TL;DR
This paper introduces a novel application of Restricted Boltzmann Machines (RBMs) to analyze aptamer selection data (SELEX). They demonstrate that RBMs can accurately predict aptamer binding affinity, generate novel aptamer sequences with high binding potential, and reveal underlying sequence features responsible for binding. This work offers a powerful new tool for designing aptamers for diagnostics and therapeutics.
What Problem Does This Paper Address?
Discovering molecules that can specifically bind to a target is crucial for diagnostics and therapeutics. Aptamers, short DNA or RNA molecules, are promising candidates due to their specificity and ease of synthesis. The SELEX (Systematic Evolution of Ligands by Exponential Enrichment) process is a common method for selecting aptamers. However, analyzing the vast amount of data generated by SELEX and predicting aptamer "fitness" (binding ability) remains a significant challenge. Existing methods struggle to reliably quantify fitness for individual sequences, hindering the design of new, improved aptamers.
Methods
The core of this work is the application of Restricted Boltzmann Machines (RBMs), a type of two-layer neural network. The authors train RBMs in an unsupervised manner on sequences obtained from SELEX experiments.
- RBM Training: They explore two training strategies:
- RBM-SC (with Counts): This method incorporates the frequency (counts) of each sequence in the SELEX rounds. This approach excels at modeling the very best binders but struggles with generalization.
- RBM-SU (Unique Sequences): This method focuses on unique sequences, discarding their counts. This approach demonstrates superior generalization capabilities and is better at capturing the broader landscape of aptamer sequences.
- Fitness Prediction: The log-likelihood assigned by the RBM to a sequence is shown to be a strong predictor of its experimental binding fitness.
- Sequence Generation: Trained RBMs are used to generate novel aptamer sequences in silico.
- Feature Extraction: The learned weights of the RBM are analyzed to identify sequence motifs (like G-quadruplexes) that contribute to binding.
- Clustering: The activity of RBM's hidden units is used to cluster aptamer sequences into families.
Datasets
The primary dataset used in this study consists of DNA sequences obtained from SELEX experiments for thrombin aptamers. Specifically, data from rounds 5, 6, 7, and 8 of the SELEX procedure were used, including the sequences and their corresponding counts. This dataset is derived from a prior publication [55].
Key Insights and Ideas
- RBM Log-likelihood as a Fitness Proxy: The log-likelihood score assigned by an RBM trained on SELEX data directly correlates with the experimental binding fitness of an aptamer. This is a crucial finding, as it provides a computational way to estimate how well an aptamer will bind.
- Generalization is Key for Design: Training RBMs on unique sequences (RBM-SU) rather than sequence counts (RBM-SC) leads to significantly better generalization. This means the model can predict the fitness of unseen sequences and generate novel, diverse aptamers, which is essential for de novo design.
- Additive Contributions of Aptamer Loops: The binding affinity of double-loop aptamers can be accurately modeled as the additive contribution of each individual loop's binding affinity. This suggests a modular mechanism for aptamer binding.
- Interpretable Features: RBM weights reveal biologically relevant sequence motifs, such as G-quadruplexes, which are known to be important for thrombin binding. This interpretability is a significant advantage over black-box models.
- "Parasite" Sequences: Early SELEX rounds can contain "parasite" sequences where one loop binds the target strongly, while the other loop is less specific. These can be identified and their impact on RBM training understood.
Key Results
- High Correlation: The RBM log-likelihood showed a very strong correlation (R² = 0.99) with experimental enrichment ratios for aptamer sequences.
- Accurate Prediction: The RBM-SU model achieved 93% accuracy in predicting whether generated sequences would bind thrombin, validated through experimental gel shift assays.
- De Novo Design Success: The RBM-SU model successfully generated 27 aptamer sequences, with 19 out of 21 predicted binders confirmed experimentally.
- Motif Identification: RBM weights revealed G-rich motifs, consistent with known G-quadruplex structures important for thrombin binding.
- Clustering of Families: Hidden unit activities in RBM-SC allowed for the clustering of aptamers into four previously identified families.
- Exosite Specificity: The study found that most generated aptamers preferentially bind to exosite I of thrombin, and RBM-SU models this preference better than RBM-SC.
Limitations and Future Work
- RBM-SC vs. RBM-SU Trade-off: While RBM-SC accurately models high-fitness peaks (excellent binders), it fails to generalize and generate diverse sequences. RBM-SU generates diverse binders but doesn't perfectly capture the highest fitness peaks. The authors suggest that RBM-SU is more appropriate for aptamer design.
- Competition Assay Nuances: While RBM-SU predicts binding well, it underestimates the strength of known high-affinity binders (ThA, ThD) in competition assays. The authors suggest this might be due to the model's focus on diversity over absolute counts and potential allosteric effects in thrombin binding.
- Supervised Learning Challenges: Standard supervised learning approaches (DNNs, random forests) struggled with the SELEX dataset, likely due to the inherent bias towards binders and the difficulty in accurately labeling non-binders.
- Future Work: The authors suggest that RBMs can be applied to other aptamer datasets and selection protocols, including competition assays for complex targets and even other molecular design problems like antibody discovery or protein engineering. Further investigation into the allosteric mechanisms of thrombin binding could also be beneficial.
Summary updated