AI-assisted adeno-associated virus (AAV) capsid engineering has transformative potential in terms of accelerating and optimizing the development of gene therapies.
That’s the view of Fangzhi Tan, PhD, first author of a recent paper, and associate researcher at Zhongda Hospital, Southeast University in China. “AI is not a replacement for existing methods, but a synergistic tool to augment experimental workflows and reduce development timelines,” he stresses to GEN. As such, AI and its subset, machine learning (ML), “can accelerate discovery of novel capsid variants, such as those with optimized tropism, reduced immunogenicity, and enhanced delivery efficacy,” he says.
He envisions human-AI collaboration, “where AI rapidly generates and prioritizes design candidates, while human expertise ensures biological relevance and contextual validation.”
A bridge between strategies
Tan says AI-assisted capsid engineering acts as a bridge between in silico modeling, rational design, and directed evolution, by complementing and enhancing them. For example, with:
“In silico methods, AI can refine computational predictions by analyzing vast datasets (e.g., structural biology, evolutionary sequences) to identify non-obvious patterns, improving the accuracy of capsid property predictions.
“Rational design, AI accelerates hypothesis generation by proposing mutations or chimeric designs based on multi-parametric optimization (e.g., targeting specificity, stability), which can then be validated experimentally.
“Directed evolution: AI reduces the experimental burden by pre-screening libraries or simulating evolutionary trajectories, enabling focused exploration of high-potential variants.”
In the near future, AI-assisted AAV capsid engineering will play a central role in helping biopharmaceutical manufacturers accelerate gene therapy development, Tan predicts.
He specifically highlights:
Optimization of capsid properties and their identification
Cost and time reductions by minimizing the need for large-scale experimental screening
Expanded therapeutic applications that may be possible after surmounting current AAV vector limitations such as payload capacity
Integration into regulatory processes, where, “as AI models become more robust and interpretable,” they may provide “predictive evidence of safety and efficacy.”
Transparency is still an issue, however, along with a penchant of AI to “excel on training data but under-perform on new data. The complexity of real biological systems also may exceed the current processing capabilities of AI models,” the scientists note.
To overcome such issues, Tan points out the need for “high-quality training datasets and transparent AI models. Manufacturers adopting AI must prioritize interpretability and human oversight to mitigate risks of model bias or ‘hallucinated’ designs.”
Within the next five to ten years, he predicts, “AI is poised to shift capsid engineering from an artisanal process to a scalable, data-driven discipline.”