode_dim_one"]],decoder_architecture    =   [model_params["decode_dim_zero"],model_params["decode_dim_one"]],n_latent                =   model_params["n_latent"],n_patterns              =   model_params["n_patterns"],convolve_patterns       =   model_params["conv_pat"],conv_decoder_size       =   model_params["d_c_size"],logit_p                 =   model_params["logit_p"],sparsity                =   model_params["sparsity"],encode_nonlinearity_type       =   model_params["encode_nonlin"],decode_nonlinearity_type       =   model_params["decode_nonlin"],final_decode_nonlinearity      =   model_params["final_decode_nonlin"],output_bias             =   model_params["output_bias"],final_pwm_scale         =   model_params["final_pwm_scale"],working_dir             =   ".")print ("Model built")file_prefix = "BLAT_ECOLX"
blat_vae_model.load_parameters(file_prefix=file_prefix)print ("Parameters loaded\n\n")blat_custom_matr_mutant_name_list, blat_custom_matr_delta_elbos \= blat_data_helper.custom_mutant_matrix("mutations/BLAT_ECOLX_Ranganathan2015.csv", \blat_vae_model, N_pred_iterations=500)generate_spearmanr(blat_custom_matr_mutant_name_list, blat_custom_matr_delta_elbos, \"mutations/BLAT_ECOLX_Ranganathan2015.csv", "2500")

Encoding sequences
Neff = 8355.0
Data Shape = (8355, 253, 20)
Model built
Parameters loadedN: 4807, Spearmanr: 0.743886370415797, p-val: 0.0