New study explores an FMQA-based optimization framework for RNA design

RNA design is central to next-generation therapeutics, yet identifying sequences that reliably fold into desired structures remains a major computational challenge, often constrained by high cost and time. A new study from Keio University explores the use of factorization machine with quadratic-optimization annealing (FMQA) for RNA inverse folding, while also examining how different encoding strategies may influence artificial intelligence (AI)-driven design performance, revealing an underexplored dimension of biomolecular engineering.

RNA has emerged as one of the most promising molecules in modern medicine, enabling advances from mRNA vaccines and gene therapies to genome editing and synthetic biology. However, designing RNA molecules that reliably fold into a desired secondary structure remains a major challenge. Even for relatively short sequences, the number of possible nucleotide combinations grows exponentially, making it difficult to identify optimal candidates. As a result, conventional computational methods often require extensive candidate evaluations, creating a significant bottleneck when experimental validation is both time-consuming and costly.

To address this challenge, researchers from Keio University, led by Project Lecturer Shuta Kikuchi of the Graduate School of Science and Technology and Professor Shu Tanaka of the Department of Applied Physics and Physico-Informatics, developed a novel RNA inverse folding framework based on Factorization Machine with Quadratic Optimization Annealing (FMQA). This machine learning and Ising machine-driven black-box optimization approach is designed to identify high-quality RNA sequence candidates with relatively few evaluations.

“We investigated a new application of FMQA in biomolecular design, where its potential remains relatively unexplored. Since RNA, DNA, and protein sequences are inherently categorical in nature, it is unclear how converting them into binary representations affects optimization performance. In this study, we examined RNA inverse folding and the influence of different encoding and assignment choices within FMQA,” says Dr. Kikuchi. The findings were published in Scientific Reports on May 3, 2026.

The researchers formulated RNA inverse folding as an optimization problem aimed at identifying sequences most likely to fold into a predefined target structure. FMQA served as the core optimization engine, and its performance was evaluated across four binary encoding methods—one-hot, domain-wall, binary, and unary—alongside all possible nucleotide-to-integer assignments for adenine (A), uracil (U), guanine (G), and cytosine (C). RNA design quality was assessed using the Normalized Ensemble Defect (NED), which measures the agreement between predicted and target structures. FMQA was benchmarked against random search, genetic algorithms, and Bayesian optimization.

The results showed that the encoding strategy plays a decisive role in artificial intelligence (AI) and Ising machine-driven RNA design. One-hot and domain-wall encodings consistently outperformed binary and unary representations, producing sequences with lower NED values and higher success rates. Importantly, domain-wall encoding introduced a search bias toward specific integer states. When guanine (G) and cytosine (C) were assigned to these favored states, G–C base pairs accumulated more frequently in stem regions, resulting in greater thermodynamic stability and improved design performance. Across benchmarks, FMQA also identified high-quality RNA designs with fewer function evaluations than competing methods, demonstrating strong efficiency in search-constrained settings.

Beyond RNA inverse folding, the findings carry broader implications for computational biology and optimization science. They demonstrate that annealing-based optimization frameworks such as FMQA can be effectively extended to life-science problems, strengthening the bridge between quantum-inspired computing and biomolecular engineering. More importantly, the study highlights that data encoding is not merely a preprocessing step, but a design variable that can fundamentally shape optimization outcomes. These insights may guide future applications of FMQA in biomolecular design, materials discovery, and polymer engineering.

Looking ahead, this approach could accelerate the design of functional biomolecules, particularly RNA systems that must reliably adopt specific structures for therapeutic or diagnostic applications. “Potential applications include biosensors, genome-editing tools, aptamers, ribozymes, and riboswitches,” notes Dr. Kikuchi. “Because DNA, RNA, and proteins are all represented by categorical biological sequences, the approach may also be extended to broader biomolecular design.” Furthermore, because FMQA is a flexible black-box optimization framework, future implementations could incorporate experimentally measured properties such as molecular stability, binding affinity, or gene-expression control, helping to bridge computational design and laboratory validation. “The insights gained from this study are not limited to RNA,” adds Prof. Tanaka. “They have a generality that allows them to be applied to discrete design problems where each evaluation is costly, including materials and molecular design.” In the long term, such evaluation-efficient optimization strategies may help reduce the experimental burden and accelerate discovery across biotechnology and medicine.

“Because FMQA formulates the learned surrogate model as a quadratic optimization problem, it can be implemented with quantum annealing machines,” says Dr. Kikuchi. “This perspective points to an exciting future direction: advancing ‘Quantum for Biology’ by exploring how next-generation quantum and quantum-inspired computing technologies can support biomolecular design.”

In conclusion, this study establishes FMQA as a powerful and evaluation-efficient framework for RNA inverse folding. It also highlights a key but often overlooked insight: the way biological sequences are encoded can be as influential as the optimization algorithm itself. Together, these findings open new directions for more efficient, scalable, and effective approaches to biomolecular design.


Reference
Title of original paper: Factorization machine with quadratic-optimization annealing for RNA inverse folding and evaluation of binary-integer encoding and nucleotide assignment
Journal: Scientific Reports
DOI: https://doi.org/10.1038/s41598-026-50891-7