ML Embedding Clustering for DNA Sequence Screening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for screening DNA, RNA, and protein sequences are inefficient due to high background noise and low detection sensitivity, requiring multiple rounds of experiments and relying solely on sequence alignment that neglects higher-order interactions and biophysical properties, making it difficult to identify desired sequences.
Innovation Solution
A method involving the preparation of a library of sequences, subjecting them to screening experiments, creating a dataset of sequence identities and outcomes, and using machine learning to extract features, with a pre-trained model from external datasets to group similar sequences and redesign a more specific library, considering 2D and 3D structural information and biophysical properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional sequence alignment methods are used to group similar sequences, then sequences can be clustered based on character-level similarity, but higher-order interactions and biophysical properties are not considered
Solution Approach 1:
The patent transitions from character-level sequence alignment to embedding space clustering, adding a new dimensional representation of sequences. By converting sequences to vector embeddings that capture higher-order interactions and biophysical properties, the system clusters sequences in this new embedding space rather than traditional sequence space, thereby recovering information lost in conventional alignment methods
Solution Approach 2:
The patent introduces machine learning models (embedders) as intermediaries between the raw sequence data and the clustering analysis. These embedder models transform sequences into embedding vectors that encode higher-order interactions and biophysical properties, serving as a mediator that preserves information that would otherwise be lost in direct sequence comparison
2Measurement precision
If multiple rounds of screening experiments are conducted to amplify true-positive sequences, then detection sensitivity is improved, but time consumption and experimental costs increase
Solution Approach 1:
The patent performs preliminary clustering analysis using embedding-based similarity before conducting screening experiments. By pre-grouping sequences into clusters based on their embedding representations, the system identifies promising candidate clusters that are more likely to contain true-positive sequences, allowing researchers to focus screening efforts on these prioritized clusters rather than screening all sequences equally across multiple rounds
Solution Approach 2:
The patent employs unsupervised learning models that automatically learn sequence representations and perform clustering without requiring manual intervention or multiple experimental rounds. The embedder models self-adjust to capture relevant features, and the clustering algorithm automatically identifies meaningful groups, reducing the need for iterative experimental screening
3Measurement precision
If conventional sequence alignment is used to compare sequences, then sequence similarity can be calculated, but biophysical properties such as cell toxicity cannot be discovered
Solution Approach 1:
The patent changes the parameters used for sequence comparison from basic character-level alignment to embedding space distance metrics. By training embedder models on data that includes biophysical properties, the resulting embedding representations encode these properties, allowing similarity comparison in the embedding space to reflect not just sequence similarity but also biophysical property similarity
Data Source
AI summary
A method of characterizing biological sequences includes: preparing a library of sequences; subjecting the sequences in the library to at least one screening experiment to obtain an experiment outcome of each of the sequences; creating a first dataset comprising identities of the sequences and the experiment outcomes of the sequences; and training a first neural network using the first dataset to extract first sequence features from the sequences in the first dataset. A second neural network may be additionally be trained using a second dataset based on an external database to generate a pre-trained model, which is used extract additional features from the first dataset.