TL;DR
Researchers have developed AptaBLE, a deep learning platform that significantly improves the prediction of aptamer-protein binding and enables the de novo generation of novel aptamers. This work addresses the limitations of traditional aptamer discovery methods, offering a faster, more cost-effective, and potentially less biased approach for developing aptamer-based therapeutics and diagnostics.
What Problem Does This Paper Address?
Discovering aptamers – short, single-stranded DNA or RNA molecules that bind to specific targets with high affinity – is crucial for developing new diagnostics and therapeutics. The current gold standard, Systematic Evolution of Ligands by Exponential enrichment (SELEX), is a laborious, multi-month process prone to experimental biases, particularly from PCR amplification, which can lead to the loss of potentially valuable aptamers. While computational methods exist, they often struggle with structural information scarcity, computational cost, or limited scope. This paper aims to bridge this gap by presenting a deep learning framework that can both predict aptamer-protein interactions and generate novel aptamers, overcoming the limitations of existing methods.
Methods
AptaBLE employs a symmetric deep learning architecture that processes protein and aptamer sequences independently using specialized pretrained encoders (ESM-2 for proteins, Nucleotide Transformer v2 for DNA aptamers, and RNA-FM for RNA aptamers). These encoded representations are then fused using a novel symmetric cross-attention mechanism. This architecture allows for bidirectional information flow, capturing complex interaction patterns without relying on explicit 3D structural information. For de novo aptamer generation, AptaBLE integrates with two methods: Monte Carlo Tree Search (MCTS) and Monte Carlo Tree Guidance (MCTG). These methods use AptaBLE's predictions as a reward signal to guide the exploration of the vast sequence space, iteratively generating and refining aptamer sequences with desired binding properties.
Datasets
The study utilized several datasets:
- Training Data: A curated dataset of validated aptamer-protein binding pairs from the UTexas Aptamer Database, Li et al. (2014), and proprietary data from Atom Bioworks. This included 1,089 DNA aptamers and 757 RNA aptamers. Negative examples were generated by shuffling sequences from positive examples.
- Benchmark Datasets:
- A benchmark set of 16 ssDNA and 39 RNA aptamers, with their protein targets omitted from the training data, to evaluate generalization.
- An auxiliary benchmark set of four known aptamer-protein pairs (SARS-CoV-2 Spike glycoprotein, gp120, and basic fibroblast growth factor) for per-complex comparison.
- Retrospective Analysis: Data from a previously performed SELEX experiment targeting human CD117 (hCD117) was used to evaluate AptaBLE's ability to identify high-affinity binders from existing SELEX libraries.
Key Insights and Ideas
- Sequence-based interaction prediction: AptaBLE demonstrates that deep learning models can effectively predict aptamer-protein binding using only sequence information, bypassing the need for complex structural data.
- Overcoming SELEX limitations: The framework can identify potentially high-affinity aptamers that might be overlooked by traditional SELEX clustering methods due to experimental biases.
- De novo generation: AptaBLE can be used not just for prediction but also for actively designing novel aptamers with specific binding affinities and selectivities, opening new avenues for rational aptamer design.
- Cross-attention for interface insights: Attention maps generated by AptaBLE can provide clues about potential protein-aptamer binding interfaces, although further investigation is needed to confirm their direct interpretability.
Key Results
- AptaBLE significantly outperforms baseline methods like AptaTrans and AlphaFold3 in predicting aptamer-protein binding for both DNA and RNA aptamers, particularly in terms of recall (minimizing false negatives), which is crucial for SELEX analysis.
- The platform successfully identified previously unclustered aptamers from a hCD117 SELEX library that exhibited high-affinity binding.
- De novo generation using AptaBLE produced novel aptamers with low-nanomolar affinities (e.g., ~31 nM for human CD25) and high selectivity.
- Runtime analysis showed AptaBLE's efficiency, with scoring a single aptamer-protein pair taking approximately 48 milliseconds on an A100 GPU.
Limitations and Future Work
- Interpretability of attention maps: While attention maps offer potential insights into binding interfaces, their direct correlation with specific atomic interactions requires further validation.
- RNA aptamer generation: The current work focuses primarily on ssDNA aptamer generation; future work aims to validate AptaBLE-designed RNA aptamers.
- Finer-grained workflows: The authors plan to explore more detailed applications, such as using AptaBLE for mutagenesis studies and optimizing aptamer sequences for specific therapeutic or diagnostic applications.
- Broader applicability: The team anticipates that AptaBLE will bridge the gap between computational tools and their widespread adoption in aptamer design and generation workflows.
Summary updated