Structure-based depth generative model for binding site descriptor extraction and re-molecule generation

Abstract

In some aspects, the disclosure describes a method of sampling a ligand. In some embodiments, the method includes receiving a target descriptor. In some embodiments, the method includes generating a potential descriptor in an engineered chemistry space based at least in part on a target descriptor. In some embodiments, the method includes generating a ligand descriptor based at least in part on a potential descriptor.

Description

[0001] Cross-references

[0002] This application claims the benefit of U.S. Provisional Application No. 63 / 315,288, filed on March 1, 2022, which is incorporated herein by reference in its entirety. Background Art

[0003] Deep generative learning can refer to a method of training a machine learning model to roughly estimate the distribution of a given data set while enabling the generation of coherent samples from the data distribution. A machine learning model trained in this way can be called a deep generative model.

[0004] Deep generative learning can be applied to chemical datasets to create deep generative models for chemistry, allowing the generation of coherent molecular structures that are not present in a given training dataset of chemical structures. Summary of the invention

[0005] In some aspects, the present disclosure describes a method for sampling ligands, comprising: receiving a target descriptor; generating a potential descriptor in an engineered chemical space based at least in part on the target descriptor; and generating a ligand descriptor based at least in part on the potential descriptor; wherein the engineered chemical space comprises a unified embedding for at least a plurality of target potential descriptors and a plurality of ligand potential descriptors.

[0006] In some embodiments, in the engineered chemical space, a given target potential descriptor for a given target has similarity to a given ligand potential descriptor for a given ligand when the given ligand targets the given target and has dissimilarity when the given ligand does not target the given target.

[0007] In some embodiments, a given target potential descriptor is different from a given ligand potential descriptor.

[0008] In some embodiments, the engineered chemical space is organized at least in part based on spatial information of multiple ligands, multiple targets, or both.

[0009] In some implementations, similarity or dissimilarity may be measured using a similarity function.

[0010] In some embodiments, the similarity function includes a distance-based similarity function, an angle-based similarity function, a set-based similarity function, or any combination thereof.

[0011] In some embodiments, the angle-based similarity function is a cosine similarity function.

[0012] In some embodiments, the method is performed at least in part by using a neural network.

[0013] In some embodiments, the generating in (b) is performed at least in part by using an encoder.

[0014] In some embodiments, the generating in (c) is performed at least in part by using a decoder.

[0015] In some embodiments, the neural network includes a bottleneck architecture that includes an encoder and a decoder.

[0016] In some embodiments, the bottleneck architecture includes a U-net.

[0017] In some embodiments, the decoder includes a language model.

[0018] In some embodiments, the language model is a long short-term memory model (LSTM).

[0019] In some embodiments, at least one hidden unit of the LSTM is initialized with a latent descriptor.

[0020] In some embodiments, the neural network includes a convolutional layer.

[0021] In some embodiments, the neural network includes densely connected layers.

[0022] In some embodiments, the neural network includes skip connections.

[0023] In some embodiments, generating latent descriptors is performed using at least an encoder.

[0024] In some embodiments, at least the encoder is trained at least in part by using contrastive learning.

[0025] In some embodiments, contrastive learning is based at least in part on a contrastive loss calculated between pairs of target latent descriptors and ligand latent descriptors in the training data.

[0026] In some embodiments, the contrast loss includes at least: L(p,f) = [α - sim(p,f) + sim(p,f')] + +[α-sim(p,f)+sim(p',f)] + , where L(p,f) refers to the contrastive loss, α is the margin parameter, p is the first target latent descriptor in the training data, f is the first ligand latent descriptor in the training data, p' is the second target latent descriptor in the training data that is at least similar to the first ligand latent descriptor, f' is the second ligand latent descriptor in the training data that is at least similar to the first target latent descriptor, sim is the similarity function calculated between its brackets, and [x] + ≡max(x,0).

[0027] In some embodiments, generating ligand descriptors is performed stochastically.

[0028] In some embodiments, the language model is trained based at least in part on a first reconstruction loss.

[0029] In some embodiments, the first reconstruction loss is a probabilistic loss function.

[0030] In some embodiments, the first reconstruction loss includes at least: L(X f ,Y f )=-∑ i n Y f,i ·log(X f,i ), where L(X f ,Y f ) refers to the first reconstruction loss, where -∑ i n refers to the negative sum of each element i among the total n elements in the operand, where X f,i is the ligand identifier element in the training data, and where Y f,i is the ligand identifier element reconstructed by the language model.

[0031] In some embodiments, a ligand descriptor comprises a ligand identifier.

[0032] In some embodiments, the ligand identifier comprises a textual identifier.

[0033] In some embodiments, the text identifier comprises SMILES, InChI, or SELFIES.

[0034] In some embodiments, the ligand identifier comprises a molecular adjacency matrix or a molecular graph.

[0035] In some embodiments, the bottleneck architecture is trained based at least in part on a second reconstruction loss that is based at least in part on a target-ligand interaction.

[0036] In some embodiments, the target-ligand interaction is ligand occupancy.

[0037] In some embodiments, the second reconstruction loss includes at least: L(X o ,Y o )=||X o -Y o || 2 2, where L(X o ,Y o ) refers to the second reconstruction loss, where X o is the ligand occupancy descriptor in the training data, and where Yo Reconstruction of ligand occupancy descriptors by autoencoder.

[0038] In some embodiments, the target descriptor comprises a protein target descriptor.

[0039] In some embodiments, the target descriptor comprises a protein pocket descriptor.

[0040] In some embodiments, the target descriptors include features describing at least one of the following: hydrophobicity, aromaticity, hydrogen bond acceptance, hydrogen bond donation, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction capability, hydrophobic interaction capability, π-π interaction capability, and halogen interaction capability.

[0041] In some embodiments, a protein pocket descriptor includes a spatial map of one or more protein targets.

[0042] In some embodiments, the spatial pattern is a grid.

[0043] In some embodiments, the spatial pattern is a rectangular grid.

[0044] In some embodiments, the spatial map includes at least 3 dimensions.

[0045] In some embodiments, the spatial pattern comprises a resolution of less than about 3.5 angstroms.

[0046] In some aspects, the present disclosure describes a method for machine learning an engineered chemical space, comprising: providing a neural network comprising: an input layer configured to receive at least a target descriptor; a latent layer configured to output at least a potential descriptor, wherein the latent layer is connected to the input layer; an output layer configured to output at least a target-ligand interaction descriptor, wherein the output layer is connected to the latent layer; at least one parameter; providing training data comprising a plurality of target descriptors, a plurality of target-ligand interaction descriptors, and a plurality of ligand potential descriptors; training the neural network by: (i) inputting at least a plurality of target descriptors at an input layer of the neural network; (ii) outputting a plurality of output potential descriptors at the latent layer and outputting a plurality of output target-ligand interaction descriptors at an output layer; and (iii) optimizing a plurality of loss functions based at least in part on the plurality of output potential descriptors and the plurality of output target-ligand interaction descriptors by updating at least one parameter of the neural network, such that the neural network learns an engineered chemical space comprising a unified embedding for at least a plurality of target descriptors and a plurality of ligand potential descriptors.

[0047] In some embodiments, the engineered chemical space is organized at least in part based on spatial information of a plurality of target descriptors and a plurality of ligand potential descriptors.

[0048] In some embodiments, in the engineered chemical space, a given target potential descriptor for a given target has similarity to a given ligand potential descriptor for a given ligand when the given ligand targets the given target and has dissimilarity when the given ligand does not target the given target.

[0049] In some embodiments, a given target potential descriptor is different from a given ligand potential descriptor.

[0050] In some embodiments, similarity or dissimilarity may be measured using a similarity function.

[0051] In some embodiments, the plurality of loss functions include contrastive loss functions calculated between a plurality of pairs formed between a plurality of target latent descriptors and a plurality of ligand latent descriptors output by the latent layer.

[0052] In some embodiments, training the neural network includes optimizing a loss function comprising a first reconstruction loss function calculated between a plurality of target-ligand interaction descriptors of the training data and a plurality of reconstructed target-ligand interaction descriptors output by the first output layer.

[0053] In some embodiments, the neural network further comprises a second output layer configured to output at least a ligand descriptor, wherein the second output layer is connected to the latent layer.

[0054] In some embodiments, the training data further includes a plurality of ligand descriptors.

[0055] In some embodiments, training the neural network further comprises outputting a plurality of output ligand descriptors, and optimizing a loss function comprising a second reconstruction loss function calculated between the plurality of ligand descriptors of the training data and the plurality of output target-ligand interaction descriptors output by the second output layer.

[0056] In some embodiments, the plurality of target descriptors of the training data include features describing at least one of: hydrophobicity, aromaticity, hydrogen bond acceptance, hydrogen bond donation, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction capability, hydrophobic interaction capability, π-π interaction capability, and halogen interaction capability.

[0057] In some embodiments, the plurality of target-ligand interaction descriptors in the training data includes ligand occupancy for a given target.

[0058] In some embodiments, the plurality of potential descriptors for ligands in the training data comprises a SMILES for a given ligand.

[0059] In some embodiments, the neural network further includes a plurality of hidden layers connecting at least two of the following: the input layer, the latent layer, and the first output layer and the second output layer.

[0060] In some aspects, the present disclosure describes a computer-implemented method of implementing any of the methods disclosed herein in a computer.

[0061] In some aspects, the present disclosure describes a computer program product comprising a computer-readable medium having computer-executable code encoded therein, the computer-executable code being adapted to be executed to implement any of the methods disclosed herein.

[0062] In some aspects, the present disclosure describes a non-transitory computer-readable storage medium encoded with a computer program comprising instructions executable by one or more processors to implement any of the methods disclosed herein.

[0063] In some aspects, the present disclosure describes a computer-implemented system comprising: a digital processing device including at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program comprising instructions executable by the digital processing device to implement any of the methods disclosed herein.

[0064] Incorporation by reference

[0065] All publications, patents, and patent applications mentioned in this specification are incorporated herein 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. In the event that the publications and patents or patent applications incorporated by reference conflict with the disclosure contained in this specification, this specification is intended to supersede and / or take precedence over any such conflicting material. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The novel features of the present disclosure are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present disclosure will be obtained by reference to the following detailed description of illustrative embodiments in which the principles of the present disclosure are utilized, and the accompanying drawings, in which:

[0067] Figure 1Pharmacophore property channels of a data set according to some embodiments are shown. A protein-ligand pair is represented as a 4D tensor having 13 pharmacophore property channels: 'hydrophobicity', 'aromaticity', 'hydrogen bond_acceptor', 'hydrogen bond_donor', 'positive_ionizable', 'negative_ionizable', 'metal', 'pocket occupancy', 'residual hydrogen bond interaction', 'residual hydrophobic interaction', 'residual π-π interaction', 'residual halogen interaction', and 'ligand occupancy'. The upper left figure shows channel 1 ('hydrophobicity'). The upper right figure shows channel 8 ('pocket occupancy'). The lower left figure shows channel 9 ('residual hydrogen bond interaction') in gray and channel 10 ('residual hydrophobic interaction') in white. The lower right figure shows channel 13 ('ligand occupancy').

[0068] Figure 2 Schematically illustrates a neural network architecture according to some embodiments. The neural network includes a U-Net and an LSTM. The output from the bottleneck layer of the U-Net (latent descriptor) is passed through 6 linear dense layers to serve as the initial memory nodes of a language model (e.g., LSTM) to guide SMILES generation towards ligands with structures and / or properties that fit the input pocket. The neural network architecture allows learning two tasks together, such as (i) outputting ligand occupancy from the U-Net decoder and (ii) SMILES from the LSTM to create a more efficient latent embedding.

[0069] Figure 3 Some concepts behind contrastive learning according to some embodiments are shown. Two encoders can be used to learn a joint embedding (e.g., in a multimodal space) for 'image' and 'string' pairs. The decoder can be used to interpret potential descriptors from the joint embedding space while incorporating new 'images' to generate new content. For example, the encoder can be used to create a potential descriptor from an image, and the decoder can be used to generate a string describing the image (e.g., a descriptive sentence). In another example, the encoder can be used to create a potential descriptor from a string (e.g., a sentence), and the decoder can be used to generate an image based on the string.

[0070] Figure 4The reasoning method according to some embodiments is shown. The reasoning method shown in the upper figure receives the target descriptor of the target as input, and outputs the ligand occupancy of the target and / or the ligand identifier of the candidate ligand predicted to be combined with the target. The reasoning method shown in the middle figure receives the target descriptor of the target as input to output the potential descriptor in the engineered chemical space, which can be navigated (e.g., by adding noise) to select different potential descriptors, and wherein the potential descriptor is used to output the candidate ligand predicted to be combined with the target. The reasoning method shown in the lower figure receives the ligand identifier as input to output the potential descriptor in the engineered chemical space, which can be navigated to select different potential descriptors, and wherein the potential descriptor is used to output a new candidate ligand, which can target a target similar to the ligand of the input ligand descriptor.

[0071] Figure 5 A non-limiting example of a computing device is shown; in this case, the computing device refers to a device having one or more processors, memory, storage, and a network interface. DETAILED DESCRIPTION

[0072] Deep generative learning can be applied to chemical datasets to create deep generative models for chemistry (which may be synonymously referred to herein as generative chemical models) that allow the generation of coherent chemical descriptors that do not exist in a given training chemical dataset. Generative chemical models can be used to obtain the potential information contained in a chemical dataset to provide chemical descriptors corresponding to substances with a specific set of properties or a specific set of purposes. For example, SMILES identifiers can be generated for drugs targeting specific proteins. In another example, chemical formulas for materials with a set of appropriate properties for industrial applications can be generated.

[0073] Considerations for creating generative chemical models involve designing engineered chemical spaces (which may be synonymously referred to herein as latent spaces). Some machine learning models involve using techniques to compress information supported in a high-dimensional space into a low-dimensional space. A simple example of this technique involves training a neural network including a bottleneck layer to automatically encode chemical structures, where the bottleneck layer has a lower number of dimensions than the input or output of the neural network.

[0074] In some aspects, the present disclosure describes systems and methods for designing engineered chemical spaces in a multi-task learning framework. When at least two chemical information sources are incorporated into a unified embedding space, the system and method of the present disclosure can be useful. At least two chemical information sources can be embedded so that they all include similar embedding manifolds, however, individual embedding points of a given data set can be embedded in different coordinates. For example, the embedding of a protein target and a ligand occupying the protein target can be embedded so that the corresponding protein and ligand pairs are closely but not necessarily embedded in the embedding space at the same coordinates. Therefore, the embedding for a protein pocket can be close to multiple embeddings of a ligand that can target a protein pocket; and the embedding of a ligand can be close to multiple protein pockets that a ligand can target. The engineered chemical space can be constructed so that the large-scale structure (e.g., topological structure) of the protein pocket embedding and the ligand embedding can be similar but not identical. In some embodiments, this is achieved by using a contrastive learning method.

[0075] When a task may be partially related to another task, the system and method of the present disclosure can be useful when training a neural network with at least two tasks. For example, the task of generating a spatial grid of a ligand occupying a protein pocket may be partially related to the task of generating a SMILES string of a ligand. The information related to generating the spatial grid may contain more information than the identifier required information for generating the ligand, and vice versa. The neural network can be constructed so that the information related to generating the spatial grid is partially incorporated into the generation SMILES string. The neural network can be constructed so that the information related to generating the SMILES string is partially incorporated into the generation spatial grid. In addition, the engineered chemical space can be constructed so that it includes both spatial representation (for example, the occupancy of the ligand to the protein pocket) and text representation (for example, the ligand SMILES). In some embodiments, this can be achieved by using a U-net neural network.

[0076] Engineering chemical space

[0077] In some embodiments, the engineered chemical space can be formed using dimensionality reduction. The terms reduction, dimensionality reduction, projection, component analysis, feature space reduction, latent space engineering, feature space engineering, representation engineering, or latent space embedding can refer to a method of transforming given data with an initial number of dimensions into another form of data with fewer dimensions than the initial number of dimensions. In some embodiments, the term can refer to the principle of reducing a set of input dimensions to a smaller set of output dimensions.

[0078] In some embodiments, the engineered chemical space includes a unified embedding for at least multiple target descriptors and multiple ligand descriptors. In some embodiments, the engineered chemical space uses contrastive learning to train to create a unified embedding for at least multiple target descriptors and multiple ligand descriptors. In some embodiments, in the engineered chemical space, the given target potential descriptor of a given target has similarities with the given ligand potential descriptor of a given ligand when the given ligand targets the given target, and has dissimilarity when the given ligand does not target the given target. In some embodiments, the engineered chemical space is at least partially organized based on the spatial information of multiple ligands, multiple targets, or both thereof. In some embodiments, the given target potential descriptor is different from the given ligand potential descriptor.

[0079] In some embodiments, similarity or dissimilarity can be measured using a similarity function. In some embodiments, the similarity function includes a distance-based similarity function, an angle-based similarity function, a set-based similarity function, or any combination thereof. In some embodiments, the angle-based similarity function is a cosine similarity function. In some embodiments, the distance-based similarity function can be based at least in part on Euclidean distance. In some embodiments, the set-based similarity function can be a clustering function.

[0080] In some embodiments, in the engineered chemical space, multiple given ligand potential descriptors for multiple given ligands can form unique clusters or can be close to each other when multiple given ligands target a given target or a group of similar targets. In some embodiments, in the engineered chemical space, multiple given target potential descriptors for multiple given targets can form unique clusters or can be close to each other when multiple targets can be targeted by a given ligand or a group of similar ligands.

[0081] In some embodiments, the embedding for multiple target descriptors is a transformation of data from one representation to another. In some embodiments, the embedding for multiple ligand descriptors is a transformation of data from one representation to another. In some embodiments, the transformation can transform one data form into another data form with a smaller dimension. In some embodiments, the transformation can include at least partially linearizing one or more meanders in the data in the high-dimensional space. In some embodiments, the transformation can be performed on data including data in Euclidean space. In some embodiments, the transformation can be performed on data including data in a graphic space. In some embodiments, the transformation can be performed on data in a discrete space. In some embodiments, the transformation can be performed on data including data in a frequency space. In some embodiments, the transformation can transform data as follows: discrete space to continuous space, continuous space to discrete space, graphic space to continuous space, continuous space to graphic space, graphic space to discrete space, discrete space to graphic space, or any combination thereof. In some embodiments, the transformation can include transforming data in a discrete space into a frequency domain. In some embodiments, the transformation can include transforming data in a continuous space into a frequency domain. In some embodiments, the transformation can include transforming data in a graphic space into a frequency domain.

[0082] In some embodiments, the method or system of the present disclosure may use clustering. The term clustering or cluster analysis may refer to a method of grouping samples in a data set by a certain measure. Samples may be grouped in a set space, for example, element 'a' is grouped in a set 'A'. Samples may be grouped in a continuous space, for example, element 'a' is a point in Euclidean space within a distance 'l' from the centroid of the elements comprising cluster 'A'. Samples may be grouped in a graphical space, for example, element 'a' is highly correlated with the elements comprising cluster 'A'. These terms may refer to the principle of organizing multiple elements into groups in a mathematical space based on a certain measure.

[0083] Clustering may include grouping any number of samples in a data set by a certain measure. In some embodiments, clustering may include K-means clustering. In some embodiments, clustering may include hierarchical clustering. In some embodiments, clustering may include using a random forest model. In some embodiments, clustering may include a boosted tree model. In some embodiments, clustering may include using a support vector machine. In some embodiments, clustering may include calculating one or more N-1 dimensional surfaces in an N-dimensional space that divides a data set into clusters. In some embodiments, clustering may include distribution-based clustering. In some embodiments, clustering may include fitting multiple prior distributions to data distributed in an N-dimensional space. In some embodiments, clustering may include using density-based clustering. In some embodiments, clustering may include using fuzzy clustering. In some embodiments, clustering may include calculating probability values ​​of data points belonging to a cluster. In some embodiments, clustering may include using restrictions. In some embodiments, clustering may include using supervised learning. In some embodiments, clustering may include using unsupervised learning.

[0084] In some embodiments, clustering may include grouping samples based on similarity. In some embodiments, clustering may include grouping samples based on quantitative similarity. In some embodiments, clustering may include grouping samples based on one or more features of each sample. In some embodiments, clustering may include grouping samples based on one or more tags of each sample. In some embodiments, clustering may include grouping samples based on Euclidean coordinates. In some embodiments, clustering may include grouping samples based on features of nodes and edges of each sample.

[0085] In some embodiments, different clusters can be completely independent, such that different clusters do not share elements (e.g., two clusters are mutually exclusive). In some embodiments, different clusters can share one or more elements and at the same time not share one or more elements (e.g., two clusters can partially overlap). In some embodiments, a cluster can be surrounded by another cluster, such that all elements in one cluster are included in the other cluster.

[0086] In some embodiments, the comparison may include a comparison between a first group and a different second group. In some embodiments, the first group or the second group may be a cluster independently of each other. In some embodiments, the first group or the second group may be a group of clusters independently of each other. In some embodiments, the comparison may include a comparison between a cluster and a group of clusters. In some embodiments, the comparison may include a comparison between a first group of clusters and a second group of clusters different from the first group. In some embodiments, a group may be a sample. In some embodiments, a group may be a group of samples. In some embodiments, the comparison may include a comparison between a sample and a group of samples. In some embodiments, the comparison may include a comparison between a group of samples and a group of samples.

[0087] Various machine learning techniques

[0088] In some cases, machine learning (ML) may generally involve identifying and recognizing patterns in existing data in order to facilitate making predictions for subsequent data. ML may include an ML model (e.g., which may include an ML algorithm). Machine learning (whether analytical or statistical in nature) may provide deductive or abductive reasoning based on actual or simulated data. The ML model may be a trained model. ML techniques may include one or more supervised, semi-supervised, self-supervised, or unsupervised ML techniques. For example, an ML model may be a trained model trained by supervised learning (e.g., various parameters are determined as weights or proportional coefficients). ML may include one or more of regression analysis, regularization, classification, dimensionality reduction, ensemble learning, meta-learning, association rule learning, cluster analysis, anomaly detection, deep learning, or ultra-deep learning. ML may include (but is not limited to): k-means, k-means clustering, k-nearest neighbors, learning vector quantization, linear regression, nonlinear regression, least squares regression, partial least squares regression, logistic regression, stepwise regression, multivariate adaptive regression splines, ridge regression, principal component regression, least absolute convergence and selection operator (LASSO), least angle regression, canonical correlation analysis, factor analysis, independent component analysis, linear discriminant analysis, multivariate scaling, non-negative matrix factorization, principal component analysis, principal coordinate analysis, projection pursuit, Sammon mapping, t-distributed random neighbor embedding, adaptive boosting (AdaBoosting), boosting, gradient boosting, bootstrap redrawing, ensemble averaging, decision tree, conditional decision tree, boosted decision tree, gradient boosted decision tree, random forest, stacked generalization (stacked generalization), Bayesian networks, Bayesian belief networks, Naive Bayes, Gaussian Naive Bayes, Multinomial Naive Bayes, Hidden Markov Model, Hierarchical Hidden Markov Model, Support Vector Machine, Encoder, Decoder, Autoencoder, Stacked Autoencoder, Perceptron, Multilayer Perceptron, Artificial Neural Network, Feedforward Neural Network, Convolutional Neural Network, Recurrent Neural Network, Long Short-Term Memory, Deep Belief Network, Deep Boltzmann Machine, Deep Convolutional Neural Network, Deep Recurrent Neural Network, or Generative Adversarial Network.

[0089] In some cases, training the ML model may include selecting one or more untrained data models to train using the training data set. The selected untrained data models may include any type of untrained ML model for supervised, semi-supervised, self-supervised, or unsupervised machine learning. The selected untrained data model is specified based on specifying relevant parameters as inputs (e.g., user inputs) for use as predicted variables or specifying other variables as potential explanatory variables. For example, the selected untrained data model may be specified to generate outputs (e.g., predictions) based on the inputs. Conditions for training the ML model from the selected untrained data model may also be selected, such as restrictions on ML model complexity or restrictions on ML model refinement beyond a certain point. The ML model may be trained (e.g., via a computer system such as a server) using the training data set. In some cases, a first subset of the training data set may be selected to train the ML model. Based on the type of the selected ML model and any conditions specified for training the ML model, the selected untrained data model may then be trained according to the first subset of the training data set using an appropriate ML technique. In some cases, due to the processing power requirements of training the ML model, the selected untrained data model may be trained using additional computing resources (e.g., cloud computing resources). In some cases, such training may continue until at least one aspect of the ML model is validated and meets selection criteria for use as a predictive model.

[0090] In some cases, one or more aspects of the ML model may be validated using a second subset of the training data set (e.g., different from the first subset of the training data set) to determine the accuracy and robustness of the ML model. Such validation may include applying the ML model to the second subset of the training data set to make predictions derived from the second subset of the training data set. The ML model may then be evaluated to determine whether the performance is sufficient based on the derived predictions. The adequacy criteria applied to the ML model may vary depending on the size of the training data set available for training, the performance of previous iterations of the trained model, or user-specified performance requirements. If the ML model does not achieve sufficient performance, additional training may be performed. The additional training may include refinement of the ML model or retraining on a different first subset of the training data set, after which the new ML model may be validated again and assessed. When the ML model has achieved sufficient performance, in some cases, the ML may be stored for present or future use. The ML model may be stored as a set of parameter values ​​or weights for use in analyzing further inputs (e.g., further relevant parameters used as further predictive variables, further explanatory variables, further user interaction data, etc.), which in some cases may also include analysis logic or indications of model effectiveness. In some cases, multiple ML models may be stored for use in generating predictions under different sets of input data conditions. In some embodiments, the ML models may be stored in a database (e.g., a database associated with a server).

[0091] Neural Network Architecture

[0092] In some embodiments, the systems and methods of the present disclosure may include a neural network or include the use of a neural network. The neural network may include various architectures, loss functions, optimization algorithms, assumptions, and various other neural network design choices. In some embodiments, the neural network includes an encoder. In some embodiments, the neural network includes a decoder. In some embodiments, the neural network includes a bottleneck architecture, which includes an encoder and a decoder. In some embodiments, the bottleneck architecture includes a U-net. In some embodiments, the bottleneck architecture includes an autoencoder.

[0093] In some embodiments, the neural network includes a language model. In some embodiments, the language model is a long short-term memory model (LSTM). In some embodiments, the language model is a convolutional neural network. In some embodiments, the language model is an autoregressive model. As used herein, a language model may refer to any neural network or algorithm configured to output a semantically correct and interpretable representation of a chemical substance, a material, or both.

[0094] In some embodiments, the neural network includes a convolutional layer. In some embodiments, the neural network includes a densely connected layer. In some embodiments, the neural network includes a skip connection. In some embodiments, the neural network may include a graph convolutional layer. In some embodiments, the neural network may include a message passing layer. In some embodiments, the neural network may include an attention layer. In some embodiments, the neural network may include a recurrent layer. In some embodiments, the neural network may include a gated recurrent unit. In some embodiments, the neural network may include a reversible layer. In some embodiments, the neural network may include a neural network with a bottleneck layer. In some embodiments, the neural network may include a residual block. In some embodiments, the neural network may include one or more dropout layers. In some embodiments, the neural network may include one or more batch normalization layers. In some embodiments, the neural network may include one or more pooling layers. In some embodiments, the neural network may include one or more upsampling layers. In some embodiments, the neural network may include one or more maximum pooling layers. Various types of layers may be used for the neural network without departing from the inventive concepts disclosed herein.

[0095] In some embodiments, the neural network includes a graphical model. In some embodiments, graphs, graphical models, and graphical models may refer to methods of modeling data in a graphical representation including nodes and edges. In some embodiments, the data may be stored in various forms and alternative forms (such as linked lists, dictionaries, spreadsheets, arrays), in permanent storage, temporary storage, etc., and are not limited to the specific embodiments disclosed herein.

[0096] In some embodiments, the neural network may include an autoencoder. In some embodiments, the neural network may include a variational autoencoder. In some embodiments, the neural network may include a generative adversarial network. In some embodiments, the neural network may include a flow model. In some embodiments, the neural network may include an autoregressive model.

[0097] The neural network may include various activation functions. In some embodiments, the activation function may be non-linear. In some embodiments, the neural network may include one or more activation functions. In some embodiments, the neural network may include ReLU, softmax, tanh, sigmoid, softplus, softsign, selu, elu, exponential, LeakyReLU, or any combination thereof. Various activation functions may be used with the neural network without departing from the inventive concepts disclosed herein.

[0098] In some embodiments, the neural network may include a regression loss function. In some embodiments, the neural network may include a logistic loss function. In some embodiments, the neural network may include a variational loss. In some embodiments, the neural network may include a prior. In some embodiments, the neural network may include a Gaussian prior. In some embodiments, the neural network may include a non-Gaussian prior. In some embodiments, the neural network may include an adversarial loss. In some embodiments, the neural network may include a reconstruction loss. In some embodiments, the neural network may be trained with an Adam optimizer. In some embodiments, the neural network may be trained with a stochastic gradient descent optimizer. In some embodiments, the neural network hyperparameters are optimized with a Gaussian process. In some embodiments, the neural network may be trained with a training / validation / test data split. In some embodiments, the neural network may be trained with a k-fold data split, where k is any positive integer. The neural network may be trained with various loss functions, the derivatives of which may be calculated to update one or more parameters of the neural network. The neural network may be trained with a hyperparameter search algorithm.

[0099] train

[0100] In some aspects, the present disclosure describes a method for machine learning engineering chemical space.

[0101] In some embodiments, the method may include providing a neural network. In some embodiments, the neural network may include an input layer configured to receive at least one target descriptor. In some embodiments, the neural network may include a potential layer configured to output at least one potential descriptor, wherein the potential layer is connected to the input layer. In some embodiments, the neural network may include an output layer configured to output at least one target-ligand interaction descriptor, wherein the output layer is connected to the potential layer. In some embodiments, the neural network may include a second output layer configured to output at least a ligand descriptor, wherein the second output layer is connected to the potential layer. In some embodiments, the neural network further includes a plurality of hidden layers connecting at least two of the following: an input layer, a potential layer, and a first output layer and a second output layer. In some embodiments, the neural network may include at least one parameter. In some embodiments, a layer in the neural network may include at least one parameter.

[0102] In some embodiments, the method may include providing training data. In some embodiments, the training data may include multiple target descriptors. In some embodiments, the training data may include multiple target-ligand interaction descriptors. In some embodiments, the training data may include multiple ligand potential descriptors. In some embodiments, the training data may include multiple ligand descriptors. In some embodiments, the multiple target descriptors of the training data include the features describing at least one of the following: hydrophobicity, aromaticity, hydrogen bond acceptance, hydrogen bond supply, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction ability, hydrophobic interaction ability, π-π interaction ability and halogen interaction ability. In some embodiments, the multiple target-ligand interaction descriptors in the training data include the ligand occupancy of a given target. In some embodiments, the multiple ligand potential descriptors in the training data include the SMILES of a given ligand.

[0103] In some embodiments, the method may include training a neural network. In some embodiments, training the neural network may include inputting at least a plurality of target descriptors at an input layer of the neural network. In some embodiments, training the neural network may include outputting a plurality of output potential descriptors at a potential layer and outputting a plurality of output target-ligand interaction descriptors at an output layer. In some embodiments, training the neural network may include optimizing a plurality of loss functions based at least in part on a plurality of output potential descriptors and a plurality of output target-ligand interaction descriptors. In some embodiments, training the neural network may include updating at least one parameter of the neural network. In some embodiments, training of the neural network may include directing the neural network to learn an engineered chemical space including a unified embedding for at least a plurality of target descriptors and a plurality of ligand potential descriptors.

[0104] In some embodiments, the engineered chemical space is at least partially organized based on the spatial information of multiple target descriptors and / or the chemical structure information from multiple ligand potential descriptors. In some embodiments, in the engineered chemical space, the given target potential descriptor of a given target has similarity with the given ligand potential descriptor of a given ligand when the given ligand targets the given target, and has dissimilarity when the given ligand does not target the given target. In some embodiments, the given target potential descriptor is different from the given ligand potential descriptor. In some embodiments, similarity or dissimilarity can be measured using a similarity function.

[0105] In some embodiments, multiple loss functions include contrast loss functions calculated between multiple pairs formed between multiple target potential descriptors and multiple ligand potential descriptors output by the potential layer. In some embodiments, at least a portion of multiple pairs can be formed between a target potential descriptor and many ligand potential descriptors. In some embodiments, at least a portion of multiple pairs can be formed between a ligand potential descriptor and many target potential descriptors. In some embodiments, multiple pairs can be formed randomly or follow a predetermined pattern or agreement (e.g., in batches of training data). In some embodiments, pairs can be formed between ligand potential descriptors and target potential descriptors that are similar to the expected. In some embodiments, pairs can be formed between ligand potential descriptors and target potential descriptors that are different from the expected.

[0106] In some embodiments, at least a portion of a neural network is trained at least in part by using a learning method that embeds two or more information sources into a unified embedding space. In some embodiments, the learning method is contrastive learning. In some embodiments, contrastive learning is based at least in part on the optimization of a loss function calculated between pairs of target potential descriptors and ligand potential descriptors in the training data. In some embodiments, the loss function comprises a contrastive loss function.

[0107] In some embodiments, the contrast loss includes at least:

[0108] L(p,f)=[α-sim(p,f)+sim(p,f')] + +[α-sim(p,f)+sim(p',f)] +

[0109] where L(p,f) refers to the contrastive loss, α is the margin parameter, p is the first target descriptor in the training data, f is the first ligand descriptor in the training data, p' is the second target descriptor in the training data that is at least similar to the first ligand descriptor, f' is the second ligand descriptor in the training data that is at least similar to the first target descriptor, sim is the similarity function calculated between its brackets, and [x] + ≡max(x,0).

[0110] In some embodiments, training the neural network comprises optimizing a loss function comprising a reconstruction loss function calculated between a plurality of target-ligand interaction descriptors of the training data and a plurality of reconstructed target-ligand interaction descriptors output by the first output layer.

[0111] In some embodiments, at least a portion of the neural network is trained based at least in part on a learning method that parameterizes a portion of the neural network to output a ligand-target interaction descriptor. In some embodiments, the learning method includes optimizing a reconstruction loss function based at least in part on ligand occupancy.

[0112] In some embodiments, reconstruction losses include at least:

[0113] L(X o ,Y o )=||X o -Y o ||2 2

[0114] Where L(X o ,Y o ) refers to the reconstruction loss, where X o is the ligand occupancy descriptor in the training data, and where Y o A reconstruction of a ligand occupancy descriptor by at least a portion of a neural network.

[0115] In some embodiments, the reconstruction loss between the data describing the interaction between the target and the ligand can be calculated. In some embodiments, the interaction can include a covalent bonding interaction, a hydrogen bonding interaction, an electrostatic interaction, or an ionic interaction. In some embodiments, the interaction can include a chemical reaction. In some embodiments, the interaction can include a conformational change induced by a ligand on the target. In some embodiments, the interaction can include a reversible binding. In some embodiments, the interaction can include an irreversible binding.

[0116] In some embodiments, training the neural network includes outputting a plurality of output ligand descriptors, and optimizing a loss function comprising a reconstruction loss function calculated between a plurality of target-ligand descriptors of the training data and a plurality of output target-ligand interaction descriptors output by a second output layer.

[0117] In some embodiments, at least a portion of the neural network is trained based at least in part on a learning method that parameterizes a portion of the neural network to output a molecular identifier. In some embodiments, the learning method is an optimization reconstruction loss. In some embodiments, the reconstruction loss is a probabilistic loss. In some embodiments, the reconstruction loss comprises at least: L(X f ,Y f )=-∑ i n Y f,i ·log(X f,i )

[0118] Where L(Xf ,Y f ) refers to the reconstruction loss, where -∑ i n refers to the negative sum of each element i among the total n elements in the operand, where X f,i is the ligand identifier element in the training data, and where Y f,i is a ligand identifier element reconstructed by at least a portion of a neural network.

[0119] Various other forms of loss functions may be used. For example, in some embodiments, the loss function may be formulated to optimize regression loss, evidence-based lower bound, maximum likelihood, KL divergence, etc., applied with various distribution functions (such as Gaussian, non-Gaussian, mixture of Gaussian, mixture of logistic functions, etc.).

[0120] In some aspects, the disclosure describes a computer-implemented method of implementing a method of training a neural network on a computer.

[0121] In some aspects, the present disclosure describes a computer program product comprising a computer readable medium having computer executable code encoded therein, the computer executable code being adapted to be executed to implement a method of training a neural network.

[0122] In some aspects, the disclosure describes a non-transitory computer-readable storage medium encoded with a computer program including instructions executable by one or more processors to train a neural network.

[0123] In some aspects, the present disclosure describes a computer-implemented system comprising: a digital processing device including at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program comprising instructions executable by the digital processing device to train a neural network.

[0124] Latent Descriptor Generation

[0125] In some embodiments, the potential descriptors can be generated using a target descriptor encoder. In some embodiments, the target descriptor encoder can be part of a neural network. In some embodiments, the target descriptor encoder receives a target descriptor and transforms the target descriptor into a potential descriptor.

[0126] In some embodiments, the potential descriptors may be generated using a ligand identifier encoder. In some embodiments, the ligand identifier encoder may perform fingerprinting based at least in part on a given ligand descriptor. In some embodiments, the fingerprinting may generate a Morgan fingerprint, a Tanimoto fingerprint, or any other fingerprint. In some embodiments, the ligand identifier encoder may be part of a neural network. In some embodiments, the ligand identifier encoder may transform a given ligand descriptor into a potential descriptor.

[0127] Potential descriptors can be embedded on a variety of supports. The support can include any sufficient number of dimensions. The sufficient number of dimensions can vary based at least in part on the complexity of the data set, the size of the data set, the type of embedding used, and other factors. The support can include a variety of geometric shapes. In some embodiments, the support can include dimensions extending from negative infinity to positive infinity. In some embodiments, the support can include periodic dimensions, for example, from zero to two π radians. Potential descriptors can be embedded in engineered chemical space to include a variety of manifold shapes.

[0128] Ligand descriptor generation

[0129] In some embodiments, generating a ligand descriptor is performed using a neural network. In some embodiments, generating a ligand descriptor is performed using a language model. In some embodiments, generating a ligand descriptor is performed randomly. In some embodiments, generating a ligand descriptor is performed with sampled noise. In some embodiments, generating a ligand descriptor is performed by receiving a potential descriptor and transforming the potential descriptor into a ligand descriptor.

[0130] In some embodiments, the ligand descriptor comprises a ligand identifier. In some embodiments, the ligand identifier comprises a text identifier. In some embodiments, the text identifier comprises SMILES, InChI (and its various subtypes) or SELFIES. In some embodiments, the ligand identifier comprises a molecular adjacent matrix or a molecular graph. In some embodiments, the ligand identifier can be any descriptor that explicitly describes the ligand.

[0131] In some embodiments, the ligand descriptor may include a molecular fingerprint. Various molecular fingerprinting methods may be used to create a molecular fingerprint. In some embodiments, the molecular fingerprint may be a Morgan fingerprint, a Tanimoto fingerprint, a multi-layer neighborhood of atoms, a spectrophore, or any combination thereof. In some embodiments, the molecular fingerprint may be generated using radial basis functions, graph convolution, latent space engineering using a neural network, or any combination thereof.

[0132] In some embodiments, the language model comprises a neural network. In some embodiments, the language model comprises a long short-term memory model (LSTM). In some embodiments, the hidden units of the LSTM are initialized using ligand descriptors. In some embodiments, the language model comprises a gated recurrent unit (GRU). In some embodiments, the hidden units of the GRU are initialized using ligand descriptors. In some embodiments, the language model comprises one or more dense layers.

[0133] Interaction descriptor generation

[0134] In some embodiments, generating a target-ligand interaction descriptor is performed using a neural network. In some embodiments, generating a target-ligand interaction descriptor is performed explicitly. In some embodiments, generating a target-ligand interaction descriptor is performed randomly. In some embodiments, generating a target-ligand interaction descriptor is performed under sampling noise. In some embodiments, generating a target-ligand interaction descriptor is performed by receiving a potential descriptor and transforming the potential descriptor into a target-ligand interaction descriptor.

[0135] In some embodiments, the target-ligand interaction descriptor can include information describing the interaction between a given ligand and a given target. In some embodiments, the information can include the ligand occupancy of the target. In some embodiments, the information can include changes in the chemical structure of the ligand, the target, or both. In some embodiments, the information can include changes in the conformation of the ligand, the target, or both. In some embodiments, the information can include a bond formed between the ligand and the target.

[0136] Training Data

[0137] Various target descriptors can be used. In some embodiments, the target descriptor includes a protein target descriptor. In some embodiments, the target descriptor includes a protein pocket descriptor. In some embodiments, the target descriptor includes a feature describing at least one of the following: hydrophobicity, aromaticity, hydrogen bond acceptance characteristics, hydrogen bond supply characteristics, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction ability, hydrophobic interaction ability, π-π interaction ability, halogen interaction ability and ligand occupancy. In some embodiments, the target descriptor includes a spatial map of one or more protein targets. In some embodiments, the spatial map is a grid. In some embodiments, the spatial map is a rectangular grid. In some embodiments, the spatial map includes at least 3 dimensions. In some embodiments, the spatial map includes at least 4 dimensions. In some embodiments, the spatial map includes a resolution of less than about 3.5 angstroms. In some embodiments, the spatial pattern comprises a resolution of less than about 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 angstroms, including increments therein. In some embodiments, the spatial pattern comprises a resolution of greater than about 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 angstroms, including increments therein.

[0138] Various ligand descriptors can be used. In some embodiments, the ligand descriptor includes a small molecule identifier. In some embodiments, the ligand descriptor includes a macromolecule identifier. In some embodiments, the ligand descriptor includes a chemical structure identifier. In some embodiments, the ligand descriptor includes SMILES. In some embodiments, the ligand descriptor includes SELFIES, InChi, a neighbor matrix, a graph, a molecular coordinate, a vibration spectrum, a molecular fingerprint, or any combination thereof. In some embodiments, the ligand descriptor includes a potential descriptor of the ligand descriptor, for example, a potential descriptor of SMILES, SELFIES, InChi, a graph, a molecular fingerprint, or any combination thereof.

[0139] In some embodiments, the target descriptor can be normalized. In some embodiments, the ligand descriptor can be normalized. In some embodiments, the term normalization can refer to a series of methods for adjusting a data set to align the data set to a common scale. In some embodiments, the normalization method can include multiplying a portion or all of the data set by a coefficient. In some embodiments, the normalization method can include adding or subtracting a constant from a portion or all of the data set. In some embodiments, the normalization method can include adjusting a portion or all of the data set to a known statistical distribution. In some embodiments, the normalization method can include adjusting a portion or all of the data set to a normal distribution. In some embodiments, the normalization method can include adjusting the data set so that the signal intensity of a portion or all of the data set is approximately the same.

[0140] reasoning

[0141] In some aspects, the present disclosure describes systems and methods for generating ligand identifiers.

[0142] In some embodiments, the ligand identifier may be generated based at least in part on the potential descriptors. In some embodiments, the ligand identifier may be generated explicitly. In some embodiments, the ligand identifier may be generated randomly. In some embodiments, the ligand identifier may be generated with added noise. In some embodiments, the ligand identifier may be generated based at least on human input. In some embodiments, the human input may include navigation of an engineered chemical space.

[0143] Figure 4 The inference method shown in the figure above receives as input a target descriptor of a target and outputs a ligand occupancy of the target and / or a ligand identifier of a candidate ligand that is predicted to bind to the target. In some embodiments, a potential descriptor can be generated based at least in part on a target descriptor. In some embodiments, a ligand descriptor can be generated based at least in part on a potential descriptor. In some embodiments, a target-ligand interaction descriptor can be generated based at least in part on a potential descriptor.

[0144] In some embodiments, the target descriptor can be a spatial map of a protein pocket. In some embodiments, the target descriptor encoder can transform the target descriptor into a potential descriptor in an engineered chemical space. In some embodiments, the potential descriptor decoder can transform the potential descriptor into a ligand descriptor. In some embodiments, the target descriptor decoder transforms the potential descriptor into a target-ligand interaction descriptor.

[0145] Figure 4The inference method shown in the middle figure receives as input a target descriptor of a target to output potential descriptors in an engineered chemical space, which can be navigated (e.g., by adding noise) to select different potential descriptors, and outputs candidate ligands predicted to bind to the target. Figure 4 The inference method shown in the figure below receives a ligand identifier as input to output potential descriptors in an engineered chemical space, which can be navigated to select different potential descriptors, and output new candidate ligands that can target a target similar to the ligand of the input ligand descriptor. In some embodiments, the engineered chemical space can have potential descriptors organized so that close potential descriptors are mapped to ligands that target similar targets.

[0146] In some embodiments, the second potential descriptor can be generated based at least in part on the first potential descriptor. In some embodiments, the second potential descriptor can be automatically generated or generated by incorporating input from humans. In some embodiments, the second descriptor can be generated by adding white or structured noise to the first potential descriptor. In some embodiments, the second descriptor can be generated based on human input to navigate the engineered chemical space.

[0147] In some aspects, the disclosure describes a computer-implemented method of implementing a method of sampling ligand descriptors or target-ligand interaction descriptors on a computer.

[0148] In some aspects, the present disclosure describes a computer program product comprising a computer readable medium having computer executable code encoded therein, the computer executable code being adapted to be executed to implement a method of sampling a ligand descriptor or a target-ligand interaction descriptor.

[0149] In some aspects, the disclosure describes a non-transitory computer-readable storage medium encoded with a computer program comprising instructions executable by one or more processors to sample a ligand descriptor or a target-ligand interaction descriptor.

[0150] In some aspects, the present disclosure describes a computer-implemented system comprising: a digital processing device comprising at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program comprising instructions executable by the digital processing device to sample a ligand descriptor or a target-ligand interaction descriptor.

[0151] Computing System

[0152] In some aspects, the present disclosure describes a computer-implemented system comprising: a digital processing device comprising at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program comprising instructions executable by the digital processing device to sample a ligand using any of the methods disclosed herein. In some aspects, the present disclosure describes a computer-implemented method of implementing any of the methods disclosed herein in a computer system. Reference Figure 5 , a block diagram is shown depicting an exemplary machine including a computer system 500 (e.g., a processing system or computing system) within which a set of instructions may be executed to cause a device to perform or implement any one or more of the aspects and / or methods to sample ligands or train a neural network to learn the engineered chemical space of the present disclosure. Figure 5 The components in the description are examples only and do not limit the scope of use or functionality of any hardware, software, embedded logic component, or combination of two or more such components to implement a particular embodiment.

[0153] The computer system 500 may include one or more processors 501, memory 503, and storage 508 that communicate with each other and with other components via a bus 540. The bus 540 may also connect a display 532, one or more input devices 533 (e.g., which may include a keypad, keyboard, mouse, stylus, etc.), one or more output devices 534, one or more storage devices 535, and various tangible storage media 536. All of these elements may be connected to the bus 540 directly or via one or more interfaces or adapters. For example, various tangible storage media 536 may be connected to the bus 540 via a storage medium interface 526. The computer system 500 may have any suitable physical form, including (but not limited to) one or more integrated circuits (ICs), printed circuit boards (PCBs), mobile handheld devices (such as mobile phones or PDAs), laptops or notebook computers, distributed computer systems, computing grids, or servers.

[0154] The computer system 500 includes one or more processors 501 (e.g., a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), or a quantum processing unit (QPU)) that perform functions. The computer system 500 can be one of a variety of high-performance computing platforms. For example, one or more processors 501 can form a high-performance computing cluster. In some embodiments, one or more processors 501 can form a distributed computing system connected by a wired and / or wireless network. In some embodiments, an array of CPUs, GPUs, QPUs, or any combination thereof can be operably connected to implement any of the methods disclosed herein. The processor 501 optionally contains a cache memory unit 502 for temporary local storage of instructions, data, or computer bit addresses. The processor 501 is configured to assist in the execution of computer-readable instructions. The computer system 500 can be Figure 5 The components depicted in provide functionality as a result of processor 501 executing non-transitory processor-executable instructions embodied in one or more tangible computer-readable storage media (such as memory 503, storage 508, storage device 535 and / or storage medium 536). The computer-readable medium can store software that implements a particular embodiment, and the processor 501 can execute the software. The memory 503 can read the software from one or more other computer-readable media (such as mass storage device 535, mass storage device 536) or from one or more other sources through a suitable interface (such as network interface 520). The software can cause the processor 501 to perform one or more processes or one or more steps of one or more processes described or illustrated herein. Executing such processes or steps can include defining data structures stored in the memory 503 and modifying the data structures as indicated by the software.

[0155] The memory 503 may include various components (e.g., machine-readable media), including (but not limited to): random access memory components (e.g., RAM 504) (e.g., static RAM (SRAM), dynamic RAM (DRAM), ferroelectric random access memory (FRAM), phase change random access memory (PRAM), etc.), read-only memory components (e.g., ROM 505), and any combination thereof. The ROM 505 may be used to communicate data and instructions unidirectionally to the processor 501, and the RAM 504 may be used to communicate data and instructions bidirectionally with the processor 501. The ROM 505 and the RAM 504 may include any suitable tangible computer-readable media described below. In one example, a basic input / output system 506 (BIOS) (including basic routines that help transfer information between elements within the computer system 500, such as during startup) may be stored in the memory 503.

[0156] Fixed storage 508 is optionally bidirectionally connected to processor 501 via storage control unit 507. Fixed storage 508 provides additional data storage capacity and may also include any suitable tangible computer-readable medium described herein. Storage 508 may be used to store operating system 509, executable files 510, data 511, applications 512 (applications), etc. Storage 508 may also include an optical drive, a solid-state memory device (e.g., a flash memory-based system), or a combination of any of the above. Information in storage 508 may be incorporated into memory 503 as virtual memory where appropriate.

[0157] In one example, storage device 535 can be removably connected to computer system 500 via storage device interface 525 (e.g., via an external port connector (not shown)). In particular, storage device 535 and associated machine-readable media can provide non-volatile and / or volatile storage of machine-readable instructions, data structures, program modules, and / or other data for computer system 500. In one example, the software can reside in whole or in part on the machine-readable medium on storage device 535. In another example, the software can reside in whole or in part on processor 501.

[0158] Bus 540 connects a variety of subsystems. Herein, reference to a bus may include one or more digital signal lines serving a common function where appropriate. Bus 540 may be any of several types of bus structures using any of a variety of bus architectures, including, but not limited to, a memory bus, a memory controller, a peripheral bus, a local bus, and any combination thereof. As examples and not limitations, such architectures include an industry standard architecture (ISA) bus, an enhanced ISA (EISA) bus, a microchannel architecture (MCA) bus, a video electronics standard association local bus (VLB), a peripheral component interconnect (PCI) bus, a PCI high-speed (PCI-X) bus, an accelerated graphics port (AGP) bus, a hypertransport (HTX) bus, a serial advanced technology attachment (SATA) bus, and any combination thereof.

[0159] The computer system 500 may also include an input device 533. In one example, a user of the computer system 500 may input commands and / or other information into the computer system 500 via the input device 533. Examples of the input device 533 include, but are not limited to: α-digital input device (e.g., keyboard), pointer device (e.g., mouse or touchpad), touchpad, touch screen, multi-touch screen, joystick, stylus, game console, audio input device (e.g., microphone, voice response system, etc.), optical scanner, video or static image capture device (e.g., camera) and any combination thereof. In some embodiments, the input device is Kinect, Leap Motion, etc. The input device 533 may be connected to the bus 540 via any of a variety of input interfaces 523 (e.g., input interface 523), and a variety of input interfaces 523 include, but are not limited to, serial, parallel, game port, USB, FIREWIRE, THUNDERBOLT, or any combination thereof. In some embodiments, the input device 533 may be used to navigate the engineered chemical space and / or manipulate its visualization. In some embodiments, the visual representation of the engineered chemical space can be enlarged, reduced, rotated, reflected along an axis or plane, etc. using human input via input device 533 .

[0160] In certain embodiments, when the computer system 500 is connected to the network 530, the computer system 500 can communicate with other devices (specifically, mobile devices and enterprise systems, distributed computing systems, cloud storage systems, cloud computing systems, etc. connected to the network 530). Communications to and from the computer system 500 can be sent through the network interface 520. For example, the network interface 520 can receive incoming communications (such as requests or responses from other devices) in the form of one or more packets (such as Internet Protocol (IP) packets) from the network 530, and the computer system 500 can store the incoming communications in the memory 503 for processing. The computer system 500 can similarly store outgoing communications (such as requests or responses to other devices) in the form of one or more packets in the memory 503 and transmit them from the network interface 520 to the network 530. The processor 501 can access these communication packets stored in the memory 503 for processing.

[0161] Examples of network interface 520 include, but are not limited to, a network interface card, a modem, and any combination thereof. Examples of network 530 or network segment 530 include, but are not limited to, a distributed computing system, a cloud computing system, a wide area network (WAN) (e.g., the Internet, an enterprise network), a local area network (LAN) (e.g., a network associated with an office, a building, a campus, or other relatively small geographic space), a telephone network, a direct connection between two computing devices, a point-to-point network, and any combination thereof. A network such as network 530 may employ wired and / or wireless communication modes. In general, any network topology may be used.

[0162] Information and data can be displayed by display 532. Examples of display 532 include, but are not limited to, cathode ray tubes (CRTs), liquid crystal displays (LCDs), thin film transistor liquid crystal displays (TFT-LCDs), organic liquid crystal displays (OLEDs) such as passive matrix OLED (PMOLED) or active matrix OLED (AMOLED) displays, plasma displays, and any combination thereof. Display 532 can be connected to processor 501, memory 503, and fixed storage 508 and other devices (such as input devices 533) via bus 540. Display 532 is connected to bus 540 via video interface 522, and the transmission of data between display 532 and bus 540 can be controlled via graphics control 521. In some embodiments, the display is a video projector. In some embodiments, the display is a head mounted display (HMD), such as a VR headset. In further embodiments, suitable VR headsets include, by way of non-limiting example, HTC Vive, Oculus Rift, Samsung Gear VR, Microsoft HoloLens, Razer OSVR, FOVE VR, Zeiss VR One, Avegant Glyph, Freefly VR headsets, etc. In still further embodiments, the display is a combination of devices such as those disclosed herein.

[0163] In addition to the display 532, the computer system 500 may also include one or more other peripheral output devices 534, including (but not limited to) audio speakers, printers, storage devices, and any combination thereof. Such peripheral output devices may be connected to the bus 540 via the output interface 524. Examples of the output interface 524 include (but not limited to) a serial port, a parallel connection, a USB port, a FIREWIRE port, a THUNDERBOLT port, and any combination thereof.

[0164] Additionally or alternatively, computer system 500 may provide functionality due to logic hardwired or otherwise embodied in circuitry that may operate in place of or in conjunction with software to perform one or more processes or one or more steps of one or more processes described or illustrated herein. References to software in this disclosure may encompass logic, and references to logic may encompass software. Furthermore, references to computer-readable media may encompass circuitry (such as an IC) storing software for execution, circuitry embodying logic for execution, or both, as appropriate. This disclosure encompasses any suitable combination of hardware, software, or both.

[0165] Those skilled in the art will appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality.

[0166] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in an alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a DSP core, or any other such configuration.

[0167] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein can be directly embodied in hardware, a software module executed by one or more processors, or a combination thereof. The software module can reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. In an alternative, the storage medium can be integrated with the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In an alternative, the processor and the storage medium can reside in a user terminal as discrete components.

[0168] As described herein, suitable computing devices include, by way of non-limiting examples, server computers, desktop computers, laptop computers, notebook computers, subnotebook computers, netbook computers, netpad computers, set-top computers, media streaming devices, handheld computers, Internet appliances, mobile smart phones, and tablet computers.

[0169] In some embodiments, the computing 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 executing applications. Those skilled in the art will recognize that, as non-limiting examples, suitable server operating systems include FreeBSD, OpenBSD, Linux, Mac OS X Windows as well as 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. Those skilled in the art will also recognize that, as non-limiting examples, suitable mobile smartphone operating systems include OS, Research In BlackBerry Windows OS, Windows OS, as well as

[0170] In some embodiments, the computer system 500 can be accessed through a user terminal to receive user commands. The user commands can include line commands, scripts, programs, etc. and various instructions that can be executed by the computer system 500. The computer system 500 can receive instructions to train a neural network, search for hyperparameters to train a neural network, generate a variety of ligand descriptors, or schedule computing jobs for the computer system 500 to execute any instructions.

[0171] Non-transitory computer readable storage medium

[0172] In some aspects, the present disclosure describes a non-transitory computer-readable storage medium encoded with a computer program, the computer program comprising instructions executable by one or more processors to sample ligands using any of the methods disclosed herein. In some embodiments, the non-transitory computer-readable storage medium may include a neural network architecture, weights for the neural network architecture, or both. In some embodiments, the non-transitory computer-readable storage medium may include a neural network training loss history, a validation loss history, test accuracy, and various measures of neural network training performance. In some embodiments, the platforms, systems, media, and methods disclosed herein include one or more non-transitory computer-readable storage media encoded with a program comprising instructions executable by an operating system of an optional networked computing device.

[0173] In further embodiments, the computer-readable storage medium is a tangible component of a computing device. In still further embodiments, the computer-readable storage medium is optionally removable from a computing device. In some embodiments, as non-limiting examples, the computer-readable storage medium includes a flash memory device, a solid-state memory, a disk drive, a tape drive, an optical drive, a distributed computing system including cloud computing systems and services, and the like. In some embodiments, the programs and instructions are permanently, substantially permanently, semi-permanently, or non-temporarily encoded on the medium.

[0174] Computer Programs

[0175] In some aspects, the present disclosure describes a computer program product comprising a computer-readable medium having computer executable code encoded therein, the computer executable code being suitable for being executed to implement any of the methods disclosed herein. In some embodiments, the platforms, systems, media, and methods disclosed herein include at least one computer program or use thereof.

[0176] A computer program includes a series of instructions that can be executed by one or more processors of a CPU of a computing device and 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), computing data structures, etc. According to the disclosure provided herein, those skilled in the art will recognize that computer programs can be written in various versions of various languages. In some embodiments, APIs can include various languages, for example, languages ​​in various versions of TensorFlow, Theano, Keras, PyTorch, or any combination thereof that can be implemented in various versions of Python, Python3, C, C#, C++, MatLab, R, Java, or any combination thereof.

[0177] The functionality of computer readable instructions can be combined or distributed in various environments as needed. In some embodiments, the computer program includes an instruction sequence. In some embodiments, the computer program includes multiple instruction sequences. In some embodiments, the computer program is provided from one location. In other embodiments, the computer program is provided from multiple locations. In various embodiments, the computer program includes one or more software modules. In various embodiments, the computer program includes one or more network applications, one or more independent applications, one or more network browser plug-ins, extensions, add-ons or add-ons or combinations thereof in part or in whole.

[0178] Web Application

[0179] In some embodiments, the computer program includes a web application. In some embodiments, a user can input a query through the web application to sample a ligand. In some embodiments, a user can add additional training data through the web application. Based on the disclosure provided herein, those skilled in the art will recognize that the web application utilizes one or more software frameworks and one or more database systems in various embodiments. In some embodiments, the web application is based on a software framework (such as .NET or Ruby on Rails (RoR). In some embodiments, the web application utilizes one or more database systems, including, as non-limiting examples, relational, non-relational, object-oriented, associative, XML, and file-oriented database systems. In further embodiments, suitable relational database systems include, as non-limiting examples, SQL Server, mySQL TM as well as Those skilled in the art will also recognize that web applications are written in one or more versions of one or more languages ​​in various embodiments. Web applications can be written in one or more markup languages, presentation definition languages, client-side scripting languages, server-side encoding and decoding languages, database query languages, or combinations thereof. In some embodiments, web applications are written to a certain extent in a markup language such as Hypertext Markup Language (HTML), Extensible Hypertext Markup Language (XHTML), or Extensible Markup Language (XML). In some embodiments, web applications are written to a certain extent in a presentation definition language such as Cascading Style Sheets (CSS). In some embodiments, web applications are written to a certain extent in a client-side scripting language such as Asynchronous JavaScript and XML (AJAX), ActionScript, JavaScript, or In some embodiments, the web application is written in part in a server-side codec language such as Active Server Pages (ASP), Perl, Java TM , JavaServer Pages (JSP), Hypertext Preprocessor (PHP), Python TM , Ruby, Tcl, Smalltalk, In some embodiments, the web application is written in a database query language (such as Structured Query Language (SQL)) to some extent. In some embodiments, the web application integrates enterprise server products such as Lotus Mobile Apps

[0180] In some embodiments, the computer program includes a mobile application provided to a mobile computing device. In some embodiments, the mobile application is provided to the mobile computing device when it is manufactured. In other embodiments, the mobile application is provided to the mobile computing device via a computer network described herein.

[0181] In view of the disclosure provided herein, mobile applications are created by techniques known to those skilled in the art, using hardware, languages, and development environments known in the art. Those skilled in the art will recognize that mobile applications are written in several languages. As non-limiting examples, suitable programming languages ​​include C, C++, C#, Objective-C, Java, and C++. TM , JavaScript, Pascal, Object Pascal, Python TM , Ruby, VB.NET, WML, and XHTML / HTML with or without CSS, or a combination of these.

[0182] Suitable mobile application development environments are available from several sources. As non-limiting examples, commercially available development environments include AirplaySDK, alcheMo, Celsius, Bedrock, Flash Lite, .NET Compact Framework, Rhomobile, and WorkLight Mobile Platform. Other freely available development environments include, as non-limiting examples, Lazarus, MobiFlex, MoSync, and Phonegap. In addition, mobile device manufacturers distribute software developer kits that include, as non-limiting examples, the iPhone and iPad (iOS) SDK, the Android TM SDK, SDK, BREW SDK, OSSDK, Symbian SDK, webOSSDK and Mobile SDK.

[0183] Standalone Application

[0184] In some embodiments, the computer program includes a stand-alone application, which is a program that runs as an independent computer process, rather than an add-on to an existing process, e.g., not a plug-in. Those skilled in the art will recognize that stand-alone applications are often compiled. A compiler is a computer program that transforms source code written in a programming language into binary object code such as assembly language or machine code. As non-limiting examples, suitable compiled programming languages ​​include C, C++, Objective-C, COBOL, Delphi, Eiffel, Java, and the like. TM , Lisp, Python TM , Visual Basic and VB.NET, or a combination thereof. Often, the computer program is compiled at least in part to create an executable program. In some embodiments, the computer program includes one or more executable compiled applications.

[0185] Software Modules

[0186] In some embodiments, the platform, system, medium and method disclosed herein include software, server and / or database module, or its use. In view of the disclosure provided herein, the software module is created by a technique known to those skilled in the art, using a machine, software and language known in the art. The software module disclosed herein is implemented in a variety of ways. In various embodiments, the software module includes a file, a code segment, a programming object, a programming structure, a distributed computing resource, a cloud computing resource or a combination thereof. In further various embodiments, the software module includes multiple files, multiple code segments, multiple programming objects, multiple programming structures, multiple distributed computing resources, multiple cloud computing resources or a combination thereof. In various embodiments, as a non-limiting example, one or more software modules include a network application, a mobile application, an independent application and a distributed or cloud computing application. In some embodiments, the software module is in a computer program or application. In other embodiments, the software module is in more than one computer program or application. In some embodiments, the software module is hosted on a machine. In other embodiments, the software module is hosted on more than one machine. In further embodiments, the software module is hosted on a distributed computing platform such as a cloud computing platform. In some embodiments, the software modules are hosted in one location on one or more machines. In other embodiments, the software modules are hosted in more than one location on one or more machines.

[0187] database

[0188] In some embodiments, the platform, system, medium and method disclosed herein include one or more databases or their uses. In view of the disclosure provided herein, those skilled in the art will recognize that many databases are suitable for storing and retrieving information about chemical data sets, potential descriptors or any combination thereof. In various embodiments, as non-limiting examples, suitable databases include relational databases, non-relational databases, object-oriented databases, object databases, entity relationship model databases, associative databases, XML databases, file-oriented databases and graphic databases. Further non-limiting examples include SQL, PostgreSQL, MySQL, Oracle, DB2, Sybase and MongoDB. In some embodiments, the database is based on the Internet. In further embodiments, the database is based on a network. In further embodiments, the database is based on cloud computing. In specific embodiments, the database is a distributed database. In other embodiments, the database is based on one or more local computer storage devices.

[0189] Example

[0190] The following examples are provided to further illustrate some embodiments of the present disclosure but are not intended to limit the scope of the present disclosure; it will be understood by their exemplary nature that other procedures, methods or techniques known to those skilled in the art may be used instead.

[0191] Example 1: Model training

[0192] data

[0193] The model was trained using the scPDB and Biolip datasets. The two datasets consist of 16,034 and 445,803 protein-ligand pairs, respectively. The protein-ligand pairs in the dataset are crystal structures. The dataset was first screened to obtain compounds with glucose-tri-phosphate / glucose-di-phosphate or natural products. Compounds with greater than 100 ng / mL were removed. The protein-ligand pairs in scPDB were randomly split into training and test sets of size 7000 and 1294, respectively. The protein-ligand pairs in Biolip were split into training and test sets of size 14000 and 3615, respectively.

[0194] Data Characterization

[0195] The data were characterized using the protein-ligand interaction analysis tool (“plip”, Adasme et al., PLIP 2021: expanding the scope of the protein-ligand interaction profiler to DNA and RNA, NAR 2021) to annotate the interacting atoms. The data were voxelized into a discretized cubic grid using the Python package HTMD. The crystal structure of the protein-ligand pair was transformed into a four-dimensional (height × width × length × channel) tensor, which is shown in Figure 1 Each of the 13 channels in the data represents one or more pharmacophore properties, one or more interaction labels (occupancy) of an atom or heavy atom, and one or more labels of an atom. Channels 1 to 13 represent 'Hydrophobicity', 'Aromaticity', 'Hydrogen bond_acceptor', 'Hydrogen bond_donor', 'Positive_ionizable', 'Negative_ionizable', 'Metal', 'Pocket occupancy', 'Residual hydrogen bond interactions', 'Residual hydrophobic interactions', 'Residual π-π interactions', 'Residual halogen interactions', and 'Ligand occupancy', respectively.

[0196] The goal of the training is to create a model that can be given an empty pocket (input) and generate good molecules to fit in the pocket. By following the data characterization steps, the protein-ligand complex PDB file is transformed into a 32×32×32×13 voxelized 4D tensor, each of which is paired with the corresponding SMILES of the ligand. The first 12 channels of the 4D tensor are used as input to the model. The last channel of the 4D tensor occupied by the ligand is used as the output of the model. The SMILES of the ligand is also used as output. The SMILES of the ligand is used to generate the Morgan fingerprint of each SMILES in the data set.

[0197] Model Architecture

[0198] Figure 2 The overall architecture of the model is shown. The architecture can be divided into two parts, U-Net and LSTM models. Information flows from the input of the U-Net that receives the empty pocket tensor (e.g., characterization data of channels without ligand occupancy) to the output of the U-Net that outputs the ligand occupancy tensor and flows to the output of the LSTM that outputs the SMILES of the ligand occupying the empty pocket. The model is a multi-task learning model (outputting ligand occupancy and ligand SMILES). By learning two tasks together, the spatial information from the U-net architecture (learned by inputting the tensor of the empty pocket and outputting the tensor of ligand occupancy) can be shared with the memory of the LSTM through the bottleneck layer.

[0199] The encoder in U-Net starts with 4 residual blocks. Each residual block consists of 2 3×3×3 convolutional layers, 2 batch normalization layers, and 2 ReLUs. A max pooling operation is applied after each residual block.

[0200] After feeding the 12×32×32×32 4D tensor input into the encoder, the target latent descriptor is obtained, which is a lower dimensional representation of the information from the empty pocket. The target latent descriptor extracted by the encoder is transformed into the initial memory state of the LSTM through 6 dense linear layers, guiding the language model to generate SMILES of molecules fitting the pocket.

[0201] The pocket descriptors are decoded using 3D transposed convolutions, concatenated with the corresponding feature maps from the infection path, and passed through the residual block mentioned in the encoder. After 3 repetitions, a convolutional layer followed by a sigmoid activation function as the last layer is used to map the 4D tensor to an output 3D tensor (i.e., ligand occupancy).

[0202] LSTM

[0203] In parallel with the existing U-Net architecture, another branch (i.e., LSTM) is added to predict the SMILES of the ligand occupying the input pocket. The target latent descriptor extracted by the encoder is passed through 6 linear layers as the initial cells and hidden states in the LSTM. Therefore, the LSTM incorporates the pocket information to generate the SMILES of the ligand fitting the input pocket.

[0204] Loss Function

[0205] The model is trained using a combination of loss functions.

[0206] Ligand occupancy loss

[0207] L(X o ,Y o )=||X o -Y o ||2 2

[0208] Where L(X o ,Y o ) refers to the first reconstruction loss,

[0209] Where X o is the ligand occupancy descriptor in the training data,

[0210] where Y o for the reconstruction of the ligand occupancy descriptor via a bottleneck architecture, and

[0211] ||x|| p ≡(x1 p +···+x n p ) 1 / p .

[0212] LSTM Loss

[0213] L(X f ,Y f )=-∑ i n Y f,i ·log(X f,i )

[0214] Where L(X f ,Y f ) refers to the second reconstruction loss,

[0215] Where -∑ i n refers to the negative sum of each element i among the total n elements in the operand,

[0216] Where X f,i is the ligand identifier element in the training data, and

[0217] where Y f,i is the ligand identifier element reconstructed by the language model.

[0218] Contrastive loss

[0219] L(p,f)=[α-sim(p,f)+sim(p,f')] + +[α-sim(p,f)+sim(p',f)] +

[0220] Where L(p,f) refers to the contrast loss,

[0221] α is the marginal parameter,

[0222] p is the first target descriptor in the training data,

[0223] f is the first ligand descriptor in the training data,

[0224] p' is a second target descriptor in the training data that is at least similar to the first ligand descriptor,

[0225] f' is a second ligand descriptor in the training data that is at least similar to the first target descriptor,

[0226] sim is the similarity function computed between its brackets, and

[0227] [x] + ≡max(x,0).

[0228] The ligand occupancy loss causes the model to extract information from the input pocket and generate ligand spatial information (ligand occupancy) accordingly.

[0229] The contrast loss embeds the target latent descriptor from the empty pocket and the fingerprint of the crystal ligand (e.g., Morgan fingerprint) into the same embedding space. Through this loss, the target latent descriptor extracted from the encoder is more closely connected to the ligand latent descriptor (e.g., Morgan fingerprint), which makes the language model make more use of pocket information.

[0230] Example 2: Model Reasoning

[0231] This example describes systems and methods for generating ligand descriptors and / or interaction descriptors according to some embodiments. Figure 4 The inference method shown in the figure above receives as input a target descriptor of a target and outputs ligand occupancy of the target and / or ligand identifiers of candidate ligands predicted to bind to the target.

[0232] The target descriptor can be a spatial map of a protein pocket. The target descriptor encoder transforms the target descriptor into a potential descriptor in an engineered chemical space. Subsequently, the ligand descriptor decoder transforms the potential descriptor into a ligand descriptor. For example, this inference pattern can be useful when generating a candidate drug for targeting a protein target. By providing a description of a protein target, an identifier for a candidate drug can be generated.

[0233] Alternatively or in combination, the target descriptor decoder transforms the potential descriptors into interaction descriptors. For example, this inference mode can be useful in studying how a protein target can interact with a candidate drug. By providing a description of a protein target, a prediction of how a target and a ligand can interact can be generated.

[0234] Figure 4 The reasoning method shown in the middle figure receives the target descriptor of the target as input to output potential descriptors in the engineered chemical space, which can be navigated (e.g., by adding noise) to select different potential descriptors, and outputs candidate ligands predicted to bind to the target. This shows that once a potential descriptor has been generated from the target descriptor, another potential descriptor can be selected automatically (e.g., random noise is added to the potential descriptor) or purposefully by humans (e.g., human experts navigate the engineered chemical space to explore other possible potential descriptors). Once a potential descriptor is selected, it can be decoded to generate a ligand descriptor.

[0235] Figure 4 The inference method shown in the figure below receives a ligand identifier as input to output potential descriptors in an engineered chemical space, which can be navigated to select different potential descriptors, and output new candidate ligands that can target a target similar to the ligand of the input ligand descriptor. In some embodiments, the engineered chemical space can have potential descriptors organized so that close potential descriptors are mapped to ligands that target similar targets.

[0236] Numbered Implementation Plan

[0237] The following list of numbered embodiments of the invention should be considered to disclose various features of the invention, which may be considered specific to the particular embodiment in which they are discussed, or may be combined with various other features as listed in other embodiments. Therefore, just because a feature is discussed in a particular embodiment, the use of that feature is not necessarily limited to that embodiment.

[0238] Embodiment 1. A method of sampling ligands, comprising: (a) receiving a target descriptor; (b) generating potential descriptors in an engineered chemical space based at least in part on the target descriptors; and (c) generating ligand descriptors based at least in part on the potential descriptors.

[0239] Embodiment 2. The method of embodiment 1, wherein the engineered chemical space comprises a unified embedding for at least a plurality of target potential descriptors and a plurality of ligand potential descriptors.

[0240] Embodiment 3. The method of embodiment 1 or 2, wherein in the engineered chemical space, a given target potential descriptor of a given target has similarity with a given ligand potential descriptor of a given ligand when the given ligand targets the given target, and has dissimilarity when the given ligand does not target the given target.

[0241] Embodiment 4. The method of embodiment 3, wherein the given target potential descriptor is different from the given ligand potential descriptor.

[0242] Embodiment 5. The method of any one of Embodiments 1 to 4, wherein the engineered chemical space is organized at least in part based on spatial information of a plurality of ligands, a plurality of targets, or both.

[0243] Embodiment 6. The method of any one of embodiments 3 to 5, wherein the similarity or dissimilarity can be measured using a similarity function.

[0244] Embodiment 7. The method of embodiment 6, wherein the similarity function comprises a distance-based similarity function, an angle-based similarity function, a set-based similarity function, or any combination thereof.

[0245] Embodiment 8. The method of embodiment 7, wherein the angle-based similarity function is a cosine similarity function.

[0246] Embodiment 9. The method of any one of embodiments 1 to 8, wherein the method is performed at least in part by using a neural network.

[0247] Embodiment 10. The method of embodiment 9, wherein the generating in (b) is performed at least in part by using an encoder.

[0248] Embodiment 11. The method of Embodiment 9 or 10, wherein the generating in (c) is performed at least in part by using a decoder.

[0249] Embodiment 12. The method of embodiment 11, wherein the neural network comprises a bottleneck architecture comprising an encoder and a decoder.

[0250] Embodiment 13. The method of embodiment 12, wherein the bottleneck architecture comprises U-net.

[0251] Embodiment 14. The method of any one of Embodiments 11 to 13, wherein the decoder comprises a language model.

[0252] Embodiment 15. A method as described in Embodiment 14, wherein the language model is a long short-term memory model (LSTM).

[0253] Embodiment 16. The method of embodiment 15, wherein at least one hidden unit of the LSTM is initialized with a latent descriptor.

[0254] Embodiment 17. The method of any one of Embodiments 9 to 16, wherein the neural network comprises a convolutional layer.

[0255] Embodiment 18. The method of any one of Embodiments 9 to 17, wherein the neural network comprises densely connected layers.

[0256] Embodiment 19. The method of any one of Embodiments 9 to 18, wherein the neural network comprises skip connections.

[0257] Embodiment 20. The method of any one of Embodiments 10 to 19, wherein generating the latent descriptors is performed at least using an encoder.

[0258] Embodiment 21. The method of any one of Embodiments 10 to 20, wherein at least the encoder is trained at least in part by using contrastive learning.

[0259] Embodiment 22. The method of embodiment 21, wherein contrastive learning is based at least in part on contrastive losses computed between pairs formed between target latent descriptors and ligand latent descriptors from the training data.

[0260] Embodiment 23. The method of embodiment 22, wherein the contrast loss comprises at least:

[0261] L(p,f)=[α-sim(p,f)+sim(p,f')] + +[α-sim(p,f)+sim(p',f)] +

[0262] Where L(p,f) refers to the contrast loss,

[0263] α is the marginal parameter,

[0264] p is the first target potential descriptor from the training data,

[0265] f is the first ligand potential descriptor from the training data,

[0266] p' is a second target potential descriptor from the training data that is at least similar to the first ligand potential descriptor,

[0267] f' is a second ligand potential descriptor from the training data that is at least similar to the first target potential descriptor,

[0268] sim is the similarity function computed between its brackets, and

[0269] [x] + ≡max(x,0).

[0270] Embodiment 24. The method of any one of embodiments 1 to 23, wherein generating the ligand descriptors is performed stochastically.

[0271] Embodiment 25. The method of embodiment 24, wherein the language model is trained at least in part based on the first reconstruction loss.

[0272] Embodiment 26. The method of embodiment 25, wherein the first reconstruction loss is a probabilistic loss function.

[0273] Embodiment 27. The method of Embodiment 26, wherein the first reconstruction loss comprises at least:

[0274] L(X f ,Y f )=-∑ i n Y f,i ·log(X f,i )

[0275] Where L(X f ,Y f ) refers to the first reconstruction loss,

[0276] Where -∑ i n refers to the negative sum of each element i among the total n elements in the operand, where X f,i is the ligand identifier element from the training data, and

[0277] where Y f,i is the ligand identifier element reconstructed by the language model.

[0278] Embodiment 28. The method of any one of embodiments 1 to 27, wherein the ligand descriptor comprises a ligand identifier.

[0279] Embodiment 29. The method of embodiment 28, wherein the ligand identifier comprises a textual identifier.

[0280] Embodiment 30. The method of embodiment 29, wherein the text identifier comprises SMILES, InChI, or SELFIES.

[0281] Embodiment 31. The method of embodiment 28 or 29, wherein the ligand identifier comprises a molecular adjacency matrix or a molecular graph.

[0282] Embodiment 32. The method of any one of Embodiments 12 to 31, wherein the bottleneck architecture is trained at least in part based on a second reconstruction loss that is at least in part based on a target-ligand interaction.

[0283] Embodiment 33. The method of embodiment 32, wherein the target-ligand interaction comprises ligand occupancy.

[0284] Embodiment 34. The method of embodiment 33, wherein the second reconstruction loss comprises at least:

[0285] L(X o ,Y o )=||X o -Y o || 2 2

[0286] Where L(X o ,Y o ) refers to the second reconstruction loss,

[0287] Where X o is the ligand occupancy descriptor in the training data, and

[0288] where Y o Reconstruction of ligand occupancy descriptors by autoencoder.

[0289] Embodiment 35. The method of any one of embodiments 1 to 34, wherein the target descriptor comprises a protein target descriptor.

[0290] Embodiment 36. The method of any one of embodiments 1 to 35, wherein the target descriptor comprises a protein pocket descriptor.

[0291] Embodiment 37. The method of any one of embodiments 1 to 36, wherein the target descriptor comprises a feature describing at least one of the following: hydrophobicity, aromaticity, hydrogen bond acceptance, hydrogen bond donation, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction ability, hydrophobic interaction ability, π-π interaction ability, and halogen interaction ability.

[0292] Embodiment 38. The method of embodiment 36 or 37, wherein the protein pocket descriptor comprises a spatial map of one or more protein targets.

[0293] Embodiment 39. The method of Embodiment 38, wherein the spatial pattern is a grid.

[0294] Embodiment 40. The method of Embodiment 39, wherein the spatial pattern is a rectangular grid.

[0295] Embodiment 41. The method of Embodiment 40, wherein the spatial pattern comprises at least 3 dimensions.

[0296] Embodiment 42. The method of Embodiment 41, wherein the spatial pattern comprises a resolution of less than about 3.5 angstroms.

[0297] Embodiment 43. A computer-implemented method of implementing any one of the methods of Embodiments 1 to 42 in a computer.

[0298] Embodiment 44. A computer program product comprising a computer readable medium having computer executable code encoded therein, the computer executable code being adapted to be executed to implement any of the methods of Embodiments 1 to 42.

[0299] Embodiment 45. A non-transitory computer readable storage medium encoded with a computer program comprising instructions executable by one or more processors to sample a ligand using any of the methods of Embodiments 1 to 42.

[0300] Embodiment 46. A computer-implemented system comprising: a digital processing device comprising at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program comprising instructions executable by the digital processing device to sample a ligand using any of the methods of Embodiments 1 to 42.

[0301] Embodiment 47. A method for machine learning an engineered chemical space, comprising: (a) providing a neural network, the neural network comprising: i. an input layer configured to receive at least a target descriptor; ii. a latent layer configured to output at least a potential descriptor, wherein the latent layer is connected to the input layer; iii. an output layer configured to output at least a target-ligand interaction descriptor, wherein the output layer is connected to the latent layer; iv. at least one parameter; (b) providing training data including a plurality of target descriptors, a plurality of target-ligand interaction descriptors, and a plurality of ligand potential descriptors; (c) training the neural network by the following steps: (i) inputting at least a plurality of target descriptors at an input layer of the neural network; (ii) outputting a plurality of output potential descriptors at the latent layer and outputting a plurality of output target-ligand interaction descriptors at the output layer; and (iii) optimizing a plurality of loss functions based at least in part on the plurality of output potential descriptors and the plurality of output target-ligand interaction descriptors by updating at least one parameter of the neural network, so that the neural network learns an engineered chemical space including a unified embedding for at least a plurality of target descriptors and a plurality of ligand potential descriptors.

[0302] Embodiment 48. The method of embodiment 47, wherein the engineered chemical space is organized at least in part based on spatial information of a plurality of target descriptors and a plurality of ligand potential descriptors.

[0303] Embodiment 49. The method of embodiment 47 or 48, wherein in the engineered chemical space, a given target potential descriptor for a given target has similarity with a given ligand potential descriptor for a given ligand when the given ligand targets the given target, and has dissimilarity when the given ligand does not target the given target.

[0304] Embodiment 50. The method of embodiment 49, wherein the given target potential descriptor is different from the given ligand potential descriptor.

[0305] Embodiment 51. A method as described in Embodiment 49 or 50, wherein the similarity or dissimilarity can be measured using a similarity function.

[0306] Embodiment 52. The method of any one of Embodiments 47 to 51, wherein the plurality of loss functions comprises contrastive loss functions calculated between a plurality of pairs formed between a plurality of target latent descriptors and a plurality of ligand latent descriptors output by the latent layer.

[0307] Embodiment 53. A method as described in any one of Embodiments 47 to 52, wherein training the neural network includes optimizing a loss function, the loss function including a first reconstruction loss function calculated between a plurality of target-ligand interaction descriptors of the training data and a plurality of reconstructed target-ligand interaction descriptors output by the first output layer.

[0308] Embodiment 54. The method of any one of Embodiments 47 to 53, wherein the neural network further comprises a second output layer configured to output at least a ligand descriptor, wherein the second output layer is connected to the latent layer.

[0309] Embodiment 55. The method of any one of Embodiments 47 to 54, wherein the training data further comprises a plurality of ligand descriptors.

[0310] Embodiment 56. A method as described in any one of Embodiments 47 to 55, wherein training the neural network further comprises outputting a plurality of output ligand descriptors, and optimizing a loss function comprising a second reconstruction loss function calculated between the plurality of ligand descriptors of the training data and the plurality of output target-ligand interaction descriptors output by the second output layer.

[0311] Embodiment 57. The method of any one of Embodiments 47 to 56, wherein the plurality of target descriptors of the training data comprises features describing at least one of: hydrophobicity, aromaticity, hydrogen bond acceptance, hydrogen bond donation, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction ability, hydrophobic interaction ability, π-π interaction ability, and halogen interaction ability.

[0312] Embodiment 58. The method of any one of Embodiments 47 to 57, wherein the plurality of target-ligand interaction descriptors in the training data comprises ligand occupancy for a given target.

[0313] Embodiment 59. The method of any one of Embodiments 47 to 58, wherein the plurality of potential descriptors for ligands in the training data comprises a SMILES for a given ligand.

[0314] Embodiment 60. The method of any one of Embodiments 47 to 59, wherein the neural network further comprises a plurality of hidden layers connecting at least two of the following: the input layer, the latent layer, and the first output layer and the second output layer.

[0315] Embodiment 61. A computer-implemented method of implementing any one of the methods of Embodiments 47 to 60 in a computer.

[0316] Embodiment 62. A computer program product comprising a computer readable medium having computer executable code encoded therein, the computer executable code being adapted to be executed to implement any of the methods of Embodiments 47 to 60.

[0317] Embodiment 63. A non-transitory computer-readable storage medium encoded with a computer program comprising instructions executable by one or more processors to train a neural network using any of the methods of Embodiments 47 to 60.

[0318] Embodiment 64. A computer-implemented system comprising: a digital processing device including at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program comprising instructions executable by the digital processing device to train a neural network using any of the methods of Embodiments 47 to 60.

[0319] Although preferred embodiments of the present disclosure 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. For those skilled in the art, many changes, modifications, and substitutions will now occur without departing from the present disclosure. It should be understood that various alternatives to the embodiments of the present disclosure may be used to practice the present disclosure. The appended claims are intended to define the scope of the present disclosure, and thus encompass methods and structures within the scope of these claims and their equivalents.

Claims

1. A method for sampling a ligand, comprising: (a) receiving a target descriptor; (b) generating potential descriptors in an engineered chemical space based at least in part on the target descriptors; as well as (c) generating a ligand descriptor based at least in part on the potential descriptor; Wherein the engineered chemical space comprises a unified embedding for at least a plurality of target potential descriptors and a plurality of ligand potential descriptors.

2. A method as claimed in claim 1, wherein in the engineered chemical space, a given target potential descriptor of a given target has similarity with a given ligand potential descriptor of a given ligand when the given ligand targets the given target, and has dissimilarity when the given ligand does not target the given target.

3. The method of claim 2, wherein the given target potential descriptor is different from the given ligand potential descriptor. The method of claim 2 , wherein the similarity or the dissimilarity can be measured using a similarity function.

5. The method of claim 1, wherein the engineered chemical space is organized at least in part based on spatial information of a plurality of ligands and a plurality of targets.

6. The method of claim 1, wherein the method is performed at least in part by using a neural network, wherein the generating in (b) is performed at least in part by using an encoder, and wherein the generating in (c) is performed at least in part by using a decoder, and wherein the neural network comprises a bottleneck architecture comprising the encoder and the decoder.

7. The method of claim 6, wherein the bottleneck architecture comprises U-net.

8. The method of claim 6, wherein the decoder comprises a language model.

9. The method of claim 8, wherein the language model is a long short-term memory model (LSTM), wherein at least one hidden unit of the LSTM is initialized with the latent descriptor.

10. The method of claim 6, wherein at least the encoder is trained at least in part by using contrastive learning.

11. The method of claim 10, wherein the contrastive learning is based at least in part on a contrastive loss calculated between pairs formed between the target latent descriptors and the ligand latent descriptors from training data.

12. The method of claim 11, wherein the contrast loss comprises at least: L(p,f)=[α-sim(p,f)+sim(p,f')] + +[α-sim(p,f)+sim(p',f)] + Where L(p,f) refers to the contrast loss, α is the marginal parameter, p is the first target potential descriptor in the training data, f is the first ligand potential descriptor in the training data, p' is a second target potential descriptor in the training data that is at least similar to the first ligand potential descriptor, f' is a second ligand potential descriptor in the training data that is at least similar to the first target potential descriptor, sim is the similarity function computed between its brackets, and [x] + ≡max(x,0)。 13. The method of claim 8, wherein the language model is trained at least in part based on a first reconstruction loss, the first reconstruction loss comprising at least: L(X f ,AND f )=-∑ i n AND f,i ·log(X f,i ) Where L(X f ,Y f ) refers to the first reconstruction loss, Where -∑ i n refers to the negative sum of each element i among the total n elements in the operand, Where X f,i is a ligand identifier element in the training data, and where Y f,i is the ligand identifier element reconstructed by the language model.

14. The method of claim 1, wherein the ligand descriptor comprises a textual identifier.

15. The method of claim 1, wherein the ligand descriptor comprises a molecular adjacency matrix or a molecular graph.

16. The method of claim 6, wherein the bottleneck architecture is trained based at least in part on a second reconstruction loss, the second reconstruction loss being based at least in part on a target-ligand interaction.

17. The method of claim 16, wherein the target-ligand interaction is ligand occupancy.

18. The method of claim 16, wherein the second reconstruction loss comprises at least: L(X o ,AND o )=||X o -AND o || 2 2 Where L(X o ,Y o ) refers to the second reconstruction loss, Where X o is the ligand occupancy descriptor in the training data, and where Y o is the reconstruction of the ligand occupancy descriptor by the autoencoder.

19. The method of claim 1, wherein the target descriptor comprises a protein target descriptor.

20. The method of claim 1, wherein the target descriptor comprises a protein pocket descriptor.

21. The method of claim 20, wherein the protein pocket descriptor comprises a spatial map of one or more protein targets.

22. The method of claim 1, wherein the target descriptors include features describing at least one of: hydrophobicity, aromaticity, hydrogen bond acceptance, hydrogen bond donation, positive ionization, negative ionization, metallicity, pocket occupancy, hydrogen bond interaction ability, hydrophobic interaction ability, π-π interaction ability, and halogen interaction ability.

23. A method for sampling a ligand, comprising: (a) receiving a target descriptor; (b) generating potential descriptors in an engineered chemical space based at least in part on the target descriptors; as well as (c) generating a ligand descriptor based at least in part on the potential descriptor; Wherein, in the engineered chemical space, a given target potential descriptor of a given target has similarity with a given ligand potential descriptor of a given ligand when the given ligand targets the given target, and has dissimilarity when the given ligand does not target the given target.

24. A method for sampling a ligand, comprising: (a) receiving a target descriptor; (b) generating potential descriptors in an engineered chemical space based at least in part on the target descriptors; as well as (c) generating a ligand descriptor based at least in part on the potential descriptor; Wherein the engineered chemical space is organized at least in part based on spatial information of a plurality of ligands, a plurality of targets, or both.

25. A method for training a machine learning algorithm to learn an engineered chemical space, comprising: (a) providing a neural network, the neural network comprising: i. an input layer, the input layer being configured to receive at least a target descriptor; ii. A latent layer configured to output at least a latent descriptor, wherein The latent layer is connected to the input layer; iii. an output layer configured to output at least a target-ligand interaction descriptor, wherein the output layer is connected to the latent layer; and iv. at least one parameter; (b) providing training data including a plurality of target descriptors, a plurality of target-ligand interaction descriptors, and a plurality of ligand potential descriptors; and (c) training the neural network by: (i) inputting at least the plurality of target descriptors at the input layer of the neural network; (ii) outputting a plurality of output latent descriptors at the latent layer and a plurality of output target-ligand interaction descriptors at the output layer; and (iii) optimizing a plurality of loss functions based at least in part on the plurality of output latent descriptors and the plurality of output target-ligand interaction descriptors by updating at least one parameter of the neural network, such that the neural network learns the engineered chemical space including a unified embedding for at least the plurality of target descriptors and the plurality of ligand latent descriptors.

Images