The invention provides systems and methods for discovering candidate therapies for genetic conditions and also for screening those therapies in vitro for evidence of neurotoxicity. Where a medical condition is a consequence of a genetic target such as a mutated gene, the disclosure provides in silico methods to generate lists of candidate sequences for antisense oligonucleotides (ASOs) that will potentially bind to the gene or transcripts from the gene in vivo and treat the associated condition by restoring a healthy phenotype of gene expression. The invention provides in vitro methods for screening candidate ASO sequences for symptoms of neurotoxicity in vivo. For example, candidate sequences that are output by the in silico analytical pipeline can be synthesized and assayed against live cells in vitro.

A system and method for designing protein sequences conditioned on a specific target fold. The system is a transformer-based generative framework for modeling a complex sequence-structure relationship. To mitigate the heterogeneity between the sequence domain and the fold domain, a Fold-to-Sequence model jointly learns a sequence embedding using a transformer and a fold embedding from the density of secondary structural elements in 3D voxels. The joint sequence-fold representation through novel intra-domain and cross-domain losses with an intra-domain loss forcing two semantically similar (where the proteins should have the same fold(s)) samples from the same domain to be close to each other in a latent space, while a cross-domain loss forces two semantically similar samples in different domains to be closer. In an embodiment, the Fold-to-Sequence model performs design tasks that include low resolution structures, structures with region of missing residues, and NMR structural ensembles.

A system and method for designing protein sequences conditioned on a specific target fold. The system is a transformer-based generative framework for modeling a complex sequence-structure relationship. To mitigate the heterogeneity between the sequence domain and the fold domain, a Fold-to-Sequence model jointly learns a sequence embedding using a transformer and a fold embedding from the density of secondary structural elements in 3D voxels. The joint sequence-fold representation through novel intra-domain and cross-domain losses with an intra-domain loss forcing two semantically similar (where the proteins should have the same fold(s)) samples from the same domain to be close to each other in a latent space, while a cross-domain loss forces two semantically similar samples in different domains to be closer. In an embodiment, the Fold-to-Sequence model performs design tasks that include low resolution structures, structures with region of missing residues, and NMR structural ensembles.

Disclosed herein are optimized guide RNAs (gRNAs) that have increased target binding specificity and reduced off-target binding. Further disclosed herein are methods of designing and using the optimized gRNAs.

Disclosed herein are optimized guide RNAs (gRNAs) that have increased target binding specificity and reduced off-target binding. Further disclosed herein are methods of designing and using the optimized gRNAs.

A computer-implemented method for inferring a 3D structure of a genome is provided. The method includes providing genome interaction data and operating an autoencoder including a structured sequence of n autoencoder units, each of which including an encoder unit and a decoder unit, each of which is implemented as a recurrent neural network unit. The method includes additionally training the autoencoder by feeding all vectors of genome interaction data to the encoder units. Thereby, the training of the auto-encoder units is performed stepwise by using inner state of respective previous autoencoder units in the cascaded sequence of autoencoder units and performing backpropagation within each of the plurality of autoencoder units after all autoencoder units have processed their respective input values, and using the output values of the encoder units for deriving a 3D model for a visualization of the genome.

A computer-implemented method for inferring a 3D structure of a genome is provided. The method includes providing genome interaction data and operating an autoencoder including a structured sequence of n autoencoder units, each of which including an encoder unit and a decoder unit, each of which is implemented as a recurrent neural network unit. The method includes additionally training the autoencoder by feeding all vectors of genome interaction data to the encoder units. Thereby, the training of the auto-encoder units is performed stepwise by using inner state of respective previous autoencoder units in the cascaded sequence of autoencoder units and performing backpropagation within each of the plurality of autoencoder units after all autoencoder units have processed their respective input values, and using the output values of the encoder units for deriving a 3D model for a visualization of the genome.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a binding prediction neural network. In one aspect, a method comprises: instantiating a plurality of structure prediction neural networks, wherein each structure prediction neural network has a respective neural network architecture and is configured to process data defining an input polynucleotide to generate data defining a predicted structure of the input polynucleotide; training each of the plurality of structure prediction neural networks; after training the plurality of structure prediction neural networks, determining a respective performance measure of each structure prediction neural network based at least in part on a prediction accuracy of the structure prediction neural network; and generating, based on the performance measures of the structure prediction neural networks, a binding prediction neural network.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a binding prediction neural network. In one aspect, a method comprises: instantiating a plurality of structure prediction neural networks, wherein each structure prediction neural network has a respective neural network architecture and is configured to process data defining an input polynucleotide to generate data defining a predicted structure of the input polynucleotide; training each of the plurality of structure prediction neural networks; after training the plurality of structure prediction neural networks, determining a respective performance measure of each structure prediction neural network based at least in part on a prediction accuracy of the structure prediction neural network; and generating, based on the performance measures of the structure prediction neural networks, a binding prediction neural network.

Methods for the top-down design of nucleic acid nanostructures of arbitrary geometry based on target shape of spherical or non-spherical topology are described. The methods facilitate 3D molecular programming of lipids, proteins, sugars, and RNAs based on a DNA scaffold of arbitrary 2D or 3D shape. Geometric objects are rendered as node-edge networks of parallel nucleic acid duplexes, and a nucleic acid scaffold routed throughout the network using a spanning tree formula. Nucleic acid nanostructures produced according to top-down design methods are also described. In some embodiments, the nanostructures include single-stranded nucleic acid scaffold, DX crossovers, and staple strands. In other embodiments, the nanostructures include single-stranded nucleic acid scaffold, PX crossovers and no staples. Modified nanostructures include chemically modified nucleotides and conjugated to other molecules are described.