Antibody Sequence Generation Using Trained Models for High Binding Affinity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for designing antibodies with high binding affinity are limited by semi-laboratory techniques and frequency-based computational methods, which fail to ensure the generated sequences have optimal binding properties.
Innovation Solution
A method and system for generating antibody sequences by creating complementarity determining regions (CDR) from framework regions (FR) using trained models, such as Autoregressive Convolutional Neural Networks, Long Short-Term Memory networks, Markov models, and GPT-2, to produce sequences with high binding affinity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If frequency-based computational methods are used to design antibodies, then the design process can be automated, but the binding affinity of generated sequences cannot be guaranteed
Solution Approach 1:
The patent applies preliminary action by pre-training deep learning models on large datasets of known antibody-antigen pairs before actual antibody design. This pre-training phase enables the model to learn binding affinity patterns in advance, so when generating new antibody sequences, the model can predict and optimize for high binding affinity from the start, rather than relying on post-generation filtering or manual validation.
Solution Approach 2:
The patent implements feedback mechanisms through the use of deep learning models that continuously predict binding affinity during the antibody sequence generation process. The model provides real-time feedback on the predicted affinity of generated sequences, allowing the system to iteratively optimize sequences to meet binding affinity thresholds, thereby ensuring reliability while maintaining automation.
2Adaptability or versatility
If deep mutational scanning and large libraries are used to explore protein sequences, then more possibilities can be covered, but the cost and complexity of synthesis and sequencing increase
Solution Approach 1:
The patent replaces the mechanical and chemical processes of deep mutational scanning and physical sequencing with a computational deep learning model. Instead of synthesizing and sequencing large libraries of antibody variants in the lab, the system uses trained neural networks to predict and generate high-affinity antibody sequences in silico, dramatically reducing the complexity and cost of exploring sequence space while maintaining or improving coverage of viable possibilities.
3Ease of manufacture
If whole antibody sequences are generated at once, then the process is simple, but there is no way to ascertain high binding affinity
Solution Approach 1:
The patent applies segmentation by dividing the antibody sequence generation process into distinct functional regions: framework regions (FR1-4) and complementarity-determining regions (CDR1-3). The deep learning model is trained to generate CDR sequences that specifically target and bind to antigen epitopes, while FR regions provide structural stability. This segmented approach allows the system to maintain simplicity in the overall generation process while achieving precise control over binding affinity through specialized CDR design.
Data Source
AI summary
A method and system for generating a plurality of antibody sequences of a target from one or more framework regions based on at least one model. The model is trained on a training dataset of high binding affinity to generate the complementarity determining regions (CDR) from the received one or more framework regions (FR). The generated complementarity determining regions (CDR) from the each of the one or more framework regions (FR) are combined with the associated one or more framework regions to generate one or more regions of the target. The generated one or more regions comprises each of the received one or more framework regions (FR) and corresponding each of the generated complementarity determining regions (CDR). The generated one or more regions are concatenated to generate the plurality of antibody sequences of the target. The generated plurality of antibody sequences of the target has high binding affinity.