Machine learning-guided peptide design

Abstract

Systems, devices, software, and methods for engineering amino acid sequences configured to possess specific protein functions or properties. The methods implement machine learning to process an input seed sequence and generate an optimized sequence with the desired function or property as output.

Description

[0001] Related applications

[0002] This application claims the benefit of U.S. Provisional Application Nos. 62 / 882,150 and 62 / 882,159, both filed on August 2, 2019. The entire teachings of the above applications are incorporated herein by reference.

[0003] Incorporate material by reference into an ASCII text file

[0004] This application incorporates by reference the Sequence Listing contained in the following concurrently filed ASCII text file:

[0005] a) File name: GBD_SeqListing_ST25.txt; created on July 29, 2020, size: 5 KB. Background Art

[0006] Proteins are essential macromolecules that perform or are involved in many functions within organisms, including, for example, catalyzing metabolic reactions, promoting DNA replication, responding to stimuli, providing structure to cells and tissues, and transporting molecules. Proteins are composed of one or more chains of amino acids and typically form a three-dimensional conformation. Summary of the Invention

[0007] Described herein are systems, devices, software, and methods for generating or modifying protein or polypeptide sequences to achieve a function and / or property or to improve upon the function and / or property thereof. These sequences can be determined computationally on a computer. Artificial intelligence or machine learning is used to provide a novel framework for rationally engineering proteins or polypeptides. Thus, novel polypeptide sequences that differ from naturally occurring proteins can be generated to possess a desired function or property.

[0008] Designing amino acid sequences (e.g., proteins) for specific functions has long been a goal in molecular biology. However, predicting protein amino acid sequences based on function or properties is highly challenging, at least in part due to the structural complexity that can arise from seemingly simple primary amino acid sequences. To date, one approach has been to use in vitro random mutagenesis followed by selection, resulting in a directed evolution process. However, this approach is time- and resource-intensive, typically requiring the generation of mutant clones, which in turn are subject to biases in library design or limited exploration of sequence space, screening these clones for the desired properties, and iteratively repeating this process. In fact, traditional methods have failed to provide accurate and reproducible methods for predicting protein function based on amino acid sequence, let alone predicting amino acid sequence based on protein function. Indeed, the conventional wisdom regarding function-based protein primary sequence prediction is that primary protein sequence cannot be directly linked to known function because so much protein function is driven by its ultimate tertiary (or quaternary) structure.

[0009] In contrast, the ability to use computational or computer methods to engineer proteins with properties or functions of interest could transform the field of protein design. Despite extensive research on this topic, little progress has been made to date. Thus, disclosed herein are innovative systems, devices, software, and methods for generating amino acid sequence codes for polypeptides or proteins configured to have specific properties and / or functions. Thus, the innovations described herein are unexpected and yield unexpected results given conventional thinking about protein analysis and protein structure.

[0010] Described herein is a method for engineering an improved biopolymer sequence evaluated by function, the method comprising: (a) calculating a change in the function relative to an embedding at a starting point according to a step size, the starting point being provided to a system comprising a supervised model for predicting the function of a biopolymer sequence and a decoder network, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network being trained to provide a probabilistic biopolymer sequence given the embedding of the biopolymer sequence in the function space, optionally wherein the starting point is an embedded seed biopolymer sequence, thereby providing a first update point in the function space; (b) optionally calculating a change in the function relative to the embedding at the first update point in the function space, and optionally iteratively calculating a change in the function relative to the embedding at further update points; (c) providing the first update point or optionally iterative further update points in the function space when the first update point or optionally iterative further update points approach a desired level of function, to the decoder network; and (d) obtaining a probabilistically improved biopolymer sequence from the decoder.

[0011] In this article, the term "function" can have a dual meaning. On the one hand, function can qualitatively represent some properties and / or abilities of a protein in the biological field (e.g., fluorescence). On the other hand, function can quantitatively represent some quality factors associated with such properties and / or abilities in the biological field, such as a measure of the intensity of a fluorescence effect.

[0012] Therefore, the term "function space" is not limited to its mathematical meaning of a set of functions that accept an input from the same space and map that input to an output in the same space or another space. Rather, a function space can include a compressed representation of a biopolymer sequence from which the value of a function—that is, a quantitative figure of merit for a desired property and / or capability—can be derived.

[0013] In particular, a compressed representation can include two or more numerical values ​​that can be interpreted as coordinates in a Cartesian vector space having two or more dimensions. However, the Cartesian vector space may not be completely filled by these compressed representations. Instead, the compressed representation may form a subspace within the Cartesian vector space. This is one meaning of the term "embedded" as used herein for compressed representations.

[0014] In some embodiments, the embedding is a continuously differentiable function space representing the function and having one or more gradients. In some embodiments, computing the change of the function relative to the embedding comprises taking a derivative of the function with respect to the embedding.

[0015] In particular, supervised model training can link embeddings to functionality in the following sense: that is, if two biopolymer sequences have similar values ​​of the quality factor in the quantitative sense of functionality, then their compressed representations are close in the feature space. This facilitates targeted updates to the compressed representation to obtain biopolymer sequences with improved quality factors.

[0016] The phrase "having one or more gradients" should not be interpreted as being limited in the sense that the gradient must be computed according to some explicit function mapping the compressed representation to a quantitative figure of merit. Rather, the dependence of the figure of merit on the compressed representation may be a learned relationship for which no explicit functional term is available. For such a learned relationship, the gradient in the functional space of the embedding may be computed, for example, by backpropagation. For example, if a first compressed representation of a biopolymer sequence in the embedding is converted by a decoder into a biopolymer sequence, and this biopolymer sequence is in turn fed to an encoder and mapped to a compressed representation, the supervised model may compute the quantitative figure of merit from such a compressed representation. The gradient of this figure of merit with respect to its value in the original compressed representation may then be obtained by backpropagation. This is inFigure 3A This is described in more detail in .

[0017] As mentioned earlier, a particular embedding space and a particular quality factor can be two sides of the same coin, since compressed representations with similar quality factors are close together in the embedding space. Therefore, an embedding space can be considered “differentiable” if there is a meaningful way to obtain the gradient of the quality factor function with respect to the numerical values ​​that make up the compressed representation.

[0018] The term "probabilistic biopolymer sequence" may include, in particular, a distribution of biopolymer sequences from which a biopolymer sequence can be obtained by sampling. For example, if a biopolymer sequence of a defined length L is sought, and the set of available amino acids at each position is fixed, then for each position and each available amino acid in the sequence, the probabilistic biopolymer sequence may indicate the probability that the position is occupied by that particular amino acid. This is in the Figure 3C This is described in more detail in .

[0019] In some embodiments, the function is a composite function of two or more component functions. In some embodiments, the composite function is a weighted sum of the two or more composite functions. In some embodiments, two or more starting points in the embedding are used simultaneously, for example, at least two starting points. In some embodiments, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 100, or 200 starting points may be used simultaneously, however this is a non-limiting list. In some embodiments, during sampling, correlations between residues in a probabilistic sequence comprising a probability distribution for residue identities are considered using conditional probabilities that take into account already generated sequence portions. In some embodiments, the method further comprises selecting a maximum likelihood improved biopolymer sequence from the probabilistic biopolymer sequence comprising a probability distribution for residue identities. In some embodiments, the method further comprises sampling a marginal distribution at each residue in the probabilistic biopolymer sequence comprising a probability distribution for residue identities. In some embodiments, the change in the function relative to the embedding is calculated by calculating the change in the function relative to the encoder, then calculating the change in the encoder relative to the change in the decoder, and calculating the change in the decoder relative to the embedding. In some embodiments, the method includes providing a first update point in the function space or a further update point in the function space to the decoder network to provide an intermediate probabilistic biopolymer sequence, providing the intermediate probabilistic biopolymer sequence to the supervised model network to predict a function of the intermediate probabilistic biopolymer sequence, and then calculating a change in an embedding of the function relative to the intermediate probabilistic biopolymer to provide a further update point in the function space.

[0020] A system is described herein, comprising a processor; and a non-transitory computer-readable medium encoded with software, the software configured to cause the processor to: (a) calculate a change in the embedding of the function relative to a starting point according to a step size, thereby providing a first update point in the function space, the starting point being provided to a system comprising a supervised model for predicting a function of a biopolymer sequence and a decoder network, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network being trained to provide a probabilistic biopolymer sequence given the embedding of the biopolymer sequence in the function space, optionally wherein the starting point is an embedded seed biopolymer sequence; (b) optionally calculating a change in the embedding of the function relative to a first update point in the function space, and optionally iteratively calculating a change in the embedding of the function relative to further update points; (c) providing the first update point or optionally iterative further update points in the function space to the decoder network when the first update point or optionally iterative further update points approach a desired level of function; and (d) obtaining a probabilistically improved biopolymer sequence from the decoder. In some embodiments, the embedding is a continuously differentiable function space representing the function and having one or more gradients. In some embodiments, calculating the change in the function relative to the embedding includes taking the derivative of the function with respect to the embedding. In some embodiments, the function is a composite function of two or more component functions. In some embodiments, the composite function is a weighted sum of the two or more composite functions. In some embodiments, two or more starting points in the embedding are used simultaneously, for example, at least two. In certain embodiments, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 100, or 200 starting points may be used, however, this is a non-limiting list. In some embodiments, during the sampling process, correlations between residues in a probabilistic sequence including a probability distribution of residue identities are considered using conditional probabilities taking into account already generated sequence portions. In some embodiments, the processor is further configured to select a maximum likelihood improved biopolymer sequence from the probabilistic biopolymer sequence including a probability distribution of residue identities. In some embodiments, the processor is further configured to sample a marginal distribution at each residue in the probabilistic biopolymer sequence including a probability distribution of residue identities. In some embodiments, the change of the function relative to the embedding is calculated by calculating the change of the function relative to the encoder, then calculating the change of the encoder relative to the decoder, and calculating the change of the decoder relative to the embedding.In some embodiments, the processor is further configured to: provide a first updated point in the function space or a further updated point in the function space to the decoder network to provide an intermediate probabilistic biopolymer sequence, provide the intermediate probabilistic biopolymer sequence to the supervised model network to predict a function of the intermediate probabilistic biopolymer sequence, and then calculate a change in an embedding of the function relative to the intermediate probabilistic biopolymer to provide a further updated point in the function space.

[0021] Described herein is a non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to: (a) calculate a change in the embedding of the function relative to a starting point according to a step size, thereby providing a first update point in the function space, wherein the starting point is provided to a system comprising a supervised model for predicting a function of a biopolymer sequence and a decoder network, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a probabilistic biopolymer sequence given the embedding of the biopolymer sequence in the function space, optionally wherein the starting point is an embedded seed biopolymer sequence; (b) optionally calculate a change in the embedding of the function relative to a first update point in the function space, and optionally iteratively calculate a change in the embedding of the function relative to further update points; (c) provide the first update point or optionally iterative further update points in the function space to the decoder network when the first update point or optionally iterative further update points approach a desired level of function; and (d) obtain a probabilistically improved biopolymer sequence from the decoder. In some embodiments, the embedding is a continuously differentiable function space representing the function and having one or more gradients. In some embodiments, calculating the change in the function relative to the embedding includes taking the derivative of the function with respect to the embedding. In some embodiments, the function is a composite function of two or more component functions. In some embodiments, the composite function is a weighted sum of the two or more composite functions. In some embodiments, two or more starting points in the embedding are used simultaneously, for example, at least two. In some embodiments, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 100, or 200 starting points may be used, but this is a non-limiting list. In some embodiments, during the sampling process, correlations between residues in the probabilistic sequence including a probability distribution for residue identity are considered using conditional probabilities that take into account the generated sequence portions. In some embodiments, the processor is further configured to select a maximum likelihood improved biopolymer sequence from the probabilistic biopolymer sequence including a probability distribution for residue identity. In some embodiments, the processor is further configured to sample a marginal distribution at each residue in the probabilistic biopolymer sequence including a probability distribution for residue identity. In some embodiments, the change of the function relative to the embedding is calculated by calculating the change of the function relative to the encoder, then calculating the change of the encoder relative to the decoder, and calculating the change of the decoder relative to the embedding.In some embodiments, the processor is further configured to: provide a first updated point in the function space or a further updated point in the function space to the decoder network to provide an intermediate probabilistic biopolymer sequence, provide the intermediate probabilistic biopolymer sequence to the supervised model network to predict a function of the intermediate probabilistic biopolymer sequence, and then calculate a change in an embedding of the function relative to the intermediate probabilistic biopolymer to provide a further updated point in the function space.

[0022] Disclosed herein is a method for engineering improved biopolymer sequences evaluated by function, the method comprising: (a) predicting a function of an embedding starting point, the starting point being provided to a system comprising a supervised model network for predicting the function of a biopolymer sequence and a decoder network, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network being trained to provide a predicted probabilistic biopolymer sequence, optionally wherein the starting point is an embedding seed biopolymer sequence; (b) calculating a change in the embedding of the function relative to the starting point according to a step size, thereby providing a first updated point in the function space; (c) at the decoder network, calculating a first intermediate probabilistic biopolymer sequence based on the first updated point in the function space; and (d) at the supervised model, predicting the first intermediate probabilistic biopolymer sequence based on the first intermediate probabilistic biopolymer sequence. (e) calculating a change in the embedding of the function relative to a first update point in the function space to provide an updated point in the function space; (f) calculating, at the decoder network, an additional intermediate probabilistic biopolymer sequence based on the updated point in the function space; (g) predicting, by the supervised model, a function of the additional intermediate probabilistic biopolymer sequence based on the additional intermediate probabilistic biopolymer sequence; (h) calculating a change in the embedding of the function relative to a further first update point in the function space to provide a further updated point in the function space, optionally iterating steps (g)-(i), wherein the further update point in the function space referenced in step (h) is considered to be the further updated point in the function space in step (f); and (i) providing the embedded point in the function space to the decoder network when approaching a desired level of function; and obtaining a probabilistically improved biopolymer sequence from the decoder. In some embodiments, the biopolymer is a protein. In some embodiments, the seed biopolymer sequence is an average of a plurality of sequences. In some embodiments, the seed biopolymer sequence has no function or has a level of function below the desired level of function. In some embodiments, the encoder is trained using a training dataset of at least 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, or 200 biopolymer sequences. In some embodiments, the encoder is a convolutional neural network (CNN) or a recurrent neural network (RNN). In some embodiments, the encoder is a transformer neural network. In some embodiments, the encoder comprises one or more convolutional layers, pooling layers, fully connected layers, normalization layers, or any combination thereof. In some embodiments, the encoder is a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional network. In some embodiments, the convolutional neural network is a two-dimensional or higher dimensional convolutional neural network.In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the encoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the encoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the encoder is optimized by a program selected from the following: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta or NAdam. In some embodiments, the encoder is trained using a transfer learning procedure. In some embodiments, the transfer learning procedure comprises: training a first model using a first training dataset of biopolymer sequences without functional labels, generating a second model comprising at least a portion of the first model, and training the second model using a second training dataset of biopolymer sequences with functional labels, thereby generating a trained encoder. In some embodiments, the decoder is trained using a training dataset of at least 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, or 200 biopolymer sequences. In some embodiments, the decoder is a convolutional neural network (CNN) or a recurrent neural network (RNN). In some embodiments, the decoder is a transform neural network. In some embodiments, the decoder comprises one or more convolutional layers, pooling layers, fully connected layers, normalization layers, or any combination thereof. In some embodiments, the decoder is a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional network. In some embodiments, the convolutional neural network is a two-dimensional or higher-dimensional convolutional neural network. In some embodiments, the convolutional neural network has a convolutional architecture selected from: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet.In some embodiments, the decoder comprises at least 10, 50, 100, 250, 500, 750, or 1000 layers. In some embodiments, the decoder employs a regularization method comprising L1-L2 regularization on one or more layers, skip connections on one or more layers, dropouts on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the decoder is optimized using a procedure selected from the group consisting of Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta, or NAdam. In some embodiments, the decoder is trained using a transfer learning procedure. In some embodiments, the transfer learning procedure comprises: training a first model using a first biopolymer sequence training dataset without function labels, generating a second model comprising at least a portion of the first model, and training the second model using a second biopolymer sequence training dataset with function labels, thereby generating a trained decoder. In some embodiments, one or more functions of the improved biopolymer sequence are improved compared to one or more functions of the seed biopolymer sequence. In some embodiments, the one or more functions are selected from fluorescence, enzyme activity, nuclease activity, and protein stability. In some embodiments, the biopolymer sequence is evaluated using a weighted linear combination of two or more functions.

[0023] A computer system is described herein that includes a processor; and a non-transitory computer-readable medium encoded with software, the software configured to cause the processor to: (a) calculate a change in the embedding of the function relative to a starting point according to a step size, thereby providing a first update point in the function space, the embedding starting point being provided to a system including a supervised model network of a function for predicting a biopolymer sequence and a decoder network, the supervised model network including an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network being trained to provide a predicted probabilistic biopolymer sequence given the embedding of the predicted biopolymer sequence in the function space, optionally wherein the starting point is an embedding seed biopolymer sequence; (b) at the decoder network, calculate a first intermediate probabilistic biopolymer sequence based on the first update point in the function space; and (c) at the supervised model, predict a first intermediate probabilistic biopolymer sequence based on the first intermediate probabilistic biopolymer sequence. (d) calculating a function of the first intermediate probabilistic biopolymer sequence, (e) calculating a change in the embedding of the function relative to a first update point in the function space to provide an updated point in the function space; (f) predicting a function of the additional intermediate probabilistic biopolymer sequence at the supervised model based on the additional intermediate probabilistic biopolymer sequence; (g) calculating a change in the embedding of the function relative to a further first update point in the function space to provide a further updated point in the function space, optionally iterating steps (f)-(g), wherein the further update point in the function space referenced in step (g) is considered the further update point in the function space in step (e); and (i) providing the embedded point in the function space to the decoder network when approaching a desired level of function; and (j) obtaining a probabilistically improved biopolymer sequence from the decoder. In some embodiments, the biopolymer is a protein. In some embodiments, the seed biopolymer sequence is an average of a plurality of sequences. In some embodiments, the seed biopolymer sequence has no function or has a function level below the desired level of function. In some embodiments, the encoder is trained using a training dataset of at least 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, or 200 biopolymer sequences. In some embodiments, the encoder is a convolutional neural network (CNN) or a recurrent neural network (RNN). In some embodiments, the encoder is a transformation neural network. In some embodiments, the encoder comprises one or more convolutional layers, pooling layers, fully connected layers, normalization layers, or any combination thereof. In some embodiments, the encoder is a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional network.In some embodiments, the convolutional neural network is a two-dimensional or higher dimensional convolutional neural network. In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the encoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the encoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the encoder is optimized using a procedure selected from the group consisting of Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta, or NAdam. In some embodiments, the encoder is trained using a transfer learning procedure. In some embodiments, the transfer learning procedure comprises training a first model using a first biopolymer sequence training dataset without labeled functions, generating a second model comprising at least a portion of the first model, and training the second model using a second biopolymer sequence training dataset with labeled functions, thereby generating a trained encoder. In some embodiments, the decoder is trained using a training dataset of at least 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, or 200 biopolymer sequences. In some embodiments, the decoder is a convolutional neural network (CNN) or a recurrent neural network (RNN). In some embodiments, the decoder is a transform neural network. In some embodiments, the decoder comprises one or more convolutional layers, pooling layers, fully connected layers, normalization layers, or any combination thereof. In some embodiments, the decoder is a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional neural network. In some embodiments, the convolutional neural network is a two-dimensional or higher-dimensional convolutional neural network.In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the decoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the decoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the decoder is optimized by a program selected from the following: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta or NAdam. In some embodiments, the decoder is trained using a transfer learning procedure. In some embodiments, the transfer learning procedure comprises: training a first model using a first biopolymer sequence training dataset without function labels, generating a second model comprising at least a portion of the first model, and training the second model using a second biopolymer sequence training dataset with function labels, thereby generating a trained decoder. In some embodiments, one or more functions of the improved biopolymer sequence are improved compared to one or more functions of the seed biopolymer sequence. In some embodiments, the one or more functions are selected from fluorescence, enzyme activity, nuclease activity, and protein stability. In some embodiments, the biopolymer sequence is evaluated using a weighted linear combination of two or more functions.

[0024] A non-transitory computer-readable medium is described herein, the computer-readable medium comprising instructions that, when executed by a processor, cause the processor to: (a) predict a function of an embedding starting point, wherein the starting point is an embedding seed biopolymer sequence, the starting point being provided to a system comprising a supervised model network for predicting the function of the biopolymer sequence and a decoder network, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network being trained to provide a predicted probabilistic biopolymer sequence given the predicted biopolymer sequence is embedded in the function space; (b) calculate a change in the embedding of the function relative to the starting point according to a step size, thereby providing a first update point in the function space; (c) provide the first update point in the function space to the decoder network to provide a first intermediate probabilistic biopolymer sequence; and (d) generate a first intermediate probabilistic biopolymer sequence based on the first intermediate probabilistic biopolymer sequence by the supervised model.

[0014] The method of claim 1, wherein the decoder network is configured to predict a function of a first intermediate probabilistic biopolymer sequence based on a sequence of sequences; (e) calculating a change in the embedding of the function relative to a first update point in the function space to provide an updated point in the function space; (f) providing, by the decoder network, an additional intermediate probabilistic biopolymer sequence based on the updated point in the function space; (g) predicting a function of the additional intermediate probabilistic biopolymer sequence and providing the additional intermediate probabilistic biopolymer sequence to the supervised model; (h) calculating a change in the embedding of the function relative to a further first update point in the function space to provide a further updated point in the function space, optionally iterating steps (f)-(h), wherein the further updated point in the function space referenced in step (h) is considered to be the further updated point in the function space in step (f); and (i) providing the embedded point to the decoder network when approaching a desired level of function in the function space; and obtaining a probabilistically improved biopolymer sequence from the decoder network. In some embodiments, the biopolymer is a protein. In some embodiments, the seed biopolymer sequence is an average of a plurality of sequences. In some embodiments, the seed biopolymer sequence has no function or a level of function below the desired level of function. In some embodiments, the encoder is trained using a training dataset of at least 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, or 200 biopolymer sequences. In some embodiments, the encoder is a convolutional neural network (CNN) or a recurrent neural network (RNN). In some embodiments, the encoder is a transformation neural network. In some embodiments, the encoder comprises one or more convolutional layers, pooling layers, fully connected layers, normalization layers, or any combination thereof. In some embodiments, the encoder is a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional network. In some embodiments, the convolutional neural network is a two-dimensional or higher dimensional convolutional neural network.In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the encoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the encoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the encoder is optimized by a program selected from the following: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta or NAdam. In some embodiments, the encoder is trained using a transfer learning procedure. In some embodiments, the transfer learning procedure comprises: training a first model using a first training dataset of biopolymer sequences without functional labels, generating a second model comprising at least a portion of the first model, and training the second model using a second training dataset of biopolymer sequences with functional labels, thereby generating a trained encoder. In some embodiments, the decoder is trained using a training dataset of at least 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, or 200 biopolymer sequences. In some embodiments, the decoder is a convolutional neural network (CNN) or a recurrent neural network (RNN). In some embodiments, the decoder is a transform neural network. In some embodiments, the decoder comprises one or more convolutional layers, pooling layers, fully connected layers, normalization layers, or any combination thereof. In some embodiments, the decoder is a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional network. In some embodiments, the convolutional neural network is a two-dimensional or higher-dimensional convolutional neural network. In some embodiments, the convolutional neural network has a convolutional architecture selected from: VGG16, VGG19, DeepResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet.In some embodiments, the decoder comprises at least 10, 50, 100, 250, 500, 750, or 1000 or more layers. In some embodiments, the decoder employs a regularization method comprising L1-L2 regularization on one or more layers, skip connections on one or more layers, dropouts on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the decoder is optimized using a procedure selected from the group consisting of Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta, or NAdam. In some embodiments, the decoder is trained using a transfer learning procedure. In some embodiments, the transfer learning procedure comprises: training a first model using a first biopolymer sequence training dataset without function labels, generating a second model comprising at least a portion of the first model, and training the second model using a second biopolymer sequence training dataset with function labels, thereby generating a trained decoder. In some embodiments, one or more functions of the improved biopolymer sequence are improved compared to one or more functions of the seed biopolymer sequence. In some embodiments, the one or more functions are selected from fluorescence, enzyme activity, nuclease activity, and protein stability. In some embodiments, the biopolymer sequence is evaluated using a weighted linear combination of two or more functions.

[0025] Disclosed herein is a computer-implemented method for engineering a biopolymer sequence having a specified protein function, the method comprising: (a) generating an embedding of an initial biopolymer sequence via an encoder method; (b) iteratively changing the embedding to correspond to the specified protein function via an optimization method by adjusting one or more embedding parameters to generate an updated embedding; and (c) processing the updated embedding via a decoder method to generate a final biopolymer sequence. In some embodiments, the biopolymer sequence comprises a primary protein amino acid sequence. In some embodiments, the amino acid sequence results in a protein configuration that produces a protein function. In some embodiments, the protein function comprises fluorescence. In some embodiments, the protein function comprises enzymatic activity. In some embodiments, the protein function comprises nuclease activity. In some embodiments, the protein function comprises a degree of protein stability. In some embodiments, the encoder method is configured to receive the initial biopolymer sequence and generate the embedding. In some embodiments, the encoder method comprises a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional neural network. In some embodiments, the convolutional neural network is a two-dimensional or higher-dimensional convolutional neural network. In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the encoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the encoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the encoder is optimized by a program selected from the following: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta or NAdam. In some embodiments, the decoder method comprises a deep convolutional neural network. In some embodiments, the biopolymer sequence is evaluated using a weighted linear combination of two or more functions. In some embodiments, the optimization method generates an updated embedding using gradient-based descent in a continuously differentiable embedding space.In some embodiments, the optimization method uses an optimization scheme selected from Adam, RMS Prop, Ada delta, AdamMAX, or SGD with momentum. In some embodiments, the final biopolymer sequence is further optimized for at least one additional protein function. In some embodiments, the optimization method generates the updated embedding based on a composite function that combines the protein function and the at least one additional protein function. In some embodiments, the composite function is a weighted linear combination of two or more functions corresponding to the protein function and the at least one additional protein function.

[0026] Disclosed herein is a computer-implemented method for engineering a biopolymer sequence having a specified protein function, comprising: (a) generating an embedding of an initial biopolymer sequence via an encoder method; (b) adjusting the embedding to achieve the specified protein function via an optimization method by modifying one or more embedding parameters to generate an updated embedding; and (c) processing the updated embedding via a decoder method to generate a final biopolymer sequence.

[0027] Described herein is a computer system comprising a processor and a non-transitory computer-readable medium encoded with software configured to cause the processor to: (a) generate an embedding of an initial biopolymer sequence via an encoder method; (b) generate an updated embedding via an optimization method by iteratively modifying the embedding to correspond to the specified protein function by adjusting one or more embedding parameters; and (c) process the updated embedding via a decoder method to generate a final biopolymer sequence. In some embodiments, the biopolymer sequence comprises a primary protein amino acid sequence. In some embodiments, the amino acid sequence results in a protein configuration that produces a protein function. In some embodiments, the protein function comprises fluorescence. In some embodiments, the protein function comprises enzymatic activity. In some embodiments, the protein function comprises nuclease activity. In some embodiments, the protein function comprises a degree of protein stability. In some embodiments, the encoder method is configured to receive the initial biopolymer sequence and generate the embedding. In some embodiments, the encoder method comprises a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional neural network. In some embodiments, the convolutional neural network is a two-dimensional or higher-dimensional convolutional neural network. In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the encoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the encoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the encoder is optimized by a program selected from the following: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta or NAdam. In some embodiments, the decoder method comprises a deep convolutional neural network. In some embodiments, the biopolymer sequence is evaluated using a weighted linear combination of two or more functions. In some embodiments, the optimization method generates an updated embedding using gradient-based descent in a continuously differentiable embedding space.In some embodiments, the optimization method uses an optimization scheme selected from Adam, RMS Prop, Ada delta, AdamMAX, or SGD with momentum. In some embodiments, the final biopolymer sequence is further optimized for at least one additional protein function. In some embodiments, the optimization method generates the updated embedding based on a composite function that combines the protein function and the at least one additional protein function. In some embodiments, the composite function is a weighted linear combination of two or more functions corresponding to the protein function and the at least one additional protein function.

[0028] Described herein is a non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to: (a) generate an embedding of an initial biopolymer sequence via an encoder method; (b) generate an updated embedding via an optimization method by iteratively adjusting one or more embedding parameters to correspond to the specified protein function; and (c) process the updated embedding via a decoder method to generate a final biopolymer sequence. In some embodiments, the biopolymer sequence comprises a primary protein amino acid sequence. In some embodiments, the amino acid sequence results in a protein configuration that produces a protein function. In some embodiments, the protein function comprises fluorescence. In some embodiments, the protein function comprises enzymatic activity. In some embodiments, the protein function comprises nuclease activity. In some embodiments, the protein function comprises a degree of protein stability. In some embodiments, the encoder method is configured to receive the initial biopolymer sequence and generate the embedding. In some embodiments, the encoder method comprises a deep convolutional neural network. In some embodiments, the convolutional neural network is a one-dimensional convolutional neural network. In some embodiments, the convolutional neural network is a two-dimensional or higher-dimensional convolutional neural network. In some embodiments, the convolutional neural network has a convolutional architecture selected from the following: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet or MobileNet. In some embodiments, the encoder includes at least 10, 50, 100, 250, 500, 750 or 1000 or more layers. In some embodiments, the encoder uses a regularization method including L1-L2 regularization on one or more layers, skip connections on one or more layers, exit on one or more layers, or a combination thereof. In some embodiments, the regularization is performed using batch normalization. In some embodiments, the regularization is performed using group normalization. In some embodiments, the encoder is optimized by a program selected from the following: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta or NAdam. In some embodiments, the decoder method comprises a deep convolutional neural network. In some embodiments, the biopolymer sequence is evaluated using a weighted linear combination of two or more functions. In some embodiments, the optimization method generates an updated embedding using gradient-based descent in a continuously differentiable embedding space.In some embodiments, the optimization method uses an optimization scheme selected from Adam, RMS Prop, Ada delta, AdamMAX, or SGD with momentum. In some embodiments, the final biopolymer sequence is further optimized for at least one additional protein function. In some embodiments, the optimization method generates the updated embedding based on a composite function that combines the protein function and the at least one additional protein function. In some embodiments, the composite function is a weighted linear combination of two or more functions corresponding to the protein function and the at least one additional protein function.

[0029] Disclosed herein is a method of making a biopolymer comprising synthesizing an improved biopolymer sequence obtainable by the method of any one of the preceding embodiments or using the system of any one of the preceding embodiments.

[0030] Disclosed herein is a fluorescent protein comprising an amino acid sequence relative to SEQ ID NO: 1 and having increased fluorescence relative to SEQ ID NO: 1, the amino acid sequence comprising a substitution at a site selected from Y39, F64, V68, D129, V163, K166, G191, or a combination thereof. In some embodiments, the fluorescent protein comprises substitutions at 2, 3, 4, 5, 6, or all 7 of Y39, F64, V68, D129, V163, K166, and G191. In some embodiments, the fluorescent protein comprises S65 relative to SEQ ID NO: 1. In some embodiments, the amino acid sequence comprises S65 relative to SEQ ID NO: 1. In some embodiments, the amino acid sequence comprises substitutions at F64 and V68. In some embodiments, the amino acid sequence comprises 1, 2, 3, 4, or all 5 of Y39, D129, V163, K166, and G191. In some embodiments, the substitutions at Y39, F64, V68, D129, V163, K166, or G191 are Y39C, F64L, V68M, D129G, V163A, K166R, or G191V, respectively. In some embodiments, the fluorescent protein comprises an amino acid sequence that is at least 80%, 85%, 90%, 92%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, 99% or more identical to SEQ ID NO: 1. In some embodiments, the fluorescent protein comprises at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, or 15 mutations relative to SEQ ID NO: 1. In some embodiments, the fluorescent protein comprises no more than 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, or 15 mutations relative to SEQ ID NO: 1. In some embodiments, the fluorescent protein has a fluorescence intensity of at least about: 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45 or 50 times that of SEQ ID NO: 1. In some embodiments, the fluorescent protein has a fluorescence intensity of at least about: 2, 3, 4 or 5 times that of superfolder GFP (AIC82357). In some embodiments, disclosed herein is a fusion protein comprising a fluorescent protein. In some embodiments, disclosed herein is a nucleic acid comprising a sequence encoding a fluorescent protein or fusion protein. In some embodiments, disclosed herein is a vector comprising a nucleic acid. In some embodiments, disclosed herein is a host cell comprising a protein, nucleic acid or vector. In some embodiments, disclosed herein is a method of visualization comprising detecting a fluorescent protein. In some embodiments, the detection is by detecting the wavelength of the emission spectrum of the fluorescent protein. In some embodiments, the visualization is in a cell. In some embodiments, the cell is in an isolated biological tissue, in vitro or in vivo.In some embodiments, disclosed herein is a method for expressing a fluorescent protein or fusion protein, comprising introducing an expression vector into a cell, the expression vector comprising a nucleic acid encoding a polypeptide. In some embodiments, the method further comprises culturing the cell to grow a batch of cultured cells and purifying the polypeptide from the batch of cultured cells. In some embodiments, disclosed herein is a method for detecting a fluorescent signal from a polypeptide within a biological cell or tissue, or within a tissue, the method comprising: (a) introducing a fluorescent protein or an expression vector comprising a nucleic acid encoding the fluorescent protein into the biological cell or tissue; (b) directing light of a first wavelength suitable for exciting the fluorescent protein onto the biological cell or tissue; and (c) detecting light of a second wavelength emitted by the fluorescent protein in response to absorption of the first wavelength of light. In some embodiments, the second wavelength of light is detected using a fluorescence microscope or fluorescence-activated cell sorting (FACS). In some embodiments, the biological cell or tissue is a prokaryotic or eukaryotic cell. In some embodiments, the expression vector comprises a fusion gene comprising the nucleic acid encoding the polypeptide fused to another gene at its N-terminus or C-terminus. In some embodiments, the expression vector comprises a promoter that controls expression of the polypeptide, the promoter being either a constitutively active promoter or an inducible expression promoter.

[0031] A method for training a supervised model for use in a method or system as described above is disclosed. The supervised model includes an encoder network configured to map a biopolymer sequence to a representation in an embedded function space. The supervised model is configured to predict a function of the biopolymer sequence based on the representations. The method includes the following steps: (a) providing a plurality of training biopolymer sequences, wherein each training biopolymer sequence is labeled with a function; (b) using the encoder to map each training biopolymer sequence to a representation in the embedded function space; (c) using the supervised model to predict a function for each training biopolymer sequence based on the representations; (d) for each training biopolymer sequence, using a predetermined prediction loss function to determine the degree of consistency of the predicted function with the function according to the label of the corresponding training biopolymer sequence; and (e) optimizing parameters characterizing the behavior of the supervised model, with the goal of improving the ranking of further training biopolymer sequences as a result of the prediction loss function when further training biopolymer sequences are processed by the supervised model.

[0032] A method for training a decoder for use in a method or system as described above is disclosed. The decoder is configured to map a representation of a biopolymer sequence from an embedding function space to a probabilistic biopolymer sequence. The method comprises the following steps: (a) providing a plurality of representations of the biopolymer sequence in the embedding function space; (b) using the decoder to map each representation to a probabilistic biopolymer sequence; (c) extracting a sample biopolymer sequence from each probabilistic biopolymer sequence; (d) using a trained encoder to map the sample biopolymer sequence to a representation in the embedding function space; (e) using a predetermined reconstruction loss function to determine the degree of consistency of each so determined representation with the corresponding original representation; and (f) optimizing parameters characterizing the behavior of the decoder with the aim of improving the rating of further representations of the biopolymer sequence from the embedding function space as they are processed by the decoder via the reconstruction loss function.

[0033] Optionally, the encoder is part of a supervised model configured to predict a function of the biopolymer sequence based on the representations generated by the decoder, and the method further comprises: (a) providing at least a portion of the plurality of representations of the biopolymer sequence to the decoder by mapping training biopolymer sequences to representations in the embedded function space using the trained encoder; (b) for a sample biopolymer sequence drawn from the probabilistic biopolymer sequence, predicting a function of the sample biopolymer sequence using the supervised model; (c) comparing the function with a function predicted by the same supervised model for a corresponding original training biopolymer sequence; (d) determining a degree of consistency between the function predicted for the sample biopolymer sequence and the function predicted for the original training biopolymer sequence using a predetermined consistency loss function; and (e) optimizing parameters characterizing the behavior of the decoder with the aim of improving a ranking by means of the consistency loss function and / or by means of a predetermined combination of the consistency loss function and the reconstruction loss function generated when further representations of the biopolymer sequence generated by the encoder from the training biopolymer sequence are processed by the decoder.

[0034] A method for training an ensemble of supervised models and decoders is disclosed. The supervised model includes an encoder network configured to map a biopolymer sequence to a representation in an embedding function space. The supervised model is configured to predict a function of the biopolymer sequence based on the representation. The decoder is configured to map the representation of the biopolymer sequence from the embedding function space to a probabilistic biopolymer sequence. The method comprises the following steps: (a) providing a plurality of training biopolymer sequences, wherein each training biopolymer sequence is labeled with a function; (b) using the encoder to map each training biopolymer sequence to a representation in the embedded function space; (c) using the supervised model to predict the function of each training biopolymer sequence based on the representations; (d) using the decoder to map each representation in the embedded function space to a probabilistic biopolymer sequence; (e) extracting sample biopolymer sequences from the probabilistic biopolymer sequence; (f) for each training biopolymer sequence, using a predetermined prediction loss function to determine the degree of consistency of the predicted function with the function according to the label of the corresponding training biopolymer sequence; (g) for each sample biopolymer sequence, using a predetermined reconstruction loss function to determine the degree of consistency of the sample biopolymer sequence with the original training biopolymer sequence from which the sample biopolymer sequence was generated; and (h) optimizing parameters characterizing the behavior of the supervised model and parameters characterizing the behavior of the decoder, with the aim of improving the rating by a predetermined combination of the prediction loss function and the reconstruction loss function.

[0035] Furthermore, a set of parameters characterizing the behavior of a supervised model, encoder or decoder obtained according to one of these training methods is another product within the scope of the present invention.

[0036] Incorporated by Reference

[0037] All publications, patents, and patent applications mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference. Specifically, U.S. Application No. 62 / 804,036 is incorporated herein by reference. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] This patent or application file contains at least one drawing drawn in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee. A better understanding of the features and advantages of the present invention will be obtained by reference to the following detailed description, which sets forth illustrative embodiments in which the principles of the present disclosure are utilized, and the accompanying drawings, in which:

[0039] Figure 1 A diagram illustrating a non-limiting embodiment of an encoder that is a neural network is shown.

[0040] Figure 2 A diagram illustrating a non-limiting embodiment of a decoder that is a neural network is shown.

[0041] Figure 3A A non-limiting overview of a gradient-based design procedure is shown.

[0042] Figure 3B A non-limiting example of one iteration of the gradient-based design procedure is shown.

[0043] Figure 3C A non-limiting example of a matrix encoding a probabilistic sequence generated by a decoder is shown.

[0044] Figure 4 A diagram illustrating a non-limiting embodiment of a decoder verification procedure is shown.

[0045] Figure 5A A plot showing the relationship between the predicted fluorescence values ​​and the true fluorescence values ​​of the GFP encoder model for the training dataset.

[0046] Figure 5B A plot showing the relationship between the predicted fluorescence values ​​and the true fluorescence values ​​of the GFP encoder model for the validation dataset.

[0047] Figures 6A to 6B An exemplary embodiment of a computing system as described herein is shown.

[0048] Figure 7 Shown is a diagram illustrating a non-limiting example of gradient-based design (GBD) for engineering GFP sequences.

[0049] Figure 8 Shown are experimental validation results of relative fluorescence values ​​of GFP sequences created using GBD.

[0050] Figure 9 Shown are pairwise amino acid sequence alignments of avGFP and the GBD-engineered GFP sequence with the highest experimentally verified fluorescence.

[0051] Figure 10 Shown is a graph illustrating the evolution of predicted resistance over rounds or iterations of gradient-based design.

[0052] Figure 11 Shown are the results of validation experiments conducted to evaluate the actual antibiotic resistance conferred by seven novel β-lactamases designed using gradient-based design.

[0053] 12A to 12F are graphs illustrating discrete optimization results for RNA optimization (12A-C) and lattice protein optimization (12D-F).

[0054] 13A to 13H is a diagram illustrating the results of gradient-based optimization.

[0055] FIG. 14A to FIG. 14B is a graph illustrating the effect of increasing the weight of the regularization term λ: a larger λ leads to a decrease in model error, but during the optimization process, the model is constrained to assign high probability p θ The sequence diversity is also reduced accordingly.

[0056] FIG. 15A to FIG. 15B The diagram shows the heuristic motivation GBD: it drives the group to Z region that can be reliably decoded.

[0057] Figure 16 The figure shows that GBD can find the optimal value farther away from the initial seed sequence than the discrete method while maintaining a fairly low error.

[0058] Figure 17 is a graph illustrating wet lab data testing the variance generated by the listed proteins, validating the affinity of the proteins generated. DETAILED DESCRIPTION

[0059] Described herein are systems, devices, software, and methods for generating predictions of amino acid sequences corresponding to properties or functions. Machine learning methods allow for the generation of models that receive input data (e.g., primary amino acid sequence) and generate modified amino acid sequences corresponding to one or more functions or features of a resulting polypeptide or protein defined at least in part by an amino acid sequence. The input data may include additional information, such as contact maps of amino acid interactions, tertiary protein structure, or other relevant information relevant to polypeptide structure. In some cases, transfer learning is used to improve the predictive power of the model when labeled training data is insufficient. The input amino acid sequence can be mapped to an embedding space, optimized within the embedding space for a desired function or property (e.g., increasing the reaction rate of an enzyme), and then decoded into a modified amino acid sequence mapped to a desired function or property.

[0060] The present disclosure incorporates the novel discovery that proteins are amenable to rational sequence design based on machine learning, such as gradient-based design using deep neural networks, which allows the use of standard optimization techniques (e.g., gradient ascent) to create amino acid sequences that perform a desired function. In an illustrative example of gradient-based design, an initial amino acid sequence is projected into a new embedding space that represents the function of the protein. An embedding of a protein sequence is a representation of the protein as a point in a D-dimensional space. In this new space, the protein can be encoded as a vector of two numbers (e.g., in the case of a 2-dimensional space), which provides the coordinates of the protein in the embedding space. One property of an embedding space is that proteins that are nearby in this space are functionally similar and related. Therefore, when a collection of proteins is embedded in this space, the functional similarity of any two proteins can be determined by calculating the distance between them using a Euclidean metric.

[0061] Computational protein design

[0062] In some embodiments, the apparatus, software, system and method disclosed herein utilize (multiple) machine learning methods as tools for protein design. In some embodiments, a continuously differentiable embedding space is used to generate a new protein or polypeptide sequence mapped to a desired function or property. In some cases, the process includes providing a seed sequence (e.g., a sequence that does not perform (multiple) desired functions or does not perform the desired function at the desired level), projecting the seed sequence into the embedding space, iteratively optimizing the sequence by making small changes in the embedding space, and then mapping these changes back into the sequence space. In some cases, the seed sequence lacks the desired function or property (e.g., a beta-lactamase that does not have antibiotic resistance). In some cases, the seed sequence has a certain function or property (e.g., a baseline GFP sequence with some fluorescence). The seed sequence can have the highest or "best" available function or property (e.g., a GFP with the highest fluorescence intensity in the literature). The seed sequence can have the function or property closest to the desired function or property. For example, a seed GFP sequence whose fluorescence intensity value is closest to the final desired fluorescence intensity value can be selected. The seed sequence can be based on the average or consensus sequence of a single sequence or multiple sequences. For example, multiple GFP sequences can be averaged to produce a consensus sequence. The averaged sequence can represent the starting point of the "best" sequence (e.g., the sequence to be optimized that has the highest or closest level of the desired function or property). The methods disclosed herein can utilize more than one method or training model. In some embodiments, two neural networks working together are provided: an encoder network and a decoder network. The encoder network can receive an amino acid sequence that can be represented as a sequence of one-hot vectors and generate an embedding for the protein. Similarly, the decoder can take the embedding and return the amino acid sequence that maps to a specific point in the embedding space.

[0063] To change the function of a given protein, an encoder network can first be used to project the initial sequence into an embedding space. Next, the function of the protein can be changed by "moving" the position of the initial sequence within the embedding space to a region of space occupied by a protein with a desired function (or level of function, such as enhanced function). Once the embedded sequence has moved to the desired region of the embedding space (and therefore achieved the desired level of function), a decoder network can be used to receive the new coordinates in the embedding space and generate the actual amino acid sequence that will encode a real protein with the desired function or level of function. In some embodiments where the encoder network and the decoder network are deep neural networks, partial derivatives can be calculated for points in the embedding space, allowing optimization methods (such as, for example, gradient-based optimization procedures) to calculate the steepest direction of improvement in that space.

[0064] A simplified, step-by-step overview of one embodiment of computational protein design described herein includes the following steps:

[0065] (1) A protein is selected as a "seed" protein. This protein is used as the base sequence to be modified.

[0066] (2) Use the encoder network to project the protein into the embedding space.

[0067] (3) The seed protein within the embedding space is iteratively improved using a gradient ascent procedure based on the derivatives of the functions with respect to the embedding provided by the encoder network.

[0068] (4) Once the desired level of functionality is obtained, the final embedding is mapped back to sequence space using a decoder network. This produces an amino acid sequence with the desired level of functionality.

[0069] Construction of embedding space

[0070] In some embodiments, the devices, software, systems and methods disclosed herein utilize encoders to generate an embedding space when given an input such as a primary amino acid sequence. In some embodiments, the encoder is constructed by training a neural network (e.g., a deep neural network) to predict the desired function based on a set of labeled training data. The encoder model can be a supervised model using a convolutional neural network (CNN) in the form of 1D convolution (e.g., a primary amino acid sequence), 2D convolution (e.g., a contact map of amino acid interactions), or 3D convolution (e.g., a tertiary protein structure). The convolutional architecture can be any of the architectures described below: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet, or MobileNet.

[0071] In some embodiments, the encoder utilizes any number of alternative regularization methods to prevent overfitting. Illustrative, non-limiting examples of regularization methods include early stopping, including dropout at at least 1, 2, 3, 4 layers up to all layers, including L1-L2 regularization at at least 1, 2, 3, 4 layers up to all layers, including jump connections at at least 1, 2, 3, 4 layers up to all layers. Here, the term "drop out" can include, in particular, randomly deactivating some neurons or other processing units of the layer during training, so that training is actually performed on a large number of slightly different network architectures. This reduces "overfitting", that is, over-adapting the network to the specific training data at hand, rather than learning general knowledge from the training data. Alternatively or in combination, batch normalization or group normalization can be used to perform regularization.

[0072] In some embodiments, the encoder is optimized using any of the following non-limiting optimization procedures: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta, or NAdam. The model can be optimized using any of the following activation functions: softmax, elu, SeLU, softplus, softsign, ReLU, tanh, sigmoid, hard_sigmoid, exponential, PReLU, and LeaskyReLU or linear.

[0073] In some embodiments, the encoder comprises 3 to 100,000 layers. In some embodiments, the encoder comprises 3 to 5 layers, 3 to 10 layers, 3 to 50 layers, 3 to 100 layers, 3 to 500 layers, 3 to 1,000 layers, 3 to 5,000 layers, 3 to 10,000 layers, 3 to 50,000 layers, 3 to 100,000 layers, 3 to 100,000 layers, 5 to 10 layers, 5 to 50 layers, 5 to 100 layers, 5 to 500 layers, 5 to 1,000 layers, 5 to 5,000 layers, 5 to 10,000 layers, 5 to 50,000 layers, 5 to 100,000 layers, 00 layers, 5 layers to 100,000 layers, 10 layers to 50 layers, 10 layers to 100 layers, 10 layers to 500 layers, 10 layers to 1,000 layers, 10 layers to 5,000 layers, 10 layers to 10,000 layers, 10 layers to 50,000 layers, 10 layers to 100,000 layers, 10 layers to 100,000 layers, 50 layers to 100,000 layers, 50 layers to 100,000 layers, 50 layers to 100,000 layers, 50 layers to 100, 000 layers, 100 layers to 500 layers, 100 layers to 1,000 layers, 100 layers to 5,000 layers, 100 layers to 10,000 layers, 100 layers to 50,000 layers, 100 layers to 100,000 layers, 100 layers to 100,000 layers, 500 layers to 1,000 layers, 500 layers to 5,000 layers, 500 layers to 10,000 layers, 500 layers to 50,000 layers, 500 layers to 100,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 10,000 layers. In some embodiments, the number of layers of the present invention is preferably from 100 layers to 50,000 layers, 1,000 layers to 100,000 layers, 1,000 layers to 100,000 layers, 5,000 layers to 10,000 layers, 5,000 layers to 50,000 layers, 5,000 layers to 100,000 layers, 10,000 layers to 50,000 layers, 10,000 layers to 100,000 layers, 10,000 layers to 100,000 layers, 50,000 layers to 100,000 layers, 50,000 layers to 100,000 layers, or 100,000 layers to 100,000 layers.In some embodiments, the encoder comprises 3 layers, 5 layers, 10 layers, 50 layers, 100 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, or 100,000 layers. In some embodiments, the encoder comprises at least 3 layers, 5 layers, 10 layers, 50 layers, 100 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, or 100,000 layers. In some embodiments, the encoder includes at most 5 layers, 10 layers, 50 layers, 100 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, or 100,000 layers.

[0074] In some embodiments, the encoder is trained to predict the function or property of a protein or polypeptide given its raw amino acid sequence. As a byproduct of learning to predict, the encoder's penultimate layer encodes the original sequence in an embedding space. Thus, to embed a given sequence, the sequence is passed through all layers of the network up to the penultimate layer, and the activation pattern of that layer is used as the embedding. Figure 1 is a diagram illustrating a non-limiting embodiment of an encoder 100 as a neural network. The encoder neural network is trained to predict a specific function 102 given an input sequence 110. The penultimate layer is a two-dimensional embedding 104 that encodes all information about the function of a given sequence. Thus, the encoder can take an input sequence, such as an amino acid sequence or a nucleic acid sequence corresponding to an amino acid sequence, and process the sequence to create an embedding or vectorized representation of the source sequence that captures the function of the amino acid sequence within an embedding space. The selection of the initial source sequence can be based on a rational approach (e.g., the protein(s) with the highest level of function) or by some other means (e.g., random selection).

[0075] However, it is not strictly required that the encoder participate in all the processes from the input sequence all the way to the specific quantitative value of the function. Instead, a layer or other processing unit different from the encoder can accept the embedding provided by the encoder and map it to the desired quantitative value of the function. Figure 3A One such embodiment is illustrated in FIG.

[0076] The encoder and decoder can be trained at least partially in tandem in an encoder-decoder arrangement. Regardless of whether the quantitative value of the function is evaluated internally or externally to the encoder, starting with an input biopolymer sequence, the compressed representation in the embedding space generated by the encoder can be fed to the decoder, and the degree of consistency of the probabilistic biopolymer sequence provided by the decoder with the original input biopolymer sequence can then be determined. For example, one or more samples can be drawn from the probabilistic biopolymer sequence, and the one or more drawn samples can be compared to the original input biopolymer sequence. Parameters characterizing the behavior of the encoder and / or decoder can then be optimized to maximize the consistency between the probabilistic biopolymer sequence and the original input biopolymer sequence.

[0077] As will be discussed later, this consistency can be measured by a predetermined loss function ("reconstruction loss"). Most importantly, the prediction of function can be trained on input biopolymer sequences labeled with known values ​​of the function that should be reproduced by the prediction. The consistency of the prediction with the actual known values ​​of the function can be measured by another loss that can be combined with the reconstruction loss in any suitable manner.

[0078] In some embodiments, the encoder is generated at least in part using transfer learning to improve performance. The starting point can be a complete first model that is frozen except for the output layer (or one or more additional layers) and trained on the target protein function or protein feature. The starting point can be a pre-trained model in which the embedding layer, the last two layers, the last three layers, or all layers are unfrozen, and the rest of the model is frozen during training on the target protein function or protein feature.

[0079] Gradient-based protein design in embedding space

[0080] In some embodiments, the devices, software, systems, and methods disclosed herein obtain an initial embedding of input data (e.g., a primary amino acid sequence) and optimize the embedding toward a specific function or property. In some embodiments, once the embedding has been created, the embedding is optimized toward a given function using mathematical methods such as a 'back propagation' method to calculate the derivative of the embedding with respect to the function to be optimized. Given an initial embedding E1, a learning rate r, and a gradient of the function F The following updates can be performed to create a new embed E2:

[0081]

[0082] The gradient of F is implicitly defined by the encoder network, and since the encoder is differentiable almost everywhere, the derivative of the embedding with respect to the function can be computed. The above update procedure can be repeated until the desired level of function is achieved.

[0083] Figure 3B is a diagram illustrating the iterations of gradient-based design (GBD). First, the source embedding 354 is fed into the GBD network 350 consisting of a decoder 356 and a supervision model 358. The gradient 364 is calculated and used to produce a new embedding, which is then fed back to the GBD network 350 via the decoder 356 to ultimately generate the function F2 382. This process can be repeated until the desired level of function is achieved or until the predicted function has saturated.

[0084] There are many possible variations of this update rule, including different step sizes for r and different optimization schemes such as Adam, RMS Prop, Ada delta, AdamMAX, and SGD with momentum. Additionally, the updates described above are examples of "first-order" methods that use only information about the first-order derivatives, but in some embodiments, higher-order methods that exploit the information contained in the Hessian can be used, such as, for example, second-order methods.

[0085] Using the embedding optimization method described herein, constraints and other required data can be combined as long as they can be incorporated into the update equation. In some embodiments, embedding is optimized for at least two, at least three, at least four, at least five, at least six, at least seven, at least eight, at least nine or at least ten parameters (e.g., desired functions and / or properties). As a non-limiting and illustrative example, the sequence is optimized for function F1 (e.g., fluorescence) and function F2 (e.g., thermal stability). In this case, the encoder has been trained to predict these two functions, so it is allowed to use a composite function F=c1F1+c2F2 to merge these two functions into the optimization process, and weight the functions as needed. Therefore, this composite function can be optimized, for example, using the gradient-based update program described herein. In some embodiments, the devices, software, systems and methods described herein utilize composite functions combined with weights that express the relative preferences for F1 and F2 under this framework (e.g., mostly maximizing fluorescence, but also combining some thermal stability).

[0086] Mapping back to protein space: the decoder network

[0087] In some embodiments, the devices, software, systems and methods disclosed herein obtain a seed embedding that has been optimized to achieve a certain desired functional level, and utilize a decoder to map the optimized coordinates in the embedding space back to the protein space. In some embodiments, a decoder such as a neural network is trained to generate an amino acid sequence based on an input including an embedding. The network essentially provides the "inverse" of the encoder and can be implemented using a deep convolutional neural network. In other words, the encoder receives an input amino acid sequence and generates an embedding of the sequence mapped to the embedding space, and the decoder receives an input (optimized) embedding coordinate and generates the resulting amino acid sequence. The decoder can be trained using labeled data (e.g., β-lactamase labeled with antibiotic resistance information) or unlabeled data (e.g., β-lactamase lacking antibiotic resistance information). In some embodiments, the overall structure of the decoder and the encoder is the same. For example, the number of variants of the decoder (architecture, number of layers, optimizer, etc.) can be the same as that of the encoder.

[0088] In some embodiments, the devices, software, systems and methods disclosed herein utilize a decoder to process an input (such as a primary amino acid sequence or other biopolymer sequence) and generate a predicted sequence (e.g., a probabilistic sequence with a distribution of amino acids at each position). In some embodiments, the decoder is constructed by training a neural network (e.g., a deep neural network) to generate a predicted sequence based on a set of labeled training data. For example, embeddings can be generated from labeled training data and then used to train the decoder. The decoder model can be a supervised model using a convolutional neural network (CNN) in the form of 1D convolution (e.g., a primary amino acid sequence), 2D convolution (e.g., a contact map of amino acid interactions), or 3D convolution (e.g., a tertiary protein structure). The convolutional architecture can be any of the architectures described below: VGG16, VGG19, DeepResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNet ResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet, or MobileNet.

[0089] In some embodiments, the decoder utilizes any number of alternative regularization methods to prevent overfitting. Illustrative, non-limiting examples of regularization methods include early stopping, including dropouts at at least 1, 2, 3, 4 layers up to all layers, including L1-L2 regularization at at least 1, 2, 3, 4 layers up to all layers, including skip connections at at least 1, 2, 3, 4 layers up to all layers. Regularization can be performed using batch normalization or group normalization.

[0090] In some embodiments, the decoder is optimized using any of the following non-limiting optimization procedures: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrop, SGD without momentum, Adagrad, Adadelta, or NAdam. The model can be optimized using any of the following activation functions: softmax, elu, SeLU, softplus, softsign, ReLU, tanh, sigmoid, hard_sigmoid, exponential, PReLU, and LeaskyReLU or linear.

[0091] In some embodiments, the decoder comprises 3 to 100,000 layers. In some embodiments, the decoder comprises 3 to 5 layers, 3 to 10 layers, 3 to 50 layers, 3 to 100 layers, 3 to 500 layers, 3 to 1,000 layers, 3 to 5,000 layers, 3 to 10,000 layers, 3 to 50,000 layers, 3 to 100,000 layers, 3 to 100,000 layers, 5 to 10 layers, 5 to 50 layers, 5 to 100 layers, 5 to 500 layers, 5 to 1,000 layers, 5 to 5,000 layers, 5 to 10,000 layers, 5 to 50,000 layers, 5 to 100,000 layers. 00 layers, 5 layers to 100,000 layers, 10 layers to 50 layers, 10 layers to 100 layers, 10 layers to 500 layers, 10 layers to 1,000 layers, 10 layers to 5,000 layers, 10 layers to 10,000 layers, 10 layers to 50,000 layers, 10 layers to 100,000 layers, 10 layers to 100,000 layers, 50 layers to 100,000 layers, 50 layers to 100,000 layers, 50 layers to 100,000 layers, 50 layers to 100, 000 layers, 100 layers to 500 layers, 100 layers to 1,000 layers, 100 layers to 5,000 layers, 100 layers to 10,000 layers, 100 layers to 50,000 layers, 100 layers to 100,000 layers, 100 layers to 100,000 layers, 500 layers to 1,000 layers, 500 layers to 5,000 layers, 500 layers to 10,000 layers, 500 layers to 50,000 layers, 500 layers to 100,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 10,000 layers. In some embodiments, the number of layers of the present invention is preferably from 100 layers to 50,000 layers, 1,000 layers to 100,000 layers, 1,000 layers to 100,000 layers, 5,000 layers to 10,000 layers, 5,000 layers to 50,000 layers, 5,000 layers to 100,000 layers, 10,000 layers to 50,000 layers, 10,000 layers to 100,000 layers, 10,000 layers to 100,000 layers, 50,000 layers to 100,000 layers, 50,000 layers to 100,000 layers, or 100,000 layers to 100,000 layers.In some embodiments, the decoder comprises 3 layers, 5 layers, 10 layers, 50 layers, 100 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, or 100,000 layers. In some embodiments, the decoder comprises at least 3 layers, 5 layers, 10 layers, 50 layers, 100 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, or 100,000 layers. In some embodiments, the decoder includes at most 5 layers, 10 layers, 50 layers, 100 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, or 100,000 layers.

[0092] In some embodiments, the decoder is trained to predict the original amino acid sequence of a protein or polypeptide given a sequence embedding. In some embodiments, the decoder is generated at least in part using transfer learning to improve performance. The starting point can be a complete first model that is frozen except for the output layer (or one or more additional layers) and trained on the target protein function or protein feature. The starting point can be a pre-trained model in which the embedding layer, the last two layers, the last three layers, or all layers are unfrozen, and the rest of the model is frozen during training on the target protein function or protein feature.

[0093] In some embodiments, the decoder is trained using a procedure similar to that used to train the encoder. For example, a set of training sequences is obtained and embeddings are created for these sequences using the trained encoder. These embeddings represent the input to the decoder, while the output is the original sequence that the decoder must predict. In some embodiments, a convolutional neural network is used for the decoder, which mirrors the architecture of the encoder in reverse. Other types of neural networks can be used, for example, recurrent neural networks (RNNs) such as long short-term memory (LSTM) networks.

[0094] The decoder can be trained to minimize a loss, namely the resident categorical cross entropy, to reconstruct a sequence mapped to a given embedding (also known as the reconstruction loss). In some embodiments, additional terms are added to the loss, which have been found to provide substantial improvements to the process. The following notation is used herein:

[0095] ax: amino acid sequence

[0096] by: the measurable property of interest in x, e.g., fluorescence

[0097] cf(x): A function that accepts x to predict y, e.g., a deep neural network

[0098] d.enc(x): a submodule of f(x) that produces an embedding (e) of the sequence (x)

[0099] e.dec(e): A separate decoder module that takes an embedding (e) and produces a reconstructed sequence (x')

[0100] f.x': the output of the decoder dec(e), i.e., the reconstructed sequence generated from the embedding(e)

[0101] In addition to the reconstruction loss, the reconstructed sequence (x') is fed back through the original supervised model f(x') to produce a prediction value (called y') using the decoder's reconstructed sequence. The predicted value of the reconstructed sequence (y') is compared with the predicted value of the given sequence (called y* and which is calculated using f(x)). Similar x and x' values ​​and / or similar y' and y* values ​​indicate that the decoder is working effectively. To enforce this, in some embodiments, an additional term is added to the network's loss function using the Kullback-Leibler divergence (KLD). The KLD between any y' and y* is expressed as:

[0102] a.KLD(y^′,y^*)=y^′*log(y^* / y′)

[0103] The loss including this term is expressed as:

[0104] a.loss = λ_1*CCE+λ_2*KLD(y^′, y^*), where CCE is the categorical cross entropy reconstruction loss, and λ_1 and λ_2 are tuning parameters.

[0105] Figure 2 is a diagram illustrating an example of a decoder as a neural network. The decoder network 200 has four layers of nodes, where the first layer 202 corresponds to an embedding layer that can receive input from the encoder described herein. In this illustrative example, the next two layers 204 and 206 are hidden layers, and the last layer 208 is the final layer that outputs the amino acid sequence "decoded" from the embedding.

[0106] Figure 3A 3 is a diagram illustrating an embodiment of an overview of a gradient-based design procedure. An encoder 310 can be used to generate a source embedding 304. The source embedding is fed to a decoder 306, which is then converted into a probabilistic sequence (e.g., an amino acid distribution at each residue). The probabilistic sequence can then be processed by a supervised model 308 including the encoder 310 to produce a predicted function value 312. The gradient 314 of the function (F) model is obtained with respect to the input embedding 304 and is calculated by using backpropagation through the supervised model and the decoder.

[0107] Figure 3CAn example of a probabilistic biopolymer sequence 390 generated by the decoder is shown. In this example, the probabilistic biopolymer sequence 390 can be illustrated by a matrix 392. The columns of the matrix 392 represent each of the 20 possible amino acids, and the rows represent the residue positions in a protein of length L. The first amino acid (row 1) is always methionine, so the probability of M (column 7) is 1, and the probabilities of the remaining amino acids are 0. For example, the next residue (row 2) may have an 80% probability of being W and a 20% probability of being G. To generate a sequence, the maximum likelihood sequence suggested by the matrix can be selected, which requires selecting the amino acid with the highest probability at each position. Alternatively, the sequence can be randomly generated by sampling each position according to the amino acid probability, for example, by randomly picking W or G at position 2 with a probability of 80% versus 20%, respectively.

[0108] Decoder Verification

[0109] In some embodiments, the apparatus, software, system and method disclosed herein provide a decoder validation framework to determine the performance of the decoder. An effective decoder can predict which sequence maps to a given embedding with very high accuracy. Therefore, the decoder can be validated by processing the same input (e.g., an amino acid sequence) using the encoder and encoder-decoder framework described herein. The encoder will generate an output indicating the desired function and / or property, which is used as a reference for evaluating the output of the encoder-decoder framework. As an illustrative example, an encoder and decoder are generated according to the method described herein. Next, the encoder is used to embed each protein in the training set and the validation set. Then, these embeddings are decoded using a decoder. Finally, the encoder is used to predict the function value of the decoded sequence, and these predicted values ​​are compared with the values ​​predicted using the original sequence.

[0110] Figure 4 An overview of one embodiment of a decoder verification process 400 is shown in FIG. Figure 4 As shown, an encoder neural network 402 is shown at the top, which receives a primary amino acid sequence (e.g., green fluorescent protein) as input and processes the sequence to output a prediction 406 of a function (e.g., fluorescence intensity). The encoder-decoder framework 408 below shows an encoder network 412 with a penultimate embedding layer that is identical to the encoder neural network 402 except that the calculation of the prediction 406 is missing. The encoder network 412 is connected or linked (or otherwise provided with input) to the decoder network 410 to decode the sequence, which is then fed back to the encoder network 402 to obtain a predicted function 416. Therefore, when the values ​​of the two predictions 406 and 416 are close, the result provides verification that the decoder 410 is effectively mapping the embedding to the sequence corresponding to the desired function.

[0111] The similarity or correspondence between the predicted values ​​can be calculated in a variety of ways. In some embodiments, the correlation between the predicted value from the original sequence and the predicted value from the decoded sequence is determined. In some embodiments, the correlation is about 0.7 to about 0.99. In some embodiments, the correlation is from about 0.7 to about 0.75, from about 0.7 to about 0.8, from about 0.7 to about 0.85, from about 0.7 to about 0.9, from about 0.7 to about 0.95, from about 0.7 to about 0.99, from about 0.75 to about 0.8, from about 0.75 to about 0.85, from about 0.75 to about 0.9, from about 0.75 to about 0.95, from about 0.75 to about 0.99, from about 0.8 to about 0.85, from about 0.8 to about 0.9, from about 0.8 to about 0.95, from about 0.8 to about 0.99, from about 0.85 to about 0.9, from about 0.85 to about 0.95, from about 0.85 to about 0.99, from about 0.9 to about 0.95, from about 0.9 to about 0.99, or from about 0.95 to about 0.99. In some embodiments, the correlation is about 0.7, about 0.75, about 0.8, about 0.85, about 0.9, about 0.95, or about 0.99. In some embodiments, the correlation is at least about 0.7, about 0.75, about 0.8, about 0.85, about 0.9, or about 0.95. In some embodiments, the correlation is at most about 0.75, about 0.8, about 0.85, about 0.9, about 0.95, or about 0.99.

[0112] Additional performance metrics can be used to validate the systems and methods disclosed herein, such as positive predictive value (PPV), F1, mean squared error, area under the receiver operating characteristic (ROC), and area under the precision-recall curve (PRC).

[0113] In some embodiments, the methods disclosed herein generate results with a positive predictive value (PPV). In some embodiments, the PPV is 0.7 to 0.99. In some embodiments, the PPV is 0.7 to 0.75, 0.7 to 0.8, 0.7 to 0.85, 0.7 to 0.9, 0.7 to 0.95, 0.7 to 0.99, 0.75 to 0.8, 0.75 to 0.85, 0.75 to 0.9, 0.75 to 0.95, 0.75 to 0.99, 0.8 to 0.85, 0.8 to 0.9, 0.8 to 0.95, 0.8 to 0.99, 0.85 to 0.9, 0.85 to 0.95, 0.85 to 0.99, 0.9 to 0.95, 0.9 to 0.99, or 0.95 to 0.99. In some embodiments, the PPV is 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, or 0.99. In some embodiments, the PPV is at least 0.7, 0.75, 0.8, 0.85, 0.9, or 0.95. In some embodiments, the PPV is at most 0.75, 0.8, 0.85, 0.9, 0.95, or 0.99.

[0114] In some embodiments, the methods disclosed herein generate results having an F1 value. In some embodiments, F1 is 0.5 to 0.95. In some embodiments, F1 is 0.5 to 0.6, 0.5 to 0.7, 0.5 to 0.75, 0.5 to 0.8, 0.5 to 0.85, 0.5 to 0.9, 0.5 to 0.95, 0.6 to 0.7, 0.6 to 0.75, 0.6 to 0.8, 0.6 to 0.85, 0.6 to 0.9, 0.6 to 0.95, 0.7 to 0.7 .95, 0.75 to 0.8, 0.75 to 0.85, 0.75 to 0.9, 0.75 to 0.95, 0.8 to 0.85, 0.8 to 0.9, 0.8 to 0.95, 0.85 to 0.9, or 0.95. In some embodiments, F1 is 0.5, 0.6, 0.7, 0.75, 0.8, 0.85, 0.9, or 0.95. In some embodiments, F1 is at least 0.5, 0.6, 0.7, 0.75, 0.8, 0.85, or 0.9. In some embodiments, F1 is at most 0.6, 0.7, 0.75, 0.8, 0.85, 0.9, or 0.95.

[0115] In some embodiments, the methods disclosed herein produce results with a mean square error. In some embodiments, the mean square error is 0.01 to 0.3. In some embodiments, the mean square error is 0.01 to 0.05, 0.01 to 0.1, 0.01 to 0.15, 0.01 to 0.2, 0.01 to 0.25, 0.01 to 0.3, 0.05 to 0.1, 0.05 to 0.15, 0.05 to 0.2, 0.05 to 0.25, 0.05 to 0.3, 0.1 to 0.15, 0.1 to 0.2, 0.1 to 0.25, 0.1 to 0.3, 0.15 to 0.2, 0.15 to 0.25, 0.15 to 0.3, 0.2 to 0.25, 0.2 to 0.3, or 0.25 to 0.3. In some embodiments, the mean square error is 0.01, 0.05, 0.1, 0.15, 0.2, 0.25, or 0.3. In some embodiments, the mean square error is at least 0.01, 0.05, 0.1, 0.15, 0.2, or 0.25. In some embodiments, the mean square error is at most 0.05, 0.1, 0.15, 0.2, 0.25, or 0.3.

[0116] In some embodiments, the method disclosed herein generates a result with an area under ROC. In some embodiments, the area under ROC is 0.7 to 0.95. In some embodiments, the area under ROC is 0.95 to 0.9, 0.95 to 0.85, 0.95 to 0.8, 0.95 to 0.75, 0.95 to 0.7, 0.9 to 0.85, 0.9 to 0.8, 0.9 to 0.75, 0.9 to 0.7, 0.85 to 0.8, 0.85 to 0.75, 0.85 to 0.7, 0.8 to 0.75, 0.8 to 0.7, or 0.75 to 0.7. In some embodiments, the area under ROC is 0.95, 0.9, 0.85, 0.8, 0.75 or 0.7. In some embodiments, the area under ROC is at least 0.95, 0.9, 0.85, 0.8 or 0.75. In some embodiments, the area under the ROC is at most 0.9, 0.85, 0.8, 0.75, or 0.7.

[0117] In certain embodiments, the method disclosed herein generates a result with an area under PRC. In certain embodiments, the area under PRC is 0.7 to 0.95. In certain embodiments, the area under PRC is 0.95 to 0.9, 0.95 to 0.85, 0.95 to 0.8, 0.95 to 0.75, 0.95 to 0.7, 0.9 to 0.85, 0.9 to 0.8, 0.9 to 0.75, 0.9 to 0.7, 0.85 to 0.8, 0.85 to 0.75, 0.85 to 0.7, 0.8 to 0.75, 0.8 to 0.7, or 0.75 to 0.7. In certain embodiments, the area under PRC is 0.95, 0.9, 0.85, 0.8, 0.75 or 0.7. In certain embodiments, the area under PRC is at least 0.95, 0.9, 0.85, 0.8 or 0.75. In some embodiments, the area under the PRC is at most 0.9, 0.85, 0.8, 0.75, or 0.7.

[0118] Peptide sequence prediction

[0119] This article describes devices, software, systems and methods for evaluating input data such as initial amino acid sequences (or nucleic acid sequences encoding amino acid sequences) to predict one or more new amino acid sequences corresponding to polypeptides or proteins configured to have specific functions or properties. Extrapolation of specific amino acid sequences (such as proteins) that can perform certain functions or have certain properties has always been a goal of molecular biology. Therefore, the devices, software, systems and methods described herein utilize artificial intelligence or machine learning techniques to predict sequence information for the ability to analyze polypeptides or proteins. Compared to standard non-machine learning methods, machine learning techniques can generate models with increased predictive power. In some cases, when there is not enough data to train the model to obtain the desired output, transfer learning can be utilized to improve prediction accuracy. Alternatively, in some cases, when there is enough data to train the model to achieve statistical parameters comparable to the model incorporated with transfer learning, transfer learning is not used.

[0120] In some embodiments, the input data comprises the primary amino acid sequence of a protein or polypeptide. In some cases, a labeled training data set comprising the primary amino acid sequence is used to train the model. For example, the data set may include the amino acid sequence of a fluorescent protein labeled based on the degree of fluorescence intensity. Therefore, a machine learning method can be used to train the model with this data set to generate a prediction of the fluorescence intensity of the amino acid sequence input. In other words, the model can be an encoder, such as a deep neural network trained to predict function based on the primary amino acid sequence input. In some embodiments, the input data also includes information other than the primary amino acid sequence, such as surface charge, hydrophobic surface area, measured or predicted solubility or other relevant information. In some embodiments, the input data comprises multidimensional input data, which includes data of multiple types or categories.

[0121] In some embodiments, the devices, software, systems and methods described herein utilize data augmentation to enhance the performance of (multiple) predictive models. Data augmentation requires training using examples or variations of similar but different training data sets. For example, in image classification, image data can be enhanced by slightly changing the orientation of the image (e.g., slightly rotating it). In some embodiments, data input (e.g., a primary amino acid sequence) is enhanced by random mutations and / or biologically known mutations, multiple sequence alignments, contact maps of amino acid interactions, and / or tertiary protein structures to the primary amino acid sequence. Additional enhancement strategies include using known isoforms and predicted isoforms from alternative splicing transcripts. For example, input data can be enhanced by including isoforms corresponding to alternative splicing transcripts with the same function or property. Therefore, data on isoforms or mutations can allow identification of those parts or features of the primary sequence that do not significantly affect the predicted function or property. This allows the model to interpret information, such as amino acid mutations that enhance, reduce, or do not affect the predicted protein properties (e.g., stability). For example, data input can include a sequence with randomly substituted amino acids at a position known to not affect function. This allows a model trained with this data to learn that the predicted functions are invariant with respect to those specific mutations.

[0122] The devices, software, systems and methods described herein can be used to generate sequence predictions based on one or more of a variety of different functions and / or properties. The predictions can relate to protein functions and / or properties (e.g., enzymatic activity, stability, etc.). Amino acid sequences can be predicted or mapped based on protein stability, which can include various indicators such as thermal stability, oxidative stability, or serum stability. In some embodiments, the encoder is configured to incorporate information related to one or more structural features such as secondary structure, tertiary protein structure, quaternary structure, or any combination thereof. The secondary structure can include whether the amino acids or amino acid sequences in a given polypeptide are predicted to have an α-helical structure, a β-pleated structure, or a disordered or loop structure. The tertiary structure can include the position or location of the amino acids or polypeptide portions in three-dimensional space. The quaternary structure can include the position or location of multiple polypeptides that form a single protein. In some embodiments, the prediction includes sequences based on one or more functions. Polypeptide or protein functions can belong to various categories, including metabolic reactions, DNA replication, providing structure, transport, antigen recognition, intracellular or extracellular signal transduction, and other functional categories. In some embodiments, the prediction includes enzymatic functions, such as catalytic efficiency (e.g., specificity constant k cat / K M ) or catalytic specificity.

[0123] In certain embodiments, sequence prediction is based on the enzyme function of protein or polypeptide.In certain embodiments, protein function is enzyme function.Enzyme can carry out various enzymatic reactions, and can be classified as migratory enzyme (for example, functional group is migrated from one molecule to another molecule), oxidoreductase (for example, catalyzes redox reaction), hydrolase (for example, cuts chemical bond via hydrolysis), lyase (for example, produces double bond), ligase (for example, connects two molecules via covalent bond) and isomerase (for example, catalyzes the structural change from one isomer to another isomer within molecule).In certain embodiments, hydrolase includes protease, such as serine protease, threonine protease, cysteine ​​protease, metalloproteinase, asparagine peptide lyase, glutamic acid protease and aspartic acid protease.Serine protease has multiple physiological effects in aspects such as coagulation, wound healing, digestion, immune response and tumor invasion and metastasis.The example of serine protease includes chymotrypsin, trypsin, elastase, factor 10, factor 11, thrombin, plasmin, C1r, C1s and C3 convertase. Threonine proteases include a family of proteases that have a threonine amino acid within their active catalytic site. Examples of threonine proteases include the subunits of the proteasome. The proteasome is a barrel-shaped protein complex composed of α and β subunits. The catalytically active β subunit may include a conserved N-terminal threonine amino acid at each catalytically active site. Cysteine ​​proteases have a catalytic mechanism that utilizes the cysteine ​​sulfhydryl group. Examples of cysteine ​​proteases include papain, cathepsins, caspases, and calpains. Aspartic proteases have two aspartic acid residues in the active site that participate in acid / base catalysis. Examples of aspartic proteases include the digestive enzyme pepsin, some lysosomal proteases, and renin. Metalloproteases include the digestive enzyme carboxypeptidases, matrix metalloproteinases (MMPs) that play a role in extracellular matrix remodeling and cell signaling, ADAMs (a disintegrin and metalloproteinase domain), and lysosomal proteases. Other non-limiting examples of enzymes include proteases, nucleases, DNA ligases, polymerases, cellulases, ligninases, amylases, lipases, pectinases, xylanases, lignin peroxidases, decarboxylases, mannanases, dehydrogenases, and other polypeptide-based enzymes.

[0124] In some embodiments, the enzymatic reaction comprises post-translational modification of the target molecule. Examples of post-translational modifications include acetylation, amidation, formylation, glycosylation, hydroxylation, methylation, myristoylation, phosphorylation, deamidation, prenylation (e.g., farnesylation, geranylation, etc.), ubiquitination, ribosylation, and sulfation. Phosphorylation can occur on amino acids (e.g., tyrosine, serine, threonine, or histidine).

[0125] In some embodiments, the protein function is luminescence, which is light emission that does not require application of heating. In some embodiments, the protein function is chemiluminescence, such as bioluminescence. For example, a chemiluminescent enzyme (e.g., luciferin) can act on a substrate (luciferin) to catalyze substrate oxidation, thereby releasing light. In some embodiments, the protein function is fluorescence, wherein a fluorescent protein or peptide absorbs light of certain (multiple) wavelengths and emits light of different (multiple) wavelengths. Examples of fluorescent proteins include green fluorescent protein (GFP) or derivatives of GFP, such as EBFP, EBFP2, Azurite, mKalama1, ECFP, Cerulean, CyPet, YFP, Citrine, Venus, or YPet. Some proteins, such as GFP, are naturally fluorescent. Examples of fluorescent proteins include EGFP, blue fluorescent proteins (EBFP, EBFP2, Azurite, mKalama1), cyan fluorescent proteins (ECFP, Cerulean, CyPet), yellow fluorescent proteins (YFP, Citrine, Venus, YPet), redox-sensitive GFP (roGFP), and monomeric GFP.

[0126] In some embodiments, protein functions include enzymatic function, binding (e.g., DNA / RNA binding, protein binding, etc.), immune function (e.g., antibodies), contraction (e.g., actin, myosin), and other functions. In some embodiments, the output includes a primary sequence associated with the protein function, such as, for example, enzymatic function or binding kinetics. As an example, such an output can be obtained by optimizing a composite function that incorporates a desired metric (e.g., any one of affinity, specificity, or reaction rate).

[0127] In some embodiments, the systems and methods disclosed herein generate biopolymer sequences corresponding to functions or properties. In some cases, the biopolymer sequences are nucleic acids. In some cases, the biopolymer sequences are polypeptides. Examples of specific biopolymer sequences include fluorescent proteins (such as GFP) and enzymes (such as β-lactamase). In one embodiment, a reference GFP sequence (e.g., avGFP) is defined by a 238 amino acid long polypeptide having the following sequence:

[0128]

[0129] GFP sequences designed using gradient-based design can include sequences having less than 100% sequence identity to a reference GFP sequence. In some cases, the GBD-optimized GFP sequence has 80% to 99% sequence identity relative to SEQ ID NO: 1. In some cases, the GBD-optimized GFP sequence has 80% to 99% sequence identity relative to SEQ ID NO: 1. NO: 1 has the following sequence identity: 80% to 85%, 80% to 90%, 80% to 95%, 80% to 96%, 80% to 97%, 80% to 98%, 80% to 99%, 85% to 90%, 85% to 95%, 85% to 96%, 85% to 97%, 85% to 98%, 85% to 99%, 90% to 95%, 90% to 96%, 90% to 97%, 90% to 98%, 90% to 99%, 95% to 96%, 95% to 97%, 95% to 98%, 95% to 99%, 96% to 97%, 96% to 98%, 96% to 99%, 97% to 98%, 97% to 99%, or 98% to 99%. In some cases, the GBD-optimized GFP sequence has a sequence identity of 80%, 85%, 90%, 95%, 96%, 97%, 98%, or 99% relative to SEQ ID NO: 1. In some cases, the GBD-optimized GFP sequence has a sequence identity of at least 80%, 85%, 90%, 95%, 96%, 97%, or 98% relative to SEQ ID NO: 1. In some cases, the GBD-optimized GFP sequence has a sequence identity of at most 85%, 90%, 95%, 96%, 97%, 98%, or 99% relative to SEQ ID NO: 1. In some cases, the GBD-optimized GFP sequence has fewer than 45 (e.g., fewer than 40, 35, 30, 25, 20, 15, or 10) amino acid substitutions relative to SEQ ID NO: 1. In some cases, the GBD-optimized GFP sequence comprises at least one, two, three, four, five, six, or seven point mutations relative to the reference GFP sequence. The GBD optimized GFP sequence can be defined by selecting one or more mutations from Y39C, F64L, V68M, D129G, V163A, K166R, and G191V, including combinations thereof, such as 1, 2, 3, 4, 5, 6, or all 7 mutations. In some cases, the GBD optimized GFP sequence does not include the S65T mutation. In certain embodiments, the GBD optimized GFP sequence provided by the invention includes an N-terminal methionine, while in other embodiments, the sequence does not include an N-terminal methionine.

[0130] In some embodiments, disclosed herein are nucleic acid sequences encoding GBD-optimized polypeptide sequences (e.g., GFP and / or β-lactamase). Also disclosed herein are vectors comprising the nucleic acid sequences, such as prokaryotic and / or eukaryotic expression vectors. The expression vectors can be constitutively active or have inducible expression (e.g., a tetracycline-inducible promoter). For example, the CMV promoter is constitutively active, but can also be regulated using a Tet operator element that allows inducible expression in the presence of tetracycline / doxycycline.

[0131] Polypeptide and the nucleic acid sequence encoding these polypeptides can be used in various imaging techniques. For example, fluorescence microscopy, cell activated cell sorting (FACS), flow cytometry and other technologies based on fluorescence imaging can utilize the fluorescent protein of the present disclosure. The GFP protein optimized by GBD can provide a brightness higher than the standard reference GFP protein. In some cases, the fluorescence brightness of the GFP protein optimized by GBD is 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45 or 50 times or more compared with the brightness of the GFP sequence (such as avGFP) that is not optimized.

[0132] In some embodiments, the machine learning method(s) described herein comprise supervised machine learning. Supervised machine learning includes classification and regression. In some embodiments, the machine learning method(s) comprise unsupervised machine learning. Unsupervised machine learning includes clustering, autoencoders, variational autoencoders, protein language models (e.g., where the model predicts the next amino acid in a sequence when the previous amino acid is accessible), and association rule mining.

[0133] Machine Learning

[0134] This article describes devices, software, systems and methods that apply one or more methods to analyze input data to generate a sequence that is mapped to one or more protein or polypeptide properties or functions. In certain embodiments, these methods utilize statistical modeling to generate a prediction or estimate about (multiple) protein or polypeptide functions or properties. In certain embodiments, methods are used to embed a primary sequence (such as an amino acid sequence) into an embedding space, optimize the embedding sequence for the desired function or property, and process the optimized embedding to generate a sequence that is predicted to have the function or property. In certain embodiments, an encoder-decoder framework is utilized, in which two models are combined to allow the use of a first model to embed an initial sequence, and then a second model is used to map the optimized embedding to the sequence.

[0135] In some embodiments, the method utilizes a predictive model such as a neural network, a decision tree, a support vector machine, or other applicable model. Using training data, the method can form a classifier for generating a classification or prediction based on the relevant features. A variety of methods can be used to classify the features selected for classification. In some embodiments, the training method comprises a machine learning method.

[0136] In some embodiments, the machine learning method uses a support vector machine (SVM), naive Bayesian classification, random forest, or artificial neural network. Machine learning techniques include binning procedures, boosting procedures, random forest methods, and combinations thereof. In some embodiments, the prediction model is a deep neural network. In some embodiments, the prediction model is a deep convolutional neural network.

[0137] In some embodiments, the machine learning method uses a supervised learning method. In supervised learning, the method generates a function from labeled training data. Each training example is a pair consisting of an input object and a desired output value. In some embodiments, the best solution allows the method to correctly determine the class label for an unseen situation. In some embodiments, the supervised learning method requires the user to determine one or more control parameters. These parameters are optionally adjusted by optimizing the performance of a subset of the training set (called a validation set). After parameter adjustment and learning, the performance of the resulting function is optionally measured with a test set separate from the training set. Regression methods are commonly used in supervised learning. Therefore, supervised learning allows the use of training data in which the expected output is known in advance to generate or train a model or classifier, such as in calculating protein function when the primary amino acid sequence is known.

[0138] In some embodiments, the machine learning method uses an unsupervised learning approach. In unsupervised learning, the method generates a function to describe the hidden structure from unlabeled data (e.g., classification or categorization is not included in the observations). Because the examples provided to the learner are unlabeled, the accuracy of the structure output by the method is not evaluated. Unsupervised learning methods include clustering, anomaly detection, and neural network-based methods, including autoencoders and variational autoencoders.

[0139] In some embodiments, machine learning methods utilize multi-class learning. Multi-task learning (MTL) is a field of machine learning in which more than one learning task is solved simultaneously in a manner that utilizes the commonalities and differences across multiple tasks. Compared to training those models separately, the advantages of this method may include improving the learning efficiency and prediction accuracy of a specific prediction model. Regularization can be provided to prevent overfitting by requiring a method to perform well on related tasks. This method may be better than regularization that applies the same penalty to all complexities. When applied to tasks or predictions with significant commonalities and / or insufficient samples, multi-class learning may be particularly useful. In some embodiments, multi-class learning is effective for tasks that do not have significant commonalities (e.g., irrelevant tasks or classifications). In some embodiments, multi-class learning is used in combination with transfer learning.

[0140] In some embodiments, the machine learning method learns in batches based on the training data set and other inputs to the batch. In other embodiments, the machine learning method performs additional learning while updating weights and error calculations (e.g., using new or updated training data). In some embodiments, the machine learning method updates the predictive model based on the new or updated data. For example, the machine learning method can be applied to new or updated data to be retrained or optimized to generate a new predictive model. In some embodiments, the machine learning method or model is periodically retrained as additional data becomes available.

[0141] In some embodiments, the classifier or training method of the present disclosure comprises a feature space. In some cases, the classifier comprises two or more feature spaces. In some embodiments, the two or more feature spaces are different from each other. In some embodiments, the accuracy of classification or prediction is improved by combining two or more feature spaces in the classifier rather than using a single feature space. Attributes typically constitute the input features of the feature space and are labeled to indicate the classification of each case for a given input feature set corresponding to that case.

[0142] In some embodiments, a model is trained using one or more sets of training data using a machine learning method. In some embodiments, the methods described herein comprise training a model using a training dataset. In some embodiments, a model is trained using a training dataset comprising a plurality of amino acid sequences. In some embodiments, the training dataset comprises at least 1 million, 2 million, 3 million, 4 million, 5 million, 6 million, 7 million, 8 million, 9 million, 10 million, 15 million, 20 million, 25 million, 30 million, 35 million, 40 million, 45 million, 50 million, 55 million, 56 million, 57 million, 58 million protein amino acid sequences. In some embodiments, the training dataset comprises at least 10,000, 20,000, 30,000, 40,000, 50,000, 60,000, 70,000, 80,000, 90,000, 100,000, 150,000, 200,000, 250,000, 300,000, 350,000, 400,000, 450,000, 500,000, 600,000, 700,000, 800,000, 900,000, or 1 million or more amino acid sequences. In some embodiments, the training data set comprises at least 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, or 10,000 or more annotations. Although exemplary embodiments of the present disclosure include machine learning methods using deep neural networks, various types of methods are contemplated. In some embodiments, the method utilizes a predictive model such as a neural network, a decision tree, a support vector machine, or other applicable model. In some embodiments, the machine learning method is selected from the group consisting of: supervised learning, semi-supervised learning, and unsupervised learning, such as support vector machines (SVM), naive Bayes classification, random forests, artificial neural networks, decision trees, K-means, learning vector quantization (LVQ), self-organizing maps (SOM), graphical models, regression methods (e.g., linear, logistic, multivariate, association rule learning, deep learning, dimensionality reduction, and set selection methods. In some embodiments, the machine learning method is selected from the group consisting of: support vector machines (SVM), naive Bayes classification, random forests, and artificial neural networks. Machine learning techniques include binning procedures, boosting procedures, random forests, and combinations thereof. Illustrative methods for analyzing data include, but are not limited to, methods that directly process a large number of variables, such as statistical methods and methods based on machine learning techniques. Statistical methods include penalized logistic regression, microarray prediction analysis (PAM), shrunken centroid-based methods, support vector machine analysis, and regularized linear discriminant analysis.

[0143] The various models described herein (both supervised and unsupervised) can have alternative regularization methods, including early stopping, including dropout at 1, 2, 3, 4 layers up to all layers, including L1-L2 regularization at 1, 2, 3, 4 layers up to all layers, including skip connections at 1, 2, 3, 4 layers up to all layers. For the first and second models, batch normalization or group normalization can be used for regularization. L1 regularization (also known as LASSO) controls the length allowed for the L1 norm of the weight vector, while L2 controls the size of the L2 norm that can be. Skip connections can be obtained from the Resnet architecture.

[0144] Various models trained using machine learning described herein can be optimized using any of the following optimizers: Adam, RMS prop, stochastic gradient descent (SGD) with momentum, SGD with momentum and Nestrov accelerated gradient, SGD without momentum, Adagrad, Adadelta, or NAdam. The model can be optimized using any of the following activation functions: softmax, elu, SeLU, softplus, softsign, ReLU, tanh, sigmoid, hard_sigmoid, exponential, PReLU, and LeaskyReLU or linear. A loss function can be used to measure the performance of a model. Loss can be understood as the cost of prediction inaccuracy. For example, the cross entropy loss function measures the performance of a classification model, the output of which is a probability value between 0 and 1 (e.g., 0 represents no antibiotic resistance and 1 represents complete antibiotic resistance). The loss value increases as the predicted probability deviates from the actual value.

[0145] In some embodiments, the methods described herein include "reweighting" the loss function that the optimizer listed above attempts to minimize, placing approximately equal weight on both positive and negative examples. For example, one of the 180,000 outputs predicts the probability that a given protein is a membrane protein. Since a protein can only be a membrane protein or not, this is a binary classification task, and the traditional loss function for binary classification tasks is "binary cross entropy": loss(p,y) = -y*log(p) - (1-y)*log(1-p), where p is the probability of being a membrane protein according to the network, and y is the "label" which is 1 if the protein is a membrane protein and 0 if the protein is not a membrane protein. If there are many more examples where y=0, problems can arise because the network may learn a pathological rule that always predicts a very low probability for that annotation because it is rarely penalized for always predicting y=0. To address this problem, in some embodiments, the loss function is modified to the following: loss(p,y) = -w1*y*log(p) -w0*(1-y)*log(1-p), where w1 is the positive class weight and w0 is the negative class weight. The method assumes that w0 = 1 and w1 = 1 / √((1-f0) / f1), where f0 is the frequency of negative examples and f1 is the frequency of positive examples. This weighting scheme "overweights" rare positive examples and "underweights" more common negative examples. Therefore, the method disclosed herein may include incorporating a weighting scheme that provides overweighting and / or underweighting into the loss function to address the uneven distribution of negative and positive examples.

[0146] In some embodiments, the training model (eg, a neural network) comprises from 10 layers to 1,000,000 layers. In some embodiments, the neural network comprises 10 to 50 layers, 10 to 100 layers, 10 to 200 layers, 10 to 500 layers, 10 to 1,000 layers, 10 to 5,000 layers, 10 to 10,000 layers, 10 to 50,000 layers, 10 to 100,000 layers, 10 to 500,000 layers, 10 to 1,000,000 layers, 50 to 100 layers, 50 to 200 layers, 50 to 500 layers, 50 to 1,000 layers, 50 to 5,000 layers, 50 to 10,000 layers, 50 to 50, 000 layers, 50 layers to 100,000 layers, 50 layers to 500,000 layers, 50 layers to 1,000,000 layers, 100 layers to 200 layers, 100 layers to 500 layers, 100 layers to 1,000 layers, 100 layers to 5,000 layers, 100 layers to 10,000 layers, 100 layers to 50,000 layers, 100 layers to 100,000 layers, 100 layers to 500,000 layers, 100 layers to 1,000,000 layers, 200 layers to 500 layers, 200 layers to 1,000 layers, 200 layers to 5,000 layers, 200 layers to 10,000 layers , 200 layers to 50,000 layers, 200 layers to 100,000 layers, 200 layers to 500,000 layers, 200 layers to 1,000,000 layers, 500 layers to 1,000 layers, 500 layers to 5,000 layers, 500 layers to 10,000 layers, 500 layers to 50,000 layers, 500 layers to 100,000 layers, 500 layers to 500,000 layers, 500 layers to 1,000,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers layers, 1,000 layers to 500,000 layers, 1,000 layers to 1,000,000 layers, 5,000 layers to 10,000 layers, 5,000 layers to 50,000 layers, 5,000 layers to 100,000 layers, 5,000 layers to 500,000 layers, 5,000 layers to 1,000,000 layers, 10,000 layers to 50,000 layers, 10,000 layers to 100,000 layers, 10,000 layers to 500,000 layers, 10,000 layers to 1,000,000 layers, 50,000 layers to 100,000 layers, 50,000 layers to 500,000 layersIn some embodiments, the neural network comprises 10 layers, 50 layers, 100 layers, 200 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, 500,000 layers, or 1,000,000 layers. In some embodiments, the neural network comprises at least 10 layers, 50 layers, 100 layers, 200 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, or 500,000 layers. In some embodiments, the neural network comprises at most 50 layers, 100 layers, 200 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, 500,000 layers, or 1,000,000 layers.

[0147] In certain embodiments, machine learning methods include testing with data not used for training to assess the training model or classifier of its predictive power. In certain embodiments, the predictive power of training model or classifier is assessed using one or more performance indicators. These performance indicators include classification accuracy, specificity, sensitivity, positive predictive value, negative predictive value, measurement area under the receiver operating curve (AUROC), mean square error, false discovery rate and the Pearson correlation between the predicted value and the actual value, and these performance indicators are determined by testing it for a group of independent cases. In some cases, for at least about 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190 or 200 independent cases (including increments therein), method has at least about 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95% or more AUROC (including increments therein). In some cases, the method has an accuracy of at least about 75%, 80%, 85%, 90%, 95% or more for at least about 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or 200 independent cases (including increments therein). In some cases, the method has a specificity of at least about 75%, 80%, 85%, 90%, 95% or more for at least about 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or 200 independent cases (including increments therein). In some cases, the method has a sensitivity of at least about 75%, 80%, 85%, 90%, 95% or more for at least about 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or 200 independent cases (including increments therein). In some cases, the method has a positive predictive value of at least about 75%, 80%, 85%, 90%, 95% or more for at least about 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or 200 independent cases (including increments therein). In some cases, the method has a negative predictive value of at least about 75%, 80%, 85%, 90%, 95% or more for at least about 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or 200 independent cases (including increments therein).

[0148] Transfer Learning

[0149] This article describes devices, software, systems, and methods for generating protein or polypeptide sequences based on one or more desired properties or functions. In some embodiments, transfer learning is used to improve prediction accuracy. Transfer learning is a machine learning technique in which a model developed for one task can be reused as the starting point for a model for a second task. By having the model learn on a related task with abundant data, transfer learning can be used to improve the prediction accuracy of tasks with limited data. The transfer learning method described in PCT application number PCT / US2020 / 01751762 / 804,036 is incorporated herein by reference. Therefore, this article describes a method for learning the general functional characteristics of proteins from a large data set of sequenced proteins and using it as the starting point of a model to predict any specific protein function, property, or feature. Therefore, the generation of an encoder can include transfer learning to improve the performance of the encoder when processing the input sequence into an embedding. Therefore, the improved embedding can improve the performance of the entire encoder-decoder framework. The present disclosure recognizes the surprising discovery that the information encoded by the first prediction model in all sequenced proteins can be migrated to design a specific protein function of interest using a second prediction model. In some embodiments, the prediction model is a neural network, such as a deep convolutional neural network.

[0150] The present disclosure can be implemented via one or more embodiments to achieve one or more of the following advantages. In some embodiments, models trained using transfer learning show improvements from a resource consumption perspective, such as showing a small memory footprint, low latency, or low computational cost. This advantage cannot be underestimated in complex analyses that may require enormous computing power. In some cases, it is necessary to use transfer learning to train a sufficiently accurate model within a reasonable time period (e.g., a few days rather than a few weeks). In some embodiments, models trained using transfer learning provide high accuracy compared to models trained without transfer learning. In some embodiments, the use of deep neural networks and / or transfer learning in systems for predicting polypeptide sequences, structures, properties, and / or functions improves computational efficiency compared to other methods or models without transfer learning.

[0151] In some embodiments, a first system is provided that includes a neural network embedder or encoder. In some embodiments, the neural network embedder includes one or more embedding layers. In some embodiments, the input to the neural network includes a protein sequence represented as a "one-hot" vector that encodes the amino acid sequence as a matrix. For example, within the matrix, each row can be configured to contain exactly one non-zero entry, corresponding to the amino acid present at the residue. In some embodiments, the first system includes a neural network predictor. In some embodiments, the predictor includes one or more output layers for generating predictions or outputs based on the input. In some embodiments, the first system is pre-trained using a first training dataset to provide a pre-trained neural network embedder. Using transfer learning, the pre-trained first system or a portion thereof can be transferred to form part of a second system. When used in the second system, one or more layers of the neural network embedder can be frozen. In some embodiments, the second system includes the neural network embedder or a portion thereof from the first system. In some embodiments, the second system includes a neural network embedder and a neural network predictor. The neural network predictor can include one or more output layers for generating a final output or prediction. The second system can be trained using a second training dataset labeled according to the protein function or property of interest. As used herein, embedders and predictors can refer to components of a prediction model, such as a neural network trained using machine learning. Within the encoder-decoder framework disclosed herein, an embedding layer can be processed for optimization and subsequently "decoded" into an updated or optimized sequence for one or more functions.

[0152] In some embodiments, transfer learning is used to train a first model, at least a portion of which is used to form a part of a second model. The input data of the first model can include a large data repository of known natural and synthetic proteins, regardless of function or other properties. The input data can include any combination of: primary amino acid sequence, secondary structure sequence, contact map of amino acid interactions, primary amino acid sequence as a function of amino acid physicochemical properties, and / or tertiary protein structure. Although these specific examples are provided herein, any additional information related to a protein or polypeptide is contemplated. In some embodiments, the input data is embedded. For example, the input data can be represented as a multidimensional tensor of binary unique heat encoding of the sequence, a real value (e.g., in the case of physicochemical properties or 3-dimensional atomic positions from tertiary structure), an adjacency matrix of pairwise interactions, or a direct embedding of the data (e.g., character embedding of the primary amino acid sequence). The first system can include a convolutional neural network architecture with an embedding vector and a linear model, which is trained using UniProt amino acid sequences and approximately 70,000 annotations (e.g., sequence tags). During transfer learning, the embedding vectors and convolutional neural network portion of the first system or model are transferred to form the core of a second system or model, which now incorporates a new linear model that is configured to predict protein properties or functions. The second system is trained using a second training dataset based on desired sequence tags corresponding to protein properties or functions. Once training is complete, the second system can be evaluated against a validation dataset and / or a test dataset (e.g., data not used in training).

[0153] In some embodiments, the data input of the first model and / or the second model is enhanced by additional data (e.g., random mutations of the primary amino acid sequence and / or biologically known mutations, contact maps of amino acid interactions, and / or tertiary protein structures). Additional enhancement strategies include the use of known isoforms and predicted isoforms from alternative splicing transcripts. In some embodiments, different types of inputs (e.g., amino acid sequences, contact maps, etc.) are processed by different parts of one or more models. After the initial processing step, information from multiple data sources can be combined at the layer of the network. For example, the network can include sequence encoders, contact map encoders, and other encoders configured to receive and / or process various types of data inputs. In some embodiments, data is converted into embeddings within one or more layers in the network.

[0154] The labels of the data input of the first model can be extracted from one or more public protein sequence annotation resources, such as: gene ontology (GO), Pfam domain, SUPFAM domain, enzyme commission (EC) number, taxonomy, extreme microorganism name, keyword, orthologous group assignment including OrthoDB and KEGG orthologs. In addition, labels can be assigned based on known structures or folding classifications specified by databases (such as SCOP, FSSP or CATH), including full α, full β, α+β, α / β, membrane, intrinsic disorder, coiled coil, small protein or designed protein. For proteins with known structure, quantitative global properties (such as total surface charge, hydrophobic surface area, measured or predicted solubility or other digital quantities) can be used as other labels fitted by predictive models (such as multi-task models). Although these inputs are described in the context of transfer learning, it is also contemplated that these inputs are applied to non-transfer learning methods. In some embodiments, the first model comprises an annotation layer that is stripped away to leave a core network consisting of an encoder. The annotation layer can include multiple independent layers, each layer corresponding to a specific annotation, such as, for example, primary amino acid sequence, GO, Pfam, Interpro, SUPFAM, KO, OrthoDB, and keywords. In some embodiments, the annotation layer comprises at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90, 100, 1000, 5000, 100000, or 150000 or more independent layers. In some embodiments, the annotation layer comprises 180000 independent layers. In some embodiments, at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90, 100, 1000, 5000, 100000, or 150000 or more annotations are used to train the model. In some embodiments, about 180000 annotations are used to train the model. In some embodiments, a plurality of annotations across multiple functional representations (e.g., one or more of GO, Pfam, Keywords, Kegg Ontology, Interpro, SUPFAM, and OrthoDB) are used to train the model. Amino acid sequences and annotation information can be obtained from various databases (e.g., UniProt).

[0155] In some embodiments, the first model and the second model comprise a neural network architecture. The first model and the second model can be supervised models using a convolutional architecture in the form of 1D convolution (e.g., primary amino acid sequence), 2D convolution (e.g., contact map of amino acid interactions), or 3D convolution (e.g., tertiary protein structure). The convolutional architecture can be one of the following described architectures: VGG16, VGG19, Deep ResNet, Inception / GoogLeNet (V1-V4), Inception / GoogLeNetResNet, Xception, AlexNet, LeNet, MobileNet, DenseNet, NASNet, or MobileNet. In some embodiments, a single model approach (e.g., non-transfer learning) utilizing any of the architectures described herein is contemplated.

[0156] The first model can also be an unsupervised model using a generative adversarial network (GAN), a recurrent neural network, or a variational autoencoder (VAE). If it is a GAN, the first model can be a conditional GAN, a deep convolutional GAN, a StackGAN, an infoGAN, a Wasserstein GAN, or a disco GANS model. In the case of a recurrent neural network, the first model can be a Bi-LSTM / LSTM, a Bi-GRU / GRU, or a converter network. In some embodiments, a single model method (e.g., non-transfer learning) is contemplated for generating an encoder and / or decoder using any architecture described herein. In some embodiments, GAN is a DCGAN, CGAN, SGAN / progressive GAN, SAGAN, LSGAN, WGAN, EBGAN, BEGAN, or infoGAN. Recurrent neural networks (RNNs) are variants of traditional neural networks built for sequential data. LSTM refers to long short-term memory (a type of neuron in RNN), and its memory allows it to model the order or time dependency in the data. GRU refers to a gated recurrent unit (a variant of LSTM) that attempts to address some of the shortcomings of LSTM. Bi-LSTM / Bi-GRU refers to "bidirectional" variants of LSTM and GRU. Typically, LSTM and GRU process sequentially in the "forward" direction, but the bidirectional versions also learn in the "backward" direction. LSTM can use hidden states to retain information from data inputs that have already passed through it. A unidirectional LSTM only retains information from the past because it only sees past inputs. In contrast, a bidirectional LSTM runs data inputs in both directions, from the past to the future, and vice versa. Therefore, a bidirectional LSTM running in both the forward and reverse directions retains information from both the future and the past.

[0157] The second model can use the first model as the starting point for training. The starting point can be the complete first model that is frozen except for the output layer, and the model is trained for the target protein function or protein property. The starting point can be the first model in which the embedding layer, the last 2 layers, the last 3 layers, or all layers are unfrozen, and the rest of the model is frozen during the training of the target protein function or protein property. The starting point can be the first model in which the embedding layer is removed and 1, 2, 3 or more layers are added, and the target protein function or protein property is trained. In some embodiments, the number of frozen layers is 1 to 10. In some embodiments, the number of frozen layers is 1 to 2, 1 to 3, 1 to 4, 1 to 5, 1 to 6, 1 to 7, 1 to 8, 1 to 9, 1 to 10, 2 to 3, 2 to 4, 2 to 5, 2 to 6, 2 to 7, 2 to 8, 2 to 9, 2 to 10, 3 to 4, 3 to 5, 3 to 6, 3 to 7, 3 to 8, 3 to 9, 3 to 10, 4 to 5, 4 to 6, 4 to 7, 4 to 8, 4 to 9, 4 to 10, 5 to 6, 5 to 7, 5 to 8, 5 to 9, 5 to 10, 6 to 7, 6 to 8, 6 to 9, 6 to 10, 7 to 8, 7 to 9, 7 to 10, 8 to 9, 8 to 10, or 9 to 10. In some embodiments, the number of frozen layers is 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10. In some embodiments, the number of frozen layers is at least 1, 2, 3, 4, 5, 6, 7, 8, or 9. In some embodiments, the number of frozen layers is at most 2, 3, 4, 5, 6, 7, 8, 9, or 10. In some embodiments, no layers are frozen during transfer learning. In some embodiments, the number of layers frozen in the first model is determined at least in part based on the number of samples available for training the second model. The present disclosure recognizes that freezing (multiple) layers or increasing the number of frozen layers can enhance the predictive performance of the second model. This effect may be more prominent when the sample size for training the second model is small. In some embodiments, all layers from the first model are frozen when the second model has no more than 200, 190, 180, 170, 160, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60, 50, 40, or 30 samples in the training set. In some embodiments, at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, or at least 100 layers in the first model are frozen for migration to the second model when the number of samples used to train the second model does not exceed 200, 190, 180, 170, 160, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60, 50, 40, or 30 samples in the training set.

[0158] The first model and the second model can have 10-100 layers, 100-500 layers, 500-1000 layers, 1000-10000 layers, or up to 1000000 layers. In some embodiments, the first model and / or the second model comprises 10 layers or 1,000,000 layers. In some embodiments, the first model and / or the second model comprises 10 to 50 layers, 10 to 100 layers, 10 to 200 layers, 10 to 500 layers, 10 to 1,000 layers, 10 to 5,000 layers, 10 to 10,000 layers, 10 to 50,000 layers, 10 to 100,000 layers, 10 to 500,000 layers, 10 to 1,000,000 layers, 50 to 100 layers, 50 to 200 layers, 50 to 500 layers, 50 to 1,000 layers, 50 to 5,000 layers, or layers, 50 layers to 10,000 layers, 50 layers to 50,000 layers, 50 layers to 100,000 layers, 50 layers to 500,000 layers, 50 layers to 1,000,000 layers, 100 layers to 200 layers, 100 layers to 500 layers, 100 layers to 1,000 layers, 100 layers to 5,000 layers, 100 layers to 10,000 layers, 100 layers to 50,000 layers, 100 layers to 100,000 layers, 100 layers to 500,000 layers, 100 layers to 1,000,000 layers, 200 layers to 500 layers, 200 layers to 1,000 layers, 200 layers to 5,000 layers, 200 layers to 10,000 layers, 200 layers to 50,000 layers, 200 layers to 100,000 layers, 200 layers to 500,000 layers, 200 layers to 1,000,000 layers, 500 layers to 1,000 layers, 500 layers to 5,000 layers, 500 layers to 10,000 layers, 500 layers to 50,000 layers, 500 layers to 100,000 layers, 500 layers to 500,000 layers, 500 layers to 1,000,000 layers, 1,000 layers to 5,000 layers, 1,000 layers to 10,000 layers, 1,000 layers to 50,000 layers, 1,000 layers to 100,000 layers, 1,000 layers to 500,000 layers, 1,000 layers to 1,000,000 layers, 5,000 layers to 10,000 layers, 5,000 layers to 50,000 layers, 5,000 layers to 100,000 layers, 5,000 layers to 500,000 layers, 5,000 layers to 1,000,000 layers, 10,000 layers to 50,000 layers, 10,000 layers to 100,000 layers, 10,000 layers to 500,000 layers, 10,000 layers to 1,000,000 layers, 50,000 layers to 100,000 layers, 50,000 layers to 500,000 layers, 50,000 layers to 1,000,000 layers, 100,000 layers to 500,000 layers, 100,000 layers to 1,000,000 layers, or 500,000 layers to 1,000,000 layers. In some embodiments, the first model and / or the second model comprises 10 layers, 50 layers, 100 layers, 200 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, 500,000 layers, or 1,000,000 layers. In some embodiments, the first model and / or the second model comprises at least 10 layers, 50 layers, 100 layers, 200 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, or 500,000 layers. In some embodiments, the first model and / or the second model comprises at most 50 layers, 100 layers, 200 layers, 500 layers, 1,000 layers, 5,000 layers, 10,000 layers, 50,000 layers, 100,000 layers, 500,000 layers, or 1,000,000 layers.

[0159] In some embodiments, a first system is described herein that comprises a neural network embedder and, optionally, a neural network predictor. In some embodiments, a second system comprises a neural network embedder and a neural network predictor. In some embodiments, the embedder comprises 10 to 200 layers. In some embodiments, the embedder comprises 10 to 20 layers, 10 to 30 layers, 10 to 40 layers, 10 to 50 layers, 10 to 60 layers, 10 to 70 layers, 10 to 80 layers, 10 to 90 layers, 10 to 100 layers, 10 to 200 layers, 20 to 30 layers, 20 to 40 layers, 20 to 50 layers, 2 ... 0 layers to 60 layers, 20 layers to 70 layers, 20 layers to 80 layers, 20 layers to 90 layers, 20 layers to 100 layers, 20 layers to 200 layers, 30 layers to 40 layers, 30 layers to 50 layers, 30 layers to 60 layers, 30 layers to 70 layers, 30 layers to 80 layers, 30 layers to 90 layers, 30 layers to 100 layers, 30 layers to 200 layers, 40 layers to 50 layers, 50 layers, 40 layers to 60 layers, 40 layers to 70 layers, 40 layers to 80 layers, 40 layers to 90 layers, 40 layers to 100 layers, 40 layers to 200 layers, 50 layers to 60 layers, 50 layers to 70 layers, 50 layers to 80 layers, 50 layers to 90 layers, 50 layers to 100 layers, 50 layers to 200 layers, 60 layers to 70 layers, 60 layers to In some embodiments, the embedder comprises 10 layers, 20 layers, 30 layers, 40 layers, 50 layers, 60 layers, 70 layers, 80 layers, 90 layers, 100 layers, or 200 layers. In some embodiments, the embedder comprises at least 10 layers, 20 layers, 30 layers, 40 layers, 50 layers, 60 layers, 70 layers, 80 layers, 90 layers, or 100 layers. In some embodiments, the embedder comprises at most 20 layers, 30 layers, 40 layers, 50 layers, 60 layers, 70 layers, 80 layers, 90 layers, 100 layers, or 200 layers.

[0160] In some embodiments, transfer learning is not used to generate the final training model. For example, when sufficient data is available, a model generated at least in part using transfer learning does not provide a significant improvement in prediction compared to a model that does not use transfer learning (e.g., when tested against a test dataset). Therefore, in some embodiments, a non-transfer learning approach is used to generate the training model.

[0161] Computing systems and software

[0162] In some embodiments, the system as described herein is configured to provide software applications, such as a polypeptide prediction engine (e.g., providing an encoder-decoder framework). In some embodiments, the polypeptide prediction engine comprises one or more models for predicting an amino acid sequence mapped to at least one function or property based on input data (e.g., an initial seed amino acid sequence). In some embodiments, the system as described herein comprises a computing device, such as a digital processing device. In some embodiments, the system as described herein comprises a network element for communicating with a server. In some embodiments, the system as described herein comprises a server. In some embodiments, the system is configured to upload to a server and / or download data from a server. In some embodiments, the server is configured to store input data, output, and / or other information. In some embodiments, the server is configured to back up data from a system or device.

[0163] In some embodiments, the system includes one or more digital processing devices. In some embodiments, the system includes multiple processing units configured to generate (multiple) training models. In some embodiments, the system includes multiple graphics processing units (GPUs), which are suitable for machine learning applications. For example, compared to a central processing unit (CPU), a GPU is generally characterized by an increase in the number of smaller logical cores consisting of an arithmetic logic unit (ALU), a control unit, and a memory cache. Therefore, the GPU is configured to process a larger number of simple and identical calculations in parallel, which are suitable for common mathematical matrix calculations in machine learning methods. In some embodiments, the system includes one or more tensor processing units (TPUs), which are AI application-specific integrated circuits (ASICs) developed by Google for neural network machine learning. In some embodiments, the method described herein is implemented on a system comprising multiple GPUs and / or TPUs. In some embodiments, the system includes at least 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 30, 40, 50, 60, 70, 80, 90, or 100 or more GPUs or TPUs. In some embodiments, the GPU or TPU is configured to provide parallel processing.

[0164] In some embodiments, the system or device is configured to encrypt data. In some embodiments, the data on the server is encrypted. In some embodiments, the system or device includes a data storage unit or memory for storing data. In some embodiments, data encryption is performed using the Advanced Encryption Standard (AES). In some embodiments, data encryption is performed using 128-bit, 192-bit, or 256-bit AES encryption. In some embodiments, data encryption includes full disk encryption of the data storage unit. In some embodiments, data encryption includes virtual disk encryption. In some embodiments, data encryption includes file encryption. In some embodiments, data transmitted or otherwise communicated between the system or device and other devices or servers is encrypted during transmission. In some embodiments, wireless communications between the system or device and other devices or servers are encrypted. In some embodiments, data in transmission is encrypted using Secure Sockets Layer (SSL).

[0165] The apparatus as described herein comprises a digital processing device comprising one or more hardware central processing units (CPUs) or general purpose graphics processing units (GPGPUs) that perform device functions. The digital processing device further comprises an operating system configured to execute executable instructions. The digital processing device is optionally connected to a computer network. The digital processing device is optionally connected to the Internet so that it accesses the World Wide Web. The digital processing device is optionally connected to a cloud computing infrastructure. Suitable digital processing devices include, by way of non-limiting example, server computers, desktop computers, laptop computers, notebook computers, small notebook computers, netbook computers, netpad computers, set-top computers, streaming media devices, handheld computers, internet devices, mobile smartphones, tablet computers, personal digital assistants, video game consoles, and communication media. Those skilled in the art will recognize that many smartphones are suitable for use in the system described herein.

[0166] Typically, a digital processing device includes an operating system configured to execute executable instructions. For example, an operating system is software, including programs and data, that manages the hardware of the device and provides services for the execution of applications. Those skilled in the art will recognize that suitable server operating systems include, by way of non-limiting example, FreeBSD, OpenBSD, Linux, Mac OS X Windows and Those skilled in the art will recognize that suitable personal computer operating systems include, by way of non-limiting example, Mac OS and UNIX-like operating systems, such as In some embodiments, the operating system is provided by cloud computing.

[0167] The digital processing device as described herein includes or is operably coupled to a storage and / or memory device. A storage and / or memory device is one or more physical devices for temporarily or permanently storing data or programs. In some embodiments, the device is a volatile memory and requires power to maintain the stored information. In some embodiments, the device is a non-volatile memory and retains the stored information when the digital processing device is not powered. In other embodiments, the non-volatile memory includes flash memory. In some embodiments, the non-volatile memory includes dynamic random access memory (DRAM). In some embodiments, the non-volatile memory includes ferroelectric random access memory (FRAM). In some embodiments, the non-volatile memory includes phase change random access memory (PRAM). In other embodiments, the device is a storage device, including, by way of non-limiting example, a CD-ROM, a DVD, a flash memory device, a disk drive, a tape drive, an optical drive, and cloud computing-based storage. In other embodiments, the storage and / or memory device is a combination of those devices disclosed herein.

[0168] In some embodiments, a system or method as described herein generates a database containing or comprising input and / or output data. Some embodiments of the systems described herein are computer-based systems. These embodiments include a CPU (including a processor and a memory), which may be in the form of a non-transient computer-readable storage medium. These system embodiments further include software typically stored in a memory (e.g., in the form of a non-transient computer-readable storage medium), wherein the software is configured to enable the processor to perform functions. The software embodiments incorporated into the systems described herein contain one or more modules.

[0169] In various embodiments, the apparatus includes a computing device or component, such as a digital processing device. In some embodiments described herein, the digital processing device includes a display to display visual information. Non-limiting examples of displays suitable for use with the systems and methods described herein include a liquid crystal display (LCD), a thin film transistor liquid crystal display (TFT-LCD), an organic light emitting diode (OLED) display, an OLED display, an active matrix OLED (AMOLED) display, or a plasma display.

[0170] In some embodiments described herein, the digital processing device includes an input device for receiving information. Non-limiting examples of input devices suitable for use with the systems and methods described herein include a keyboard, a mouse, a trackball, a trackpad, or a stylus. In some embodiments, the input device is a touch screen or multiple touch screens.

[0171] The systems and methods described herein typically include one or more non-transitory computer-readable storage media encoded with a program that includes instructions that can be executed by an operating system of an optionally networked digital processing device. In some embodiments of the systems and methods described herein, the non-transitory storage medium is a component of a digital processing device that is a component of the system or used in the method. In still other embodiments, the computer-readable storage medium is optionally removable from the digital processing device. In some embodiments, the computer-readable storage medium, by way of non-limiting example, includes a CD-ROM, a DVD, a flash memory device, a solid-state memory, a disk drive, a tape drive, an optical drive, a cloud computing system, and a server, etc. In some cases, the program and instructions are permanently, substantially permanently, semi-permanently, or non-transitory encoded on the medium.

[0172] Typically, the systems and methods described herein include at least one computer program or its use. A computer program comprises a series of instructions that can be executed in a CPU of a digital processing device and is written to perform a specified task. Computer-readable instructions can be implemented as program modules that perform specific tasks or implement specific abstract data types, such as functions, objects, application programming interfaces (APIs), data structures, and the like. Based on the disclosure provided herein, one skilled in the art will recognize that computer programs can be written in various versions of various languages. The functionality of the computer-readable instructions can be combined or distributed as needed in various environments. In some embodiments, the computer program comprises a single instruction sequence. In some embodiments, the computer program comprises multiple instruction sequences. In some embodiments, the computer program is provided from a single location. In other embodiments, the computer program is provided from multiple locations. In various embodiments, the computer program comprises one or more software modules. In various embodiments, the computer program may comprise, in part or in whole, one or more web applications, one or more mobile applications, one or more stand-alone applications, one or more web browser plug-ins, extensions, add-ons, or additional components, or a combination thereof. In various embodiments, a software module comprises a file, a code segment, a programming object, a programming structure, or a combination thereof. In other various embodiments, a software module comprises multiple files, multiple code segments, multiple programming objects, multiple programming structures, or a combination thereof. In various embodiments, by way of non-limiting example, the one or more software modules include a web application, a mobile application, and a standalone application. In some embodiments, the software module is within a single computer program or application. In other embodiments, the software module is within more than one computer program or application. In some embodiments, the software module resides on a single machine. In other embodiments, the software module resides on more than one machine. In yet another embodiment, the software module resides on a cloud computing platform. In some embodiments, the software module resides on one or more machines in a single location. In other embodiments, the software module resides on one or more machines in more than one location.

[0173] Typically, the systems and methods described herein include and / or utilize one or more databases. In view of the disclosure provided herein, those skilled in the art will recognize that many databases are suitable for storing and retrieving baseline data sets, files, file systems, objects, object systems, and the data structures and other types of information described herein. In various embodiments, by way of non-limiting examples, suitable databases include relational databases, non-relational databases, object-oriented databases, object databases, entity relationship model databases, associated databases, and XML databases. Other non-limiting examples include SQL, PostgreSQL, MySQL, Oracle, DB2, and Sybase. In some embodiments, the database is based on the Internet. In other embodiments, the database is based on a network. In still other embodiments, the database is based on cloud computing. In other embodiments, the database is based on one or more local computer storage devices.

[0174] Figure 6A The diagram illustrates a computer network or similar digital processing environment in which embodiments of the present invention may be implemented.

[0175] The client computer(s) / device(s) 50 and the server computer(s) 60 provide processing devices, storage devices, and input / output devices for executing application programs, etc. The client computer(s) / device(s) 50 may also be linked to other computing devices, including other client devices / processes 50 and server computer(s) 60, via a communication network 70. The communication network 70 may be a remote access network, a global network (e.g., the Internet), a global collection of computers, a local or wide area network, and a network currently using the corresponding protocols (TCP / IP, Other electronic equipment / computer network architectures are also applicable.

[0176] Figure 6B yes Figure 6A5 . A diagram of an example internal structure of a computer (e.g., client processor / device 50 or server computer 60) in a computer system of FIG. Each computer 50, 60 includes a system bus 79, where a bus is a set of hardware lines used to transfer data between components of a computer or processing system. The system bus 79 is essentially a shared channel that connects the different components of the computer system (e.g., processor, disk storage, memory, input / output ports, network ports, etc.), enabling information to be transferred between the components. Attached to the system bus 79 is an I / O device interface 82 for connecting various input and output devices (e.g., keyboard, mouse, display, printer, speakers, etc.) to the computer 50, 60. A network interface 86 allows the computer to connect to various other devices attached to a network (e.g., network 70 of FIG. 5 ). Memory 90 provides volatile storage for computer software instructions 92 and data 94 used to implement embodiments of the present invention (e.g., the neural network, encoder, and decoder detailed above). Disk storage 95 provides non-volatile storage for computer software instructions 92 and data 94 used to implement embodiments of the present invention. Central processor unit 84 is also attached to system bus 79 and provides execution of computer instructions.

[0177] In one embodiment, the processor routine 92 and data 94 are computer program products (generally referred to as 92) comprising a non-transitory computer-readable medium (e.g., a removable storage medium such as one or more DVD-ROMs, CD-ROMs, floppy disks, magnetic tapes, etc.) that provides at least a portion of the software instructions for the system of the present invention. The computer program product 92 can be installed by any suitable software installation program as is known in the art. In another embodiment, at least a portion of the software instructions can also be downloaded via wired communication and / or wireless connection. In other embodiments, the program of the present invention is a computer program propagation signal product embodied on a propagation medium (e.g., radio waves, infrared waves, laser waves, acoustic waves, or electric waves propagated on a global network (e.g., the Internet or other network(s))). Such a carrier medium or signal can be used to provide at least a portion of the software instructions for the routine / program 92 of the present invention.

[0178] Certain definitions

[0179] As used herein, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. For example, the term "a sample" includes a plurality of samples, including mixtures thereof. Any reference to "or" herein is intended to encompass "and / or" unless otherwise indicated.

[0180] As used herein, the term "nucleic acid" generally refers to one or more core bases, nucleosides or nucleotides. For example, a nucleic acid can include one or more nucleotides selected from adenosine (A), cytosine (C), guanine (G), thymine (T) and uracil (U) or variants thereof. Nucleotides generally include nucleosides and at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 or more phosphate (PO3) groups. Nucleotides can include core bases, five-carbon sugars (ribose or deoxyribose) and one or more phosphate groups. Ribonucleotides include nucleotides in which the sugar is ribose. Deoxyribonucleotides include nucleotides in which the sugar is deoxyribose. Nucleotides can be nucleoside monophosphates, nucleoside diphosphates, nucleoside triphosphates or nucleoside polyphosphates. Adenine, cytosine, guanine, thymine and uracil are referred to as typical or primary core bases. Nucleotides with non-primary or non-typical core bases include modified bases, such as modified purines and modified pyrimidines. Modified purine nucleobases include hypoxanthine, xanthine, and 7-methylguanine, which are part of the nucleosides inosine, xanthine, and 7-methylguanosine, respectively. Modified pyrimidine nucleobases include 5,6-dihydrouracil and 5-methylcytosine, which are part of the nucleosides dihydrouridine and 5-methylcytosine, respectively. Other atypical nucleosides include pseudouridine (Ψ), which is commonly found in tRNA.

[0181] As used herein, the terms "polypeptide," "protein," and "peptide" are used interchangeably and refer to a polymer of amino acid residues linked via peptide bonds, and which may be composed of two or more polypeptide chains. The terms "polypeptide," "protein," and "peptide" refer to a polymer of at least two amino acid monomers linked together by an amide bond. Amino acids may be L optical isomers or D optical isomers. More specifically, the terms "polypeptide," "protein," and "peptide" refer to a molecule composed of two or more amino acids in a specific order; for example, an order determined by the base sequence of nucleotides in a gene or RNA encoding the protein. Proteins are essential for the structure, function, and regulation of body cells, tissues, and organs, and each protein has a unique function. Examples are hormones, enzymes, antibodies, and any fragments thereof. In some cases, a protein may be a portion of a protein, such as a domain, subdomain, or motif of a protein. In some cases, a protein may be a variant (or mutation) of a protein, in which one or more amino acid residues are inserted into, deleted from, and / or substituted into a naturally occurring (or at least known) protein amino acid sequence. A protein or its variant may be naturally occurring or recombinant. A polypeptide can be a single linear polymer chain of amino acids held together by peptide bonds between the carboxyl and amino groups of adjacent amino acid residues. For example, polypeptides can be modified by the addition of carbohydrates, phosphorylation, and the like. A protein can contain one or more polypeptides. Amino acids include the typical amino acids arginine, histidine, lysine, aspartic acid, glutamic acid, serine, threonine, asparagine, glutamine, cysteine, glycine, proline, alanine, valine, isoleucine, leucine, methionine, phenylalanine, tyrosine, and tryptophan. Amino acids can also include atypical amino acids such as selenocysteine ​​and pyrrolysine. For example, polypeptides can be modified by the addition of carbohydrates, lipids, phosphorylation, and the like, such as through post-translational modification, as well as combinations thereof. A protein can contain one or more polypeptides. Amino acids include the typical L-amino acids arginine, histidine, lysine, aspartic acid, glutamic acid, serine, threonine, asparagine, glutamine, cysteine, glycine, proline, alanine, valine, isoleucine, leucine, methionine, phenylalanine, tyrosine, and tryptophan. Amino acids can also include atypical amino acids, such as the D-isomers of typical amino acids, and additional atypical amino acids, such as selenocysteine ​​and pyrrolysine. Amino acids also include atypical β-alanine, 4-aminobutyric acid, 6-aminocaproic acid, sarcosine, statin, citrulline, homocitrulline, homoserine, norleucine, valerine, and ornithine.Polypeptide can also include post-translational modifications, which include one or more of the following: acetylation, amidation, formylation, glycosylation, hydroxylation, methylation, myristoylation, phosphorylation, deamidation, prenylation (e.g., farnesylation, geranylation, etc.), ubiquitination, ribosylation and sulfation, including the aforementioned combination. Therefore, in some embodiments, the polypeptide provided by the invention or used in the method or system provided by the invention can contain in different embodiments: only typical amino acids, only atypical amino acids or a combination of typical amino acids and atypical amino acids, such as one or more D-amino acid residues in other L-amino acid containing polypeptides.

[0182] As used herein, the term "neural network" refers to an artificial neural network. An artificial neural network has a general structure of interconnected groups of nodes. The nodes are typically organized into multiple layers, where each layer contains one or more nodes. Signals can propagate from one layer to the next through the neural network. In some embodiments, the neural network comprises an embedder. The embedder may include one or more layers, such as an embedding layer. In some embodiments, the neural network comprises a predictor. The predictor may include one or more output layers that generate an output or result (e.g., a predicted function or property based on the primary amino acid sequence).

[0183] As used herein, the term "artificial intelligence" generally refers to a machine or computer that can perform tasks in an "intelligent" or non-repetitive or rote or pre-programmed manner.

[0184] As used herein, the term "machine learning" refers to a type of learning in which a machine (eg, a computer program) can learn on its own without being programmed.

[0185] As used herein, the phrase "at least one of a, b, c, and d" means a, b, c, or d, and any and all combinations comprising two or more of a, b, c, and d.

[0186] Example

[0187] Example 1: Engineering Green Fluorescent Protein Using Gradient-Based Design in Computer

[0188] A computational machine learning approach was used to convert non-luminescent proteins into fluorescent proteins. The source data for this experiment consisted of 50,000 publicly available GFP sequences that had been assayed for fluorescence. An encoder neural network was generated with the help of transfer learning by using a model pre-trained on the UniProt database and then training this model to predict fluorescence from sequences. The 80% of proteins with the lowest brightness were selected as the training dataset, while the 20% with the highest brightness were retained as the validation dataset. The mean squared error for both the training and validation sets was <0.001, demonstrating high accuracy in predicting fluorescence directly from sequence. Figure 5Aand Figure 5B The data graphs of the actual fluorescence values ​​and the predicted fluorescence values ​​in the training set and the validation set are shown in FIG.

[0189] Figure 7 A diagram illustrating gradient-based design (GBD) for engineering GFP sequences is shown. An embedding 702 is optimized based on the gradient. A decoder 704 is used to determine the GFP sequence based on the embedding, which can then be evaluated by a GFP fluorescence model 706 to obtain a predicted fluorescence 708. Figure 7 As shown, the process of generating GFP sequences using gradient-based design involves taking a step in the embedding space guided by the gradient, making a prediction 710, re-evaluating the gradient 712, and then repeating the process.

[0190] After training the encoder, a currently non-fluorescent sequence is selected as a seed protein and projected into an embedding space (e.g., a 2-dimensional space) using the trained encoder. A gradient-based update procedure is run to improve the embedding, thereby optimizing the embedding of the seed protein. Next, derivatives are calculated and used to move through the embedding space toward regions of higher functionality. The optimized embedding coordinates are improved in terms of fluorescence functionality. Once the desired level of functionality is achieved, the coordinates in the embedding space are projected back into protein space, resulting in an amino acid sequence with the desired functionality.

[0191] A selection of 60 GBD-designed sequences with the highest predicted brightness was chosen for experimental validation. Figure 8 The results of experimental validation of sequences created using GBD are shown in . The Y axis is the fold change in fluorescence relative to avGFP (WT). Figure 8 From left to right, the following are shown: (1) the brightness of WT—avGFP, which serves as the control for all GFP sequences trained on the supervised model; (2) engineered: artificially designed GFPs, referred to as "super folders" (sfGFP); and (3) GBD: novel sequences created using a gradient-based design program. As can be seen, in some cases, sequences designed by GBD are approximately 50 times brighter than wild-type and training sequences, and 5 times brighter than the well-known human-engineered sfGFP. These results demonstrate that GBD is capable of engineering polypeptides that have superior functionality to that of human-engineered polypeptides.

[0192] Figure 9 A pairwise amino acid sequence alignment of avGFP and the GBD-engineered GFP sequence with the highest experimentally verified fluorescence (approximately 50-fold higher than avGFP) is shown 900. A period '.' indicates no mutation relative to avGFP, while mutations or pairwise differences are indicated by the single-letter amino acid code representing the GBD-engineered GFP amino acid residue at the indicated position in the alignment. Figure 9As shown, pairwise alignment revealed 7 amino acid mutations or residue differences between avGFP (which is SEQ. NO. 1) and the GBD-engineered GFP polypeptide sequence (which can be referred to as SEQ. NO. 2).

[0193] avGFP is a 238-amino acid polypeptide having the following sequence: SEQ ID NO: 1. The GBD-engineered GFP polypeptide has seven amino acid mutations relative to the avGFP sequence: Y39C, F64L, V68M, D129G, V163A, K166R, and G191V.

[0194] The decoder achieved >99.9% residue accuracy on both the training and validation data, meaning that on average the decoder made 0.5 errors per GPF sequence (assuming a GFP length of 238 amino acids). Next, the decoder's performance was evaluated for protein design. First, the encoder was used to embed each protein in the training and validation sets. Next, the decoder was used to decode these embeddings. Finally, the encoder was used to predict fluorescence values ​​for the decoded sequences, and these predictions were compared with those predicted using the original sequences. Figure 4 An overview of the process is shown in .

[0195] The correlation between the predictions from the original sequence and the predictions from the decoded sequence was calculated. A high degree of consistency was observed in both the training and validation datasets. Table 1 summarizes these observations.

[0196] Table 1

[0197] data Correlation train 0.99 verify 0.77

[0198] Example 2: Engineering a β-lactamase gene using gradient-based design in silico

[0199] A computational machine learning approach was used to convert β-lactamases to confer resistance to antibiotics to which they were previously resistant. Using a training set of 662 publicly available β-lactamase sequences for which resistance to 11 antibiotics had been measured, a multi-task deep learning model was developed to predict resistance to these antibiotics based on amino acid sequence.

[0200] Next, 20 β-lactamases that were not resistant to the test antibiotic were selected from the training set with the goal of designing new sequences that were resistant to this antibiotic. Gradient-based design (GBD) was applied to these sequences for a total of 100 iterations. Figure 10 A visualization of this process is shown in . As mentioned before, the initial sequence is used as a seed for mapping into the embedding space and subsequently optimized over 100 iterations. Figure 10The predicted resistance of the designed sequences to the test antibiotics is shown as a function of iterations of gradient-based design. The y-axis represents the resistance predicted by the model, and the x-axis represents the rounds or iterations of gradient-based design when optimizing the embedding. Figure 10 The figure shows how predicted resistance increases with rounds or iterations of GBD. The seed sequence starts with low resistance (round 0) and is iteratively improved to have high predicted resistance (probability > 0.9) after several rounds. As shown in the figure, it seems that predicted resistance peaks at around round 25 and then reaches a stable level.

[0201] Unlike GFP, β-lactamases have variable lengths, so in this example, GBD is able to control the length of the protein.

[0202] Seven sequences were selected for experimental verification, as shown in Table 2 below.

[0203] Table 2. Seven sequences designed by GBD were selected for experimental validation. These seven sequences were chosen based on a combination of factors, including a high probability of resistance to the test antibiotic (ResistanceProb), low sequence identity with sequences in the training data that are resistant to the test antibiotic (ClassPercentID), and low mutual sequence identity. The longest β-lactamase in the training data is 400 amino acids, and several GBD-designed β-lactamase peptide sequences exceed this length.

[0204]

[0205]

[0206] Validation experiments were performed for seven novel β-lactamases designed using GBD. Bacteria transformed with vectors expressing β-lactamases were serially diluted 10-fold and grown on agar plates in the presence of 8 μg / ml test antibiotics plus 1 mM IPTG. Figure 11 This is a diagram illustrating an antibiotic resistance test. The representative β-lactamase TEM-1 is shown in the last column. Clearly, several of the designed sequences exhibited greater resistance to the test antibiotic than TEM-1. The β-lactamases in columns 14-1 and 14-2 had colonies with five spots pointing downward. Column 14-3 had colonies with seven spots pointing downward. Columns 14-4, 14-6, and 14-7 had colonies with four spots pointing downward. Column 14-5 had colonies with three spots pointing downward. Meanwhile, TEM-1 had only two spots pointing downward.

[0207] Example 3 – Comprehensive Experiment Using Gradient-Based Design on Simulated Terrain

[0208] The goal of this disclosure is to use machine learning to computationally design biological sequences with specific functional properties. A common strategy is model-based optimization: a model that maps sequences to functions is trained on labeled data and then optimized to produce sequences with the desired function. However, naive optimization methods cannot avoid out-of-distribution inputs, where the model error is high. To address these issues, explicit and implicit methods constrain the target to inputs within the distribution, allowing for efficient generation of new biological sequences.

[0209] Protein engineering involves the generation of novel proteins with desired functional properties. This field has numerous applications, including the design of protein therapeutics, agricultural proteins, and industrial biocatalysts. Identifying amino acid sequences that encode proteins with a desired function is challenging, in part because while the subset of functional sequences is very small, the space of candidate sequences is vast.

[0210] One family of approaches that has achieved success is directed evolution: an iterative process that alternates between sampling from a library of genetic variants and screening for variants with improved function from which to build the next round of candidates. Even with the development of high-throughput assays, this process is time- and resource-intensive, requiring multiple iterations and screening of large numbers of variants. In many applications, designing high-throughput assays for the desired functional properties is challenging or infeasible.

[0211] Recent approaches utilize machine learning methods to design libraries more efficiently and obtain sequences with higher fitness with fewer iterations / screening. One such approach is model-based optimization. In this setting, a model that maps sequences to functions is fitted to labeled data. The model then computationally screens variants and designs a library with higher fitness. In embodiments, the systems and methods of the present disclosure ameliorate issues encountered in naive approaches to model-based optimization and improve the resulting sequences.

[0212] In this example, let X represent the space of protein sequences and f be a real-valued mapping on the protein space that encodes a property of interest (e.g., fluorescence, activity, expression, solubility). The task of designing a novel protein with a specified function can then be reformulated as finding the following solution:

[0213]

[0214] where f is usually unknown. This type of problem is called model-based optimization. This problem can be restricted to a static setting where f cannot be directly queried but a labeled dataset is available. where label y i May be noisy: y i ≈f(x i ).

[0215] The naive approach is to use D to fit a model f that approximates f. θ , then solve:

[0216]

[0217] This tends to produce poor results because the optimizer may find θ Wrongly large points. The key problem is that the space of possible amino acid sequences has very high dimensionality, but data are usually sampled from a much lower dimensional subspace. In practice, θ is high dimensional, and f θ This is exacerbated by the fact that the output is highly nonlinear (e.g. due to phenomena like epistasis in biology). Therefore, the output must be constrained in some way to limit the search to f θ A class of acceptable sequences that are good approximations of f.

[0218] One approach is to transform the probability model p θ Fit to (x i ) N So that p θ (x) is the probability that the sequence x is sampled from the data distribution. Some examples of model classes for which the likelihood can be explicitly calculated (or lower bounded) are first-order / site-wise models, hidden Markov models, conditional random fields, variational autoencoders (VAEs), autoregressive models, and flow-based models. In an embodiment, the method optimizes the following function:

[0219]

[0220] where λ>0 is a fixed hyperparameter. Labeled data are often expensive or scarce, but unlabeled protein examples from the family of interest are readily available. In practice, p θ A larger data set of unlabeled proteins from this series can be fitted.

[0221] One challenge of optimizing directly in sequence space is that sequence space is discrete and therefore not amenable to gradient-based methods. θ The fact that is a smooth function of the learned continuous sequence space representation allows for more efficient use of gradients and optimization. To this end, f θ =a θ e θ , where f θ is an L-layer neural network, e θ :Z, called the encoder, is the first K layers, and a θ : Z→R, called the annotator, is the last LK layer. This allows us to move the optimization to space Z and utilize the gradient. The non-regularized analogy is to solve:

[0222]

[0223] Then fit the probabilistic decoder Its mapping So that for x' sampled from the data distribution

[0224]

[0225] It can return One might expect the problem to be complicated here, since the gradient may shift z * Pull in not only a θ and The method is motivated by the observation that due to a θ and is trained on the same data manifold, so The reconstruction error is often related to a θ The mean absolute error is related to .

[0226] The following objective function is proposed:

[0227]

[0228] This adds an implicit constraint to the optimization. The stable solution of (5) corresponds to Z regions with low entropy and low reconstruction error. A heuristic way to think about this regularization is that since the decoder is trained to output a distribution centered on points in the data distribution, the mapping can be thought of as projecting onto the data manifold. While the earlier f θ is a map on X, and the equation shows that f θ is a mapping on p(). However, the following describes the f fitted by equation (5) θ Finally, with p in equation (3) θ Similarly, if gradient ascent is used as gradient-based design (GBD) via Equation (5), the decoder Can be fitted to a larger dataset of unlabeled proteins from the series of interest.

[0229] Results - Comprehensive Experiment

[0230] Evaluating model-based optimization methods requires querying the ground truth function f. In practice, this can be slow and / or expensive. To aid in developing and evaluating methods, the method is tested through comprehensive experiments in two settings: a lattice protein optimization task and an RNA optimization task. In both tasks, the ground truth f is highly nonlinear and approximates nontrivial biophysical properties of real biological sequences.

[0231] Lattice proteins refer to the simplifying assumption that a protein of length L is constrained to take on conformations located on a two-dimensional lattice with no self-intersections. This assumption allows the enumeration of all possible conformations and the exact computation of the partition function, enabling the efficient calculation of many thermodynamic properties. The ground-truth fitness f is defined as the free energy of the amino acid chain relative to a fixed conformation sf. Optimizing a sequence for this fitness is equivalent to finding a sequence that is stable relative to a fixed structure, a long-standing goal of sequence design.

[0232] The free energy of a nucleotide sequence relative to a fixed conformation can be efficiently calculated without the many simplifying assumptions made in a 2D lattice protein model. In the RNA optimization setting, f is defined in nucleotide sequence space as the free energy relative to a fixed conformation sf of known tRNA structures.

[0233] For both tasks, after defining f, a fitness landscape is generated from which training data is selected via modified Metropolis-Hastings sampling. Under Metropolis-Hastings, the probability of a sequence x being included in the landscape is asymptotically proportional to f(x). The data is split according to fitness: validation data is uniformly sampled from sequences with higher fitness and training data is uniformly sampled from sequences with lower fitness, to evaluate the method's ability to generate sequences with fitness greater than those seen during training, a desirable property in real-world applications.

[0234] Convolutional Neural Network θ and point-by-point p θ is fitted to the data. A set of 192 seed sequences is drawn from the training data and optimized according to discrete optimization objectives (2) and (3) and gradient-based optimization objectives (4) and (5). The discrete objectives are optimized using a greedy local search algorithm, where at each step, multiple candidate mutations are sampled from the empirical distribution given by the training data, and the best mutation is selected for each sequence in the cohort according to the objective.

[0235] Naive optimization quickly drives the group into regions of space where model error is high and fails to improve the group's average fitness in either experiment. Regularization reduces this effect, allowing the group's average fitness to increase while model error remains low. Few (<1%) of the generated sequences exceed the fitness values ​​seen during training on either task.

[0236] 12A to 12F are graphs illustrating discrete optimization results for RNA optimization (12A-C) and lattice protein optimization (12D-F). Figure 12A and Figure 12DThe figure shows the fitness of the entire population (μ±σ) during optimization. Naive optimization does not lead to a significant improvement in the average fitness in any environment, while the regularized objective does. Figure 12B and Figure 12E The fitness of subgroups consisting of the top 10 percentiles of fitness are shown (shaded area represents the minimum to maximum performance in the subgroup). In the RNA sandbox, no method was able to find a sequence with a significantly higher fitness than that seen during training. Figure 12C and Figure 12F The graph shows the f θ The absolute deviation from f (μ±σ). The naive objective fails to improve group performance because the group moves to parts of the space where the model is unreliable.

[0237] Figure 14 illustrates the effect of increasing the weight of the regularization term λ in Equation (3): a larger λ leads to a decrease in model error, but during the optimization process, the model is constrained to assign high probability p θ sequences, the sequence diversity also decreases accordingly. For all experiments testing this system, λ was set to 5 if not specified otherwise. However, other tests can also use other values. The left figure illustrates that as λ in the objective (3) increases, the average model error (μ±σ) of the entire group decreases, while the right figure illustrates that the sequence diversity in the group also decreases. The data were acquired from the Lattice Protein Sandbox environment. Gradient-based methods enter the space faster than discrete methods. GBD is able to explore regions of sequence space farther away from the initial seed while maintaining a low model error comparable to discrete regularization methods.

[0238] 13A to 13H The results of gradient-based optimization are shown in Figure 2. The problems highlighted above when optimizing are only exacerbated when working in Z: without regularization, not only is the group driven to the point z where a θ (z) has an unrealistically (and incorrectly) high predicted fitness value, and the decoded sequence Nor will it be f θ is predicted to have high fitness. In both settings, naive optimization fails to improve the average fitness of the entire population, nor finds sequences with fitness exceeding that seen during training. GBD does not exhibit this behavior: it successfully optimizes f θ d * 、a θ and In both cases, GBD improves the average fitness of the group, and the fitness of the top 10% of sequences in the group consistently exceeds the fitness seen during training.

[0239] 13A to 13DThe gradient-based optimization results for RNA optimization are shown, and 13E to 13H Lattice protein optimization is illustrated. Figure 13A and Figure 13E Pictured The maximum likelihood decoded sequence true fitness of the entire population during optimization. Naive optimization did not lead to a significant improvement in average fitness in the RNA sandbox, but resulted in a significant decrease in population fitness in the lattice protein environment. GBD was able to successfully improve average population fitness during optimization. Figure 13B and Figure 13F The fitness of a subgroup consisting of the top 10 percentiles of fitness is shown (the shaded area is the minimum to maximum performance in the subgroup). GBD reliably finds sequences with fitness values ​​that exceed the fitness values ​​seen during training. Figure 13C and Figure 13G This diagram shows the group during optimization. Panel 2, predicted fitness of the decoded sequence at the current point in Z. Figure 13D and Figure 13H The graph shows the a of the group during optimization θ (z)(μ±σ), i.e., the predicted fitness of the current representation in Z. Naive objective fast super-optimization a θ , thereby pushing the group into Z space where it cannot be The GBD target successfully prevents this anomaly.

[0240] FIG. 15A to FIG. 15B The diagram shows the heuristic motivation GBD: it drives the group to The Z region can be reliably decoded. Observed in X, this means Nearly identical (right), or observed in Z Very small, so The data show that f θ It is also reliable in this region of space because f θ and are trained on the same distribution.

[0241] Figure 15A is the optimal protein landscape for all steps and all sequences in the group for a θ (z) with The deviation of a θ (z) with A scatter plot of the deviations. Figure 15B It's a picture The accuracy of a is plotted against the same data. θ (z) with The maximum likelihood decoding of the points in Z is drawn by the deviation of Reliably decoded Z regions implicitly provide regularization. Since f θ and The fits are on the same distribution, so the predicted fitness in this region is reliable.

[0242] In comprehensive experiments, GBD was able to meet or exceed the performance of Monte Carlo optimization methods that explored based on group fitness (average and maximum). In practice, GBD is much faster: discrete methods involve generating and evaluating K candidate mutations in each iteration. This requires K forward passes through the model per sequence per iteration. GBD requires one forward pass and one backward pass per sequence per iteration.

[0243] in addition, Figure 16 The number of mutations (μ±σ) from the initial seeds in the cohort during optimization of various targets in the lattice protein is shown. Figure 16 The figure shows that GBD can find the optimal value farther away from the initial seed sequence than the discrete method while maintaining a fairly low error.

[0244] Table 3 provides a comparison of all the methods discussed, along with a random search baseline. On the RNA Sandbox, GBD was the only method explored that generated sequences with greater fitness than seen across the entire landscape generated by Metropolis Hastings (which ran for several orders of magnitude more iterations than the optimization). The Python package LatticeProteins enumerates all possible non-self-crossing conformations of a 16-amino acid chain. This enumeration is used to calculate the free energy of a 16-amino acid chain at a fixed conformation, sf. The fitness function, f, is defined over the space of sequences of length 32 amino acids as follows:

[0245] f(x)=E(x1)+E(x2)-R(x1,x2) (6)

[0246] Here, E(x1) is the free energy of the chain formed by the first 16 amino acid residues relative to sf, and E(x2) is the free energy of the chain formed by the last 16 amino acid residues relative to sf.

[0247] R(x1, x2) = c((x1) i , (x2) i ) (7)

[0248] and c(α,β) is a constant interaction term sampled from the standard normal of all amino acids α,β.

[0249] RNA structural fitness function

[0250] Let sf be a fixed tRNA structure. With the help of the Python package ViennaRNA, the fitness function f is defined on the space of sequences of length 70 nucleotides as:

[0251] f(x)=E(x)-min(exp(βd(s f , s x )), 20) (8) where d represents the Hamming distance, β = 0.3 is a hyperparameter, and s x represents the minimum energy conformation of x, and E(x) represents the sequence in conformation s x The free energy in .

[0252] Greedy Monte Carlo Search Optimization

[0253] This method optimizes objectives 2 and 3 using a greedy Monte Carlo search algorithm. x is a sequence of length L, and at each iteration, K mutations are sampled from a prior distribution given by the training data. More precisely, K positions are uniformly sampled from 1…L by substitution, and for each position, an amino acid (or nucleotide, in the case of RNA optimization) is sampled from the marginal distribution given by the data at that position. The objective is then evaluated for each variant in the library (including the original sequence), and the best variant is selected. This process continues for M steps.

[0254] D. Generation of Fitness Landscapes

[0255] In the case of fitness function f, we want to obtain the fitness function for training the supervised model f θ Since X is high-dimensional, uniform sampling is not feasible, intuitively because the probability of a sequence being randomly selected is high, which leads to extremely low fitness. The goal is to obtain samples from a distribution with density proportional to f. For each inner loop in the process, a set of M sequences is randomly initialized. For each sequence, the N mutations sampled are randomly uniformly drawn and include all MN sequences in the landscape. (x ij ) N Denotes the N variants of sequence i, and the method updates by sampling mutations from a categorical distribution [1...N], where logits are given by (f(x ij )) N The inner loop runs J steps, and the outer loop runs C times, as described further below.

[0256] Gradient-based design

[0257] Gradient-based design refers to optimizing the objective by gradient ascent (4). Given f θ 、 and the initial point z0, set An iteration of GBD consists of K steps of a gradient-based optimizer (such as Adam) to maximize h, followed by a decoding step where In practice, an effective learning rate is key to good performance, and a value of 0.05 was used throughout the experiments with K of 20.

[0258] Model architecture and training

[0259] This method is used to θ =a θ e θ The convolutional encoder e is used in all experiments. θ , which consists of alternating stacks of convolutional blocks and average pooling layers. A block consists of two layers wrapped in a residual connection. Each layer includes 1d convolution, layer normalization, dropout, and ReLU activation. A 2-layer fully connected feedforward network is used throughout. θ .Decoder Network It consists of an alternating stack of residual blocks and transposed convolutional layers, followed by a 2-layer fully connected feed-forward network.

[0260] Parameter estimation is done sequentially rather than jointly: first fit f θ , then freeze the parameters θ and fit Learning is done by stochastic gradient descent to minimize f using the ADAM optimizer respectively. θ 、 The maximum learning rate used is 10 -4 A single-cycle learning rate annealing schedule, f θ was fitted for 20 epochs, was fitted for 40 epochs. At each epoch, the model parameters were saved and after training, the best parameters measured by the validation loss were selected for generation. In all experiments, a site-wise p-value fit by maximum likelihood was used. θ .

[0261] A variational autoencoder is fitted to the data by maximizing the lower bound of the evidence. The encoder and decoder parameters are learned jointly via reparameterization (amortization). -3 A constant learning rate of is used for 50 epochs, and early stopping and patience parameters of 10 are set. For 20 iterations, N = 5000 sequences are sampled from the standard normal prior and passed through the decoder, which is determined by f θDistribute predicted fitness. The VAE was fine-tuned on these sequences for 10 epochs, reweighted to generate sequences with higher predicted fitness. The results in Table I report the iteration corresponding to the maximum mean true fitness for both methods, as both generative models collapse to the Δ quality function before the completion of 20 iterations. Therefore, the reported metrics summarize the peak performance of the methods.

[0262]

[0263] Table 3: Comparison of lattice protein optimization and RNA optimization methods: for random search, naive Monte Carlo, regularized Monte Carlo, naive gradient-based, and gradient-based design methods: the entire population being optimized, the true fitness (μ ± σ) of the top 10% of the population, and the maximum fitness sequence in the population at the end of optimization. The optimization consists of 20 iterations applied to 192 sequences sampled from the training data (which remain constant throughout the method).

[0264] Example 4 Engineering Antibodies Using Gradient-Based Design in Computer

[0265] The above describes the use of gradient-based design to generate antibodies that bind to fluorescein isothiocyanate (FITC) and have an improved dissociation constant (KD). The model is trained on a publicly available KD estimation dataset containing 2825 unique antibody sequences measured using fluorescence-activated cell sorting followed by next-generation sequencing, as described in "Measuring the sequence-affinity landscape of antibodies with massively parallel titration curves" (2016) by Adams RM, Mora T, Walczak AM, Kinney JB, Elife, the entire text of which is incorporated herein by reference. The dataset of sequences and KD pairs mapping antibody sequences to KD is split in three ways. The first split is by retaining the top 6% for performing sequence validation (so the model is trained on the bottom 94%). The second split is by retaining the top 15% for performing sequence validation (so the model is trained on the bottom 85%). The third split was performed by uniformly sampling (iid) 20% of the sequences to be retained for validation.

[0266] For each split, a supervised model consisting of an encoder (mapping sequences to embeddings) and an annotator (mapping embeddings to KD) was jointly fitted. A decoder that maps embeddings back to sequences was then fitted on the same training set. For each model, 128 seeds were uniformly sampled from the training set and optimized in two ways. The first was by performing 5 rounds of GBD, each consisting of 20 GBD steps, and then projecting back through the decoder. The second was by performing 5 rounds of GBD+ (where the objective was augmented with first-order regularization), each consisting of 20 GBD steps, and then projecting back through the decoder. GBD+ uses additional regularization, including the use of MSA (multiple sequence alignment) to constrain the method. Therefore, each model produces two sets of candidates (one for each method, GBD and GBD+). The final sequences to be ranked are selected from each group by first labeling each candidate with the predicted expression (from an independently trained expression model fitted to a dataset of sequence and expression data split in an iid (independent and identically distributed) manner). The cohort was filtered in two ways: if a sequence was predicted to be underexpressed, it was removed, and if a sequence's predicted fitness was lower than its seed's initial predicted fitness, it was removed. Of the remaining sequences, the highest predicted fitness sequence was selected for measurement in the laboratory.

[0267] Figure 17 is a graph 1700 illustrating wet lab data measuring the Kd of the listed protein variants, validating the affinity of the generated proteins.

[0268] The methods illustrated in this figure include regularized and non-regularized CDE, regularized and non-regularized GBD, and a baseline process. Figure 17 The underlying dataset is shown in Table 4 below, which lists the experimentally measured Kd values ​​for the generated proteins.

[0269] Table 4

[0270]

[0271]

[0272] A wet lab experiment to measure the Kd of our GBD-generated variants was performed as follows. Yeast cells were transformed with cloned plasmids expressing variants of unique anti-FITC scFv designs, formatted for surface display and including a cMyc tag for expression quantification. Following culture and scFv expression, yeast cells were stained with fluorescein antigen and a fluorescently conjugated anti-cMyc antibody at various concentrations. After reaching equilibrium, cells stained at each concentration were measured by flow cytometry. The median fluorescence intensity of fluorescein antigen binding was calculated after gating on expressing cells. The median fluorescence data were fit to a standard single binding affinity curve to determine the approximate binding affinity Kd (dissociation constant) for each cloned scFv variant. These results demonstrate that GBD outperforms other design methods for designing FITC antibodies.

[0273] Although preferred embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Without departing from the present invention, it will now be clear to those skilled in the art that many variations, changes, and substitutions are possible. It should be understood that the various alternatives of the embodiments of the present invention described herein can be used to practice the present invention. The accompanying claims are intended to define the scope of the present invention, and thus cover methods and structures within the scope of these claims and their equivalents.

[0274] The disclosure of this application also includes the following illustrative examples:

[0275] Illustrative Example 1: A method for engineering an improved biopolymer sequence through functional evaluation, the method comprising:

[0276] (a) providing an embedding starting point to a system comprising a supervised model for predicting a function of a biopolymer sequence and a decoder network, optionally wherein the starting point is an embedding seed biopolymer sequence, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a probabilistic biopolymer sequence given the embedding of the biopolymer sequence in the function space;

[0277] (b) computing the change in the embedding of the function relative to the starting point according to the step size, thereby providing a first update point in the function space;

[0278] (c) optionally computing a change in the function relative to the embedding at a first update point in the function space, and optionally iteratively computing a change in the function relative to the embedding at further update points;

[0279] (d) providing the first update point, or optionally the iterative further update point, in the function space to the decoder network when the first update point, or optionally the iterative further update point, in the function space approaches the desired function level; and

[0280] (e) Obtaining probabilistically improved biopolymer sequences from the decoder.

[0281] Illustrative Example 2: A method for engineering an improved biopolymer sequence through functional evaluation, the method comprising:

[0282] (a) providing an embedding starting point to a system comprising a supervised model network that predicts a function of a biopolymer sequence and a decoder network, optionally wherein the starting point is an embedding seed biopolymer sequence, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a predicted probabilistic biopolymer sequence given the embedding of the predicted biopolymer sequence in the function space;

[0283] (b) predicting the function of the embedding starting point;

[0284] (c) computing the change in the embedding of the function relative to the starting point according to the step size, thereby providing a first update point in the function space;

[0285] (d) providing a first updated point in the function space to the decoder network to provide a first intermediate probabilistic biopolymer sequence;

[0286] (e) providing the first intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the first intermediate probabilistic biopolymer sequence,

[0287] (f) computing a change in the embedding of the function relative to the first update point in the function space to provide an updated point in the function space;

[0288] (g) providing updated points in the function space to the decoder network to provide additional intermediate probabilistic biopolymer sequences;

[0289] (h) providing the additional intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the additional intermediate probabilistic biopolymer sequence;

[0290] (i) then calculating a change in the embedding of the function relative to the first further update point in the function space to provide a further update point in the function space, optionally iterating steps (g)-(i), wherein the further update point in the function space referenced in step (i) is taken to be the further update point in the function space in step (g); and

[0291] (j) providing the embedded point to the decoder network when approaching a desired level of functionality in the functionality space; and obtaining a probabilistically improved biopolymer sequence from the decoder.

[0292] Illustrative embodiment 3: A non-transitory and / or non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to:

[0293] (a) providing an embedding starting point to a system comprising a supervised model for predicting a function of a biopolymer sequence and a decoder network, optionally wherein the starting point is an embedding seed biopolymer sequence, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a probabilistic biopolymer sequence given the embedding of the biopolymer sequence in the function space;

[0294] (b) computing the change in the embedding of the function relative to the starting point according to the step size, thereby providing a first update point in the function space;

[0295] (c) optionally computing a change in the function relative to the embedding at a first update point in the function space, and optionally iteratively computing a change in the function relative to the embedding at further update points;

[0296] (d) providing the first update point, or optionally the iterative further update point, in the function space to the decoder network when the first update point, or optionally the iterative further update point, in the function space approaches the desired function level; and

[0297] (e) Obtaining probabilistically improved biopolymer sequences from the decoder.

[0298] Illustrative Embodiment 4: A system comprising a processor and a non-transitory and / or non-transitory computer readable medium comprising instructions that, when executed by the processor, cause the processor to:

[0299] (a) providing an embedding starting point to a system comprising a supervised model for predicting a function of a biopolymer sequence and a decoder network, optionally wherein the starting point is an embedding seed biopolymer sequence, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a probabilistic biopolymer sequence given the embedding of the biopolymer sequence in the function space;

[0300] (b) computing the change in the embedding of the function relative to the starting point according to the step size, thereby providing a first update point in the function space;

[0301] (c) optionally computing a change in the function relative to the embedding at a first update point in the function space, and optionally iteratively computing a change in the function relative to the embedding at further update points;

[0302] (d) providing the first update point, or optionally the iterative further update point, in the function space to the decoder network when the first update point, or optionally the iterative further update point, in the function space approaches the desired function level; and

[0303] (e) Obtaining probabilistically improved biopolymer sequences from the decoder.

[0304] Illustrative Embodiment 5: A system comprising a processor and a non-transitory and / or non-transitory computer readable medium comprising instructions that, when executed by the processor, cause the processor to:

[0305] (a) providing an embedding starting point to a system comprising a supervised model network that predicts a function of a biopolymer sequence and a decoder network, optionally wherein the starting point is an embedding seed biopolymer sequence, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a predicted probabilistic biopolymer sequence given the embedding of the predicted biopolymer sequence in the function space;

[0306] (b) predicting the function of the embedding starting point;

[0307] (c) computing the change in the embedding of the function relative to the starting point according to the step size, thereby providing a first update point in the function space;

[0308] (d) providing a first updated point in the function space to the decoder network to provide a first intermediate probabilistic biopolymer sequence;

[0309] (e) providing the first intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the first intermediate probabilistic biopolymer sequence,

[0310] (f) computing a change in the embedding of the function relative to the first update point in the function space to provide an updated point in the function space;

[0311] (g) providing updated points in the function space to the decoder network to provide additional intermediate probabilistic biopolymer sequences;

[0312] (h) providing the additional intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the additional intermediate probabilistic biopolymer sequence;

[0313] (i) then calculating a change in the embedding of the function relative to the first further update point in the function space to provide a further update point in the function space, optionally iterating steps (g)-(i), wherein the further update point in the function space referenced in step (i) is taken to be the further update point in the function space in step (g); and

[0314] (j) providing the embedded point to the decoder network when approaching a desired level of functionality in the functionality space; and obtaining a probabilistically improved biopolymer sequence from the decoder.

[0315] Illustrative Embodiment 6: A non-transitory and / or non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to:

[0316] (a) providing an embedding starting point to a system comprising a supervised model network that predicts a function of a biopolymer sequence and a decoder network, optionally wherein the starting point is an embedding seed biopolymer sequence, the supervised model network comprising an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function, and the decoder network trained to provide a predicted probabilistic biopolymer sequence given the embedding of the predicted biopolymer sequence in the function space;

[0317] (b) predicting the function of the embedding starting point;

[0318] (c) computing the change in the embedding of the function relative to the starting point according to the step size, thereby providing a first update point in the function space;

[0319] (d) providing a first updated point in the function space to the decoder network to provide a first intermediate probabilistic biopolymer sequence;

[0320] (e) providing the first intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the first intermediate probabilistic biopolymer sequence,

[0321] (f) computing a change in the embedding of the function relative to the first update point in the function space to provide an updated point in the function space;

[0322] (g) providing updated points in the function space to the decoder network to provide additional intermediate probabilistic biopolymer sequences;

[0323] (h) providing the additional intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the additional intermediate probabilistic biopolymer sequence;

[0324] (i) then calculating a change in the embedding of the function relative to the first further update point in the function space to provide a further update point in the function space, optionally iterating steps (g)-(i), wherein the further update point in the function space referenced in step (i) is taken to be the further update point in the function space in step (g); and

[0325] (j) providing the embedded point to the decoder network when approaching a desired level of functionality in the functionality space; and obtaining a probabilistically improved biopolymer sequence from the decoder. Sequence Listing <110> Flagship Pioneering InnovationsVI, LLC Feala, Jacob D. Beam, Andrew L. Gibson, Molly K. Cabral, Bernard J. <120> Machine learning-guided peptide design <130> 5708.1007-002 <150> US 62 / 882,150 <151> 2019-08-02 <150> US 62 / 882,159 <151> 2019-08-02 <160> 2 <170> PatentIn version 3.5 <210> 1 <211> 238 <212> PRT <213> Aequorea victoria <400> 1 Met Ser Lys Gly Glu Glu Leu Phe Thr Gly Val Val Pro Ile Leu Val 1 5 10 15 Glu Leu Asp Gly Asp Val Asn Gly His Lys Phe Ser Val Ser Gly Glu 20 25 30 Gly Glu Gly Asp Ala Thr Tyr Gly Lys Leu Thr Leu Lys Phe Ile Cys 35 40 45 Thr Thr Gly Lys Leu Pro Val Pro Trp Pro Thr Leu Val Thr Thr Phe 50 55 60 Ser Tyr Gly Val Gln Cys Phe Ser Arg Tyr Pro Asp His Met Lys Gln 65 70 75 80 His Asp Phe Phe Lys Ser Ala Met Pro Glu Gly Tyr Val Gln Glu Arg 85 90 95 Thr Ile Phe Phe Lys Asp Asp Gly Asn Tyr Lys Thr Arg Ala Glu Val 100 105 110 Lys Phe Glu Gly Asp Thr Leu Val Asn Arg Ile Glu Leu Lys Gly Ile 115 120 125 Asp Phe Lys Glu Asp Gly Asn Ile Leu Gly His Lys Leu Glu Tyr Asn 130 135 140 Tyr Asn Ser His Asn Val Tyr Ile Met Ala Asp Lys Gln Lys Asn Gly 145 150 155 160 Ile Lys Val Asn Phe Lys Ile Arg His Asn Ile Glu Asp Gly Ser Val 165 170 175 Gln Leu Ala Asp His Tyr Gln Gln Asn Thr Pro Ile Gly Asp Gly Pro 180 185 190 Val Leu Leu Pro Asp Asn His Tyr Leu Ser Thr Gln Ser Ala Leu Ser 195 200 205 Lys Asp Pro Asn Glu Lys Arg Asp His Met Val Leu Leu Glu Phe Val 210 215 220 Thr Ala Ala Gly Ile Thr His Gly Met Asp Glu Leu Tyr Lys 225 230 235 <210> 2 <211> 238 <212> PRT <213> Artificial sequence <220> <223> Synthetic <400> 2 Met Ser Lys Gly Glu Glu Leu Phe Thr Gly Val Val Pro Ile Leu Val 1 5 10 15 Glu Leu Asp Gly Asp Val Asn Gly His Lys Phe Ser Val Ser Gly Glu 20 25 30 Gly Glu Gly Asp Ala Thr Cys Gly Lys Leu Thr Leu Lys Phe Ile Cys 35 40 45 Thr Thr Gly Lys Leu Pro Val Pro Trp Pro Thr Leu Val Thr Thr Leu 50 55 60 Ser Tyr Gly Met Gln Cys Phe Ser Arg Tyr Pro Asp His Met Lys Gln 65 70 75 80 His Asp Phe Phe Lys Ser Ala Met Pro Glu Gly Tyr Val Gln Glu Arg 85 90 95 Thr Ile Phe Phe Lys Asp Asp Gly Asn Tyr Lys Thr Arg Ala Glu Val 100 105 110 Lys Phe Glu Gly Asp Thr Leu Val Asn Arg Ile Glu Leu Lys Gly Ile 115 120 125 Gly Phe Lys Glu Asp Gly Asn Ile Leu Gly His Lys Leu Glu Tyr Asn 130 135 140 Tyr Asn Ser His Asn Val Tyr Ile Met Ala Asp Lys Gln Lys Asn Gly 145 150 155 160 Ile Lys Ala Asn Phe Arg Ile Arg His Asn Ile Glu Asp Gly Ser Val 165 170 175 Gln Leu Ala Asp His Tyr Gln Gln Asn Thr Pro Ile Gly Asp Val Pro 180 185 190 Val Leu Leu Pro Asp Asn His Tyr Leu Ser Thr Gln Ser Ala Leu Ser 195 200 205 Lys Asp Pro Asn Glu Lys Arg Asp His Met Val Leu Leu Glu Phe Val 210 215 220 Thr Ala Ala Gly Ile Thr His Gly Met Asp Glu Leu Tyr Lys 225 230 235

Claims

1. A method for engineering an improved biopolymer sequence through functional evaluation, the method comprising: (a) providing an embedding starting point, the starting point being an embedding seed biopolymer sequence, the starting point being provided to a system comprising (i) a supervised model of a prediction function and (ii) a decoder network, the function being a function of the biopolymer sequence, wherein: The supervised model includes an encoder network that provides an embedding of the biopolymer sequence in a function space representing the function. The decoder network is trained to provide a probabilistic biopolymer sequence given the biopolymer sequence is embedded in the function space; and The embedding is a continuously differentiable function space that represents the function and has one or more gradients; (b) computing a change of the function relative to the embedding at the starting point according to a step size, the computed change providing a first update point in the function space, wherein the change of the function relative to the embedding is computed by computing a change of the function relative to the encoder network, then computing a change of the encoder network relative to a change of the decoder network, and computing a change of the decoder network relative to the embedding; (c) providing the decoder network with a first update point in the function space when the desired function level is achieved within a specified threshold at the first update point; and (d) obtaining a probabilistically improved biopolymer sequence from the decoder network, wherein the functionality of the probabilistically improved biopolymer sequence is improved compared to the functionality of the seed biopolymer sequence.

2. The method of claim 1, further comprising: computing a second change of the feature relative to the embedding at the first update point in the feature space; as well as The process of iteratively computing the second change of the function relative to the embedding at further update points.

3. The method according to claim 2, wherein: Providing the first update point may be performed when the required functionality level is achieved within said certain threshold at an optionally iterative further update point, and providing the further update point comprises providing the iterative further update point to the decoder network.

4. The method according to claim 1, wherein Computing the change of the function with respect to the embedding includes taking the derivative of the function with respect to the embedding.

5. The method according to claim 1, wherein The function is a composite function of two or more composite functions.

6. The method according to claim 5, wherein: The composite function is a weighted sum of the two or more composite functions.

7. The method of claim 1, wherein: Two or more starting points in this embedding are used simultaneously.

8. The method of claim 1, wherein: During sampling, correlations between residues in a probabilistic biopolymer sequence, including a probability distribution of residue identities, are taken into account using conditional probabilities taking into account portions of the probabilistic biopolymer sequence that have been generated.

9. The method of claim 1, further comprising selecting a maximum likelihood improved biopolymer sequence from among the probabilistic biopolymer sequences comprising a probability distribution of residue identities.

10. The method of claim 1, further comprising sampling a marginal distribution at each residue of a probabilistic biopolymer sequence comprising a probability distribution of residue identities.

11. The method of claim 1 , further comprising: providing a first updated point in the function space or a further updated point in the function space to the decoder network to provide an intermediate probabilistic biopolymer sequence, providing the intermediate probabilistic biopolymer sequence to the supervised model to predict a function of the intermediate probabilistic biopolymer sequence, and The change in the function relative to the embedding of the intermediate probabilistic biopolymer is calculated to provide further updated points in the function space.

Images