All current drug discovery approaches suffer from several common issues. First of all, they require a huge amount of funding to take a drug candidate from pre-clinical development to clinical trials, thereby crippling further funding for optimization or re-discovery.
Second, they lack efficacy due to each patient’s unique needs and their specific genetic background. Third, pharmaceutical companies need a timeline of about 15 years for a successful drug development, which costs more than $1 billion.
The fourth issue is that only 1 in 10 small molecule projects become candidates for clinical trials, and only about 1 in 10 of those compounds will then pass successfully through clinical trials. Finally, many disease conditions cannot be successfully addressed through the traditional drug development process.
The good news is that the recent advancements in artificial intelligence (AI) and machine learning have the potential to overcome these issues and transform the drug development process, making it both more efficient and more effective.
There are many components, directly and indirectly, related to the drug discovery and development pipeline that AI can enhance. These include but are not limited to computer-aided organic synthesis, compound discovery, assay development, biomarker, and target discovery. In general, AI aims to automate and optimize slow processes to substantially speed up the R&D drug discovery process.
Traditional vs. Al-based drug discovery
| Traditional | Al-based |
| • Target-driven | • Data-driven |
| • Work well for easily druggable targets that have a well-defined structure and whose interactions inside the cell are understood in detail | • Complex algorithms and machine learning can extract meaningful information from a large dataset |
| • Extremely limited due to the complex nature of cellular interactions & limited knowledge of intricate cellular pathways | • Identify compounds that could bind to ‘undruggable targets,’ i.e., proteins whose structures are not defined |
Benefits of applying AI to drug discovery
The application of AI to drug discovery can revolutionize the current time scale and scope of drug discovery.
- AI does not rely on predetermined targets for drug discovery. Therefore, subjective bias and existing knowledge are not a factor in this drug development process.
- AI utilizes the latest advances in biology and computing to develop state-of-the-art algorithms for drug discovery. With the rapid increase in processing power and reduction in processing cost, AI has the potential to level the playing field in drug development.
- AI has a higher predictive power to define meaningful interactions in a drug screen. Therefore, the potential for false positives can be reduced by carefully designing the assay parameters in question.
- Most importantly, AI has the potential to move drug screening from the bench to a virtual lab, where results of a screen can be obtained with greater speed, and promising targets can be shortlisted without the need for extensive experimental input and manpower hours.
Drawbacks of applying AI to drug discovery
As is the case with any advance that brings a paradigm shift in our understanding of existing technology, AI still cannot replace a human scientist entirely in the process of drug discovery.
- AI predictions are as good as the algorithms used to investigate a dataset. The algorithm should clearly lay out the criteria used to parse out meaningful information when the results are in the ‘gray zone’ of interpretation.
- AI can suffer from algorithm bias, where the creators’ own bias manifests itself in the way information is processed to generate predictions. Therefore, the process is not entirely objective.
- While the cost of supercomputing and high-throughput screening has decreased appreciably over the past decade, establishing these pipelines still requires significant investment.
- Ultimately, a computer’s predictions have to be verified by scientists to make sure they are valid.
Challenges of AI in drug discovery
AI has shown great promise in drug discovery, but not without challenges. AI faces many challenges, such as lack of data, lack of interoperability, and the curse of dimensionality.
The lack of data is a recurring problem throughout all industries implementing AI. In a traditional biological study, the minimum number of samples is five to be valid. Meanwhile, most machine learning algorithms must be trained on hundreds and thousands of data points or samples to perform well.
Another challenge is the lack of interpretability. It is often difficult to explain how a model makes certain predictions and performs. It is more likely an occurrence with deep learning in which each layer adds complexity to the model. The explanation of each layer’s outputs can become exponentially complex as the number of layers increases.
Another hindrance in using AI to predict drug targets remains to translate traditional basic research conducted in labs worldwide into a language that a computer can understand. Machine learning programs rely on data presented in a format where patterns can be identified, and the machine can be trained. This often requires a sophisticated experimental design where human error is kept at a minimum, and multiple different iterations of an experiment can be performed in nearly identical conditions.
Machine learning algorithms convert data into pathways detection, 3D protein structure, metabolite mass measure, etc. These AI transformations can occur with unprecedented speed. However, in many cases, the data being used is not of optimal quality (e.g., resolution of images) or is not balanced (i.e., samples from rare diseases are under-represented in the dataset).
