Systems and methods for generating and training convolutional neural networks using biological sequences and relevance scores derived from structural, biochemical, population and evolutionary data

Inventors

XIONG, Hui Yuan • Frey, Brendan

Assignees

Deep Genomics Inc

Abstract

We describe systems and methods for generating and training convolutional neural networks using biological sequences and relevance scores derived from structural, biochemical, population and evolutionary data. The convolutional neural networks take as input biological sequences and additional information and output molecular phenotypes. Biological sequences may include DNA, RNA and protein sequences. Molecular phenotypes may include protein-DNA interactions, protein-RNA interactions, protein-protein interactions, splicing patterns, polyadenylation patterns, and microRNA-RNA interactions, which may be described using numerical, categorical or ordinal attributes. Intermediate layers of the convolutional neural networks are weighted using relevance score sequences, for example, conservation tracks. The resulting molecular phenotype convolutional neural networks may be used in genetic testing, to identify drug targets, to identify patients that respond similarly to a drug, to ascertain health risks, or to connect patients that have similar molecular phenotypes.

Core Innovation

The invention relates to molecular phenotype convolutional neural networks (MPCNNs) in which convolutional layer operations are weighted using relevance score sequences. A biological sequence with a plurality of positions is processed by a first layer and subsequent layers of the MPCNN, including convolutional filters that link convolutional layer input positions to convolutional layer output positions. The last layer represents a molecular phenotype.

Each weighting unit is associated with a relevance score sequence comprising relevance score sequence positions, and each relevance score sequence position is associated with a numerical value that quantifies biological relevance of a corresponding position in the biological sequence. The numerical value is defined with respect to the convolutional filters to which the weighting unit is linked. The associated relevance score sequence is used to weight operations of the associated convolutional filter.

The relevance score sequences are derived from structural, biochemical, population, and evolutionary data tracks aligned to biological sequence positions, such as conservation, accessibility, nucleosome, RNA, and protein structural tracks. Relevance score sequences are also generated by a relevance neural network, optionally trained jointly with the MPCNN or trained separately, using backpropagation with modified gradients. The MPCNN uses an encoder and sequence encoding and includes training via gradient-based optimizers.

Claims Coverage

The partial content includes two independent claims: a system claim and a method claim. Both claims cover the same core concept: weighting convolutional filter operations in an MPCNN using relevance score sequences with numerical values that quantify biological relevance per position relative to convolutional filters. Overall, the claims include at least one key inventive feature in each independent claim, with refinements provided by dependent claims.

Across both independent claims, the inventive coverage centers on using relevance score sequences that assign numerical biological relevance to biological sequence positions relative to convolutional filters, and applying those relevance scores as weighting operations within an MPCNN to produce molecular phenotype outputs.

Stated Advantages

Documented Applications