In the drug discovery race, pharmaceutical companies face a common problem: traditional candidate screening is slow, costly, and labor-intensive. To address these challenges, many have adopted automation and machine learning (ML). Together, these technologies optimize efficiency, reduce costs, and increase drug discovery success rates.
Applications of automation and AI in drug development
The core steps in developing a new drug have remained consistent. First, a target within the body is selected for the drug to interact with. Next, researchers design a molecule to affect that target. Once synthesized in the lab, the molecule undergoes lab testing to confirm it works as intended and does not produce unwanted effects. Finally, it is tested in humans to assess its safety and effectiveness.
Automation and artificial intelligence (AI) can optimize each of these steps. However, most companies are focusing on three failure points in the drug development pipeline:
1. Target selection and validation;
2. Drug design;
3. Patient stratification for treatment benefit.
Target identification and validation
ML enables the analysis of vast genomic, proteomic, and transcriptomic datasets. For example, it can analyze patient data to identify genes and proteins that are consistently overexpressed or mutated in certain conditions, providing new leads for drug development. Once a target is identified, ML assists in validating its role in disease, determining how a drug can modulate it, and predicting potential off-target effects.
Designing a drug that effectively interacts with the target is a primary focus of today's innovation.
Koon Mook Kang and colleagues recently described an application of ML in drug discovery.1 They developed a deep-learning model that predicts where drugs can bind to proteins. Using this method, they identified a new binding site on the P2X3 receptor, which allowed them to screen for potential drug candidates. The model improved hit rates tenfold and significantly accelerated the discovery of novel P2X3-targeting compounds, which is important for treating chronic pain and respiratory diseases.
Drug design
Designing a drug that effectively interacts with the target is a primary focus of today's innovation. ML can generate molecules with 3D structures tailored to interact with biological targets. Moreover, scientists can use ML to redesign existing drugs, enhancing their binding to disease-related proteins or repurposing them for new therapeutic applications. After making adjustments in simulation, researchers can synthesize and test the most promising designs.
A notable example of AI and automation in drug design is the work of King-Smith and colleagues at the University of Cambridge.2 They developed a platform that combines automated experiments with ML to predict how chemicals will interact. Their approach, validated on a dataset of over 39,000 pharmaceutically relevant reactions, is called the chemical "reactome". This approach can significantly accelerate the drug design process because it reduces the need for time-consuming trial-and-error approaches in the lab.
Patient stratification
ML can uncover patterns that predict treatment outcomes by analyzing large datasets, including clinical trial results, genetic information, and patient health records. For example, in a study of 2,538 cancer patients treated with atezolizumab, an ML model used clinical and biological factors to group patients into high- and low-risk categories for mortality, providing more precise predictions for treatment success.3
This type of ML-driven stratification is valuable in optimizing treatment decisions and improving overall outcomes in oncology treatments.
From challenges to change
Automation and ML have accelerated drug discovery by reducing costs and improving processes. However, limitations remain, such as algorithmic biases, ethical concerns, and the need for high computational demands. As these technologies advance, it will be interesting to see how they will redefine the operational, strategic, and competitive landscapes of drug discovery.
References:
1. https://pubmed.ncbi.nlm.nih.gov/35849939/
