Key Highlights

• Artificial intelligence (AI) is rapidly being implemented across the drug development process from target discovery to managing clinical trials. 
• AI is being used for RNA interference (RNAi) to identify optimal targets and improve the design of siRNAs through the selection of modified bases. 
• AI is improving the identification of neoantigens for personalized cancer therapies, the stability and translation of RNAs, and generation of novel LNP components for effective RNA-based medicine drug delivery. 

Introduction 

Artificial intelligence (AI) has and is changing the world, including how drugs are developed. Although AI has great potential for improving the speed and success of drug development and design, there has been limited clinical trial success due to several outstanding challenges. In this eBlog, we review the impact of AI on the development of RNA-based and RNA-targeting medicines. 

General approaches for AI-powered drug development 

There are three main approaches to using AI for drug discovery¹:

• Target discovery (also called target identification) uses AI to identify the right biological molecule or pathway to target for therapeutic success².
• Drug design creates an optimal molecule either by improving the chemistry, sequence, or manufacturing process.
• End-to-end drug development uses AI for both target discovery and drug design.

Each approach can use different AI models and carries different degrees of risk¹. 

AI for siRNA drug development 

siRNAs are small RNAs that bind to target genes, leading to their degradation by the RISC complex. AI is being applied to siRNA drug development in three main ways: target discovery, synthesis decisions, and off-target prediction. 

An effective siRNA needs to bind to an accessible target that will improve disease outcomes. Since siRNAs bind to their targets via complementary sequence pairing, their design is well suited for several types of AI models^4,5. An outstanding challenge in using AI for target discovery is that, while sequence is a major driver of siRNA efficacy, the cellular environment can greatly affect binding. Factors such as structural changes and competition from miRNAs and proteins greatly complicate this process. 

siRNAs typically incorporate modified bases to improve stability, uptake and target specificity. Historically, base modification choice has been through known chemistry and rational design, although now, AI is being used to predict the effectiveness of different chemical modifications^6,7. An AI-based approach allows for the possibility of novel chemistry that can significantly improve siRNA efficacy, though extensive validation will be needed to confirm any algorithmic predictions. 

While siRNAs effectively downregulate gene expression, they pose a high risk of off-target effects due to partial miRNA-like binding (as discussed in one of our previous eBlogs). AI approaches are being developed to predict potential off-target effects, but direct validation of their predictions will be crucial for improving model accuracy^3,8. 

AI for RNA-based medicines 

RNA-based medicines use cellular machinery to create a therapeutically effective protein. AI is currently being applied across the development process for RNA therapies, from identifying optimal RNA sequences to optimizing the manufacturing process. 

RNA-based medicines lead to the generation of a therapeutic protein in cells, this can be an antigen to invoke an immune response or a replacement for an ineffective protein due to a genetic disorder. One application of AI is to identify the optimal protein to be encoded into an RNA therapeutic. For example, AI models are being used to predict optimal neoantigens for personalized cancer therapies. This can improve patient outcomes by generating immune responses against their specific tumor. 

To be an effective therapeutic, RNA needs to be stable and well-translated. This can be achieved by intelligent sequence design: specific sequences can enable more efficient translation through codon optimization¹⁰, enhance base pairing to improve RNA stability¹¹, or fine-tune regulation by endogenous miRNAs or RNA-binding proteins. AI algorithms are being applied to design RNAs that optimize for each of these features^12,13,14, although many of these effects are cell-type specific. This means that successful AI-generated sequences require extensive datasets capturing key features of structure, translation, and regulatory factor binding¹⁵. 

RNA-based medicines can’t be delivered on their own, they need to be encapsulated into a drug delivery vehicle such as a lipid nanoparticle (LNP). In addition to improving sequence design, AI is also being applied to develop optimal LNPs that have improved RNA release and targeting capabilities^17,18.  

Finally, AI is also being used to improve manufacturing processes to improve yields and reduce contaminants. One example of this is from a recent paper by Merck where they used a machine learning approach to optimize in vitro transcription to improve RNA yield and purity¹⁹.  

Outstanding challenges 

AI-driven drug discovery rests on three pillars: high-quality data, effective algorithms, and robust computational infrastructure. As AI becomes more and more embedded in drug development processes it is becoming clear that there are two main challenges for the successful use of AI.  

The first is that high-quality, reproducible data can be hard to find. Although there is a wealth of data available in public repositories, the data is unstructured and often requires extensive data processing to ensure that only reproducible and useful data is incorporated into models. In addition, diseases are multiomic. It is not just the increase or decrease of an RNA that can lead to a disease, it can be how the RNA folds to affect protein binding or how translation proceeds. It is critical that plentiful data across all dimensions of RNA are available to ensure that models are trained on the features that matter most for target identification or drug design²⁰. 

The second is that there are currently limited methods for validating an AI algorithm for identifying targets or designing effective medicines. To build on early successes, it is essential that AI predictions are thoroughly tested so that feedback can be provided for model refinement¹. 

Conclusions 

We are in the midst of an AI revolution transforming medicine development through novel target selection and therapeutic design. At Eclipsebio, we help solve two of the major challenges for AI-powered drug discovery: the generation of multiomic models for effective AI model training and empirical validation of AI-designed drugs through our full-service characterization solutions. Contact us today to discover how we can support your successful AI drug development.  

References 

1. 1. Wilczok and Zhavoronkov 
2. 2. Pun et al. 
3. 3. Lee 
4. 4. Monopoli et al. 
5. 5. Long et al. 
6. 6. Shinohara et al. 
7. 7. Martinelli 
8. 8. Liu et al. 
9. 9. Cai et al. 
10. 10. Li et al.  
11. 11. Zhang et al. 
12. 12. Cetnar et al. 
13. 13. Castillo-Hair et al. 
14. 14. Kim et al. 
15. 15. Morrow et al. 
16. 16. Schlusser et al. 
17. 17. Bae et al. 
18. 18. Witten et al. 
19. 19. McMinn et al. 
20. 20. Hwang et al.