Date

2022-8-31

Author

Çıray, Fulya

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

309
views

303
downloads

Cite This

Drug development is a costly process, especially in terms of the required time and money. Many promising drug candidates are eliminated at late development stages, e.g., phase II or III of clinical trials, due to insufficient efficacy or unexpected adverse health related affects. Lately, pharmaceutical companies are evaluating computational approaches, to increase the efficiency of this process. In this thesis study, we investigated the computational prediction of the approval of drug candidate compounds by regulatory bodies (i.e., approved for an official use to treat the indicated disease) while the trial process is still continuing, using relevant information from previous discovery and development stages and machine learning. As a preliminary analysis, we examined drug substructures to observe whether the presence of specific molecular structures in drug candidates lead to undesirable outcomes (i.e., unapproved). In the main part of the study, we employed a wider and more heterogeneous set of features including molecular and physicochemical properties of drugs, together with clinical trial and patent related features, to represent each drug-indication pair as a heterogeneous numerical vector. Following data gathering, manual curation and imputation procedures, our finalized feature vectors are processed by random forest (RF) classifiers to train independent drug approval prediction models for 14 different disease groups. We achieved high prediction scores in our cross validation-based performance evaluation, varying in ranges of; accuracy: 0.67-0.81, precision: 0.77-0.82, recall: 0.77-0.96, F1-score: 0.77-0.88 and MCC: 0.45-0.62. Furthermore, by conducting a temporal analysis, we showed that our method is also capable of producing successful results in a prospective manner. We also carried out a performance comparison against a baseline model and a state-of-the-art method from literature, the results of which indicated both robustness and the generalization capability of our approach. Additionally, we identified the most important features for accurately predicting drug approvals, which heavily includes clinical trial and patent related features. Within a use-case study, we showed that our method can successfully predict regulatorily approved (phase IV) drugs that are later withdrawn from the market due to severe side effects. Finally, we used pre-trained models to predict the approval of drug candidates that are currently in clinical trial phases I/II/III and presented prediction results. We hope that the results of our study and the computational tool we presented will contribute to the literature in terms of evaluating and improving the drug development process. All of the datasets, source code, results and pre-trained models of this study are freely available at https://github.com/HUBioDataLab/DrugApp.

Subject Keywords

Approval of drugs, clinical trials, drug patents, machine learning, predictive modeling

URI

https://hdl.handle.net/11511/99462

Collections

Graduate School of Informatics, Thesis