Generative and interpretable machine learning for aptamer design and analysis of in vitro sequence selection

Andrea Di Gioacchino, Jonah Procyk, Marco Molari, John S. Schreck, Yu Zhou, Yan Liu, Rémi Monasson, Simona Cocco, Petr Sulc

Introduction

Aptamers are short DNA or RNA sequences capable of binding specific targets with high affinity, offering a cost-effective and easily produced alternative to antibodies. The standard discovery process, Systematic Evolution of Ligands by Exponential Enrichment (SELEX), involves iterative rounds of selecting and amplifying binders from a massive random pool. However, traditional analysis of SELEX data often relies on simple sequence counting, which fails to capture the full complexity of the binding landscape or explain the underlying structural motifs. Furthermore, the lack of "negative" examples (sequences that failed to bind) makes standard supervised machine learning difficult to apply.

Machine Learning for Generative Aptamer Design

To overcome these limitations, researchers have employed Restricted Boltzmann Machines (RBMs)—a class of unsupervised, two-layer generative stochastic neural networks. Unlike supervised models, RBMs learn the probability distribution of the "winner" sequences that survive selection, effectively modeling what a "good" sequence looks like based on the ensemble of survivors.

Overview of the RBM-integrated SELEX workflow, from sequence selection to generative design and experimental validation.

The RBM architecture consists of a "visible" layer representing the DNA sequence and a "hidden" layer that learns latent features or motifs. The study utilized two training protocols: RBM-SC (Sequence Counts), which models high-fitness peaks by weighting sequence frequency, and RBM-SU (Unique Sequences), which encourages the learning of broader structural rules for de novo design. This approach allows the model to generate entirely new, functional aptamers that were not present in the original experimental data.

Learning the Fitness Landscape

A profound finding of this research is that the log-likelihood score assigned by the RBM correlates linearly with experimental enrichment ratios. This indicates that the model effectively learns the physical fitness landscape of binding from nothing but a list of survivors.

By modeling bivalent aptamers (two loops on a DNA nanotile), the authors identified "parasite" sequences—aptamers where one loop binds strongly while the other is essentially random "dead weight." They also discovered that the total fitness of multi-loop aptamers is roughly additive, suggesting that individual loops bind to protein exosites independently.

Results and Motif Discovery

The RBM-based approach demonstrated high predictive accuracy, correctly identifying the binder/non-binder status of 93% (25 out of 27) of synthesized test sequences in lab assays. Crucially, the model automatically discovered functional motifs, such as G-quadruplexes, without any prior structural knowledge.

Evolution of sequence log-likelihood distributions across SELEX rounds and visualization of discovered binding motifs.

As shown in the figure above, the distribution of sequence log-likelihoods shifts toward higher values in later SELEX rounds, reflecting the enrichment of high-affinity binders. The associated sequence logos reveal specific conservation patterns responsible for thrombin binding, including a single-nucleotide "switch" at position 17 (Adenine vs. Thymine) that determines which part of the thrombin protein the aptamer prefers to bind.

Conclusion

By transforming SELEX data into a generative tool, RBMs allow for the design of novel, functional aptamers that extend beyond the sequences observed in experimental pools. While the model is currently limited by the quality of the initial SELEX library and does not explicitly model 3D folding, its ability to outperform standard supervised learning approaches marks a significant step forward in computational aptamer discovery. Future work may involve integrating these generative models with 3D structural prediction tools to further refine design capabilities for complex targets.

Summary updated

Abstract

Abstract Selection protocols such as SELEX, where molecules are selected over multiple rounds for their ability to bind to a target molecule of interest, are popular methods for obtaining binders for diagnostic and therapeutic purposes. With the increasing amount of such high-throughput experimental data available, machine learning techniques have become increasingly popular for molecular datasets analysis. Here, we show that Restricted Boltzmann Machines (RBMs), a two-layer neural network architecture, can successfully be trained on sequence ensembles from SELEX experiments for thrombin aptamers, and used to estimate the fitness of the sequences obtained through the experimental protocol. As a direct consequence, we show that trained RBMs can be exploited to classify as well as generate novel molecules. To confirm our findings, we experimentally verify the generated sequences from RBM.
BibTeX
@article{gioacchino2022generative,
  author = {Andrea Di Gioacchino and Jonah Procyk and Marco Molari and John S. Schreck and Yu Zhou and Yan Liu and Rémi Monasson and Simona Cocco and Petr Sulc},
  title = {{Generative and interpretable machine learning for aptamer design and analysis of in vitro sequence selection}},
  journal = {PLOS Computational Biology},
  year = {2022},
  doi = {10.1371/journal.pcbi.1010561},
}
Figure