Autoregressive Protein Sequence Generation via Latent Space Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The exploration of protein sequence space is challenging due to its vastness and the difficulty in accurately predicting protein fold and function, making it difficult to generate proteins with desired properties.
Innovation Solution
A method using autoregressive neural networks, comprising an encoder and a decoder with an autoregressive module, generates protein sequences by determining amino acid probabilities based on previous locations in the sequence, trained end-to-end with an encoder that encodes protein sequences in a latent space.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If systematic exploration of protein variants is performed, then functional variants can be obtained, but the enormous sequence space and difficulty of predicting protein fold and function make it extremely challenging
Solution Approach 1:
The patent introduces an intermediary system (the neural network model comprising encoder, latent space, and decoder) that mediates between the input protein sequence and the output modified sequence. This intermediary learns the complex mapping between sequence and structure/function through training, avoiding the need for explicit prediction of protein fold and function during the modification process itself.
Solution Approach 2:
The patent replaces traditional mechanical/computational methods for protein design (such as explicit energy minimization, molecular dynamics simulations, and structure prediction algorithms) with a data-driven neural network approach. The system learns patterns from training data and directly generates modified sequences without explicitly modeling the physical chemistry of protein folding and function.
2Adaptability or versatility
If random mutations are generated in protein sequence, then variants can be obtained, but the vast number of possible sequences (20^300 for length 300) makes it impossible to explore meaningful functional variants
Solution Approach 1:
The system incorporates feedback through the training process where the neural network learns from training examples the relationship between sequence modifications and functional outcomes. During generation, the decoder uses feedback from the encoder and the learned latent representations to guide the creation of meaningful variants, ensuring they fall within the manifold of functional sequences rather than exploring the entire theoretical sequence space.
Solution Approach 2:
The system performs preliminary action by pre-training on a large dataset of known functional protein sequences to learn the underlying structure-function relationships. This preliminary learning phase enables the system to generate meaningful variants without having to explore the vast sequence space during the actual modification process, as the constraints and patterns are already encoded in the trained model.
Data Source
AI summary
The invention relates to a method for generating protein sequences in an autoregressive neural network comprising an encoder and a decoder, the decoder comprising an autoregressive module, the method comprising:obtaining a latent code and inputting the latent code into the autoregressive module;obtaining probabilities from the autoregressive module, the obtained probabilities representing probabilities for amino acids to be selected at locations of the sequence; andgenerating an ordered sequence of amino acids from the obtained probabilities,wherein the autoregressive module is such that, for a location of the sequence, the probabilities associated with amino acids to be selected for the location in the sequence are determined as a function of probabilities of amino acids to be selected for previous locations in the sequence andwherein the autoregressive neural network has been trained end-to-end with the encoder, the encoder making it possible to encode protein sequences in a latent space as latent codes.