RNA-based medicines have developed into a broad therapeutic platform. Clinically established modalities include antisense oligonucleotides, small interfering RNAs, messenger RNAs that direct transient protein production for vaccines, and aptamers that recognize molecular targets through sequence-dependent three-dimensional folding 1. Machine learning is moving RNA drug development from empirical screening toward programmable engineering. Across messenger RNA, antisense oligonucleotides, small interfering RNA drugs, machine learning models can prioritize sequences for activity, predict target-site accessibility, structural stability, low off-target binding, reduced innate immune activation, and favorable delivery. In the exciting field of cancer vaccines, AI models can even help optimize the codon usage, degradation rate, and innate immune stimulation and its lipid-nanoparticle formulation, biodistribution, and storage stability [5,6]. Aptamers present a different problem because their function depends on a three-dimensional fold that forms a selective binding pocket rather than primarily on Watson–Crick recognition. The engineered RNA needs to be able to recognize and deplete a metabolite that sustains malignant growth or therapy resistance. Translation remains limited, with sparse training data, imperfect RNA-structure prediction, nuclease susceptibility, pharmacokinetics, and tissue delivery.
Most aptamers have traditionally been discovered by systematic evolution of ligands by exponential enrichment, or SELEX. SELEX begins with a very large randomized oligonucleotide pool (of approximately 10¹²–10¹⁵ sequences) for incubation with the target. Bound sequences are recovered and enriched through PCR amplification through iterative repeated incubation cycles. Counterselection against related molecules or background components are introduced at various intervening steps on the way to individual candidates through a labor-intensive process 2. However, enrichment is influenced by PCR efficiency, reverse-transcription bias, nonspecific adsorption, partitioning efficiency, sequence abundance, and structural refolding. Rare high-affinity candidates may disappear, whereas readily amplified but mediocre binders may dominate. However, the iterative process is ripe for AI-based selection rounds associated with folding, ligand docking, enrichment steps.
By substituting the initial blind search through a random sequence space with AI-enabled constrained inverse design informed by structural features and known ligand-binding parameters, computational effort can be focused on a biologically plausible search space, accelerating convergence toward high-affinity functional candidates. Although many contemporary aptamers are developed as protein-binding sensors, inhibitors, or targeting ligands, bacterial riboswitch aptamer domains provide a more biologically matched starting point for engineering binders to metabolites and other small molecules. These RNA domains were evolutionarily selected to recognize metabolites, vitamins, and metal ions with high chemical specificity; accordingly, the glnA riboswitch was chosen as a template because of its conserved structure, stability, and intrinsic specificity for L-glutamine in a recent study 3. Starting from such a scaffold imports an already organized ligand-binding pocket, conserved tertiary contacts, and ion-dependent recognition geometry that most importantly can achieve target selectivity that randomized SELEX library must discover de novo. By limiting the search space and preserving discrimination among closely related metabolites, in the glnA study, iterative modification of the evolutionarily conserved scaffold, followed by structural modeling and docking, produced an optimized aptamer with substantially increased glutamine affinity, maintained the selectivity over related amino acids 3. This example illustrated why riboswitch-derived templates are particularly well suited to metabolite-directed aptamer development, maintaining the evolutionarily acquired target selectivity. Consistent with recent deep-learning–assisted small-molecule aptamer studies, structural priors can improve library design and reduce the number of selection cycles required relative to conventional random-library searches.
In order to not disrupt conserved regions of the glutamine-targeting aptamer from riboswitches of 1,410 unique species, only the variable regions were considered modifiable, altering stem length and nucleotide identity, introducing truncations and stabilizing elements 3. The crystal structure-based models of the resulting aptamers were subjected to docking-derived total energy () was used as a relative estimate of binding free energy and equilibrium behavior 3. An automated pipeline could first retrieve naturally occurring riboswitches known to recognize the metabolite or a structurally related ligand. Evolutionarily conserved residues, long-range base pairs, ion-coordination sites, and experimentally defined pocket nucleotides would be fixed or assigned low mutation probabilities. Less-conserved stems and loops would form the design space. Every candidate would then undergo automated secondary-structure prediction, ensemble-fold analysis, tertiary modeling, and docking against the intended metabolite. The same library needs to be negatively selected against stereoisomers, chemically related metabolites, pathway intermediates, and abundant extracellular molecules. The goal being the lowest target , coupled to a high off-target . Candidates with favorable target energy but poor discrimination would be removed. Because docking is a model-dependent ranking score rather than an experimentally measured thermodynamic free energy, it should be calibrated against measured dissociation constants and treated as one component of a multi-objective score. Bayesian optimization or uncertainty-guided active learning could then propose the next variants expected either to improve binding or to provide the most informative data. Local sequence changes, stem stabilization, pocket mutations, loop shortening, terminal extensions, and chemically modified nucleotides could therefore be explored systematically rather than manually.
Experimental testing showed that the optimized construct had substantially greater affinity than the original riboswitch and that a single pocket mutation markedly impaired binding 3. A machine-learning implementation would learn these sequence–structure–energy relationships after each cycle and automatically identify epistatic combinations that are difficult to infer through one-position-at-a-time mutagenesis. However, for metabolite depletion, the objective must extend beyond affinity. Passive sequestration is stoichiometric: one aptamer molecule generally removes one metabolite molecule. This may be insufficient for rapidly replenished metabolites. More effective designs could incorporate multivalent scaffolds, slow dissociation, tumor-localized retention, or coupling to metabolite-catabolizing enzymes. Unexpectedly, the optimized aptamer not only bound glutamine but increased extracellular glutaminase activity and raised its maximum reaction velocity, as an allosteric enhancer creating a mechanism that combined sequestration with catalytic depletion 3. While in this case the two distinct interfaces were likely always present in the original riboswitches from which the ultimate aptamer was derived, suggesting a novel role for bacterial riboswitches originally discovered as metabolic sensors for gene regulation. However, now one could therefore optimize two interfaces simultaneously through AI if sufficient examples could be provided to teach the requirements for a aptamer–metabolite pocket and the aptamer–enzyme interaction. Such metabolite depletion could complement chemotherapy mechanistically by transiently reducing substrates that support nucleotide synthesis, redox buffering, bioenergetics, or DNA repair, without defining a disease-specific clinical indication.
Agentic AI could coordinate the entire aptamer-development workflow rather than provide isolated predictions. Systems such as ChemCrow, Coscientist, LLM-RDF, and ChemAgents demonstrate how specialized agents can integrate literature retrieval, computational analysis, experimental design, instrument control, and data interpretation 4-7, although human oversight remains necessary because tool use can propagate structural, force-field, or energy-ranking errors and does not uniformly improve performance 8. Complementary aptamer-design approaches already illustrate key components of this strategy: RhoDesign performs structure-to-sequence generation [8], DeepAptamer identifies high-affinity candidates from early SELEX data 9, DL-SELEX applies variational autoencoders to library design and selection analysis 10, and AIoptamer combines sequence generation, affinity prediction, structural reconstruction, and molecular-dynamics refinement 11. A riboswitch-guided, -minimizing active-learning framework could integrate these capabilities to focus directly on metabolite recognition and depletion.
References
- Wang, S., Weissman, D. & Dong, Y. (2025) RNA chemistry and therapeutics. Nat Rev Drug Discov 24: 828.
- Tuerk, C. & Gold, L. (1990) Systematic evolution of ligands by exponential enrichment: RNA ligands to bacteriophage T4 DNA polymerase. Science 249: 505.
- Cruz-Hernandez, C.D., Smith, B., Billet, S., Thiruvalluvan, M., Gonzales, G., Underhill, D.M., Seki, E., Lu, S.C. & Bhowmick, N.A. (2025) A Cyanobacteria-Derived RNA Aptamer Resensitizes Prostate Cancer to Hormone Therapy. Cancer Res 85: 2714.
- Song, T., Luo, M., Zhang, X., Chen, L., Huang, Y., Cao, J., Zhu, Q., Liu, D., Zhang, B., Zou, G., Zhang, G., Zhang, F., Shang, W., Fu, Y., Jiang, J. & Luo, Y. (2025) A Multiagent-Driven Robotic AI Chemist Enabling Autonomous Chemical Research On Demand. J Am Chem Soc 147: 12534.
- Ruan, Y., Lu, C., Xu, N., He, Y., Chen, Y., Zhang, J., Xuan, J., Pan, J., Fang, Q., Gao, H., Shen, X., Ye, N., Zhang, Q. & Mo, Y. (2024) An automatic end-to-end chemical synthesis development platform powered by large language models. Nat Commun 15: 10160.
- A, M.B., Cox, S., Schilter, O., Baldassari, C., White, A.D. & Schwaller, P. (2024) Augmenting large language models with chemistry tools. Nat Mach Intell 6: 525.
- Boiko, D.A., MacKnight, R., Kline, B. & Gomes, G. (2023) Autonomous chemical research with large language models. Nature 624: 570.
- Yu, B., Baker, F.N., Chen, Z., Herb, G., Gou, B., Adu-Ampratwum, D., Ning, X. & Sun, H. (2025) Tooling or Not Tooling? The Impact of Tools on Language Agents for Chemistry Problem Solving. Find ACL NAACL 2025: 7620.
- Yang, X., Chan, C.H., Yao, S., Chu, H.Y., Lyu, M., Chen, Z., Xiao, H., Ma, Y., Yu, S., Li, F., Liu, J., Wang, L., Zhang, Z., Zhang, B.T., Zhang, L., Lu, A., Wang, Y., Zhang, G. & Yu, Y. (2025) DeepAptamer: Advancing high-affinity aptamer discovery with a hybrid deep learning model. Mol Ther Nucleic Acids 36: 102436.
- Zhao, Z., Lin, H., Lau, H.Y., Chen, H. & Hsing, I.M. (2025) Structure-enhanced deep learning accelerates aptamer selection for small molecule families like steroids. Brief Bioinform 26.
- Gupta, T., Sharma, P., Malik, S. & Pant, P. (2025) AIoptamer: Artificial Intelligence-Driven Aptamer Optimization Pipeline for Targeted Therapeutics in Healthcare. Mol Pharm 22: 4076.
