User prediction

Let’s see how a user can run maxsmi for compound prediction.

[1]:

from pathlib import Path
import ast
import os
import pandas as pd
from IPython.display import Image

[2]:

# Path to this notebook
HERE = Path(_dh[-1])

[3]:

%%bash
cd ../..
python maxsmi/prediction_unlabeled_data.py --help

usage: prediction_unlabeled_data.py [-h] [--task TASK]
                                    [--smiles_prediction USER_SMILES]

optional arguments:
  -h, --help            show this help message and exit
  --task TASK           data to be used
  --smiles_prediction USER_SMILES
                        SMILES for prediction

Let’s predict lipophilicity for the semaxanib drug given by the SMILES O=C2C(\c1ccccc1N2)=C/c3c(cc([nH]3)C)C:

[4]:

%%bash
cd ../..
python maxsmi/prediction_unlabeled_data.py --task=lipophilicity --smiles_prediction="O=C2C(\c1ccccc1N2)=C/c3c(cc([nH]3)C)C"

Script completed. Output can be found at maxsmi/user_prediction/lipophilicity_O=C2C(%5Cc1ccccc1N2)=C%2Fc3c(cc([nH]3)C)C/

Note: the backslash has a special significance in both the python programming language and SMILES notation, and therefore should be treated with care.

[5]:

from maxsmi.utils.utils_smiles import smiles_to_folder_name
smiles_to_folder_name?

Signature: smiles_to_folder_name(smiles)
Docstring:
Encodes the SMILES containing special characters as URL encoding.

Parameters
----------
smiles : str
    SMILES string describing a compound.

Returns
-------
str :
    The standard URL encoding of the SMILES if it contains special characters.

Notes
-----
: -> %3A
\ -> %5C (here backslash)
/ -> %2F
* -> %2A

Taken from:
https://www.degraeve.com/reference/urlencoding.php
File:      ~/Documents/github/maxsmi/maxsmi/utils/utils_smiles.py
Type:      function

[6]:

path = (f"{HERE}/../../maxsmi/user_prediction/"
        f"lipophilicity_O=C2C(%5Cc1ccccc1N2)=C%2Fc3c(cc([nH]3)C)C")

Let’s see what was stored during the execution.

[7]:

list_output_files = os.listdir(path)
list_output_files

[7]:

['user_prediction_table.csv', '2D_molecule.png', 'user_prediction_output.log']

There are three files:

user_prediction_table.csv: The results of the prediction in a csv table,
2D_molecule.png: A PNG figure representing the molecular graph,
user_prediction_output.log: Additional information on the prediction in a log file.

Let’s have a closer look at each of them.

The log file contains detailed information on the model used, augmentation strategy as well as the results.

[8]:

log_file = [string for string in list_output_files
            if string.endswith("log")][0]

with open(f"{path}/{log_file}") as f:
    for _ in range(15):  # show first 15 lines
        line = next(f).strip()
        print(line)

INFO:root:Start at 2021-11-11 08:37:59.404143
INFO:root:Data and task: lipophilicity
INFO:root:CUDA available: False
INFO:root:Shape of training data set before processing: (4200, 2)
INFO:root:Shape of training data set after processing: (4199, 3)
INFO:root:SMILES in training data set: False
INFO:root:Augmentation strategy: augmentation_without_duplication
INFO:root:Augmentation number: 80
INFO:root:Time for augmentation 0:00:00.006678
INFO:root:Longest smiles in training data set: 268
INFO:root:Summary of ml model used for the prediction: Convolutional1DNetwork(
(convolution): Conv1d(48, 300, kernel_size=(10,), stride=(1,))
(fully_connected_1): Linear(in_features=77700, out_features=100, bias=True)
(fully_connected_out): Linear(in_features=100, out_features=1, bias=True)
)

Let’s plot the molecular graph of the semaxanib drug.

[9]:

image_file = [string for string in list_output_files
              if string.endswith("png")][0]
Image(filename=f"{path}/{image_file}")

[9]:

../_images/example_user_prediction_16_0.png

Let’s have a closer look at the prediction.

[10]:

results_csv_file = [string for string in list_output_files
                    if string.endswith("csv")][0]
results = pd.read_csv(f"{path}/{results_csv_file}")
results

[10]:

	user_smiles	smiles_in_training	canonical_smiles	augmented_smiles	per_smiles_prediction	average_prediction	std_prediction
0	O=C2C(\c1ccccc1N2)=C/c3c(cc([nH]3)C)C	False	Cc1cc(C)c(/C=C2\C(=O)Nc3ccccc32)[nH]1	['C(=C1\\C(=O)Nc2ccccc21)\\c1[nH]c(C)cc1C', 'c...	[3.2781868, 2.9332523, 2.9482965, 3.290605 , 3...	3.108885	0.442194

[11]:

print(f"SMILES given by the user: \t{results.user_smiles[0]}\n"
      f"Part of the training set: \t{results.smiles_in_training[0]}\n"
      f"Its lipophilicity value is predicted to be: \t"
      f"{results.average_prediction[0]:.3f} "
      f"with a confidence of {results.std_prediction[0]:.3f}.")

SMILES given by the user:       O=C2C(\c1ccccc1N2)=C/c3c(cc([nH]3)C)C
Part of the training set:       False
Its lipophilicity value is predicted to be:     3.109 with a confidence of 0.442.

The first \(n=7\) random SMILES are given by:

[12]:

random_smiles = ast.literal_eval(results.augmented_smiles[0])[0:7]
random_smiles

[12]:

['C(=C1\\C(=O)Nc2ccccc21)\\c1[nH]c(C)cc1C',
 'c1c2c(ccc1)/C(C(=O)N2)=C/c1c(C)cc(C)[nH]1',
 'C1(=O)Nc2c(cccc2)/C1=C/c1c(cc(C)[nH]1)C',
 '[nH]1c(/C=C2\\C(Nc3c2cccc3)=O)c(C)cc1C',
 'C(=C1/c2c(cccc2)NC1=O)\\c1c(cc(C)[nH]1)C',
 'Cc1cc(C)[nH]c1/C=C1\\C(=O)Nc2c1cccc2',
 'N1c2c(cccc2)/C(C1=O)=C/c1[nH]c(cc1C)C']

The predictions associated to these random SMILES are:

[13]:

per_smiles_prediction = ast.literal_eval(results.per_smiles_prediction[0])[0:7]
for prediction in per_smiles_prediction:
    print(f"{prediction:.3f}")