Molecule screening system

Abstract

A sequence-to-fixed-length transformation is used to process a sequence representation of a molecule structure to yield a fixed-length representation. A transformation of the fixed-length representation yields a prediction of one or more properties of the molecular structure. For example, the molecular structure is a candidate drug and the process described above is part of a screening procedure.

Description

MOLECULE SCREENING SYSTEMCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 535,458, filed on August 30, 2023, which is incorporated herein by reference.STATEMENT AS TO FEDERALLY SPONSORED RESEARCH OR DEVELOP ENT

[0002] This invention was made with government support under R01MH124606 awarded by the National Institutes of Health. The government has certain rights in the invention.BACKGROUND OF THE INVENTION

[0003] This invention relates to using sequence representations of molecular structures in a machine- learning based search for molecules with desired properties.

[0004] Virtual screening of ultra-large (i.e., billion-member) make-on-demand chemical libraries has emerged as a promising strategy for drug discovery. One popular make-on-demand library is the Enamine REAL library, which includes about 5.5 billion drug-like compounds. The compounds are assembled using hundreds of well- validated parallel synthesis protocols, and hundreds of thousands of diverse synthetic building blocks, thus creating a growing chemical space that is ultra-large and structurally diverse, yet synthetically accessible. In general, as more compounds are screened, the true positive rate increases, and the quality of hits (e.g., the binding affinity of a drug candidate to a protein target) improves, highlighting the benefits of screening ultra-large chemical libraries. Despite these advances, most virtual screens done to date are docking-based, and thus have been limited to discovering ligands to known proteins, with accurate structures.

[0005] Machine learning (ML) screens provide a strategy complementary to docking, for tasks beyond protein ligand discovery. ML models can be trained with high-quality experimental data to predict molecular properties, and then the trained models can be used in virtual screening to prioritize molecules for later experimental validation and testing. Compared to docking models, ML models directly mapchemical structures to diverse molecular properties and even complex functions, and therefore have the potential to accelerate many aspects of drug discovery and development. Although many advances have been made to increase the accuracy of ML models for molecular property prediction, no ML approaches have been reported to screen the entire REAL database of ~5.5 billion compounds, potentially because current ML approaches do not scale well to billion-member libraries in terms of cost and time. For example, a graph neural network-based ML method has been reported to screen -107 million molecules in four days If scaled linearly to a 5.5 billion compound library, with the same hardware setup, such a virtual screen would take more than 200 days.

[0006] One existing ML-based screening approach maps simplified molecular- input line-entry system (SMILES) representations of molecules to fixed-length representations (“embeddings”). This approach has previously been used to (1) evaluate a given SMILES representation to directly predict a chemical property and (2) generate candidate SMILES representations for molecules that are expected to have a desired property.SUMMARY OF THE INVENTION

[0007] Approaches described herein make use of a sequence- to-fixed- length (or reduced length) transformation to process a sequence representation to yield a fixed- length (i.e., vector “embedding”) representation, followed by a transformation of the fixed-length representation to yield a prediction of one or more properties of the molecular structure. For example, the molecular structure is a candidate drug and the process described above is part of a screening procedure. In some examples, the particular sequence representation used is a SMILES representation, and the desired properties that are described include therapeutic activity such as inhibition of bacterial growth, inhibition of viral infection, and utility for drug delivery. One example of an approach for screening molecules includes a representation generation step, a taskspecific modeling step, and a task-specific prediction step, as described below.

[0008] Representation Generation: In one approach, a first recurrent neural network (RNN)-based deep learning model is trained to predict the A*hcharacter in a SMILES string based on the string of characters up to the (A-7)thcharacter. For example, for the SMILES string “CC (=0 ) Ocl ccccclC (=0) O”, one step of thetraining uses string “CC (=0 ) ” to predict the character “0”. For each molecule, we enumerate predictions starting from N equals 2 (i.e., predicting the 2ndcharacter based on the 1stcharacter) until N equals the length of the SMILES string (i.e., predicting the last character based on the SMILES string up to the second to the last character). Further details and embodiments of the training of the first RNN-based deep learning model can be found in the attached Appendices.

[0009] Once the first RNN-based deep learning model is trained, it can he used to form a “representation matrix” for any SMILES string. In some examples, the representation matrix for a given SMILES string is formed by extracting and combining (e.g., by averaging) the outputs of RNN layers (e.g., the hidden and / or cell states of the long short-term memory (LSTM) layers).

[0010] Task-Specific Modeling: The first RNN-based deep learning model can then be leveraged to train a task-specific model that is able to predict a biological or chemical property of a molecule based on the molecule’s representation matrix. To do so, a dataset is curated, where each molecule in the dataset has measured values for a biological or chemical property of interest. The first RNN-based deep learning model is then used to generate a representation matrix for each molecule in the dataset. The representation matrices for the molecules in the dataset and their corresponding measured values for the biological or chemical property of interest are used to train a task-specific machine-learning model that is able to predict a value of the biological or chemical property of a target molecule based on the target molecule’s representation matrix.

[0011] Task-Specific Prediction: Finally, to predict a value of the biological or chemical property for a target molecule, a SMILES string for the target molecule is processed in the first RNN-based deep learning model to generate the representation matrix for the target molecule. The representation matrix is then processed in the taskspecific model to predict a value of the biological or chemical property for the target module.

[0012] Predicted Molecules: The system described above has been successfully used to predict molecules with described properties.

[0013] In another general aspect, a method for screening molecules based on a desired property includes forming a representation of a target molecule using arepresentation model, the representation model configured to recurrently process successive parts of a sequential representation of the target molecule to form a sequence of states, wherein forming the representation of the target molecule includes combining the sequence of states and determining the desired property for the target module by processing the representation of the target molecule in a property model trained with respect to the desired property.

[0014] Aspects may include one or more of the following features.

[0015] The representation model may be trained on generative criteria to generate valid molecules. The property model may be trained using representations of molecules known to have the desired property. The representation model may be trained independently of the desired property. The representation model may be configured to generate simplified molecular-input line-entry system (SMILES) representations of molecules. Combining the sequence of states may include averaging the states of the sequence of states.

[0016] In another general aspect, a method for configuring a molecule screening system includes training a representation model on generative criteria to generate valid molecules, the representation model configured to recurrently process successive elements of a sequential representation of a target molecule to form a sequence of states and training a property model with respect to a desired property including using representations of molecules known to have the desired property.

[0017] Other features and advantages of the invention are apparent from the following description and from the claims. A number of embodiments of the invention have been described. Nevertheless, it is to be understood that the foregoing description is intended to illustrate and not to limit the scope of the invention, which is defined by the scope of the following claims. Accordingly, other embodiments are also within the scope of the following claims. For example, various modifications may be made without departing from the scope of the invention. Additionally, some of the steps described above may be order independent and thus can be performed in an order different from that described.DESCRIPTION OF DRAWINGS

[0018] FIG. 1 is a block diagram of a molecule property predictor;

[0019] FIG. 2 is detailed block diagram of a molecule sequence to matrix representation block;

[0020] FIG. 3 is a first example of a recurrent neural network block; and

[0021] FIG. 4 is a second example of a recurrent neural network block.DETAILED DESCRIPTION

[0022] Referring to FIG. 1, a number of approaches described in this document make use of a two-step procedure to transform a representation .S' 102 of a molecule (or in some examples, a set or sequence of molecules) to yield a prediction or estimate of a quantity Z 222 associated with that molecule. In a first step, the representation .S' 102 is transformed using a parameterized transformation, referred to in FIG. 1 as a “SMILES-to-Matrix” transformation 100 (recognizing that not all examples use a SMILES molecule representation). Generally, the molecule representation .S' is in the form of a sequence (s0, , . . . , sN) where the elements of the sequence elements are from a relatively small set of possible symbols. As one example, the molecule representation is a SMILES representation, and the elements of the sequence are individual letters or typographic symbols. This transformation is configured with a set of parameters 0 118. These parameters are determined largely or completely independently of the nature of the estimate or prediction that is the output Z of the two-step procedure. The output of the first step is a “matrix” representation M 122 of the molecule. In general examples, this representation is a multi-dimensional data structure, which can be considered to be a matrix or tensor and for the sake of exposition is referred to as a matrix representation.

[0023] The second step of the two-step transformation illustrated in FIG. 1 takes as input the matrix representation M , and processes it with a second parameterized transformation to yield the desired prediction or estimate Z . This transformation is task-specific, and the structure of the transformation and / or the values of the parameters 228 are specific to the task.

[0024] Referring to FIG. 2, a number of implementations of the approach illustrated in FIG. 1 implement the SMILES-to-Matrix block 100 as two-step procedure in which the molecule representation S 102 is processed by a recurrent computation element 110 (referred to below, as “a RNN block”, recurrent neuralnetwork block) to form an intermediate output X 115, and then that intermediate output is processed by a second part, a matrix encoder 120, to transform the molecule representation S to form the matrix representation M 122. The RNN block 110 is parameterized by the set of parameters 0 118 introduced above. In the illustrated implementation, the matrix encoder 120 is fixed rather than being configured with configurable parameters.

[0025] The upper portion of FIG. 2 shows the RNN 1 10 as an expanded illustration to show repeated uses of a single RNN block 112. That is, while there are N separate RNN blocks 112 shown, these represent repeated uses of the same computational block (i.e. with the same structure and parameters 0 ) in succession. In the illustrated example, the molecule representation S = sl, s2, s3,..., sN) , for example, representing a string starting with “CC(“ and ending in “O”, is processed in sequence. The first symbol s1is processed by the RNN block 112 to yield an internal “state” ] , which is a set of numerical values representing an internal result of the processing of the first symbol. In general, this state comprising a collection of numerical values may be considered to be an array or vector. The RNN block also provides an output 116, which may be a prediction of the next symbol, represented as or alternatively a probability (or other form of) distribution P(s2) of the next symbol (i.e., a list of values corresponding to the different possible next symbols, with the values representing the likelihood or confidence that the corresponding symbols occur next in the molecule representation). This output is not used in the processing of the input S , and may not be computed. As discussed below, this output is used in the process of determining the configurable parameters 0 118. The state .q resulting from processing the first symbol .s' is passed to the RNN block 112 used to process the second symbol. This process is repeated (i.e., as an iterative procedure) until all the input symbols are processed.

[0026] The result of the processing of molecule representation S 102 by the RNN 110 is a sequence of states X = (xl,x2,...,xN) . This sequence may be considered to be a two-dimensional data structure, with one dimension being the dimension of the states q . The size of the other dimension may depend on the length of the molecule representation, or alternatively be fixed, for example, by setting the elements to zero (“zero padding”) beyond the length of the molecule representation.

[0027] Continuing to refer to FIG. 2, the matrix encoder 120 processes the intermediate output X 115 to form the matrix representation M 122. A number of different examples of such an encoder may be used. For instance, the encoder may merely take the average of the inputs ; through xn, or form a concatenation (with or without zero padding) of those inputs. These and other alternatives to the form of the matrix encode are discussed later in this document.

[0028] Referring to FIG. 3, and example of an RNN block 1 12 has two inputs and two outputs. One input 104 is an input sn. This input is processed by a block 310, which may be referred to as an “embedding layer”, for example, a feedforward neural network the output of which may be referred to as an “embedding” of sn. This embedding is passed to a recurrent block 320, which also accepts the state 114 from the previous RNN block 112. The recurrent block 320 outputs the updated state 114 which is passed to the matrix encoder as shown in FIG. 2. A further output of the recurrent block 320 passed to an output block 300, which is used to determine the prediction of the next symbol sn^ . For example, the output block is a multi-layer neural network where the last layer is a “soft-max” layer that produces a probability distribution over the possible next symbols, P(sn+i) . Alternatively, the output block 320 directly generates the prediction of the next symbol, Sn+1.

[0029] Referring to FIG. 4, a specific version of the structure shown in FIG. 3 uses a long short-term memory (LSTM) architecture. In this architecture, the input state 114A and resulting output state has parts, the upper part that is considered to be the “state” of the LSTM block, and the lower which is the output of that block — these, which are combined according to the shown arrangement of neural network block and connecting multiplication (i.e., gating) modules and addition modules.

[0030] Training of the RNN 110 (see FIG. 2) makes use of a large database on molecule representations. For example, millions of molecules of the ChEMBL database are used to update the settable parameters (“weights”) of the neural network blocks shown in FIGS. 3-4. For example, with the output being the probability distribution of next symbols, a cross-entropy (e.g., maximum likelihood) criterion is used with an iterative incremental updating procedure, for example, using a backpropagation procedure. Note that this training procedure can be performed once for a variety of tasks (i.e., it is task-independent). However, there may optionally be situations in which, although independent of a specific task, the training set ofmolecules may be selected according to a particular type on which a variety of tasks may be performed.

[0031] Referring back to FIG. 1, having trained the SMILES-to-matrix module 100 in a task-independent manner, the task-specific predictor 200 is now trained maintaining the setting of the parameter (weights) of the RNN of the SMILES-to- matrix module. For this training, a relatively smaller set of chemicals are selected and labelled according to the property of interest ( ), for example a binary (Boolean) indicator of presence of absence of a property (e.g., toxicity, delivery activity, inhibition of bacterial growth), a numerical indicator of a degree of such a property. In some examples, the predictor is a feedforward deep neural network (e.g., fully connect dense interconnected three or more layers) that treats the matrix representation M 122 as a vector of quantities as input to the first layer without regard to the distinction between a time (i.e., sequence) dimension and a state dimension of the matrix. In other examples, a convolutional structure, for example, implementing convolution in the time dimension. In any case, the configurable parameters of the predictor are trained to predict the labels of this training set.

[0032] In an optional feature, the matrix representation may be augmented (e.g., concatenate with) addition features that may be determined by rule or a machinelearning technique. For example, a chemical descriptor matrix may be calculated using cheminformatics tools such as RDKit for each molecule, and concatenated to the RNN-base matrix described above.

[0033] The approaches described above may be used to predict properties of sets or sequences of molecules. For example, for a set of potentially reacting molecules, their SMILES representations may be concatenated, optionally including a separator symbol between them. Then this concatenation is used in the place of the single molecule smiles representation discussed above. For instance, SMILES representations of a ketone, amine, and isocyanide that are used in a corresponding lipid synthesis reaction are concatenated in arbitrary order of ketone-amine- isocyanide. In this way, rather than predicting the property of a molecule resulting from a reaction, the property is predicted from the precursors of the molecule (i.e., the molecules that react to form the molecule of interest).

[0034] An optional feature of the predictor block is to include an uncertainty model, for instance focal loss, model ensemble, or temperature scaling. Inexperimental evaluations, including such an uncertainty model provided increased robustness in the task-specific models.

[0035] Example applications of the approaches described above for the case of predicting a property of a single molecule were evaluated on task including predicting a molecule’s activity to inhibit bacteria growth, and inhibition of SARS-Cov2 virus infection of mammalian cells. Examples of multiple molecule inputs (i.e., using concatenation) were evaluated on tasks including inhibition of a protein-protein interaction involved in SARS-Cov2 virus infection, and delivery of mRNAs or siRNA into human cells.

[0036] The description above focusses on a screening application in which candidate molecules (or sets of precursors) are evaluated and sorted or a subset selected for further evaluation, for example, by more computationally expensive techniques or by physical evaluation (e.g., chemical synthesis and in vitro experimentation). For instance, screening of a set of target molecules according to the determined indicators of a desired property can include selecting a subset of the set of target molecules for further evaluation and providing said subset for further computational or physical evaluation of the desired property. The task of screening is a technical problem that for which the computational techniques above are a practical solution. The benefits of accurate screening includes reducing the time to discovery of a desired molecule and / or reduction of human or computational resources required to make such a discovery.

[0037] Examples of the approaches described above may be implemented in software, in hardware, or in a combination of software and hardware. Hardware may include application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) and the like. Software includes instructions, which are generally stored on non-transitory machine-readable media. Such instructions may be executed by a physical (i.e., circuit-implemented) processor, such as a general purpose computer processor (e.g., a central processing unit, CPU), a special purpose or application-specific processor, or an attached processor, such as a numerical accelerator, graphics processing unit (GPU). The processor may be hosted in a variety of devices. Some or all of the procedures may be performed at a central computer, for example at a network accessible server in a data center (which may be referred to asbeing in the “cloud”). Training procedures can be performed on different computers than the runtime prediction steps.

[0038] A number of embodiments of the invention have been described.Nevertheless, it is to be understood that the foregoing description is intended to illustrate and not to limit the scope of the invention, which is defined by the scope of the following claims. Accordingly, other embodiments are also within the scope of the following claims. For example, various modifications may be made without departing from the scope of the invention. Additionally, some of the steps described above may be order independent, and thus can be performed in an order different from that described.

Claims

WHAT IS CLAIMED IS:

1. A method for computational screening of molecules based on a desired property, the method comprising, for each target molecule of a set of target molecules: forming a representation of the target molecule using a computational representation model, the representation model configured to recurrently process successive parts of a sequential representation of the target molecule to form a sequence of states, wherein forming the representation of the target molecule includes combining the sequence of states; and determining an indicator of the desired property for the target molecule by processing the representation of the target molecule in a property model trained with respect to the desired property.

2. The method of claim 1 , screening the set of target molecules according to the determined indicators of the desired property, including selecting a subset of the set of target molecules for further evaluation and providing said subset for further computational or physical evaluation of the desired property.

3. The method of claim 2, wherein the desired property includes a property selected from the group consisting of predicting an activity to inhibit bacteria growth, an inhibition of SARS-Cov2 virus infection of mammalian cells, inhibition of a protein-protein interaction involved in SARS-Cov2 virus infection, and delivery of mRNAs or siRNA into human cells.

4. The method of claim 1 , further comprising storing the formed representations of the molecules of the set of target molecules in a database, and performing the step of determining the indicator a plurality of times for different desired property, including for each of the different desired properties retrieving the stored representations of the molecules for processing with different respective models trained with respect to the different desired properties.

5. The method of claim 1, wherein the sequence representation comprises a simplified molecular- input line-entry system (SMILES) representations of molecules.

6. The method of claim 1 , wherein forming the representation model includes inputting the sequential representation into a recurrent neural network (RNN) model, and the sequence of states comprises a sequence of states of the RNN model determined during processing to the sequential representation.

7. The method of claim 6, wherein the RNN model comprises a long short-term memory (LSTM) model.

8. The method of claim 6, further comprising selecting a first corpus of modules, and training parameters of the RNN model to match sequential characteristics of sequence representations of molecules of the first corpus, said first corpus being selected independently of presence or absences of the desired property of the molecules of said corpus.

9. The method of claim 8, wherein the processing of the representation of the target molecule comprises inputting said representation into a neural network, said neural network comprising a plurality of fully connected layers, and outputting from said neural network the indicator of the desired property.

10. The method of claim 9, further comprising selecting a second corpus of molecules, each molecule of said second corpus having an associated indicator of the desired property, and training parameters of the neural network model to match the indicators of the desired property of molecules of the second corpus, said training being conducted for a fixed configuration of the RNN model.

11. The method of claim 1, wherein forming the representation of the target model includes concatenating sequence representations of precursor molecules of said target molecule.

12. The method of claim 11, wherein the indicator of the desired property represents an indicator that a reaction of the precursor molecules yields a target molecule with said desired property.

13. The method of claim 1, wherein forming the representation of the target molecule further includes concatenating a descriptor matrix of the target molecule to the combination of the sequence of states.

14. A software-implemented molecule screening system comprising a non- transitory machine-readable medium having instructions stored thereon, said instructions when executed by a processor of said system causing the system to perform operations including: forming a representation of the target molecule using a representation model, the representation model configured to recurrently process successive parts of a sequential representation of the target molecule to form a sequence of states, wherein forming the representation of the target molecule includes combining the sequence of states; and determining an indicator of the desired property for the target molecule by processing the representation of the target molecule in a property model trained with respect to the desired property.

15. A software-implemented molecule screening system comprising: means for forming a representation of the target molecule using a representation model, the representation model configured to recurrently process successive parts of a sequential representation of the target molecule to form a sequence of states, wherein forming the representation of the target molecule includes combining the sequence of states; and means for determining an indicator of the desired property for the target molecule by processing the representation of the target molecule in a property model trained with respect to the desired property.

Images