Fireside chat: AI’s emerging role in oligonucleotide development and manufacturing

Nucleic Acid Insights 2026; 3(9), 645–653

DOI: 10.18609/nai.2026.085

Published: 13 November
Interview
Srdan Masirevic, Bao Zhong Cai

Nucleic Acid Insights convened Srđan Maširević (Computational Biologist and Machine Learning Scientist, A*STAR’s Genome Institute of Singapore) and Bao Zhong Cai (VP, CMC, GondolaBio) to discuss where AI and machine learning earn a place in oligonucleotide production today, and how far the field is from digital twins and autonomous synthesis. 

Portrait of Srđan Maširević, Computational Biologist and Machine Learning Scientist, A*STAR’s Genome Institute of SingaporePortrait of Bao Zhong Cai, VP, CMC, GondolaBio

Artificial intelligence (AI) is discussed with increasing frequency across the nucleic acid therapeutics field: it is already productive in early molecular design and in data-rich analytical tasks, and considerably less mature in the steps that connect synthesis to final product quality.

In this Fireside Chat, Nucleic Acid Insights convened Srđan Maširević (Computational Biologist and Machine Learning Scientist, A*STAR’s Genome Institute of Singapore) and Bao Zhong Cai (VP, CMC, GondolaBio) to discuss where AI and machine learning (ML) earn a place in oligonucleotide production today, the data and process foundations they depend on, and how far the field is from digital twins, real-time impurity characterization, and autonomous synthesis. This discussion was a spin-off from the Nucleic Acid Insights Special Interest Group held in April 2026 on enhancing the efficiency and sustainability of oligonucleotide manufacturing.

Where does AI genuinely add value in oligonucleotide development today, and where is the promise still ahead of the reality?

SMTo understand where AI adds value today, it helps to consider how the two fields converged. The first concepts of artificial neural networks were proposed in the 1940s and 1950s, while the major advances in nucleic acid biology came in the early 2000s with the Human Genome Project and with the discovery of the functional diversity of RNA. At that point, RNAs were regarded largely as transcripts that carry information; we now know that they perform many other functions, which matters a great deal from a therapeutic standpoint.

ML models of that era were essentially flat and blind to physics. Three developments changed that over the past 10–15 years: greater graphics processing unit (GPU) capacity; experimental techniques such as high-throughput screening, which generate very large data sets in a single experiment; and the introduction of transformer architectures. Together, these allowed us to interpret DNA and RNA sequences in context, which marked the beginning of generative biology.

Today, the clearest value lies in early-stage molecular design. Current tools perform well at optimizing mRNA therapeutics, antisense oligonucleotides, and small interfering RNAs. They are effective at suggesting which chemical modification patterns to apply, which is important because those constructs must remain stable once they enter cells, and at narrowing the candidate pool for whichever property the model was trained on.

The promise remains ahead of the reality as soon as we move from computational screens to living systems. Models perform well at the in vitro scale, where data are abundant. At the cell, tissue, and organ level, where we need to know how toxic a newly synthesized oligonucleotide is and whether it can be delivered to a specific tissue, the blind spots remain. Those blind spots are precisely where the therapeutic and commercial value sits.