A Novel Approach to Drug Discovery: Integrating Deep Learning with Biological and Chemical Data

Through computational modeling and machine learning approaches, drug discovery can be expedited and can save time and cost compared to experimental methods. An integrated framework is proposed to predict the biological activity using the data taken from Philippines database. The RDKit toolkit was used to process molecular structures. Molecular weight, lipophilicity (LogP), and hydrogen bond donor and acceptor counts were extracted as molecular descriptors representing chemical properties. The team developed a deep-layered Long-Short Term Memory (LSTM) neural network, which was capable of capturing complex non-linear relationships between molecular features and biological activity after data preprocessing and standardization. Compounds were classified active or inactive using the extracted features by the model training. Using the AUC-ROC metric, proposed model score ۰.۹۱ which shows strong ability to distinguish between classes. Even if the results show a good prediction performance, the training and validation curves shows a slight overfitting, which will lead to a less generalization to unseen data. Future works should include a larger and more diverse dataset to improve robustness. Also, optimize hyperparameters and use better regularization techniques like dropout or early stopping. In addition, using advanced pre-trained models such as ChemBERTa and DeepChem frameworks may improve predictions. In conclusion, an API-based design and real-time drug screening system could increase actionability for pharmaceutical research and improve drug discovery.

کلیدواژه ها:

Drug Discovery ، Molecular Descriptors ، Deep Learning ، Machine Learning