Predicting the yield of Suzuki Miyaura HTE reactions
import logging
import sklearn
import torch
from sklearn.metrics import mean_squared_error, r2_score, mean_absolute_error
import pandas as pd
from rxnfp.models import SmilesClassificationModel
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
import numpy as np
import os
from rdkit import Chem
from rdkit.Chem import rdChemReactions
from IPython.display import display
%matplotlib inline
torch.cuda.is_available()
NAME_SPLIT = [
('random_split_0', 4032),
('random_split_1', 4032),
('random_split_2', 4032),
('random_split_3', 4032),
('random_split_4', 4032),
('random_split_5', 4032),
('random_split_6', 4032),
('random_split_7', 4032),
('random_split_8', 4032),
('random_split_9', 4032),
]
def load_model_from_results_folder(name, split, epoch=15, results_folder='../trained_models/suzuki_miyaura', model_type='bert'):
import os
models_folder = os.path.join(results_folder, f"{name}_split_{str(split).replace('-','_')}")
model_path = [os.path.join(models_folder, o) for o in os.listdir(models_folder)
if os.path.isdir(os.path.join(models_folder,o)) and o.endswith(f'epoch-{epoch}')][0]
model = SmilesClassificationModel(model_type, model_path,
num_labels=1, args={
"regression": True
}, use_cuda=torch.cuda.is_available())
return model
def make_plot(y_test, y_pred, rsme, r2_score, mae, name):
fontsize = 16
fig, ax = plt.subplots(figsize=(8,8))
r2_patch = mpatches.Patch(label="R2 = {:.3f}".format(r2_score), color="#5402A3")
rmse_patch = mpatches.Patch(label="RMSE = {:.1f}".format(rmse), color="#5402A3")
mae_patch = mpatches.Patch(label="MAE = {:.1f}".format(mae), color="#5402A3")
plt.xlim(-5,105)
plt.ylim(-5,105)
plt.scatter(y_pred, y_test, alpha=0.2, color="#5402A3")
plt.plot(np.arange(100), np.arange(100), ls="--", c=".3")
plt.legend(handles=[r2_patch, rmse_patch, mae_patch], fontsize=fontsize)
ax.set_ylabel('Measured', fontsize=fontsize)
ax.set_xlabel('Predicted', fontsize=fontsize)
ax.set_title(name, fontsize=fontsize)
return fig
y_predictions = []
y_tests = []
r2_scores = []
rmse_scores = []
for (name, split) in NAME_SPLIT:
df = pd.read_csv(f'../data/Suzuki-Miyaura/random_splits/{name}.tsv', sep='\t')
train_df = df.iloc[:split][['rxn', 'y']]
test_df = df.iloc[split:][['rxn', 'y']]
train_df.columns = ['text', 'labels']
test_df.columns = ['text', 'labels']
mean = train_df.labels.mean()
std = train_df.labels.std()
model = load_model_from_results_folder(name, split)
y_test = test_df['labels'].values * 100
y_preds = model.predict(test_df.text.values)[0]
y_preds = y_preds * std + mean
y_preds = y_preds * 100
y_preds = np.clip(y_preds, 0, 100)
r_squared = r2_score(y_test, y_preds)
rmse = mean_squared_error(y_test, y_preds) ** 0.5
mae = mean_absolute_error(y_test, y_preds)
r2_scores.append(r_squared)
print(name, r_squared, rmse)
fig = make_plot(y_test, y_preds, rmse, r_squared, mae, name)
display(fig)
# fig.savefig(f'images/mlst_perera_results/{name}_{split}.pdf')
