Artificial intelligence (AI)-enabled RNA therapeutic design: from molecular optimization to system-level evaluation

RNA therapeutics are increasingly moving from empirical screening towards rational, computation-guided design. Recent advances in RNA language models, generative models and structure-aware learning have improved sequence representation, candidate generation and structure-constrained optimization. However, favourable molecular properties do not necessarily translate into desirable biological performance, because therapeutic outcomes also depend on modality-specific mechanisms, delivery and exposure, molecular interactions, cellular state, immune responses and network-level adaptation. Here, we review AI-enabled RNA therapeutic design through the specific lens of the interface between molecular optimization and system-level evaluation. We compare the design constraints of siRNAs, antisense oligonucleotides, aptamers, mRNAs and translation-oriented circRNAs, and examine the capabilities and current limitations of sequence-, generative- and structure-based computational models. We further discuss the data resources and perturbation representations required to connect molecular candidates with broader cellular-response models. In this context, virtual cells are considered as approximate computational frameworks for candidate prioritization and risk-oriented system-level assessment rather than complete digital replicas of living cells. Finally, we outline a conceptual generation-simulation-feedback-optimization framework for integrating molecular design with progressively higher-level evaluation. This perspective emphasizes the need for better interfaces among RNA design, pharmacological context and cellular-response modelling, rather than relying on molecular optimization alone.