Drug Repurposing (Recommendation package)

Experimental Package for building models and stacking them for recommendation score. We can also integrate them in MLflow for logging parameters.

To install and use package from github repo

!pip install git+https://github.com/btarun13/drug_repurposing.git@feature_branch

You can use above line in colab should work fine.

In case you are using VScode(Under testing), .toml files will have everything. If you have uv package manager, just run

uv sync

With this, you would be in the same enviroment setting used for the package.

Example code:

import torch
from drugrepo.model import LightGCN
from drugrepo.data_loader import prepare_data
from drugrepo.predictor import get_pair_score
from drugrepo.model_train import train_and_evaluate

At the moment there is only one type of architecture specified at in drugrepo.model . We can have more architectures defined within and use for training models with different architecture. For now, we will continue with LightGCN(output can be y-> 0,1 or y-> any float). For y-> float the loss functions will be changed. refer notebooks. These work as link prediction

How edge data(grouth truth) should be, PS: note that they are csv and not | separated.

source     |target      |y
Aspirin    |Diabetes    |1
Ibuprofen  |Cancer      |1
Paracetamol|Asthma      |1
Amoxicillin|Influenza   |1
Metformin  |Hypertension|1

How node data(embedding or inital features) can look like? PS: should use comma instead of |.

id       |name                 |embedding_0|embedding_1|embedding_2 |embedding_3
ID_type_1|somatization disorder|-0.325601  |0.076047   |0.299131    |-0.2558
ID_type_2|Down syndrome        |-0.325601  |0.076047   |0.299131    |-0.2558
ID_type_3|GRP                  |-0.325601  |0.076047   |0.299131    |-0.2558
ID_type_4|burn                 |-0.325601  |0.076047   |0.299131    |-0.2558

This is an example on how you can send in the data

input_data = prepare_data("/content/Ground Truth.csv",        #### path for edge data
             "/content/recomendation_pipeline_initial_node_embeddings.csv",  ### oath for node data
             test_size=0.3,   ### test size and random states can vary, in cause you have a hold out validation set we can use multiple train/test spilts with a lot of different seeds to measure performance on validation set
             random_state= 123)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

Training with specific hyperparameters

layer = 2
model = LightGCN(
    disease_embeddings=input_data['disease_embeddings'],
    drug_embeddings=input_data['drug_embeddings'],
    num_layers=layer).to(device)

train_and_evaluate(model, input_data, num_epochs=20, batch_size=50, lr = 0.01, device=device)  ### change hyperparameters

Epoch 10: Loss = 0.0556, Test AUC-ROC = 0.9604, Test AUC-PR = 0.9348
Epoch 20: Loss = 0.0564, Test AUC-ROC = 0.9617, Test AUC-PR = 0.9330

Results

get_pair_score(model,
               'CHEMBL.COMPOUND:CHEMBL30',  ## ID_type_random2
               'MONDO:0007362',             ## ID_type_random1
               input_data,
               device)

{'drug_id': 'CHEMBL.COMPOUND:CHEMBL30',   ## ID_type_random2
 'disease_id': 'MONDO:0007362',           ## ID_type_random1
 'raw_score': 1.5422860383987427,
 'probability': 0.8237967491149902,
 'disease_emb': [tensor([-5.9911e-02, -1.2041e-01,  5.0482e-02, -3.2367e-03,  9.9549e-02,
           7.9791e-02, -2.2457e-01,  3.2794e-01,  1.5935e-01, -1.9004e-01,
          -3.4828e-03, -1.3589e-01,  2.1126e-01,  5.3988e-01,  1.8098e-01,
           2.1025e-02, -6.7591e-02,  1.2305e-01,  6.2050e-02, -2.1232e-01,
           4.6071e-02,  1.8825e-01, -1.5292e-01,  3.0699e-01, -1.1182e-01,
           6.2287e-02,  1.0639e-01,  3.0332e-01, -2.0131e-01,  8.1552e-02,
           3.6927e-01,  1.5867e-01,  1.3609e-01,  4.8282e-02, -3.2262e-02,
          -6.1631e-02, -4.8148e-02,  1.4393e-01,  7.9521e-02, -1.3832e-01,
           2.2115e-01,  3.3398e-02,  2.3921e-01,  1.4900e-01, -1.3909e-01,
           1.5112e-01, -3.5351e-01, -7.8104e-02, -1.2823e-02,  1.3334e-01,
          -7.3039e-02, -2.5402e-01,  1.5269e-01,  2.0441e-01,  4.7075e-02,
          -1.9494e-01,  1.1327e-01, -2.3754e-01, -1.5045e-01, -2.8113e-01,
           2.0229e-01, -2.3288e-01, -6.3663e-02,  5.2580e-02,  4.5360e-02,
          -2.4321e-02,  3.8664e-01,  1.1374e-01, -2.9992e-02, -4.5821e-02,
          -1.2341e-01,  3.0121e-04, -3.1351e-01, -2.1253e-02,  7.1962e-02,
          -1.3010e-01,  2.4236e-01,  3.7549e-02,  1.1319e-01, -2.4052e-01,
           4.5178e-02,  4.5833e-02, -7.6962e-02, -8.9253e-02,  3.1187e-01,
          -1.9854e-01,  1.3689e-01, -2.7576e-01,  1.1196e-01,  5.2103e-02,
          -1.6629e-01,  9.4572e-02,  6.4623e-03, -3.1499e-01, -6.8299e-03,
          -1.4604e-01, -2.0032e-01, -1.9492e-01, -2.8079e-03, -4.1250e-02,
          -4.3154e-01, -1.1046e-01,  2.9363e-01,  1.6861e-02, -1.5691e-01,
          -4.0218e-02,  1.5355e-01,  3.2216e-01,  1.2064e-01,  3.6648e-02,
           9.5431e-02,  1.7660e-01, -6.4549e-02, -1.4671e-01, -4.7712e-02,
          -1.4081e-02,  4.6027e-02,  9.9893e-02, -4.8742e-03,  2.2973e-01,
           5.4576e-01, -2.1284e-01,  5.7996e-02,  1.0800e-01, -3.6115e-02,
          -1.6547e-01, -2.2726e-01, -5.1840e-02], device='cuda:0')],
 'drug_emb': [tensor([-2.6009e-01,  1.3955e-01,  3.4041e-01, -2.9036e-01,  2.3295e-01,
           1.8259e-01,  4.3216e-01,  1.7057e-01,  4.2607e-01,  4.9822e-01,
           2.3535e-01, -2.9875e-01,  2.2269e-01, -3.2226e-01,  4.0209e-01,
          -2.2594e-01, -8.0527e-02, -1.2607e-01, -2.5287e-01, -1.4696e-01,
           3.8881e-01,  1.5083e-01, -9.1018e-02,  1.9653e-01,  3.4487e-01,
           2.9561e-01,  5.7577e-02,  1.6204e-01, -1.7187e-01,  2.3540e-01,
           1.9444e-01,  1.9209e-01, -2.6476e-01,  3.4033e-01,  1.5154e-01,
          -2.7733e-01,  1.5178e-01, -2.1976e-01, -3.6034e-01, -7.7930e-02,
           1.5717e-01,  2.1479e-01,  3.2736e-01,  7.0454e-02, -7.2268e-02,
          -5.2160e-01, -1.8327e-01,  2.9841e-02, -3.4221e-01, -1.2182e-01,
           3.6185e-01,  6.9399e-02,  1.6537e-01, -6.5987e-02,  3.4278e-01,
          -2.4307e-01,  6.2978e-02, -3.2228e-01, -2.1620e-01,  1.1588e-02,
          -1.3403e-01,  4.1289e-02,  7.8002e-02, -4.5184e-01,  3.4374e-01,
           4.3088e-01,  9.7138e-01,  1.2647e-01, -3.5435e-01, -3.4411e-01,
          -9.2454e-04,  2.2160e-01, -5.5603e-01,  4.0592e-02, -4.9680e-02,
          -8.2038e-02,  2.5924e-01, -3.8464e-01,  6.3594e-02, -3.4363e-01,
           3.4359e-01,  3.4412e-01,  3.4924e-01, -1.4866e-01,  2.0214e-01,
          -1.6499e-01,  1.7197e-01, -1.5671e-01,  4.3623e-02,  3.2187e-01,
          -1.3201e-01,  3.8828e-02, -1.3456e-01,  7.3604e-01,  6.0945e-02,
          -2.6232e-01,  1.3944e-01, -1.4865e-01,  3.6801e-01, -8.6408e-02,
          -7.9694e-01, -1.3776e-01,  6.4576e-01, -4.2333e-02, -1.3449e-01,
           1.5805e-01,  1.6498e-01, -3.2644e-03,  8.6759e-02, -2.8934e-01,
           4.1675e-01,  3.7614e-01, -1.1607e-01,  2.0894e-01, -1.5386e-01,
          -2.0994e-01,  3.4077e-01,  3.0947e-02, -2.1975e-01, -1.1292e-01,
          -4.9965e-01,  1.2497e-01, -2.1930e-01,  1.5485e-02,  2.6300e-01,
          -2.5049e-02, -7.2167e-02,  4.0921e-01], device='cuda:0')]}