DNAformer Neural Network for DNA Storage Error Reconstruction

The patent replaces traditional mechanical error correction methods (such as Reed-Solomon codes or simple majority voting) with a deep neural network system. The DNN learns complex patterns of synthesis and sequencing errors including deletions and insertions, enabling accurate reconstruction of original DNA sequences despite these errors. This substitution allows the system to handle the complexity of biological errors that traditional algorithms cannot effectively correct.

The patent transforms the error correction problem from a discrete algebraic problem into a continuous probabilistic problem. By using soft estimates (probability distributions over possible nucleotides) instead of hard decisions, the DNN can incorporate uncertainty and gradually resolve ambiguities through multiple processing layers. This parameter transformation enables the system to handle deletion and insertion errors more effectively by modeling them as probabilistic events rather than deterministic errors.

The patent performs preliminary error modeling and feature extraction before the main clustering and reconstruction tasks. The DNN is pre-trained to recognize patterns of synthesis and sequencing errors, and to generate soft estimates that capture the likelihood of different nucleotide sequences. This preliminary processing simplifies the subsequent clustering task by providing the algorithm with pre-processed, error-aware representations of the DNA strands, reducing the computational burden of exact clustering.

The patent introduces soft estimates as an intermediary representation between the raw sequencing data and the final clustered results. Instead of directly clustering hard sequence assignments, the system clusters probabilistic soft estimates that encode uncertainty and error information. This intermediary layer acts as a buffer that simplifies the clustering problem by working with smoothed, continuous representations rather than discrete, noisy sequences.

The patent implements a feedback mechanism where the DNN processes multiple sequence copies and uses the collective information to reconstruct the original sequence. The soft estimates from multiple copies are combined and fed back through the neural network, allowing the system to iteratively improve reconstruction accuracy. This feedback loop enables effective error correction using fewer copies than traditional methods would require, as the DNN learns to extract maximum information from each copy and intelligently combine them.

The patent enables the system to self-correct errors by training the DNN on simulated error patterns that mimic real synthesis and sequencing errors. The model learns to recognize and correct these errors autonomously without requiring external intervention or complex post-processing. This self-service capability reduces the need for excessive redundancy in synthesized copies, as the system can effectively correct errors from a smaller number of copies through its learned error models.