Welcome to the Deep Learning Example with CoderData¶

Here we will be using a small subset of the BeatAML data to predict the AUC of drug response values.
Our predictive variables will be transcriptomics, proteomics, and drug structure.
This small subset does not have high predictive power and as such, this is just meant as a guide for creating your own models.

Import coderdata, pandas and numpy¶

In [1]:
import pandas as pd
import numpy as np
import coderdata as cd

Import keras, the deep learning framework¶

We will be using the keras functional API. This allows us to use multiple inputs in our model.
You may need to install a couple of these dependancies to procceed.

In [109]:
import keras
from numpy import loadtxt
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import train_test_split
from rdkit import Chem
from rdkit.Chem import AllChem
from keras.models import Model
from keras.layers import Input, Dense, concatenate
from keras.optimizers import Adam
from keras.losses import MeanSquaredError
from keras.metrics import MeanAbsoluteError
from keras import layers
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
import matplotlib.pyplot as plt
from sklearn.linear_model import LinearRegression

In this example, we will use a subset of the BeatAML dataset.¶

The first step is to download the data. We will then import the data into the DatasetLoader object.
If there are issues with the download process, please ensure that your connection is not blocked by a VPN.

In [7]:
cd.download_data_by_prefix('beataml')
beataml = cd.DatasetLoader("beataml")

Downloaded beataml_drugs.tsv.gz to local repository.
Downloaded beataml_proteomics.csv.gz to local repository.
Downloaded beataml_mutations.csv.gz to local repository.
Downloaded beataml_experiments.csv.gz to local repository.
Downloaded beataml_samples.csv to local repository.
Downloaded beataml_transcriptomics.csv.gz to local repository.
Processing Data...

Check Data Information.¶

This info command allows us to see which data types are present in the beataml DatasetLoader object.
We will be using the Transcriptomics, Proteomics, Drugs, and Experiments Data in our example.
In other examples you may want to include the samples data to associate cancer and model type with each sample. However, in the BeatAML data, all data is from Acute Myeloid Leukemia patients.

In [5]:
beataml.info()

Dataset Type: beataml
Beat acute myeloid leukemia (BeatAML) data was collected though GitHub and Synapse.

Available Datatypes and Their Formats:
- drugs: Format not specified
- experiments: Format not specified
- mutations: long format
- proteomics: long format
- samples: long format
- transcriptomics: long format
In [6]:
beataml.drugs
Out[6]:
improve_drug_id chem_name formula weight inCHIKey canSMILES isoSMILES
0 SMI_43 17-AAG C31H43N3O8 585.70 AYUNIORJHRXIBJ-TXHRRWQRSA-N CC1CC(C(C(C=C(C(C(C=CC=C(C(=O)NC2=CC(=O)C(=C(C... C[C@H]1C[C@@H]([C@@H]([C@H](/C=C(/[C@@H]([C@H]...
1 SMI_3222 A-674563 C22H22N4O 358.40 BPNUQXPIQBZCMR-IBGZPJMESA-N CC1=C2C=C(C=CC2=NN1)C3=CC(=CN=C3)OCC(CC4=CC=CC... CC1=C2C=C(C=CC2=NN1)C3=CC(=CN=C3)OC[C@H](CC4=C...
2 SMI_146 ABT-737 C42H45ClN6O5S2 813.40 HPLNQCPCUACXLM-PGUFJCEWSA-N CN(C)CCC(CSC1=CC=CC=C1)NC2=C(C=C(C=C2)S(=O)(=O... CN(C)CC[C@H](CSC1=CC=CC=C1)NC2=C(C=C(C=C2)S(=O...
3 SMI_57721 AKT Inhibitor IV C31H27IN4S 614.50 NAYRELMNTQSBIN-UHFFFAOYSA-M CC[N+]1=C(N(C2=C1C=C(C=C2)C3=NC4=CC=CC=C4S3)C5... CC[N+]1=C(N(C2=C1C=C(C=C2)C3=NC4=CC=CC=C4S3)C5...
4 SMI_57735 AMPK Inhibitor C24H25N5O 399.50 XHBVYDAKJHETMP-UHFFFAOYSA-N C1CCN(CC1)CCOC2=CC=C(C=C2)C3=CN4C(=C(C=N4)C5=C... C1CCN(CC1)CCOC2=CC=C(C=C2)C3=CN4C(=C(C=N4)C5=C...
... ... ... ... ... ... ... ...
35420 SMI_414 Venetoclax C45H50ClN7O7S 868.40 LQBVNQSMGBZMKD-UHFFFAOYSA-N CC1(CCC(=C(C1)C2=CC=C(C=C2)Cl)CN3CCN(CC3)C4=CC... CC1(CCC(=C(C1)C2=CC=C(C=C2)Cl)CN3CCN(CC3)C4=CC...
35421 SMI_1062 Vismodegib C19H14Cl2N2O3S 421.30 BPQMGSKTAYIVFO-UHFFFAOYSA-N CS(=O)(=O)C1=CC(=C(C=C1)C(=O)NC2=CC(=C(C=C2)Cl... CS(=O)(=O)C1=CC(=C(C=C1)C(=O)NC2=CC(=C(C=C2)Cl...
35422 SMI_3146 Volasertib C34H50N8O3 618.80 SXNJFOWDRLKDSF-XKHVUIRMSA-N CCC1C(=O)N(C2=CN=C(N=C2N1C(C)C)NC3=C(C=C(C=C3)... CC[C@@H]1C(=O)N(C2=CN=C(N=C2N1C(C)C)NC3=C(C=C(...
35423 SMI_1137 XAV-939 C14H11F3N2OS 312.31 KLGQSVMIPOVQAX-UHFFFAOYSA-N C1CSCC2=C1N=C(NC2=O)C3=CC=C(C=C3)C(F)(F)F C1CSCC2=C1N=C(NC2=O)C3=CC=C(C=C3)C(F)(F)F
35424 SMI_181 YM-155 C20H19BrN4O3 443.30 QBIYUDDJPRGKNJ-UHFFFAOYSA-M CC1=[N+](C2=C(N1CCOC)C(=O)C3=CC=CC=C3C2=O)CC4=... CC1=[N+](C2=C(N1CCOC)C(=O)C3=CC=CC=C3C2=O)CC4=...

35425 rows × 7 columns

In [7]:
beataml.experiments
Out[7]:
source improve_sample_id improve_drug_id study auc ic50 ec50 ec50se r2fit einf hs aac1 auc1 dss1
0 synapse 4634 SMI_43 BeatAML 0.5716 NaN 11.430 0.000 0.0000 0.5716 3.525 0.4284 0.5716 0.3649
1 synapse 4634 SMI_3222 BeatAML 0.6310 NaN 10.890 0.000 -0.0000 0.6310 3.171 0.3690 0.6310 0.2989
2 synapse 4634 SMI_146 BeatAML 0.7496 NaN 11.590 0.000 0.0000 0.7496 3.594 0.2504 0.7496 0.1671
3 synapse 4634 SMI_57721 BeatAML 0.7264 NaN 7.525 3.249 0.0277 0.5343 4.000 0.4609 0.5391 0.4010
4 synapse 4634 SMI_57735 BeatAML 0.5588 NaN 11.540 0.000 -0.0000 0.5588 3.611 0.4412 0.5588 0.3791
... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
38775 synapse 4346 SMI_414 BeatAML 0.4791 10.94 11.360 0.000 -0.0000 0.4791 3.294 0.5209 0.4791 0.4677
38776 synapse 4346 SMI_1062 BeatAML 0.9319 NaN 9.805 0.000 -0.0000 0.8638 0.000 0.0681 0.9319 0.0000
38777 synapse 4346 SMI_3146 BeatAML 0.4195 10.81 11.060 0.000 -0.0000 0.4195 3.247 0.5805 0.4195 0.5339
38778 synapse 4346 SMI_1137 BeatAML 0.8347 NaN 0.036 0.000 -0.0000 0.6693 0.000 0.1653 0.8347 0.0726
38779 synapse 4346 SMI_181 BeatAML 0.7405 NaN 11.370 0.000 -0.0000 0.7405 3.878 0.2595 0.7405 0.1773

38780 rows × 14 columns

Merge Drug and Experiments¶

This merge allows us to place isoSMILES, sample ID, drug ID and auc together.

In [8]:
merged_df = pd.merge(beataml.experiments[['improve_sample_id', 'improve_drug_id', 'auc']],
                     beataml.drugs[['improve_drug_id', 'isoSMILES']],
                     on='improve_drug_id',
                     how='inner').drop_duplicates()

merged_df = merged_df.dropna(subset=['isoSMILES'])
In [9]:
merged_df
Out[9]:
improve_sample_id improve_drug_id auc isoSMILES
0 4634 SMI_43 0.5716 C[C@H]1C[C@@H]([C@@H]([C@H](/C=C(/[C@@H]([C@H]...
284 4634 SMI_3222 0.6310 CC1=C2C=C(C=CC2=NN1)C3=CC(=CN=C3)OC[C@H](CC4=C...
569 4634 SMI_146 0.7496 CN(C)CC[C@H](CSC1=CC=CC=C1)NC2=C(C=C(C=C2)S(=O...
781 4634 SMI_57721 0.7264 CC[N+]1=C(N(C2=C1C=C(C=C2)C3=NC4=CC=CC=C4S3)C5...
846 4634 SMI_57735 0.5588 C1CCN(CC1)CCOC2=CC=C(C=C2)C3=CN4C(=C(C=N4)C5=C...
... ... ... ... ...
10668667 4346 SMI_414 0.4791 CC1(CCC(=C(C1)C2=CC=C(C=C2)Cl)CN3CCN(CC3)C4=CC...
10668944 4346 SMI_1062 0.9319 CS(=O)(=O)C1=CC(=C(C=C1)C(=O)NC2=CC(=C(C=C2)Cl...
10669230 4346 SMI_3146 0.4195 CC[C@@H]1C(=O)N(C2=CN=C(N=C2N1C(C)C)NC3=C(C=C(...
10669505 4346 SMI_1137 0.8347 C1CSCC2=C1N=C(NC2=O)C3=CC=C(C=C3)C(F)(F)F
10669748 4346 SMI_181 0.7405 CC1=[N+](C2=C(N1CCOC)C(=O)C3=CC=CC=C3C2=O)CC4=...

34850 rows × 4 columns

Using the RDKit package, we can convert IsoSMILES into a machine readable format¶

The chemical fingerprint is then mapped to improve_drug_id in the fingerprint_map. These values will be retrieved later. The merged_df is also subsetted to remove redundant information.

In [10]:
# Convert SMILES to Morgan fingerprints
def smiles_to_fingerprint(smiles):
    mol = Chem.MolFromSmiles(smiles)
    fingerprint = AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=1024)  # Adjust radius and nBits as needed
    fingerprint_array = np.array(fingerprint)
    return fingerprint_array

# smiles_to_fingerprint(merged_df.isoSMILES[0])
# # Apply the function to your SMILES column
merged_df['fingerprint'] = merged_df['isoSMILES'].apply(smiles_to_fingerprint)

fingerprint_map = merged_df[['fingerprint',"improve_drug_id"]]
merged_df = merged_df[["improve_sample_id","improve_drug_id","auc"]]
In [11]:
merged_df
Out[11]:
improve_sample_id improve_drug_id auc
0 4634 SMI_43 0.5716
284 4634 SMI_3222 0.6310
569 4634 SMI_146 0.7496
781 4634 SMI_57721 0.7264
846 4634 SMI_57735 0.5588
... ... ... ...
10668667 4346 SMI_414 0.4791
10668944 4346 SMI_1062 0.9319
10669230 4346 SMI_3146 0.4195
10669505 4346 SMI_1137 0.8347
10669748 4346 SMI_181 0.7405

34850 rows × 3 columns

In [12]:
fingerprint_map
Out[12]:
fingerprint improve_drug_id
0 [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, ... SMI_43
284 [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ... SMI_3222
569 [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, ... SMI_146
781 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ... SMI_57721
846 [0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, ... SMI_57735
... ... ...
10668667 [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, ... SMI_414
10668944 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ... SMI_1062
10669230 [0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, ... SMI_3146
10669505 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ... SMI_1137
10669748 [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, ... SMI_181

34850 rows × 2 columns

In [13]:
fingerprint_dict = fingerprint_map.set_index('improve_drug_id')['fingerprint'].to_dict()
fingerprint_dict
Out[13]:
{'SMI_43': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_3222': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_146': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_57721': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57735': array([0, 0, 1, ..., 0, 0, 0]),
 'SMI_1356': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_949': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_990': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_224': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_274': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_388': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_398': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_98': array([0, 0, 0, ..., 1, 0, 0]),
 'SMI_381': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_145': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_331': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_97': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1302': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_341': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_378': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_225': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_988': array([0, 0, 0, ..., 0, 0, 1]),
 'SMI_57': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_337': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_4071': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1303': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_58': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1003': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_230': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1134': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1397': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1393': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_384': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57737': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_52446': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_3033': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_342': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_95': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57722': array([0, 0, 1, ..., 0, 0, 0]),
 'SMI_52331': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1412': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_1842': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_37': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_57732': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_57724': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57728': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_64': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_338': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_56627': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_41054': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_376': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1060': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_221': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_755': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_125': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57730': array([0, 0, 0, ..., 1, 0, 0]),
 'SMI_3276': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_227': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1057': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_3125': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_318': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_231': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_184': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57723': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1312': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57726': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_148': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_233': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57731': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_53380': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_1108': array([1, 0, 0, ..., 0, 0, 0]),
 'SMI_283': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1403': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1056': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_339': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_63': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_262': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1063': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_276': array([0, 1, 1, ..., 0, 0, 0]),
 'SMI_144': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_55709': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_52447': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_50': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_272': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_232': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57736': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_236': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_223': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_365': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_102': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_51': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_107': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_374': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_369': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_373': array([0, 1, 1, ..., 0, 0, 0]),
 'SMI_188': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_54925': array([0, 0, 1, ..., 0, 0, 0]),
 'SMI_1012': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_995': array([0, 0, 0, ..., 0, 1, 0]),
 'SMI_57733': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_101': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_3148': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1062': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1137': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_181': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_57727': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1963': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_375': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_261': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_278': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1155': array([0, 1, 1, ..., 0, 0, 0]),
 'SMI_1109': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_10': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_52922': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_2582': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_75': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1777': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_1107': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_55184': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_3166': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_410': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_256': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_265': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1327': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_106': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_140': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_1055': array([0, 0, 0, ..., 0, 1, 0]),
 'SMI_60': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_281': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57729': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_1952': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_285': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_414': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_3146': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_1061': array([0, 0, 1, ..., 0, 0, 0]),
 'SMI_1104': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57720': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_349': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_167': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_28': array([1, 1, 0, ..., 0, 0, 0]),
 'SMI_165': array([1, 0, 0, ..., 0, 0, 0]),
 'SMI_1859': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_54398': array([0, 0, 1, ..., 0, 0, 0]),
 'SMI_1984': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_182': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57734': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_53718': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_2064': array([0, 1, 0, ..., 0, 0, 0]),
 'SMI_299': array([0, 0, 0, ..., 0, 0, 0]),
 'SMI_57725': array([0, 0, 0, ..., 0, 0, 0])}

Now we merge in our Proteomics and Transcriptomics Data.¶

This method of merging drops all samples that do not have both transcriptomics and proteomics data.
This reduces our sample size to 15 and allows the tutorial to run much faster on low powered machines.

In [14]:
# Merge merged_df with transcriptomics_df based on improve_sample_id
tp = pd.merge(beataml.transcriptomics,
                     beataml.proteomics[['improve_sample_id', 'proteomics', 'entrez_id']],
                     on=['improve_sample_id','entrez_id'],
                     how='inner').drop_duplicates()

merged_df = pd.merge(merged_df,
                     tp[['improve_sample_id', 'transcriptomics','proteomics', 'entrez_id']],
                     on='improve_sample_id',
                     how='inner').drop_duplicates()
In [15]:
merged_df
Out[15]:
improve_sample_id improve_drug_id auc transcriptomics proteomics entrez_id
0 3986 SMI_236 0.0000 7.068854 -0.135 8813.0
1 3986 SMI_236 0.0000 3.859919 -0.300 6359.0
2 3986 SMI_236 0.0000 3.859919 -0.300 57147.0
3 3986 SMI_236 0.0000 9.634914 1.140 2268.0
4 3986 SMI_236 0.0000 -1.401864 -0.174 3075.0
... ... ... ... ... ... ...
8148099 4007 SMI_57727 0.9307 1.519607 -0.192 80755.0
8148100 4007 SMI_57727 0.9307 2.088872 -0.260 55957.0
8148101 4007 SMI_57727 0.9307 2.154935 0.865 6023.0
8148102 4007 SMI_57727 0.9307 7.165478 0.472 11165.0
8148103 4007 SMI_57727 0.9307 -0.435228 0.365 6047.0

8148104 rows × 6 columns

Impute Missing Values for Proteomics and Transcriptomics.¶

Here we recommend to modify the imputation method. Taking the global mean may not be apppropriate for your data. This is just simple method and reminder that this may need to be done.
In this example, no transcriptomics values are actually imputed, and less than 3500 / 8148104 (.04%) of proteomics values are imputed.

In [ ]:
#Impute missing proteomics (and transcriptomics if there are any) based on global mean. 
columns_to_fill = ['transcriptomics', 'proteomics']

for i in columns_to_fill:
    if i in merged_df.columns[merged_df.isnull().any(axis=0)]:
        merged_df[i].fillna(merged_df[i].mean(), inplace=True)
In [17]:
merged_df
Out[17]:
improve_sample_id improve_drug_id auc transcriptomics proteomics entrez_id
0 3986 SMI_236 0.0000 7.068854 -0.135 8813.0
1 3986 SMI_236 0.0000 3.859919 -0.300 6359.0
2 3986 SMI_236 0.0000 3.859919 -0.300 57147.0
3 3986 SMI_236 0.0000 9.634914 1.140 2268.0
4 3986 SMI_236 0.0000 -1.401864 -0.174 3075.0
... ... ... ... ... ... ...
8148099 4007 SMI_57727 0.9307 1.519607 -0.192 80755.0
8148100 4007 SMI_57727 0.9307 2.088872 -0.260 55957.0
8148101 4007 SMI_57727 0.9307 2.154935 0.865 6023.0
8148102 4007 SMI_57727 0.9307 7.165478 0.472 11165.0
8148103 4007 SMI_57727 0.9307 -0.435228 0.365 6047.0

8148104 rows × 6 columns

In [18]:
merged_df['fingerprint'] = merged_df['improve_drug_id'].map(fingerprint_dict)

Add the Chemical Structure (fingerprint) Data¶

Now all of our data is gathered in one place. However, we will need to do some restructuring to get this into a sutiable input for our deep learning model

In [19]:
merged_df
Out[19]:
improve_sample_id improve_drug_id auc transcriptomics proteomics entrez_id fingerprint
0 3986 SMI_236 0.0000 7.068854 -0.135 8813.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
1 3986 SMI_236 0.0000 3.859919 -0.300 6359.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
2 3986 SMI_236 0.0000 3.859919 -0.300 57147.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
3 3986 SMI_236 0.0000 9.634914 1.140 2268.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
4 3986 SMI_236 0.0000 -1.401864 -0.174 3075.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
... ... ... ... ... ... ... ...
8148099 4007 SMI_57727 0.9307 1.519607 -0.192 80755.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
8148100 4007 SMI_57727 0.9307 2.088872 -0.260 55957.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
8148101 4007 SMI_57727 0.9307 2.154935 0.865 6023.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
8148102 4007 SMI_57727 0.9307 7.165478 0.472 11165.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
8148103 4007 SMI_57727 0.9307 -0.435228 0.365 6047.0 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...

8148104 rows × 7 columns

Scaling the Data (optional)¶

Scale the transcriptomics and proteomics data to the same 0-1 scale if you wish them to have equal weights in the Deep Learning Model.

In [20]:
scaler = MinMaxScaler()
merged_df[["transcriptomics","proteomics"]] = scaler.fit_transform(merged_df[["transcriptomics","proteomics"]])

Restructure the Transcriptomics and Proteomics Data¶

These are restructured into lists within the Pandas Dataframe by Sample. Each array must be ordered in the same way be entrez id. In this case, we sued a Hashmap for time reduction. This can be done through however method you wish to use. Just make sure that each list of proteomics / transcriptomics values is linked to an entrez_id list.

In [21]:
#Create A Hashmap of {improve_sample_id: {improve_drug_id: { entrez_id: transcriptomics, proteomics, auc]}}

data_dict = {}

# Iterate over the DataFrame rows to populate the nested dictionary
for index, row in merged_df.iterrows():
    improve_sample_id = row['improve_sample_id']
    improve_drug_id = row['improve_drug_id']
    auc = row['auc']
    transcriptomics = row['transcriptomics']
    proteomics = row['proteomics']
    entrez_id = row['entrez_id']
    # print(improve_sample_id,improve_drug_id,auc,transcriptomics,proteomics,entrez_id)

    # Initialize improve_sample_id key if not present
    if improve_sample_id not in data_dict:
        data_dict[improve_sample_id] = {}
    # print(data_dict)
#     # Initialize improve_drug_id key if not present
    if improve_drug_id not in data_dict[improve_sample_id]:
        data_dict[improve_sample_id][improve_drug_id] = {}
#     # Append data to the nested dictionary
    if entrez_id not in data_dict[improve_sample_id][improve_drug_id]:
        data_dict[improve_sample_id][improve_drug_id][entrez_id] = {
            'transcriptomics': int,
            'proteomics': int,
            'auc': auc
        }
    data_dict[improve_sample_id][improve_drug_id][entrez_id]['transcriptomics'] = transcriptomics 
    data_dict[improve_sample_id][improve_drug_id][entrez_id]['proteomics'] = proteomics

Write [transcriptomics], [proteomics], and auc data to aggregated_df.¶

The key takeaway here is that each of the input variables should be in a list format.

In [22]:
#Write [transcriptomics], [proteomics], and auc data to aggregated_df. These are all written in the same order as the unique_entrez_ids set.

# # Extract the list of unique entrez_ids
unique_entrez_ids = set(merged_df.entrez_id.unique())

# # Construct the final DataFrame
final_data = []
for improve_sample_id, sample_id_data in data_dict.items():
    for improve_drug_id, drug_id_data in sample_id_data.items():
        transcriptomics = []
        proteomics = []
        for entrez_id in unique_entrez_ids:
            if entrez_id in drug_id_data:
                transcriptomics.append(drug_id_data[entrez_id]['transcriptomics'])
                proteomics.append(drug_id_data[entrez_id]['proteomics'])
                auc = drug_id_data[entrez_id]['auc']
        final_data.append([auc, transcriptomics, proteomics, improve_sample_id,improve_drug_id])

# # Create the DataFrame
aggregated_df = pd.DataFrame(final_data, columns=['auc', 'transcriptomics', 'proteomics', 'improve_sample_id', "improve_drug_id"])
aggregated_df
Out[22]:
auc transcriptomics proteomics improve_sample_id improve_drug_id
0 0.0000 [0.23608062994353704, 0.5037978202291047, 0.48... [0.3384677419354839, 0.37193548387096775, 0.38... 3986 SMI_236
1 0.7431 [0.23608062994353704, 0.5037978202291047, 0.48... [0.3384677419354839, 0.37193548387096775, 0.38... 3986 SMI_223
2 0.6135 [0.23608062994353704, 0.5037978202291047, 0.48... [0.3384677419354839, 0.37193548387096775, 0.38... 3986 SMI_365
3 0.7425 [0.18000067030285077, 0.4920917394967491, 0.47... [0.43870967741935485, 0.3573387096774194, 0.38... 3987 SMI_54925
4 0.0000 [0.18000067030285077, 0.4920917394967491, 0.47... [0.43870967741935485, 0.3573387096774194, 0.38... 3987 SMI_1012
... ... ... ... ... ...
988 0.8771 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... 4007 SMI_3148
989 1.0000 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... 4007 SMI_1062
990 0.5961 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... 4007 SMI_1137
991 0.9997 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... 4007 SMI_181
992 0.9307 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... 4007 SMI_57727

993 rows × 5 columns

The Fingerprint Variable dropped out. It is added back in here¶

In [23]:
# Map in the fingerprint by improve_drug_id. Remove improve_drug_id.
aggregated_df['fingerprint'] = aggregated_df['improve_drug_id'].map(fingerprint_dict)
aggregated_df = aggregated_df[["auc","transcriptomics","proteomics","fingerprint"]]

View the Training Data Format.¶

AUC will be trained, validated, evaluated by the transcriptomics, proteomics and fingerprint variables

In [24]:
# Training/Test/Validation Data is all under the name training_data.
training_data = aggregated_df
training_data
Out[24]:
auc transcriptomics proteomics fingerprint
0 0.0000 [0.23608062994353704, 0.5037978202291047, 0.48... [0.3384677419354839, 0.37193548387096775, 0.38... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
1 0.7431 [0.23608062994353704, 0.5037978202291047, 0.48... [0.3384677419354839, 0.37193548387096775, 0.38... [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 1, ...
2 0.6135 [0.23608062994353704, 0.5037978202291047, 0.48... [0.3384677419354839, 0.37193548387096775, 0.38... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, ...
3 0.7425 [0.18000067030285077, 0.4920917394967491, 0.47... [0.43870967741935485, 0.3573387096774194, 0.38... [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
4 0.0000 [0.18000067030285077, 0.4920917394967491, 0.47... [0.43870967741935485, 0.3573387096774194, 0.38... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...
... ... ... ... ...
988 0.8771 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
989 1.0000 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
990 0.5961 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
991 0.9997 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, ...
992 0.9307 [0.29431256693950825, 0.5401532345082279, 0.47... [0.34040322580645166, 0.3787903225806452, 0.37... [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...

993 rows × 4 columns

Split The Data into Training, Validating, and Testing Data.¶

The data here is split into 80% training, 10% validating, and 10% testing sets.

In [25]:
# Splitting the merged_data DataFrame into features (X) and target variable (y)
X = training_data[['transcriptomics', 'proteomics',"fingerprint"]]  # These columns are your features
y = training_data['auc']  # Using 'auc' column as the target variable
In [26]:
# Splitting data into training, validation, and testing sets
X_train_val, X_test, y_train_val, y_test = train_test_split(X, y, test_size=0.1, random_state=42)
X_train, X_val, y_train, y_val = train_test_split(X_train_val, y_train_val, test_size=0.1111, random_state=42) # 0.1111 * 90 = 10%

Set the Model Parameters¶

Here you may modify the inputs, layers, activation functions, and outputs of the model.
See if you can improve upon our basic setup.
A note - running some of these steps multiple times in the jupyter notebook may result in odd issues, please be careful with this behaivor.

In [27]:
#These are the array lengths of each feature set.
num_proteomics = 7284
num_transcriptomics = 7284  
num_fingerprint = 1024  

#Define inputs
transcriptomics_input = keras.Input(
    shape=(num_transcriptomics,), name="transcriptomics"
)  

proteomics_input = keras.Input(
    shape=(num_proteomics,), name="proteomics"
)  

fingerprint_input = keras.Input(
    shape=(num_fingerprint,), name="fingerprint"
)  

# You may adjust these layers and activation functions to get better (or worse) results.
# Merge all available features into a single large vector via concatenation
x = layers.concatenate([transcriptomics_input, proteomics_input, fingerprint_input])
x = layers.Dense(16, activation='relu')(x)
x = layers.Dense(64, activation="relu")(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(32, activation='relu')(x)
x = layers.Dense(12, activation='relu')(x)

# Priority prediction is here. You may add layers (and an output) after this too.
# This just allows for an early prediction in case you have a large datset and want preliminary results.
priority_pred = layers.Dense(1, name="priority",activation='relu')(x)

# Instantiate an end-to-end model predicting both priority and department
model = keras.Model(
    inputs=[transcriptomics_input, proteomics_input, fingerprint_input],
    outputs={"priority": priority_pred},
)
In [28]:
keras.utils.plot_model(model, "multi_input_model.png", show_shapes=True)

You must install pydot (`pip install pydot`) and install graphviz (see instructions at https://graphviz.gitlab.io/download/) for plot_model to work.

Compile the Model¶

You may wish to test out different optimizers, learning rates, and metrics.

In [ ]:
model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=0.0001),
    loss={
        "priority": keras.losses.MeanSquaredError(),
    },
    metrics=[MeanAbsoluteError()],
    loss_weights={"priority": 1},
)

Run your Model !¶

The model does not allow for lists in the inputs, so these are converted to Arrays.
Also ensure that the length of each array is the same on an input variable basis.
For example, if the proteomics variable has two different length arrays, you get an uninformative error message.

  • In our case, we subsetted out many samples and imputed some values to avoid this issue. You may need to add padding (empty values) to some lists to ensure that this issue is avoided. Just make sure that this padding does not shift your proteomics/transcriptomics values out of order of the entrez_id mapping list.

Feel free to change the epochs and batch size.

In [30]:
history = model.fit(
    [np.array(X_train['transcriptomics'].tolist()),
     np.array(X_train['proteomics'].tolist()),
     np.array(X_train['fingerprint'].tolist())
    ], 
    y_train, 
    epochs=100, 
    batch_size=32, 
    validation_data=(
        [
            np.array(X_val['transcriptomics'].tolist()), 
            np.array(X_val['proteomics'].tolist()), 
            np.array(X_val['fingerprint'].tolist())
        ], 
        y_val)
)

Epoch 1/100

2024-03-07 22:06:39.578569: W tensorflow/tsl/platform/profile_utils/cpu_utils.cc:128] Failed to get CPU frequency: 0 Hz

25/25 [==============================] - 1s 16ms/step - loss: 0.3187 - mean_absolute_error: 0.4867 - val_loss: 0.2245 - val_mean_absolute_error: 0.4414
Epoch 2/100
25/25 [==============================] - 0s 5ms/step - loss: 0.2099 - mean_absolute_error: 0.3757 - val_loss: 0.1138 - val_mean_absolute_error: 0.3021
Epoch 3/100
25/25 [==============================] - 0s 5ms/step - loss: 0.1783 - mean_absolute_error: 0.3399 - val_loss: 0.0712 - val_mean_absolute_error: 0.2290
Epoch 4/100
25/25 [==============================] - 0s 4ms/step - loss: 0.1419 - mean_absolute_error: 0.3088 - val_loss: 0.0421 - val_mean_absolute_error: 0.1584
Epoch 5/100
25/25 [==============================] - 0s 5ms/step - loss: 0.1281 - mean_absolute_error: 0.2879 - val_loss: 0.0451 - val_mean_absolute_error: 0.1690
Epoch 6/100
25/25 [==============================] - 0s 5ms/step - loss: 0.1105 - mean_absolute_error: 0.2707 - val_loss: 0.0426 - val_mean_absolute_error: 0.1616
Epoch 7/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0958 - mean_absolute_error: 0.2483 - val_loss: 0.0388 - val_mean_absolute_error: 0.1455
Epoch 8/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0988 - mean_absolute_error: 0.2488 - val_loss: 0.0382 - val_mean_absolute_error: 0.1433
Epoch 9/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0867 - mean_absolute_error: 0.2319 - val_loss: 0.0379 - val_mean_absolute_error: 0.1409
Epoch 10/100
25/25 [==============================] - 0s 5ms/step - loss: 0.0809 - mean_absolute_error: 0.2240 - val_loss: 0.0400 - val_mean_absolute_error: 0.1426
Epoch 11/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0820 - mean_absolute_error: 0.2302 - val_loss: 0.0414 - val_mean_absolute_error: 0.1450
Epoch 12/100
25/25 [==============================] - 0s 7ms/step - loss: 0.0767 - mean_absolute_error: 0.2233 - val_loss: 0.0391 - val_mean_absolute_error: 0.1558
Epoch 13/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0804 - mean_absolute_error: 0.2246 - val_loss: 0.0366 - val_mean_absolute_error: 0.1444
Epoch 14/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0736 - mean_absolute_error: 0.2140 - val_loss: 0.0364 - val_mean_absolute_error: 0.1438
Epoch 15/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0742 - mean_absolute_error: 0.2169 - val_loss: 0.0420 - val_mean_absolute_error: 0.1664
Epoch 16/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0803 - mean_absolute_error: 0.2277 - val_loss: 0.0417 - val_mean_absolute_error: 0.1655
Epoch 17/100
25/25 [==============================] - 0s 6ms/step - loss: 0.0665 - mean_absolute_error: 0.2101 - val_loss: 0.0380 - val_mean_absolute_error: 0.1364
Epoch 18/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0641 - mean_absolute_error: 0.2001 - val_loss: 0.0376 - val_mean_absolute_error: 0.1515
Epoch 19/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0650 - mean_absolute_error: 0.2043 - val_loss: 0.0381 - val_mean_absolute_error: 0.1360
Epoch 20/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0579 - mean_absolute_error: 0.1924 - val_loss: 0.0430 - val_mean_absolute_error: 0.1476
Epoch 21/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0610 - mean_absolute_error: 0.1939 - val_loss: 0.0606 - val_mean_absolute_error: 0.2099
Epoch 22/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0576 - mean_absolute_error: 0.1909 - val_loss: 0.0352 - val_mean_absolute_error: 0.1401
Epoch 23/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0554 - mean_absolute_error: 0.1864 - val_loss: 0.0348 - val_mean_absolute_error: 0.1357
Epoch 24/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0578 - mean_absolute_error: 0.1866 - val_loss: 0.0401 - val_mean_absolute_error: 0.1606
Epoch 25/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0553 - mean_absolute_error: 0.1847 - val_loss: 0.0380 - val_mean_absolute_error: 0.1538
Epoch 26/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0481 - mean_absolute_error: 0.1711 - val_loss: 0.0354 - val_mean_absolute_error: 0.1427
Epoch 27/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0528 - mean_absolute_error: 0.1798 - val_loss: 0.0383 - val_mean_absolute_error: 0.1354
Epoch 28/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0516 - mean_absolute_error: 0.1753 - val_loss: 0.0355 - val_mean_absolute_error: 0.1435
Epoch 29/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0506 - mean_absolute_error: 0.1744 - val_loss: 0.0349 - val_mean_absolute_error: 0.1393
Epoch 30/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0567 - mean_absolute_error: 0.1873 - val_loss: 0.0591 - val_mean_absolute_error: 0.2049
Epoch 31/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0544 - mean_absolute_error: 0.1824 - val_loss: 0.0346 - val_mean_absolute_error: 0.1352
Epoch 32/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0459 - mean_absolute_error: 0.1687 - val_loss: 0.0368 - val_mean_absolute_error: 0.1491
Epoch 33/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0447 - mean_absolute_error: 0.1643 - val_loss: 0.0348 - val_mean_absolute_error: 0.1342
Epoch 34/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0425 - mean_absolute_error: 0.1626 - val_loss: 0.0347 - val_mean_absolute_error: 0.1354
Epoch 35/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0422 - mean_absolute_error: 0.1596 - val_loss: 0.0381 - val_mean_absolute_error: 0.1530
Epoch 36/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0454 - mean_absolute_error: 0.1644 - val_loss: 0.0350 - val_mean_absolute_error: 0.1382
Epoch 37/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0432 - mean_absolute_error: 0.1594 - val_loss: 0.0375 - val_mean_absolute_error: 0.1358
Epoch 38/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0435 - mean_absolute_error: 0.1617 - val_loss: 0.0348 - val_mean_absolute_error: 0.1345
Epoch 39/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0390 - mean_absolute_error: 0.1518 - val_loss: 0.0349 - val_mean_absolute_error: 0.1369
Epoch 40/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0416 - mean_absolute_error: 0.1586 - val_loss: 0.0365 - val_mean_absolute_error: 0.1353
Epoch 41/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0380 - mean_absolute_error: 0.1496 - val_loss: 0.0352 - val_mean_absolute_error: 0.1354
Epoch 42/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0406 - mean_absolute_error: 0.1561 - val_loss: 0.0373 - val_mean_absolute_error: 0.1367
Epoch 43/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0365 - mean_absolute_error: 0.1461 - val_loss: 0.0351 - val_mean_absolute_error: 0.1364
Epoch 44/100
25/25 [==============================] - 0s 6ms/step - loss: 0.0397 - mean_absolute_error: 0.1514 - val_loss: 0.0364 - val_mean_absolute_error: 0.1448
Epoch 45/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0391 - mean_absolute_error: 0.1487 - val_loss: 0.0363 - val_mean_absolute_error: 0.1364
Epoch 46/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0402 - mean_absolute_error: 0.1519 - val_loss: 0.0361 - val_mean_absolute_error: 0.1366
Epoch 47/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0392 - mean_absolute_error: 0.1544 - val_loss: 0.0362 - val_mean_absolute_error: 0.1411
Epoch 48/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0412 - mean_absolute_error: 0.1554 - val_loss: 0.0357 - val_mean_absolute_error: 0.1373
Epoch 49/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0386 - mean_absolute_error: 0.1504 - val_loss: 0.0372 - val_mean_absolute_error: 0.1377
Epoch 50/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0357 - mean_absolute_error: 0.1443 - val_loss: 0.0370 - val_mean_absolute_error: 0.1373
Epoch 51/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0387 - mean_absolute_error: 0.1513 - val_loss: 0.0429 - val_mean_absolute_error: 0.1456
Epoch 52/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0415 - mean_absolute_error: 0.1567 - val_loss: 0.0508 - val_mean_absolute_error: 0.1599
Epoch 53/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0410 - mean_absolute_error: 0.1550 - val_loss: 0.0362 - val_mean_absolute_error: 0.1399
Epoch 54/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0349 - mean_absolute_error: 0.1401 - val_loss: 0.0371 - val_mean_absolute_error: 0.1441
Epoch 55/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0441 - mean_absolute_error: 0.1634 - val_loss: 0.0546 - val_mean_absolute_error: 0.1954
Epoch 56/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0404 - mean_absolute_error: 0.1574 - val_loss: 0.0385 - val_mean_absolute_error: 0.1388
Epoch 57/100
25/25 [==============================] - 0s 7ms/step - loss: 0.0332 - mean_absolute_error: 0.1395 - val_loss: 0.0406 - val_mean_absolute_error: 0.1418
Epoch 58/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0370 - mean_absolute_error: 0.1474 - val_loss: 0.0523 - val_mean_absolute_error: 0.1635
Epoch 59/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0356 - mean_absolute_error: 0.1432 - val_loss: 0.0447 - val_mean_absolute_error: 0.1492
Epoch 60/100
25/25 [==============================] - 0s 5ms/step - loss: 0.0360 - mean_absolute_error: 0.1446 - val_loss: 0.0370 - val_mean_absolute_error: 0.1425
Epoch 61/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0357 - mean_absolute_error: 0.1438 - val_loss: 0.0379 - val_mean_absolute_error: 0.1387
Epoch 62/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0361 - mean_absolute_error: 0.1469 - val_loss: 0.0368 - val_mean_absolute_error: 0.1416
Epoch 63/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0341 - mean_absolute_error: 0.1407 - val_loss: 0.0395 - val_mean_absolute_error: 0.1411
Epoch 64/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0364 - mean_absolute_error: 0.1440 - val_loss: 0.0432 - val_mean_absolute_error: 0.1468
Epoch 65/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0333 - mean_absolute_error: 0.1380 - val_loss: 0.0367 - val_mean_absolute_error: 0.1389
Epoch 66/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0346 - mean_absolute_error: 0.1425 - val_loss: 0.0389 - val_mean_absolute_error: 0.1404
Epoch 67/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0374 - mean_absolute_error: 0.1482 - val_loss: 0.0396 - val_mean_absolute_error: 0.1413
Epoch 68/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0340 - mean_absolute_error: 0.1386 - val_loss: 0.0366 - val_mean_absolute_error: 0.1403
Epoch 69/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0351 - mean_absolute_error: 0.1398 - val_loss: 0.0378 - val_mean_absolute_error: 0.1390
Epoch 70/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0330 - mean_absolute_error: 0.1348 - val_loss: 0.0376 - val_mean_absolute_error: 0.1392
Epoch 71/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0339 - mean_absolute_error: 0.1389 - val_loss: 0.0383 - val_mean_absolute_error: 0.1402
Epoch 72/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0322 - mean_absolute_error: 0.1342 - val_loss: 0.0375 - val_mean_absolute_error: 0.1396
Epoch 73/100
25/25 [==============================] - 0s 5ms/step - loss: 0.0333 - mean_absolute_error: 0.1385 - val_loss: 0.0410 - val_mean_absolute_error: 0.1439
Epoch 74/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0312 - mean_absolute_error: 0.1330 - val_loss: 0.0374 - val_mean_absolute_error: 0.1432
Epoch 75/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0328 - mean_absolute_error: 0.1367 - val_loss: 0.0448 - val_mean_absolute_error: 0.1494
Epoch 76/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0318 - mean_absolute_error: 0.1333 - val_loss: 0.0399 - val_mean_absolute_error: 0.1425
Epoch 77/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0304 - mean_absolute_error: 0.1323 - val_loss: 0.0549 - val_mean_absolute_error: 0.1687
Epoch 78/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0343 - mean_absolute_error: 0.1414 - val_loss: 0.0498 - val_mean_absolute_error: 0.1582
Epoch 79/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0340 - mean_absolute_error: 0.1409 - val_loss: 0.0525 - val_mean_absolute_error: 0.1639
Epoch 80/100
25/25 [==============================] - 0s 4ms/step - loss: 0.0377 - mean_absolute_error: 0.1504 - val_loss: 0.0481 - val_mean_absolute_error: 0.1555
Epoch 81/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0335 - mean_absolute_error: 0.1385 - val_loss: 0.0416 - val_mean_absolute_error: 0.1448
Epoch 82/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0305 - mean_absolute_error: 0.1309 - val_loss: 0.0388 - val_mean_absolute_error: 0.1411
Epoch 83/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0328 - mean_absolute_error: 0.1383 - val_loss: 0.0377 - val_mean_absolute_error: 0.1402
Epoch 84/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0319 - mean_absolute_error: 0.1346 - val_loss: 0.0377 - val_mean_absolute_error: 0.1406
Epoch 85/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0313 - mean_absolute_error: 0.1343 - val_loss: 0.0402 - val_mean_absolute_error: 0.1432
Epoch 86/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0295 - mean_absolute_error: 0.1294 - val_loss: 0.0370 - val_mean_absolute_error: 0.1422
Epoch 87/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0299 - mean_absolute_error: 0.1296 - val_loss: 0.0372 - val_mean_absolute_error: 0.1411
Epoch 88/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0315 - mean_absolute_error: 0.1334 - val_loss: 0.0368 - val_mean_absolute_error: 0.1437
Epoch 89/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0320 - mean_absolute_error: 0.1355 - val_loss: 0.0401 - val_mean_absolute_error: 0.1433
Epoch 90/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0295 - mean_absolute_error: 0.1295 - val_loss: 0.0377 - val_mean_absolute_error: 0.1409
Epoch 91/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0287 - mean_absolute_error: 0.1268 - val_loss: 0.0381 - val_mean_absolute_error: 0.1412
Epoch 92/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0321 - mean_absolute_error: 0.1355 - val_loss: 0.0456 - val_mean_absolute_error: 0.1517
Epoch 93/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0311 - mean_absolute_error: 0.1331 - val_loss: 0.0415 - val_mean_absolute_error: 0.1456
Epoch 94/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0310 - mean_absolute_error: 0.1324 - val_loss: 0.0411 - val_mean_absolute_error: 0.1450
Epoch 95/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0296 - mean_absolute_error: 0.1299 - val_loss: 0.0429 - val_mean_absolute_error: 0.1474
Epoch 96/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0299 - mean_absolute_error: 0.1307 - val_loss: 0.0374 - val_mean_absolute_error: 0.1414
Epoch 97/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0296 - mean_absolute_error: 0.1288 - val_loss: 0.0449 - val_mean_absolute_error: 0.1504
Epoch 98/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0317 - mean_absolute_error: 0.1353 - val_loss: 0.0381 - val_mean_absolute_error: 0.1505
Epoch 99/100
25/25 [==============================] - 0s 3ms/step - loss: 0.0303 - mean_absolute_error: 0.1298 - val_loss: 0.0413 - val_mean_absolute_error: 0.1455
Epoch 100/100
25/25 [==============================] - 0s 2ms/step - loss: 0.0305 - mean_absolute_error: 0.1282 - val_loss: 0.0408 - val_mean_absolute_error: 0.1447

Evaluate your Model¶

Here we use the test data to evaluate our model

In [31]:
losses = model.evaluate(
    [np.array(X_test['transcriptomics'].tolist()),
     np.array(X_test['proteomics'].tolist()),
     np.array(X_test['fingerprint'].tolist())
    ], 
    y_test, 
    return_dict=True) 
print(losses)

4/4 [==============================] - 0s 1ms/step - loss: 0.0218 - mean_absolute_error: 0.1116
{'loss': 0.021844929084181786, 'mean_absolute_error': 0.11160688102245331}

Plot Loss Function and Error of the Model¶

These show how your model (hopefully) improve during training.

In [111]:
# summarize history for accuracy
plt.plot(history.history['val_mean_absolute_error'])
plt.plot(history.history['mean_absolute_error'])
plt.title('Mean Absolute Error of Deep Learning Model')
plt.ylabel('Mean Absolute Error')
plt.xlabel('epoch')
plt.legend(['Validation Set','Training Set',], loc='upper right')
plt.show()
# summarize history for loss
plt.plot(history.history['val_loss'])
plt.plot(history.history['loss'])
plt.title('Loss Function of Deep Learning Model')
plt.ylabel('Loss')
plt.xlabel('epoch')
plt.legend(['Validation Set','Training Set'], loc='upper right')
plt.show()

Predict Results using Your Model¶

Here we use our test data again, but you'd likely want to apply this to other data. The inputs (and thier format) must match the model that you trained above.

Below we provide some plots and summary statistics for evaluating true vs predicted values.
Try to create a model better than ours!

In [33]:
# make a prediction
predict = model.predict([np.array(X_test['transcriptomics'].tolist()),
     np.array(X_test['proteomics'].tolist()),
     np.array(X_test['fingerprint'].tolist())
    ])

4/4 [==============================] - 0s 1ms/step
In [93]:
new_df = pd.DataFrame({
    'Predicted Values': predict["priority"].tolist(),
    'True Values': y_test
})
new_df['Predicted Values'] = new_df['Predicted Values'].apply(lambda x: x[0] if isinstance(x, list) and len(x) == 1 else x)
sorted_df = new_df.sort_values(by='True Values', ascending=True)  # Change 'ascending' to False for descending order
sorted_df
Out[93]:
Predicted Values True Values
830 0.689598 0.0000
198 0.613777 0.3066
10 0.576904 0.3380
355 0.569279 0.3572
656 0.576972 0.4010
... ... ...
485 0.907541 0.9939
371 0.961458 0.9997
323 0.957583 1.0000
430 0.911885 1.0000
362 0.750975 1.0000

100 rows × 2 columns

In [113]:
true_values = np.array(sorted_df["True Values"])
predicted_values = np.array(sorted_df["Predicted Values"])

# Plot the true values and predicted values
plt.plot(true_values)
plt.plot(predicted_values)

# Fit trendlines
true_values_trendline = LinearRegression().fit(np.arange(len(true_values)).reshape(-1, 1), true_values.reshape(-1, 1)).predict(np.arange(len(true_values)).reshape(-1, 1))
predicted_values_trendline = LinearRegression().fit(np.arange(len(predicted_values)).reshape(-1, 1), predicted_values.reshape(-1, 1)).predict(np.arange(len(predicted_values)).reshape(-1, 1))

# Plot trendlines
plt.plot(true_values_trendline, linestyle='--', color='blue', alpha=0.3)
plt.plot(predicted_values_trendline, linestyle='--', color='orange', alpha=0.3)

# Customize plot
plt.title('Predicted AUC vs True AUC')
plt.ylabel('Drug Response AUC')
plt.xlabel('Values')
plt.legend(['True Values', 'Predicted Values', 'True Values Trendline', 'Predicted Values Trendline'], loc='lower right')
plt.show()
In [112]:
true_values = np.array(new_df["True Values"])
predicted_values = np.array(new_df["Predicted Values"])

# Plot the true values and predicted values
plt.plot(true_values)
plt.plot(predicted_values)

# Fit trendlines
true_values_trendline = LinearRegression().fit(np.arange(len(true_values)).reshape(-1, 1), true_values.reshape(-1, 1)).predict(np.arange(len(true_values)).reshape(-1, 1))
predicted_values_trendline = LinearRegression().fit(np.arange(len(predicted_values)).reshape(-1, 1), predicted_values.reshape(-1, 1)).predict(np.arange(len(predicted_values)).reshape(-1, 1))


# Customize plot
plt.title('Predicted AUC vs True AUC')
plt.ylabel('Drug Response AUC')
plt.xlabel('Values')
plt.legend(['True Values', 'Predicted Values', 'True Values Trendline', 'Predicted Values Trendline'], loc='lower right')
plt.show()
In [102]:
# Calculate Summary Statistics


# Calculate Mean Absolute Error (MAE)
mae = mean_absolute_error(new_df['True Values'], new_df['Predicted Values'])

# Calculate Root Mean Squared Error (RMSE)
rmse = np.sqrt(mean_squared_error(new_df['True Values'], new_df['Predicted Values']))

# Calculate R-squared (R2) score
r2 = r2_score(new_df['True Values'], new_df['Predicted Values'])

summary_statistics = new_df.describe()

# Print the statistics
print("Mean Absolute Error (MAE):", mae)
print("Root Mean Squared Error (RMSE):", rmse)
print("R-squared (R2) Score:", r2)
print("\n")
# Print summary statistics
print(summary_statistics)

Mean Absolute Error (MAE): 0.11160688844966887
Root Mean Squared Error (RMSE): 0.14780030079805684
R-squared (R2) Score: 0.37366931237133216


       Predicted Values  True Values
count        100.000000   100.000000
mean           0.731800     0.723970
std            0.128604     0.187696
min            0.477577     0.000000
25%            0.639343     0.616050
50%            0.733571     0.736050
75%            0.838911     0.869575
max            0.961458     1.000000
In [108]:
#Side by side comparison for first 50 values.
new_df[0:50]
Out[108]:
Predicted Values True Values
918 0.777755 0.7367
525 0.716599 0.6711
567 0.684600 0.8426
656 0.576972 0.4010
915 0.833463 0.8296
429 0.644341 0.6900
855 0.477577 0.6126
711 0.530543 0.4982
174 0.477577 0.4917
604 0.572155 0.6209
865 0.881899 0.8678
449 0.695015 0.6786
777 0.683768 0.4146
580 0.851610 0.9087
76 0.796451 0.8499
371 0.961458 0.9997
884 0.799178 0.9071
136 0.515090 0.5763
158 0.729075 0.7231
290 0.787732 0.6100
673 0.748977 0.6931
321 0.702108 0.5028
757 0.899316 0.8691
70 0.666315 0.7865
355 0.569279 0.3572
359 0.961239 0.8362
107 0.886554 0.9438
265 0.640555 0.5975
825 0.701618 0.6212
139 0.563990 0.6810
184 0.626397 0.7691
708 0.688292 0.6542
622 0.725480 0.5147
941 0.828256 0.9047
305 0.676587 0.6286
809 0.726858 0.6427
306 0.818704 0.8063
767 0.721371 0.7445
863 0.902120 0.8096
323 0.957583 1.0000
59 0.477577 0.9167
298 0.494210 0.4433
668 0.631534 0.7201
617 0.635709 0.8713
370 0.789941 0.7637
23 0.700036 0.7924
30 0.774137 0.9460
816 0.760774 0.8721
10 0.576904 0.3380
514 0.685097 0.6305

Good luck creating your own Model!¶