RNA Secondary Structure Prediction with Deep Learning

Why RNA structure matters

RNA is far more than a passive messenger between DNA and protein. Non-coding RNAs regulate gene expression, catalyse reactions and form the working machinery of the cell, and their function is determined by how the single strand folds back on itself. That folding — the pattern of base pairs that gives RNA its secondary structure — governs whether a molecule is stable, what it interacts with, and how it behaves as a drug target or a therapeutic in its own right. Understanding structure is therefore a prerequisite for RNA-based medicine, from mRNA vaccines to antisense and small-molecule approaches.

Determining structure experimentally, through crystallography or chemical probing, is accurate but slow and expensive. Computational prediction fills the gap by inferring the most likely base-pairing pattern directly from sequence, at a fraction of the cost — provided the predictions are reliable enough to trust.

The limits of classical prediction

The established computational approach minimises a thermodynamic free-energy model using dynamic programming. These methods are principled and fast, but they rest on approximate energy parameters and, in their classical form, cannot represent important real structures such as pseudoknots — the crossed base pairs that appear in many functional RNAs. Their accuracy also degrades as sequences grow longer. The result is a well-understood but imperfect baseline that leaves room for data-driven methods to help.

A deep-learning approach

This thesis investigates how deep learning can improve on those existing solutions by learning the mapping from sequence to structure directly from data rather than from a fixed energy model. Framed this way, secondary-structure prediction becomes a task neural networks are well suited to: capture the long-range dependencies between distant bases that pair up, generalise across families of RNA, and represent patterns that hand-crafted thermodynamic rules struggle to encode.

The work examines how neural architectures can be trained on curated RNA datasets, how predicted base-pairing can be evaluated against known structures, and where learned models complement — rather than simply replace — the strengths of the classical thermodynamic methods. The emphasis throughout is on improving existing pipelines pragmatically, combining the reliability of established tools with the flexibility of representation learning.

A concrete example: reading a stem-loop

To make the task tangible, consider the most common RNA motif, the stem-loop (or hairpin). A stretch of the strand pairs with a complementary stretch further along — A with U, G with C — forming a double-helical "stem," while the unpaired bases in between bulge out as a "loop." Predicting this correctly means recognising that two regions which are far apart when you read the sequence left-to-right actually belong together in space. Multiply that across dozens of interacting stems, loops, bulges and junctions, and the combinatorics explain why the problem is hard. A good model has to reason about the whole sequence at once, weighing every candidate pairing against all the others, because closing one stem changes which other pairings are still possible. This is precisely the kind of global, long-range reasoning that modern neural architectures are built for, and why framing the output as a full base-pairing matrix — rather than a left-to-right label sequence — matches the structure of the biology.

Data, training and evaluation

Any learned predictor is only as good as the data and the yardstick used to judge it. RNA structure prediction is evaluated by comparing the predicted set of base pairs against experimentally determined reference structures, typically using precision, recall and F1 over the predicted pairs — a natural fit because the task reduces to deciding, for every possible pair of positions, whether the two bases bond. Careful dataset curation matters just as much: models must be tested on RNA families they were not trained on, otherwise apparent accuracy simply reflects memorised homologues rather than a genuinely learned folding prior. This thesis treats that train/test discipline as central, so that reported improvements over the classical baseline are meaningful rather than an artefact of data leakage.

The technical stack

The experiments are implemented in Python using standard deep-learning and scientific-computing tooling — neural-network frameworks for model definition and training, alongside NumPy and pandas for handling sequence datasets and structural annotations, and established RNA benchmarks for evaluation. The project is structured as a reproducible experimental pipeline covering data preparation, model training and quantitative comparison against classical predictors.

From prediction to therapeutics

Better structure prediction is not an academic nicety — it feeds directly into RNA-based medicine. The design of mRNA vaccines depends on choosing sequences that fold into stable, well-translated structures; small interfering RNAs and antisense oligonucleotides must reach their targets without being trapped in unintended folds; and riboswitches and other structured RNAs are increasingly viewed as druggable targets for small molecules. In each case, a fast and reliable computational estimate of how a candidate sequence will fold lets researchers screen and refine designs in silico long before committing to synthesis and assays, shortening iteration cycles across the life sciences.

Connection to my later research

RNA secondary structure prediction was the entry point into a longer line of work on structure-aware machine learning for the life sciences. The same core intuition — that biological molecules are governed by their shape, and that models should learn from structure rather than raw sequence — carries directly into my later Master's research on Aptamer Predictive Triage, where geometric and generative deep learning are applied to designing molecular binders. Both sit within a broader focus on applied AI for healthcare and medicine.