Generation of protein sequences using machine learning techniques
Inventors
Amimeur, Tileli • Ketchem, Randal Robert • Shaver, Jeremy Martin • Clark, Rutilio H. • Taylor, John Alex
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11587645-B2
Publication Date
2023-02-21
Expiration Date
Abstract
Amino acid sequences of antibodies can be generated using a generative adversarial network that includes a first generating component that generates amino acid sequences of antibody light chains and a second generating component generates amino acid sequences of antibody heavy chains. Amino acid sequences of antibodies can be produced by combining the respective amino acid sequences produced by the first generating component and the second generating component. The training of the first generating component and the second generating component can proceed at different rates. Additionally, the antibody amino acids produced by combining amino acid sequences from the first generating component and the second generating component may be evaluated according to complentarity-determining regions of the antibody amino acid sequences. Training datasets may be produced using amino acid sequences that correspond to antibodies have particular binding affinities with respect to molecules, such as binding affinity with major histocompatibility complex (MHC) molecules.
Core Innovation
The document describes an Antibody-GAN framework in which a computing system generates antibody amino-acid sequences using a generative adversarial network. It generates a plurality of first amino acid sequences corresponding to antibody light chains and a plurality of second amino acid sequences corresponding to antibody heavy chains, then combines the light-chain and heavy-chain sequences to produce a third amino acid sequence corresponding to an antibody including a light chain and a heavy chain.
The framework uses separate generating components for antibody light chains and antibody heavy chains within the generative adversarial network, enabling training that proceeds at different rates for the two components. Discriminator/critic feedback is used during training, including similarity scoring versus training data. The system also describes transfer learning and continued training on filtered datasets biased toward desired biophysical properties and reduced MHC binding or immunogenicity, while using structure-based position mapping for sequence encoding.
The document evaluates generated antibodies using metrics that include germline agreement, complementarity-determining region characteristics including CDR H3 length, hydrophobic and charged amino-acid counts, expression level, melting temperature, self-aggregation, and immunogenicity measures including MHC Class II binding. Proof-of-concept results include training a GAN on more than 400k antibody sequences to generate about 100k antibody libraries, followed by expression of proof-of-concept antibodies using phage display and stable CHO expression, and analyses showing control over the CDR H3 distribution and reduced predicted MHCII binding.
Claims Coverage
The independent claims cover a method and a system for generating paired antibody amino-acid sequences using a generative adversarial network with separate light-chain and heavy-chain generation, sequence combining, and similarity-based analysis against training-data sequences. Across the independent claims, the inventive features include dual generating components for light and heavy chains, combining to form an antibody light/heavy pairing, and producing a similarity-based output using training data.
Paired generative adversarial generation of light and heavy chains
Generating, by the computing system and using a generative adversarial network, a plurality of first amino acid sequences corresponding to antibody light chains and a plurality of second amino acid sequences corresponding to antibody heavy chains.
Combining light and heavy chain sequences to produce an antibody sequence
Combining, by the computing system and using the generative adversarial network, a first amino acid sequence of the plurality of first amino acid sequences with a second amino acid sequence of the plurality of second amino acid sequences to produce a third amino acid sequence corresponding to an antibody including a light chain corresponding to the first amino acid sequence and a heavy chain corresponding to the second amino acid sequence.
Similarity output based on training data
Analyzing, by the computing system and using the generative adversarial network, the third amino acid sequence with respect to an additional plurality of amino acid sequences included in training data for the generative adversarial network to produce an output indicating a measure of similarity between the third amino acid sequence and at least a portion of the additional plurality of amino acid sequences.
Training separate generating components at different rates
Training a first model of a first generating component of a generative adversarial network using a first training dataset including a first number of amino acid sequences of light chains to produce a first trained model, and training a second model of a second generating component using a second training dataset including a second number of amino acid sequences of heavy chains to produce a second trained model, wherein training the second generating component proceeds at a first rate different from a second rate of training the first generating component.
Generating additional light and heavy sequences and combining into a paired antibody sequence
Generating, using the first generating component, a first additional number of amino acid sequences corresponding to antibody light chains and generating, using the second generating component, a second additional number of amino acid sequences corresponding to antibody heavy chains, and combining using the generative adversarial network a first amino acid sequence from the first additional amino acid sequences with a second amino acid sequence from the second additional amino acid sequences to produce a third amino acid sequence corresponding to an antibody including a light chain and a heavy chain.
Across the independent claims, the inventive coverage centers on using a generative adversarial network with separate light-chain and heavy-chain generating components to generate amino-acid sequences, combining one light-chain sequence with one heavy-chain sequence to form a paired antibody amino-acid sequence, and then using training-data-related similarity analysis or a paired generation pipeline that trains the two generating components at different rates.
Stated Advantages
Provides an output indicating a measure of similarity between generated antibody third amino acid sequences and training-data sequences.
Uses separate generating components for antibody light chains and antibody heavy chains that are trained at different rates.
Documented Applications
Antibody-GAN sequence generation for antibody discovery, including generating paired antibody light chains and heavy chains and producing output measures based on similarity to training-data antibody sequences.
Interested in licensing this patent?