Written by Evien Cheng | Edited by Rose Enos
Photo by Google DeepMind
Because the discovery of a new drug requires billions of dollars and 10–15 years of exhaustive trial and error, pharmaceutical researchers have adopted machine learning (ML) models to improve decision-making in the drug development process [1]. ML is the process of “teaching” a machine to recognize patterns across millions of data points, usually recorded by humans experimentally. Using what it learned from these examples, an ML model can make a preliminary decision about new data, which can improve the efficacy of decisions made to decrease the cost of development and the time spent in clinical trials [2]. As an efficient tool for scientists to improve the pharmaceutical industry, ML will be groundbreaking in the near future.
ML has proven able to discover novel drugs and predict their effectiveness against a selected target. One of the metrics used to determine how this effectiveness is a drug’s binding affinity to protein. Binding affinity refers to how well a drug can fit with a target protein’s shape. So far, most ML models have been used to predict these types of interactions between a protein and a target. Predicting binding affinities allows scientists to optimize candidate drugs for effectiveness and stability [1]. Since these optimizations are made in silico, or on a computer, many different options and possibilities can be explored without having to test anything experimentally. Using ML to make predictions with computer programs saves time and money, which can then be allocated to trials downstream in the discovery process.
Building an ML model can be a relatively simple process, but creating one that performs well in a generalized environment is a difficult task. How well a model performs depends on the characteristics of the data provided and the learning algorithm utilized. In the pharmaceutical industry, most data comes from traditional experimental methods and laboratory data. ML offers many ways to interpret and use data depending on what suits scientists’ needs [3]. Looking at binding affinity, ML models classify a target’s binding affinity as either 0 (non-binding) or 1 (binding) [4]. While it may seem simplistic, using only a 0 or 1 value for binding affinity can make a complicated dataset much easier to interpret while still yielding a significant amount of information. After a model is trained on the dataset for the affinity values, its prediction accuracy is determined by the closeness of the predicted value to actual experimental data. Once a model is accurate enough, it can predict affinity of novel targets without having to rely on new experimental data.
ML is a promising tool for making predictions about novel drugs and decreasing the time and money required for drug development. However, the current problem facing ML is a lack of sufficient data to train the models [5]. Scientists and pharmaceutical companies can solve this problem by working together to release experimental data for public use and by building more robust databases, such as resources like PubChem and chEMBL. If such initiatives succeed, a new age of drug discovery may be on the horizon.
References:
- Dara, S., Dhamercherla, S., Jadav, S., Babu, C.H., Ahsan, M. (2022). Machine Learning in Drug Discovery: A Review. Artificial intelligence review, 55:1947-1999.
- “The Role of Machine Learning in Drug Discovery.” MRL, 30 Oct. 2023, www.mrlcg.com/resources/blog/machine-learning-for-drug-discovery.
- Sarker, I. H. (2021). Machine Learning: Algorithms, Real-World Applications and Research Directions. SN Computer Science, 160.
- Kundu, I., Paul, G., Banarjee, R. (2018). A machine learning approach towards the prediction of protein-ligand binding affinity based on fundamental molecular properties. RSC advances, 8:12127-12137.
- “The Role of Machine Learning in Drug Design: Advancements and Challenges.” VIAL, vial.com/blog/articles/the-role-of-machine-learning-in-drug-design-advancements-and-challenges.