Toward a lab-in-the-loop AI platform for RNA therapeutics
Abstract
RNA therapeutics, spanning messenger RNA vaccines, small interfering RNAs, antisense oligonucleotides, and aptamers, are advancing rapidly, yet the design space that governs whether an RNA drug succeeds remains difficult to navigate. Its principal determinants, namely sequence, chemical modification, secondary structure, delivery vehicle, and tissue tropism, are strongly coupled, and together they constitute a combinatorial space far too large for human heuristics or isolated experiments to explore effectively. Here, we envisage a platform that brings two complementary advances into concert. First, RNA and molecular foundation models can compress this design space into a learnable representation that generalizes across RNA molecules and the carriers that deliver them. Second, a lab-in-the-loop architecture narrows the gap between prediction and biological function by coupling model-generated proposals to automated wet-laboratory experiments, allowing the model to be continually recalibrated against domain-specific feedback. Calibrated uncertainty connects these components: it enables the loop to prioritize candidates whose measurement is expected to be most informative. A platform built on these principles could support joint optimization of sequence and modification, co-design of payload and delivery vehicle, off-target prediction and in silico candidate ranking, and could ultimately bring the rational design of personalized therapies within reach.