api_encoding_train_test.py

"""
An example script for using PyPEF as an API for encoding sequences
to train and test ML models and the hybrid model MERGE.

Encoding of sequences is not parallelized in this script.
"""

import os
import pandas as pd
import numpy as np
from scipy.stats import spearmanr
from tqdm import tqdm
import matplotlib.pyplot as plt
import urllib.request
from hashlib import sha256

from sklearn.model_selection import KFold, train_test_split
from pypef.utils.variant_data import remove_nan_encoded_positions, get_seqs_from_var_name
from pypef.ml.regression import (
    OneHotEncoding, AAIndexEncoding, get_regressor_performances,
    path_aaindex_dir, full_aaidx_txt_path
)
from pypef.dca.hybrid_model import DCAHybridModel, remove_gap_pos, get_delta_e_statistical_model
from pypef.dca.plmc_encoding import PLMC
from pypef.dca.gremlin_inference import GREMLIN


use_gremlin = True  # if False uses plmc (requires plmc-generated .params file)

n_aaindices_to_test = 10  # all would be 566 AAindex indices, only testing 10 here for shorter run time
# avGFP wild type sequence
wt_sequence = 'MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTL' \
              'VTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLV' \
              'NRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLAD' \
              'HYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK'

if not os.path.isdir('AVGFP'):
    os.mkdir('AVGFP')

# 1st example: A. DCA-based encoding
# -------------------------------------------------------------------------------
# 1st example, since this approach reduces the amount of data (variant-fitness pairs),
# as not all variants can necessarily be encoded using the DCA approach. To compare
# the test performance with the other encoding techniques to be tested, we start with
# DCA and split the variant fitness data so that sizes of the data sets are same for
# all encoding techniques tested.
print(f'\nRunning script... which takes ~ 2 h (GREMLIN) - 4 h in total (PLMC) when using all 566 AAindex-indices '
      f'for encoding and model testing (only {n_aaindices_to_test} AAindinces will be tested)...'
      f'\n\n(1/4) Testing DCA-based sequence encoding...\n' + "=" * 50)
if not use_gremlin:  # PLMC params file-based encoding
    try:
        dca_encoder = PLMC(
            params_file='AVGFP/uref100_avgfp_jhmmer_119_plmc_42.6.params',
            verbose=False
        )
    except (ValueError, FileNotFoundError):
        print('Did not find required files for DCA-based encoding. Downloading required files...')
        # Single substituted encoded variants: CSV files including variant name and true fitness
        url = 'https://github.com/niklases/PyPEF/raw/main/datasets/AVGFP/avGFP.csv'
        urllib.request.urlretrieve(url, 'AVGFP/avGFP.csv')
        with open('AVGFP/avGFP.csv', 'rb') as f:
            sha256_hash = sha256(f.read()).hexdigest()
        if not sha256_hash == 'be4623f35a5ba05d33a29ae6e69dc3c2e994e3c9092cd5880a8d0bbc12f187b1':
            raise SystemError("Hash of downloaded CSV file not correct, terminating further running.")
        # Getting plmc parameter file
        url = 'https://github.com/niklases/PyPEF/raw/main/datases/' \
              'AVGFP/uref100_avgfp_jhmmer_119_plmc_42.6.params'
        urllib.request.urlretrieve(
            url, 'AVGFP/uref100_avgfp_jhmmer_119_plmc_42.6.params'
        )  # 71.2 MB file size
        with open('AVGFP/uref100_avgfp_jhmmer_119_plmc_42.6.params', 'rb') as f:
            sha256_hash = sha256(f.read()).hexdigest()
        if not sha256_hash == '8baa30bc7d568906d4b587d4c6babd025041bf2f967f303fa38070d2df339830':
            raise SystemError("Hash of downloaded DCA parameter file not correct, terminating further running.")
        print('Successfully downloaded all required files!')
else:
    url = 'https://github.com/niklases/PyPEF/raw/main/datasets/AVGFP/uref100_avgfp_jhmmer_119.a2m'
    urllib.request.urlretrieve(url, 'AVGFP/uref100_avgfp_jhmmer_119.a2m')
    with open('AVGFP/uref100_avgfp_jhmmer_119.a2m', 'rb') as f:
        sha256_hash = sha256(f.read()).hexdigest()
    if not sha256_hash == 'a7be875cd6b8c81c14abae0efb3d307aa0fa8956d15ebc67c4cab492d4443777':
        raise SystemError("Hash of downloaded CSV file not correct, terminating further running.")

url = 'https://github.com/niklases/PyPEF/raw/main/datasets/AVGFP/avGFP.csv'
urllib.request.urlretrieve(url, 'AVGFP/avGFP.csv')
with open('AVGFP/avGFP.csv', 'rb') as f:
    sha256_hash = sha256(f.read()).hexdigest()
if not sha256_hash == '34ba042830f1297463bb0bac155a043cd87c9e6fd0bbbd45312eb833b170ab0f':
    raise SystemError("Hash of downloaded CSV file not correct, terminating further running.")
variant_fitness_data = pd.read_csv('AVGFP/avGFP.csv', sep=';')  # loading the avGFP dataset

variants = variant_fitness_data.iloc[:2000, 0].values  # "just" using 2000 variants for faster processing
fitnesses = variant_fitness_data.iloc[:2000, 1].values
variants_split = []
for variant in variants:
    variants_split.append(variant.split('/'))
variants, fitnesses, sequences = get_seqs_from_var_name(wt_sequence, variants_split, fitnesses)

# Splitting in sets for training (fitting and hyperparameter validation) and testing
# Change number of applied splits for training and testing here, default: n_splits = 5
n_splits = 5
kf = KFold(n_splits=n_splits, random_state=42, shuffle=True)
train_val_splits_indices, test_splits_indices = [], []

if use_gremlin:
    dca_encoder = GREMLIN(
        alignment='AVGFP/uref100_avgfp_jhmmer_119.a2m',
        wt_seq=wt_sequence,
        opt_iter=100,
        optimize=True
    )
    dca_encoder.plot_correlation_matrix(matrix_type='apc')
    gaps = dca_encoder.gaps
    variants, sequences, fitnesses = remove_gap_pos(gaps, variants, sequences, fitnesses)
    x_dca = dca_encoder.collect_encoded_sequences(sequences)
    x_wt = dca_encoder.x_wt
else:  # plmc
    dca_encoder = PLMC(
        params_file='AVGFP/uref100_avgfp_jhmmer_119_plmc_42.6.params',
        verbose=False
    )
    x_dca = dca_encoder.collect_encoded_sequences(variants)
    # removing not DCA-encodable positions (and also fitnesses, variants, and sequences)
    # from the 2000 initial variants, 1427 remain
    x_dca, fitnesses, variants, sequences = remove_nan_encoded_positions(x_dca, fitnesses, variants, sequences)
    x_wt = dca_encoder.x_wt

# Statistical model performance
y_pred = get_delta_e_statistical_model(x_dca, x_wt)
print(f'Statistical DCA model performance on all (2000) datapoints: {spearmanr(fitnesses, y_pred)[0]:.3f}')
# Split double and higher substituted variants to multiple single substitutions separated by '/'

print(f'N Variants remaining after excluding non-DCA-encodable positions: {len(x_dca)}')
assert len(x_dca) == len(fitnesses) == len(variants) == len(sequences)

for train_val_indices, test_indices in kf.split(variants):  # several variants are not included in the data anymore
    train_val_splits_indices.append(train_val_indices)
    test_splits_indices.append(test_indices)
print(f'Total number of variant-fitness data for training and '
      f'validation-based hyperparameter tuning: {len(train_val_indices)}'  # indices only vary by +- 1 for the splits
      f'\nVariants for testing: {len(test_indices)}\n')

ten_split_performance_ml, ten_split_performance_hybrid = [], []
for i, indices in enumerate(train_val_splits_indices):
    x_train_val = np.array(x_dca)[indices]
    y_train_val = np.array(fitnesses)[indices]
    x_test = np.array(x_dca)[test_splits_indices[i]]
    y_test = np.array(fitnesses)[test_splits_indices[i]]
    # get_regressor_performances() already splits train_val data in data for fitting (training) and validation
    # and after each CV-round shifts to the next hyperparameter of the regressor hyperparameter grid
    performances = get_regressor_performances(x_train_val, x_test, y_train_val, y_test, regressor='ridge')
    # performances[4] is Spearmans correlation between y_true and y_pred
    ten_split_performance_ml.append(performances[4])
    print(f'Split {i + 1}/{len(train_val_splits_indices)}:\nSpearmans rho (ML) = {performances[4]:.3f}')
    # B. Hybrid modeling
    # -------------------------------------------------------------------------------
    hybrid_model = DCAHybridModel(x_train=x_train_val, y_train=y_train_val, x_wt=x_wt)
    beta_1, beta_2, regressor = hybrid_model.settings(x_train=x_train_val, y_train=y_train_val)
    y_test_pred = hybrid_model.hybrid_prediction(x=x_test, reg=regressor, beta_1=beta_1, beta_2=beta_2)
    ten_split_performance_hybrid.append(spearmanr(y_test, y_test_pred)[0])
    print(f'Spearmans rho (Hybrid) = {spearmanr(y_test, y_test_pred)[0]:.3f}')
print('-'*80 + f'\n{n_splits}-fold mean Spearmans rho (ML)= {np.mean(ten_split_performance_ml):.3f} '
      f'+- {np.std(ten_split_performance_ml, ddof=1):.3f}\n'
      f'{n_splits}-fold mean Spearmans rho (Hybrid)= {np.mean(ten_split_performance_hybrid):.3f} '
      f'+- {np.std(ten_split_performance_ml, ddof=1):.3f}')


# 2nd example: AAindex encoding over all 566 amino acid descriptor sets
# -------------------------------------------------------------------------------
print(f'\n\n(2/4) Testing AAindex-based sequence encoding (using {n_aaindices_to_test} indices)...\n' + "=" * 75)
spearmans_rhos_aaidx, aa_index = [], []
# e.g., looping over the 566 AAindex entries, encode with each AAindex and test performance
# which can be seen as a AAindex hyperparameter search on the test set, i.e.,
# not totally fair comparison to onehot- and DCA-based encoding techniques.
# Limiting indices for testing to N = n_aaindices_to_test...
aa_indices = sorted([file for file in os.listdir(path_aaindex_dir()) if file.endswith('.txt')])[:n_aaindices_to_test]
mean_performances, ten_split_performance_std_dev, aa_indices_collected = [], [], []
for index, aaindex in enumerate(aa_indices):
    aaidx_encoder = AAIndexEncoding(full_aaidx_txt_path(aaindex), sequences)
    # two encoding options possible, FFT-ed and 'raw' encoded sequences (here we keep using 'raw' encoded sequences)
    x_aaidx_fft, x_aaidx_no_fft = aaidx_encoder.collect_encoded_sequences()
    ten_split_performance = []
    if x_aaidx_no_fft == 'skip':
        print(f'Skipped AAindex {aaindex}')
        continue
    for i, indices in enumerate(train_val_splits_indices):
        print(f'\rSplit {i+1}/{len(train_val_splits_indices)}', end='')
        x_train_val = np.array(x_aaidx_no_fft)[indices]
        y_train_val = np.array(fitnesses)[indices]
        x_test = np.array(x_aaidx_no_fft)[test_splits_indices[i]]
        y_test = np.array(fitnesses)[test_splits_indices[i]]
        performances = get_regressor_performances(x_train_val, x_test, y_train_val, y_test, regressor='ridge')
        ten_split_performance.append(performances[4])
    print(f'\r{index + 1}/{len(aa_indices)}: AAindex {aaindex}, '
          f'{n_splits}-fold mean Spearmans rho = {np.mean(ten_split_performance):.3f} '
          f'+- {np.std(ten_split_performance, ddof=1):.3f}')
    mean_performances.append(np.mean(ten_split_performance))

    ten_split_performance_std_dev.append(np.std(ten_split_performance, ddof=1))
    aa_indices_collected.append(aaindex)
max_value = max(mean_performances)
max_idx = mean_performances.index(max_value)
max_value_std = ten_split_performance_std_dev[max_idx]
max_value_aaidx = aa_indices_collected[max_idx]
print('-'*80 + f'\nBest {n_splits}-fold mean Spearmans rho = {max_value:.3f} '
      f'+- {max_value_std:.3f}, AAindex descriptor set = {max_value_aaidx}\n')


# 3rd example: OneHot encoding
# -------------------------------------------------------------------------------
print('\n(3/4) Testing OneHot sequence encoding...\n' + "=" * 50)
onehot_encoder = OneHotEncoding(sequences)
x_onehot = onehot_encoder.collect_encoded_sequences()
ten_split_performance = []
for i, indices in enumerate(train_val_splits_indices):
    x_train_val = np.array(x_onehot)[indices]
    y_train_val = np.array(fitnesses)[indices]
    x_test = np.array(x_onehot)[test_splits_indices[i]]
    y_test = np.array(fitnesses)[test_splits_indices[i]]
    performances = get_regressor_performances(x_train_val, x_test, y_train_val, y_test, regressor='ridge')
    ten_split_performance.append(performances[4])
    print(f'Split {i + 1}/{len(train_val_splits_indices)}: Spearmans rho = {performances[4]:.3f}')
print('-'*80 + f'\n{n_splits}-fold mean Spearmans rho = {np.mean(ten_split_performance):.3f} '
      f'+- {np.std(ten_split_performance, ddof=1):.3f}\n')


# 4th example: Low-N and plotting using all encoding techniques and all data
# -------------------------------------------------------------------------------
print('Lastly, encoding all variants and performing "low-N" protein engineering task.\n'
      'This could require some time... < 1 (GREMLIN DCA encoding) to ~ 2 hours (PLMC '
      'single core DCA encoding) left...')

variants = variant_fitness_data.iloc[:, 0]
fitnesses = variant_fitness_data.iloc[:, 1].tolist()
variants_split = []
for variant in variants:
    variants_split.append(variant.split('/'))
variants, fitnesses, sequences = get_seqs_from_var_name(wt_sequence, variants_split, fitnesses)
print(f'N Total variants = {len(variants)}.\nEncoding sequences...')

if use_gremlin:
    variants, sequences, fitnesses = remove_gap_pos(dca_encoder.gaps, variants, sequences, fitnesses)
    x_dca = dca_encoder.get_score(sequences, encode=True)
else:
    x_dca = dca_encoder.collect_encoded_sequences(variants)
    # removing not DCA-encodable positions (and also reduce fitnesses, variants, and sequences accordingly)
    x_dca, fitnesses, variants, sequences = remove_nan_encoded_positions(x_dca, fitnesses, variants, sequences)

print(f'N Variants remaining after excluding non-DCA-encodable positions = {len(x_dca)}')

# using the best identified index, i.e., AURR980106 for 10 tested indices or QIAN880130 for all 566 indices
print(f'AAIndex-based encoding of the {len(x_dca)} variants (using index {max_value_aaidx})...')
x_aaindex_fft, x_aaindex_no_fft = AAIndexEncoding(
    full_aaidx_txt_path(max_value_aaidx), sequences).collect_encoded_sequences()
print(f'One-hot encoding of the {len(x_dca)} variants...')
x_onehot = OneHotEncoding(sequences).collect_encoded_sequences()

assert len(x_dca) == len(x_aaindex_no_fft) == len(x_onehot) == len(fitnesses) == len(variants) == len(sequences)

all_mean_performances_dca_statistical, all_mean_performances_dca_ml, all_mean_performances_hybrid, \
    all_mean_performances_aaidx, all_mean_performances_onehot = [], [], [], [], []
all_stddevs_dca_statistical, all_stddevs_dca_ml, all_stddevs_hybrid, all_stddevs_aaidx, all_stddevs_onehot = [], [], [], [], []
low_n_train = np.arange(50, 1001, 50)
print(f'\n(4/4) Testing low N performance (with N_train = {list(low_n_train)})...\n' + "=" * 50)
pbar = tqdm(low_n_train)
for n_train in pbar:
    pbar.set_description(f'N_train = {n_train}, N_test = {len(variants) - n_train}')
    performances_dca_statistical, performances_dca_ml, performances_hybrid, performances_aaidx, performances_onehot = [], [], [], [], []
    for rnd_state in [42, 213, 573, 917, 823]:
        x_dca_train, x_dca_test, y_train, y_test = train_test_split(
            x_dca, fitnesses, train_size=n_train, random_state=rnd_state)
        # Delta E is just the sum of the encoding (embedding) relative to the WT sum
        performances_dca_statistical.append(spearmanr(y_test, np.sum(x_dca_test, axis=1) - np.sum(x_wt))[0])
        performances_dca_ml.append(get_regressor_performances(
            x_dca_train, x_dca_test, y_train, y_test, regressor='ridge')[4])  # [4] defines spearmanr correlation

        hybrid_model = DCAHybridModel(x_train=x_dca_train, y_train=y_train, x_wt=x_wt)
        beta_1, beta_2, hybrid_regressor = hybrid_model.settings(x_train=x_dca_train, y_train=y_train)
        y_hybrid_pred = hybrid_model.hybrid_prediction(x=x_dca_test, reg=hybrid_regressor, beta_1=beta_1, beta_2=beta_2)
        performances_hybrid.append(spearmanr(y_test, y_hybrid_pred)[0])

        x_aaidx_train, x_aaidx_test, y_train, y_test = train_test_split(
            x_aaindex_no_fft, fitnesses, train_size=n_train, random_state=rnd_state)
        performances_aaidx.append(get_regressor_performances(
            x_aaidx_train, x_aaidx_test, y_train, y_test, regressor='ridge')[4])


        x_onehot_train, x_onehot_test, y_train, y_test = train_test_split(
            x_onehot, fitnesses, train_size=n_train, random_state=rnd_state)
        performances_onehot.append(get_regressor_performances(
            x_onehot_train, x_onehot_test, y_train, y_test, regressor='ridge')[4])
        
    all_mean_performances_dca_statistical.append(np.mean(performances_dca_statistical))
    all_stddevs_dca_statistical.append(np.std(performances_dca_statistical, ddof=1))
    all_mean_performances_dca_ml.append(np.mean(performances_dca_ml))
    all_stddevs_dca_ml.append(np.std(performances_dca_ml, ddof=1))
    all_mean_performances_hybrid.append(np.mean(performances_hybrid))
    all_stddevs_hybrid.append(np.std(performances_hybrid, ddof=1))
    all_mean_performances_aaidx.append(np.mean(performances_aaidx))
    all_stddevs_aaidx.append(np.std(performances_aaidx, ddof=1))
    all_mean_performances_onehot.append(np.mean(performances_onehot))
    all_stddevs_onehot.append(np.std(performances_onehot, ddof=1))

# Plotting all the achieved "low-N" performances
plt.plot(low_n_train, all_mean_performances_dca_statistical, 'o--', color='tab:cyan', label='DCA (statistical model)')
plt.fill_between(low_n_train,
                 np.array(all_mean_performances_dca_statistical) - np.array(all_stddevs_dca_statistical),
                 np.array(all_mean_performances_dca_statistical) + np.array(all_stddevs_dca_statistical),
                 color='tab:cyan', alpha=0.2)
plt.plot(low_n_train, all_mean_performances_dca_ml, 'o--', color='tab:orange', label='DCA encoding (pure ML)')
plt.fill_between(low_n_train,
                 np.array(all_mean_performances_dca_ml) - np.array(all_stddevs_dca_ml),
                 np.array(all_mean_performances_dca_ml) + np.array(all_stddevs_dca_ml),
                 color='tab:orange', alpha=0.2)
plt.plot(low_n_train, all_mean_performances_hybrid, 'o--', color='tab:blue', label='Hybrid DCA model (MERGE)')
plt.fill_between(low_n_train,
                 np.array(all_mean_performances_hybrid) - np.array(all_stddevs_hybrid),
                 np.array(all_mean_performances_hybrid) + np.array(all_stddevs_hybrid),
                 color='tab:blue', alpha=0.2)
plt.plot(low_n_train, all_mean_performances_aaidx, 'o--', color='tab:green',
         label=f"AAindex ({max_value_aaidx.split('.')[0]})")
plt.fill_between(low_n_train,
                 np.array(all_mean_performances_aaidx) - np.array(all_stddevs_aaidx),
                 np.array(all_mean_performances_aaidx) + np.array(all_stddevs_aaidx),
                 color='tab:green', alpha=0.2)
plt.plot(low_n_train, all_mean_performances_onehot, 'o--', color='tab:grey', label='One-hot')
plt.fill_between(low_n_train,
                 np.array(all_mean_performances_onehot) - np.array(all_stddevs_onehot),
                 np.array(all_mean_performances_onehot) + np.array(all_stddevs_onehot),
                 color='tab:grey', alpha=0.2)
plt.legend()
plt.xlabel(r'$N_\mathrm{Train}$')
plt.ylabel(r'$\rho$')
plt.savefig('AVGFP/low_N_avGFP_extrapolation.png', dpi=500)
print('\nDone! Created figure \'AVGFP/low_N_avGFP_extrapolation.png\'.')