## Script to sanitize and split AggregatorAdvisor dataset

# 1. Load modules

pip install rdkit
pip install molvs
import pandas as pd
import numpy as np
import urllib.request
import tqdm
import rdkit
from rdkit import Chem
import molvs

standardizer = molvs.Standardizer()
fragment_remover = molvs.fragment.FragmentRemover()

# 2. Load the original dataset into a pandas DataFrame

# PLEASE download the 'raw_data.csv', which is the dataset from the paper, and run the code
# https://huggingface.co/datasets/maomlab/AggregatorAdvisor/blob/main/raw_data.csv

AA = pd.read_csv('raw_data.csv') # AA is an abbreviation of Aggregator Advisor

#3. Resolve SMILES parse error

# Smiles is 'None', found the compound on ChemSpider
# Smiles displayed 'Explicit valence for atom # 2 O, 3, is greater than permitted'
# https://www.chemspider.com/Chemical-Structure.17588253.html?rid=026abd00-5d7b-4c7a-b279-3ba43ab46203
AA.loc[AA['smiles'] == '[O-][N+](=[O-])C1=CC=CC(=C1)C2=NC(CO2)C3=CC=CC=C3' , 'smiles'] = 'c1ccc(cc1)C2COC(=N2)c3cccc(c3)[N+](=O)[O-]'

# Smiles is 'None', found the compound on ChemSpider
# Smiles displayed 'Explicit valence for atom # 2 O, 3, is greater than permitted'
# https://www.chemspider.com/Chemical-Structure.17588254.html?rid=0f94ced5-dee6-4274-b0c5-796500b40be7
AA.loc[AA['smiles'] == '[O-][N+](=[O-])C1=CC=CC(=C1)C2=NCCC(O2)C3=CC=CC=C3', 'smiles'] = 'c1ccc(cc1)C2CCN=C(O2)c3cccc(c3)[N+](=O)[O-]'

#4. Sanitize with MolVS and print problems

AA['X'] = [ \
    rdkit.Chem.MolToSmiles(
        fragment_remover.remove(
        standardizer.standardize(
        rdkit.Chem.MolFromSmiles(
        smiles))))
    for smiles in AA['smiles']]

problems = []
for index, row in tqdm.tqdm(AA.iterrows()):
    result = molvs.validate_smiles(row['X'])
    if len(result) == 0:
        continue
    problems.append( (row['substance_id'], result) )

# most are because it includes the salt form and/or it is not neutralized
for substance_id, alert in problems:
    print(f"substance_id: {substance_id}, problem: {alert[0]}")

# Result interpretation
#  - Can't kekulize mol: The error message means that kekulization would break the molecules down, so it couldn't proceed
#  It doesn't mean that the molecules are bad, it just means that normalization failed     

#5. Select columns and rename the dataset

AA.rename(columns={'X': 'SMILES'}, inplace=True)
newAA = AA[['SMILES', 'substance_id', 'aggref_index', 'logP', 'reference']]

#6. Import modules to split the dataset

import sys
from rdkit import DataStructs
from rdkit.Chem import AllChem as Chem
from rdkit.Chem import PandasTools
import scipy.spatial.distance as ssd
from scipy.cluster import hierarchy

#7. Split the dataset into test and train

#7. Split the dataset into test and train

class MolecularFingerprint:
    def __init__(self, fingerprint):
        self.fingerprint = fingerprint

    def __str__(self):
        return self.fingerprint.__str__()

def compute_fingerprint(molecule):
    try:
        fingerprint = Chem.GetMorganFingerprintAsBitVect(molecule, 2, nBits=1024)
        result = np.zeros(len(fingerprint), np.int32)
        DataStructs.ConvertToNumpyArray(fingerprint, result)
        return MolecularFingerprint(result)
    except:
        print("Fingerprints for a structure cannot be calculated")
        return None

def tanimoto_distances_yield(fingerprints, num_fingerprints):
    for i in range(1, num_fingerprints):
        yield [1 - x for x in DataStructs.BulkTanimotoSimilarity(fingerprints[i], fingerprints[:i])]

def butina_cluster(fingerprints, num_points, distance_threshold, reordering=False):
    nbr_lists = [None] * num_points
    for i in range(num_points):
        nbr_lists[i] = []

    dist_fun = tanimoto_distances_yield(fingerprints, num_points)
    for i in range(1, num_points):
        dists = next(dist_fun)

        for j in range(i):
            dij = dists[j]
            if dij <= distance_threshold:
                nbr_lists[i].append(j)
                nbr_lists[j].append(i)

    t_lists = [(len(y), x) for x, y in enumerate(nbr_lists)]
    t_lists.sort(reverse=True)

    res = []
    seen = [0] * num_points
    while t_lists:
        _, idx = t_lists.pop(0)
        if seen[idx]:
            continue
        t_res = [idx]
        for nbr in nbr_lists[idx]:
            if not seen[nbr]:
                t_res.append(nbr)
                seen[nbr] = 1
        if reordering:
            nbr_nbr = [nbr_lists[t] for t in t_res]
            nbr_nbr = frozenset().union(*nbr_nbr)
            for x, y in enumerate(t_lists):
                y1 = y[1]
                if seen[y1] or (y1 not in nbr_nbr):
                    continue
                nbr_lists[y1] = set(nbr_lists[y1]).difference(t_res)
                t_lists[x] = (len(nbr_lists[y1]), y1)
            t_lists.sort(reverse=True)
        res.append(tuple(t_res))
    return tuple(res)

def hierarchal_cluster(fingerprints):

      num_fingerprints = len(fingerprints)

      av_cluster_size = 8
      dists = []

      for i in range(0, num_fingerprints):
          sims = DataStructs.BulkTanimotoSimilarity(fingerprints[i], fingerprints)
          dists.append([1 - x for x in sims])

      dis_array = ssd.squareform(dists)
      Z = hierarchy.linkage(dis_array)
      average_cluster_size = av_cluster_size
      cluster_amount = int(num_fingerprints / average_cluster_size)
      clusters = hierarchy.cut_tree(Z, n_clusters=cluster_amount)

      clusters = list(clusters.transpose()[0])
      cs = []
      for i in range(max(clusters) + 1):
          cs.append([])

      for i in range(len(clusters)):
          cs[clusters[i]].append(i)
      return cs

def cluster_fingerprints(fingerprints, method="Auto"):
    num_fingerprints = len(fingerprints)

    if method == "Auto":
        method = "TB" if num_fingerprints >= 10000 else "Hierarchy"

    if method == "TB":
        cutoff = 0.56
        print("Butina clustering is selected. Dataset size is:", num_fingerprints)
        clusters = butina_cluster(fingerprints, num_fingerprints, cutoff)

    elif method == "Hierarchy":
        print("Hierarchical clustering is selected. Dataset size is:", num_fingerprints)
        clusters = hierarchal_cluster(fingerprints)

    return clusters

def split_dataframe(dataframe, smiles_col_index, fraction_to_train, split_for_exact_fraction=True, cluster_method="Auto"):
    try:
        import math
        smiles_column_name = dataframe.columns[smiles_col_index]
        molecule = 'molecule'
        fingerprint = 'fingerprint'
        group = 'group'
        testing = 'testing'

        try:
            PandasTools.AddMoleculeColumnToFrame(dataframe, smiles_column_name, molecule)
        except:
            print("Exception occurred during molecule generation...")

        dataframe = dataframe.loc[dataframe[molecule].notnull()]
        dataframe[fingerprint] = [compute_fingerprint(m) for m in dataframe[molecule]]
        dataframe = dataframe.loc[dataframe[fingerprint].notnull()]

        fingerprints = [Chem.GetMorganFingerprintAsBitVect(m, 2, nBits=2048) for m in dataframe[molecule]]
        clusters = cluster_fingerprints(fingerprints, method=cluster_method)

        dataframe.drop([molecule, fingerprint], axis=1, inplace=True)

        last_training_index = int(math.ceil(len(dataframe) * fraction_to_train))
        clustered = None
        cluster_no = 0
        mol_count = 0

        for cluster in clusters:
            cluster_no = cluster_no + 1
            try:
                one_cluster = dataframe.iloc[list(cluster)].copy()
            except:
                print("Wrong indexes in Cluster: %i, Molecules: %i" % (cluster_no, len(cluster)))
                continue

            one_cluster.loc[:, 'ClusterNo'] = cluster_no
            one_cluster.loc[:, 'MolCount'] = len(cluster)

            if (mol_count < last_training_index) or (cluster_no < 2):
                one_cluster.loc[:, group] = 'training'
            else:
                one_cluster.loc[:, group] = testing

            mol_count += len(cluster)
            clustered = pd.concat([clustered, one_cluster], ignore_index=True)

        if split_for_exact_fraction:
            print("Adjusting test to train ratio. It may split one cluster")
            clustered.loc[last_training_index + 1:, group] = testing

        print("Clustering finished. Training set size is %i, Test set size is %i, Fraction %.2f" %
              (len(clustered.loc[clustered[group] != testing]),
               len(clustered.loc[clustered[group] == testing]),
               len(clustered.loc[clustered[group] == testing]) / len(clustered)))

    except KeyboardInterrupt:
        print("Clustering interrupted.")

    return clustered


def realistic_split(df, smile_col_index, frac_train, split_for_exact_frac=True, cluster_method = "Auto"):
  return split_dataframe(df.copy(), smile_col_index, frac_train, split_for_exact_frac, cluster_method=cluster_method)

def split_df_into_train_and_test_sets(df):
    df['group'] = df['group'].str.replace(' ', '_')
    df['group'] = df['group'].str.lower()
    train = df[df['group'] == 'training']
    test = df[df['group'] == 'testing']
    return train, test

# 8. Test and train datasets have been made

smiles_index = 0  # Because smiles is in the first column
realistic = realistic_split(newAA.copy(), smiles_index, 0.8, split_for_exact_frac=True, cluster_method="Auto")
realistic_train, realistic_test = split_df_into_train_and_test_sets(realistic)

#9. Select columns and name the datasets

selected_columns = realistic_train[['SMILES', 'substance_id', 'aggref_index', 'logP', 'reference']]
selected_columns.to_csv("AggregatorAdvisor_train.csv", index=False)
selected_columns = realistic_test[['SMILES', 'substance_id', 'aggref_index', 'logP', 'reference']]
selected_columns.to_csv("AggregatorAdvisor_test.csv", index=False)