Background: Why RNA sequences and delivery vehicles must be considered together

Deep learning is being used to predict RNA structure and gene-editing activity and to design delivery vehicles, but limited data and differences in experimental conditions restrict its applicability. A review published in ACS Nano examined existing models in these three areas and summarized the strengths and limitations of different neural-network architectures. It should be distinguished from a study reporting the development of a new integrated platform or therapeutic effects in patients.

RNA therapeutic development involves both the RNA sequence responsible for the therapeutic function and the vehicle that carries it into cells and tissues. Changing codon combinations, RNA secondary structure, or the design of specific regions expands the candidate space, making it difficult to explore every combination through experiments alone. Deep learning's role is to learn relationships among sequence, structure, and function from existing experimental data and narrow down candidates for follow-up experiments.

The information required also depends on the prediction task. Structure prediction involves capturing relationships between distant nucleotides, whereas delivery-vehicle design involves assessing combinations of lipid components. The input data and model architecture therefore need to be examined for each task.

Key findings: Model strengths depend on the input information and prediction task

The review explains that transformers are well suited to learning relationships between nucleotides that are far apart in a sequence for RNA structure prediction. Graph neural networks (GNNs) represent nucleotides as nodes and relationships such as base pairing as edges, incorporating structural information. However, high performance does not necessarily extend to every RNA. The review notes that NuFold, for example, has difficulty predicting structures for new RNA families with few training examples.

In CRISPR gene editing, prediction of guide RNA on-target and off-target activity is distinguished from prediction of base-editing and prime-editing outcomes. Convolutional neural networks (CNNs) extract local sequence patterns, whereas transformers learn relationships across a broader range. Models combining the two architectures have also emerged, but the conditions of the training experiments limit their applicability. For example, according to the review, BE-DICT was trained on sequences inserted using lentiviruses, making it difficult to generalize its predictions to endogenous genomic sites with different chromatin environments.

For RNA delivery, a key distinction is which components of lipid nanoparticles (LNPs) a model represents. AGILE is pretrained on molecular graphs of ionizable lipids and fine-tuned with experimental measurements to predict delivery preferences across cell types. One limitation is that it does not account for interactions with other components, such as helper lipids, sterols, and polyethylene glycol (PEG) lipids.

COMET, by contrast, uses the components and proportions of the entire LNP formulation as transformer inputs. The review explains that this approach requires training data with more experimental measurements and greater computational resources than models considering a single component. The two models are examples from individual studies addressing different scopes of delivery-vehicle design.

Implications and outlook: A development process that feeds predictions back into experiments

The authors identify combining high-throughput screening with new model architectures as a future direction. They also discuss approaches that explore RNA sequences and delivery vehicles together. These prospects should not be interpreted as an already completed integrated development system.

The current bottlenecks are the shortage of high-quality data and limited generalizability. A model trained on data from a particular tissue or animal species may perform differently in another tissue or species, and results from in vitro experiments do not directly describe in vivo delivery or clinical conditions. This is why differences in experimental conditions must be considered when combining data from different experiments.

When evaluating a model, its training data should be checked for relevance to the target tissue and evaluation conditions, alongside its prediction scores. The review also explains that experiments must follow to confirm the actual function of selected candidates. Deep learning's potential can be assessed within a development process that connects candidate exploration with experimental validation. See the full ACS Nano review for details.