AI and the design of mRNA Therapeutics
mRNA Therapeutics are emerging as one of the most promising new medical modalities: their ability to introduce proteins of choice into cells of the human body can be used to treat infections, genetic disorders and cancers. Our company, Jantomarna Therapeutics, was founded because we believe in the enormous potential of this modality, and we are keen to use our founders’ deep knowledge of mRNA Biology to help make the promise of new medical treatments for currently untreatable diseases a reality.
Generating mRNA sequences useful for treating patients requires good sequence design, a fact that is widely recognised judging by the many different mRNA design approaches and tools developed by research groups from all over the world. The majority of these tools are AI-based: this is not surprising given the now ubiquitous presence of AI in all areas of human endeavour. While AI-based design tools can be powerful they come with specific issues that can be avoided if AI is combined with alternative, non-AI based solutions.
Why is sequence design important?
mRNAs are chains (“sequences”) of nucleic acids that encode information used by cells to make specific proteins. mRNA Therapeutics encode proteins that have specific therapeutic effects in the human body. Physical mRNAs for therapeutic use are produced in biochemical reactors, formulated, and administered to specific patient tissues where the encoded proteins with their therapeutic properties are being produced by the patient’s gene expression machinery, the same machinery that also processes the natural complement of human genes.
The general relationship between nucleic acids and their encoded proteins, which is known as the genetic code, is well understood from seminal work in the 1960s. A hallmark of the genetic code is that the same protein can be encoded by many different mRNA sequences, for example the SARS-CoV2 spike protein which is delivered by the Covid vaccines could be encoded by more than 10632 different sequences. While all these different sequences carry instructions to make the same protein, they differ greatly in their functionality: some will be difficult to make using industrial biochemical processes, some will produce a lot of protein whereas others will produce less, some will be very stable whereas others will decay quickly, some are prone to producing random proteins which may cause unintended biological effects whereas others are not. Making an efficient Therapeutic means controlling all of these aspects, and finding the small number of mRNAs that have just the right combination of properties that lead to optimal function.
How does AI approach this problem?
The term “AI” covers uniquely powerful pattern-learning approaches, usually relying on computationally implemented neural networks that “learn” patterns from (often very large) datasets. Chatbots apply this ability to deciphering the patterns of human language, in mRNA design the same ability is applied to learning patterns that make efficient Therapeutic sequences. In its simplest form, neural networks are “shown” large collections of different mRNA sequences for which efficiency has been measured, thereby allowing the network to “learn” which features make an mRNA sequence efficient. This information can then be used to make new mRNAs which contain efficiency-enhancing patterns in their sequences, while avoiding efficiency-reducing ones.
What other approaches are there?
The alternative to AI-based approaches is to learn the rules for making efficient sequences through old-fashioned cycles of experimentation, observation, and analysis which are at the heart of all of science. mRNA has been studied using this approach since its discovery in the early 1960s, and we now have detailed biophysical models available for many aspects of mRNA function. For example, we can follow the process of protein synthesis atom-by-atom, despite the enormous complexity of the protein synthesis machinery (the 2009 Nobel Prize for Chemistry was awarded for related discoveries). Like AI-models, mechanistic models can be applied to the design of well-performing therapeutic sequences.
How do AI models and mechanistic models compare when it comes to mRNA Therapeutics design?
In a nutshell, AI models contain information about patterns but not about function (they can identify which patterns are good but not why they are good), whereas mechanistic models contain information about function which can be applied to analysing the effect of specific patterns. The focus on patterns makes AI models inherently well suited for design purposes, whereas generating design principles from mechanistic models can be more challenging: deep understanding of how a system works is not the same as the ability to design good components for that system. On the other hand mechanistic models are much easier to transfer from one purpose to another: if you understand a system very well, you will probably be able to understand very similar systems, too. In contrast, AI models do not easily transfer to even closely related new settings because the absence of mechanistic information makes it difficult to judge which patterns discovered in the old setting are also relevant in the new one.
This last constraint is very relevant in the field of mRNA Therapeutics: AI models typically need large training datasets which, at least in the public domain, do not exist for therapeutic mRNAs. Instead, AI is usually trained on natural gene sequences. This is despite the fact that mRNA Therapeutics differ substantially from native genes, including in the sequence patterns that determine efficacy. Thus, using AI models trained on normal gene sequences for designing therapeutic sequences risks that the designs are based on irrelevant information: it's a bit like Bixonimania (the purposeful mis-training of AI language models with a non-existent disease that became accidentally accepted as fact).
So how do we design good mRNA Therapeutics at Jantomarna?
At Jantomarna we see the issues with AI models as more fundamental obstacles to good design than issues with mechanistic models. We have therefore opted to base our core approach on mechanistic models, supplemented with information from AI models for design parameters where we do not have useful mechanistic models available yet (for example, for mRNA stability in the cell). Our ability to apply mechanistic models for sequence design is enabled by our genetic computational algorithms that evaluate pools of random sequences against target parameters using information from both mechanistic and AI models, and “breed” high performing sequences through computational cycles of mutation and selection. A large portfolio of collaborative therapeutics developments has validated that this approach gives very good results, usually resulting in high performing sequences with minimal need for cycles of testing, improvement and re-testing.-testing.
Conclusion
In our view the future of mRNA Therapeutics lies in navigating trade-offs between AI approaches and design based on functional mechanistic insights. By integrating the design efficiency of AI with the robustness of mechanistic models, we aim to transcend the limitations of both methodologies in isolation. This hybrid paradigm not only addresses the specific challenges of therapeutic sequence design but also establishes a reliable framework for developing the next generation of effective, safe, and highly efficient medical treatments.
About the author
Tobias von der Haar is Professor of Systems Biology at the University of Kent, and CEO and one of four co-founders of Jantomarna Therapeutics. Tobias has combined computational and experimental approaches for studying mRNA Biology since the beginning of his PhD in 1995.
About the company
Jantomarna Therapeutics Limited is a private company limited by law registered in the UK with company number 16717863. Jantomarna’s mission is to harness the long-standing track records of its founders in mRNA Biology and Immunology for the design of mRNA Therapeutics with superior efficacy, patient safety, and manufacturability. We offer contract design of therapeutic sequences, co-development of therapeutics, and licensing of our design software. If you are interested in mRNA Therapeutics design contact us at info@jantomarna.co.uk .