RaptGen

RaptGen generates aptamer sequences using a variational autoencoder (VAE) with a profile hidden Markov model (HMM) decoder to expand and explore sequence space for nucleic acid aptamer discovery.

Key Features:

• Variational Autoencoder (VAE): RaptGen employs a VAE to map high-dimensional sequence data into a low-dimensional latent space that captures aptamer motif information.
• Profile Hidden Markov Model (HMM) Decoder: A profile HMM decoder represents and decodes motif sequences from the latent space to produce biologically relevant sequences.
• Latent Space Exploration: Embedding simulation sequence data into a motif-informed latent space enables exploration of sequence variants beyond experimentally observed high-throughput sequencing datasets.
• Sequence Embedding and Generation: Using two independent SELEX datasets, RaptGen embeds sequences into latent space and generates new aptamer candidates, including truncated aptamers produced by a short learning model.
• Activity-Guided Generation: RaptGen applies Bayesian optimization in latent space for activity-guided aptamer generation and iterative refinement of functional properties.

Scientific Applications:

• Aptamer discovery from limited data: Expand candidate diversity when SELEX or high-throughput sequencing datasets are constrained.
• Novel sequence generation: Produce new full-length and truncated aptamer candidates not present in existing sequencing datasets.
• Activity-directed optimization: Identify aptamers with desired functional properties or binding affinities via Bayesian optimization in latent space.
• Motif-informed design: Leverage captured motif information to maintain structural and functional relevance in generated sequences.

Methodology:

Variational autoencoder maps sequences to a low-dimensional latent space; a profile HMM decoder decodes motif sequences from the latent space; simulation sequence data are embedded into the latent space; two independent SELEX datasets are used for sequence embedding and generation; a short learning model produces truncated aptamers; Bayesian optimization is applied for activity-guided generation.