Molecular directed evolution by iterative experiment and machine learning

Abstract

Described herein are platforms, systems, media, and methods for machine learning driven iterative drug discovery processes. Some aspects include receiving a first input data set including first binding interaction information between a target molecule and a library of compounds; processing the first input data set using a machine learning module to generate a model representation of the binding interaction, wherein the representation is configured to predict binding between the target molecule and the input compound; determining an updated library of compounds using the model representation of the binding interactions, wherein the updated library of compounds comprises one or more new compounds predicted to bind to the target molecule; receiving a second input data set comprising second binding interaction information between the target molecule and the updated library of compounds; and processing the second input data set using the machine learning module to update the model representation in conjunction with the interaction, wherein the prediction accuracy of the updated model representation is improved.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Application No. 63 / 320,137 filed on March 15, 2022 and U.S. Provisional Application No. 63 / 320,143 filed on March 15, 2022, the entire contents of which are incorporated herein by reference. Background Art

[0003] Drug discovery is an expensive and time-consuming process in which potentially millions of candidate small molecule compounds are screened for potential drug applications. Candidates can be identified by a number of methods, including high-throughput screening (HTS), in which compounds are tested collectively using biochemical assays individually, and in silico screening, in which target binding is evaluated by modeling and / or calculating molecular interactions. HTS is practically limited to a maximum of 2-3 million compounds, but fewer are typically used. At the same time, computational screening requires prior knowledge of the crystal structure of the target and information about where the compound should bind on the target, and often fails to accurately describe how the compound will behave in the real world. Summary of the invention

[0004] Disclosed herein is a platform, system and method for improving drug discovery using a DNA-encoded chemical library by directed molecular evolution using an iterative machine learning process. Therefore, clinically useful small molecules can be identified as targets and then rapidly optimized through an iterative process until the final candidate molecules are ready for clinical testing. The advantages of the present disclosure include the ability to identify molecules of various target classes, find molecules for targets that are difficult to deal with (i.e., undrugable) for others, rapidly optimize compounds to confirm effects in vivo, and effectively bring multiple compounds into the clinical preparation stage. Rather than optimizing compounds by slowly iterating dozens of compounds per month, the present disclosure is able to create large data sets at every stage of the drug discovery cycle. By using large experimentally analyzed data sets (rather than on a computer), multi-parameter optimization (e.g., in a single experiment under a wide range of conditions) can be effectively performed on complex multi-conditional tasks (such as ADME / Tox).

[0005] Advantages of the present disclosure combine DNA encoded libraries (DELs) with machine learning to provide an improved screening process. The use of DELs provides advantages over high throughput screening (HTS) and computer calculation methods. For example, DELs allow the processing of compounds that are 1,000 times that of competitive methods (billions of compounds, while HTS and traditional screening process millions of compounds at most). DELs can be effectively processed in parallel mini-experiments ("conditions") that are 10-100× of competitive methods, while HTS usually tests 1-2 conditions at a time due to the large amount of resources required to set each additional condition. In some cases, every combination of available building blocks is completed in DELs, allowing exhaustive searches for compounds, while HTS is heavily biased towards methods that have been effective in the past, limiting the search space. For example, DELs do not require crystal structures or specific binding modes of action (unlike computer virtual / computational screening). Instead, DELs can detect compounds that bind anywhere on proteins, such as allosteric binders, hidden binding pockets, and compounds that may lay a good foundation for bispecific molecules (e.g., PROTACs and other proximity inducers, such as molecular glues).

[0006] Therefore, DEL can be used to identify compounds that bind to completely new targets that are currently considered "undruggable" because HTS does not cast a wide enough net to find them. As a result, the DEL approach generates rich, dense datasets filled with internal controls that are well suited for machine learning (ML) while having lower costs and shorter screening times than traditional HTS.

[0007] Compared to conventional HTS and DEL approaches that typically track dozens of compounds at a time, the just-in-time platforms, systems, and methods offer further advantages by integrating machine learning to generate ML-designed DELs.

[0008] In one aspect, the present invention provides a method comprising: (a) a computer-implemented method comprising: (i) receiving a first data set, the first data set comprising: a first compound descriptor for each compound in a first compound library and a compound fitness score for each compound in the first compound library; (ii) training a prediction model on the first data set; (iii) inputting a second data set into the model, the second data set comprising a second compound descriptor for each compound in a second compound library; and (iv) generating a compound fitness score for each compound in the second compound library from the prediction model using at least one or more compound descriptors of the first compound library and / or one or more compound descriptors of the second compound library, and (b) selecting a third compound library based on information including one or more compound fitness scores of the second compound library and / or one or more compound fitness scores of the first compound library. In some embodiments, the third compound library includes: (i) compounds from the second compound library, (ii) compounds from the first compound library, (iii) compounds including two or more compounds from the second compound library, (iv) compounds including two or more compounds from the first compound library, (v) compounds including compounds from the second compound library and compounds from the first compound library, (vi) compounds not present in the first compound library and the second compound library, (vii) compounds including compounds from the second compound library and compounds not present in the first compound library and the second compound library, (viii) compounds including compounds from the second compound library and compounds not present in the first compound library and the second compound library, or (ix) (i) to (viii) two or more combinations of. In some embodiments, the first library is a first DNA-encoded library (DEL) and / or the second library is a second DNA-encoded library. In some embodiments, step (b) is part of a computer-implemented method. In some embodiments, wherein step (b) is not part of a computer-implemented method. In some embodiments, step (b) includes a first sub-step and a second sub-step, the first sub-step is part of the computer-implemented method, and the second sub-step is not part of the computer-implemented method, wherein the first step and the second step are performed sequentially, and the first sub-step is performed first or the first sub-step is performed second.

[0009] In some embodiments, the information also includes an evaluation score (sometimes referred to as an external fitness score) of the compound in the second compound library and / or an evaluation score (sometimes referred to as an external fitness score) of the compound in the first compound library. In some embodiments, the evaluation score of the compound in the second compound library is a second fitness score generated independently of the compound fitness score generated by the computer-implemented method. In some embodiments, the evaluation score of the compound in the first compound library is a first fitness score that is different from the compound fitness score of the compound in the first compound library.

[0010] In some embodiments, one or more compounds in the first library are the first test compound (sometimes referred to as all products or all products compound), the building block of the first test compound, the first byproduct (sometimes referred to as byproduct or byproduct) generated during the synthesis of the first test compound, or the intermediate (sometimes referred to as intermediate) generated during the synthesis of the first test compound, or a combination of two or more thereof. In some embodiments, the first test compound is the desired product of the synthesis reaction of two or more building blocks in the building block of the first test compound. In some embodiments, the first byproduct is the undesirable product of the synthesis reaction of two or more building blocks including the first test compound. In some embodiments, the one or more compounds in the second library (or optionally, the subsequent library applicable in the iterative method) are the second test compound (sometimes referred to as all products or all products compound), the building block of the second test compound, the second byproduct (sometimes referred to as byproduct or byproduct) generated during the synthesis of the second test compound, or the intermediate (sometimes referred to as intermediate) generated during the synthesis of the second test compound, or a combination of two or more thereof. In some embodiments, the second test compound is the desired product of the synthesis reaction of two or more building blocks in the building block of the second test compound. In some embodiments, the second byproduct is the undesirable product of the building-up reaction of two or more building blocks including the second test compound.In some embodiments, one or more compounds in the third storehouse are the third test compound (sometimes referred to as full product or full product compound), the building block of the third test compound, the third byproduct (sometimes referred to as byproduct or byproduct) generated during the synthesis of the third test compound, or the intermediate (sometimes referred to as intermediate) generated during the synthesis of the third test compound, or two or more combinations thereof.In some embodiments, the third test compound is the desired product of the building-up reaction of two or more building blocks in the building block including the third test compound.In some embodiments, the third byproduct is the undesirable product of the building-up reaction of two or more building blocks including the third test compound.In some embodiments, full product includes three synthons (trisynthon), and intermediate product includes two synthons (disynthon) and / or single synthons (monosynthon).

[0011] In some embodiments, the first compound descriptor includes data or information associated with a compound in the first compound library, wherein the data or information includes binding affinity of the compound to a target molecule (optionally, wherein the target molecule is a drug target such as a protein or nucleic acid), activity of the compound, physical properties of the compound (e.g., lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing reads associated with the abundance of the compound in an experiment, compound structure, information associated with the synthesis of the compound, labeling data, process quality control data, yield associated with the synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuity data, structural data, or building block validation data, or a combination of two or more thereof. In some embodiments, the second compound descriptor includes data or information associated with a compound in the second compound library, wherein the data or information includes the binding affinity of the compound to a target molecule (optionally, wherein the target molecule is a drug target such as a protein or nucleic acid), the activity of the compound, the physical property of the compound (e.g., lipophilicity), the toxicity of the compound, the stability of the compound, the permeability of the compound, sequencing reads associated with the abundance of the compound in the experiment, the compound structure, information associated with the synthesis of the compound, labeling data, process quality control data, yield associated with the synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuity data, structural data, or building block verification data, or a combination of two or more thereof. In some embodiments, the compound is an all-product compound, an intermediate compound, or a by-product compound.

[0012] In some embodiments, the method comprises testing one or more of the compounds in the first compound library in an in vitro or in vivo assay. In some embodiments, the method comprises testing one or more of the compounds in the third compound library in an in vitro or in vivo assay.

[0013] In some embodiments, each compound in the third compound library includes a nucleic acid tag or is synthesized to include a nucleic acid tag, and the method further includes sequencing the third compound library to generate sequencing data associated with the third compound library. In some embodiments, the information of step (b) includes sequencing data associated with a compound external library (e.g., a library including nucleic acid tags from each compound in the first library and / or the second library). In some embodiments, the compound fitness score of each compound in the first compound library is generated from data including sequencing data associated with the first compound library. In some embodiments, the sequencing data includes read counts, quality scores associated with read counts, and / or includes scores calculated according to sequencing read counts or read count sets of different experimental conditions from the first compound library and / or the second compound library. In some embodiments, the score includes read counts, or read counts divided by the total number of reads in the compound selection or the average number of reads in the compound selection, or (directly or indirectly) using a similar mathematical function of read counts.

[0014] In some embodiments, at least one compound fitness score of each compound in the first compound library is generated from data including a first compound descriptor (e.g., sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuous data, structural data, or building block verification data, or a combination of two or more thereof). In some embodiments, the prediction model uses a probabilistic framework to process the first data set and the second data set, and outputs a compound fitness score for each compound in the second compound library. In some embodiments, the fitness score is generated at least in part from data including non-target counts, target counts, and / or product ratio adjustment values ​​from all product compounds. In some embodiments, the fitness score is generated at least in part from data including non-target control counts, target counts, and / or product ratio adjustment values ​​from intermediate product compounds. In some embodiments, the method includes utilizing sequencing data associated with sequencing a third compound library to generate a compound fitness score for each compound in a third compound library.

[0015] In some embodiments, the method includes determining a third compound library. In some embodiments, determining includes binding the third compound library to a target. In some embodiments, determining includes sequencing a third compound library or a subset of the third compound library (e.g., wherein the subset is a subset of compounds that bind to a target).

[0016] In some embodiments, the fitness score of any of the compounds comprises a binding and / or activity score of the compound.In some embodiments, the third library comprises one or more compounds from the second library having a compound fitness score greater than a threshold score.

[0017] In some embodiments, the method includes preprocessing the first data set and / or the second data set. In some embodiments, the preprocessing step is performed before step i and / or step iii of the computer-implemented method. In some embodiments, the method includes refining the fitness score generated from the prediction model, optionally, wherein the refinement is performed by the prediction model, and / or optionally, wherein the refinement includes merging information from an external library (e.g., a nucleic acid tag library associated with the first compound library and / or the second compound library).

[0018] In some embodiments, the second library includes one or more compounds that are different from the first library. In some embodiments, the second library includes one or more compounds that are different from the first library.

[0019] In some embodiments, the method includes repeating steps ii-vi to update the model. In some embodiments, steps (a)-(b) are iteratively repeated to identify a set of potential compounds having one or more desired properties. In some embodiments, steps (a)-(b) are iteratively repeated to identify a set of potential compounds having one or more desired compound fitness scores.

[0020] In some embodiments, the compound suitability score is related to oral drug solubility, intestinal absorption, permeability, hERG toxicity, CYP inhibition, blood-brain barrier permeability, P-glycoprotein activity, plasma protein binding and / or binding affinity of any of the compounds.

[0021] In some embodiments, the first compound descriptor input into the model includes compound structure and / or experimental data. In some embodiments, the predictive model is a machine learning model.

[0022] In some embodiments, the machine learning model includes a neural network. In some embodiments, the neural network is a graph neural network. In some embodiments, the machine learning model includes a graph neural network and an attention layer. In some embodiments, the neural network is a graph attention network.

[0023] In some embodiments, the method includes performing a validation assay on at least one compound in the third compound library. In some embodiments, the method includes performing a low-throughput analysis on at least one compound in the third compound library. In some embodiments, the method includes inputting a third data set including a compound descriptor and a compound fitness score for each compound in the third compound library into a secondary system in the validation assay. In some embodiments, the method also includes inputting data from the validation assay into a prediction model. In some embodiments, the validation assay includes binding or biochemically active proxies (proxy), binding or biochemical activity includes absorbance, fluorescence, luminescence, radioactivity, NMR, crystallography, microscopy including cryo-electron microscopy, mass spectrometry or Raman scattering One or more, for example, surface plasmon resonance (SPR) measures the reflection of polarized light, which can detect changes in the reflection angle (refractive index), and the ligand (compound) is fixed or bound to a surface (which includes a fixed target protein) to affect the quality or thickness of the surface, which changes the refractive index.

[0024] In some embodiments, the predictive model generates a predicted compound descriptor for each compound in the first compound library and / or the second compound library. In some embodiments, the compound fitness score is generated at least in part from the predicted compound descriptor for each of the compounds.

[0025] In some embodiments, the first compound library is about 10,000 compounds to about 100 billion compounds or about 10,000 compounds to about 10 billion compounds; and wherein the second compound library is about 10,000 compounds to about 100 billion compounds or about 10,000 compounds to about 10 billion compounds.

[0026] In one aspect, the present invention provides 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 perform the method described herein.

[0027] In one aspect, provided herein is a non-transitory computer-readable storage medium encoded with a computer program comprising instructions executable by one or more processors to perform the methods described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] 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 referring to the following detailed description and accompanying drawings which set forth illustrative embodiments in which the principles of the present disclosure are utilized, in which:

[0029] Figure 1 A flow chart showing a method for iteratively screening a DNA-encoded compound library to determine target binding affinity and using the data to train a machine learning model to generate updated compound predictions for the compound library according to one or more embodiments disclosed herein;

[0030] Figure 2 shows a schematic representation of three synthetic steps in a combinatorial chemistry approach for library generation according to one or more embodiments disclosed herein;

[0031] Figure 3 A comparison between the traditional compound screening process and DNA-encoded libraries that allow for large-scale screening is shown;

[0032] Figure 4 and Figure 5 A schematic diagram of the pooling and splitting process used to further expand the size of the DNA-encoded library is shown;

[0033] Figure 6 An illustrative process for a binding assay experiment for identifying compounds that bind to a target molecule within a combinatorial library of DNA-encoded compounds is shown;

[0034] Figure 7 An illustrative example of data from a binding assay experiment that can be converted into input for a machine learning model to generate a representation of the binding interaction between a compound and a target molecule is shown;

[0035] Figure 8 An illustrative diagram is shown that maps the compounds in the library to known products, known intermediates, and unknown compounds, which can be used for quality control;

[0036] Fig. 9 shows mass spectrometric data of compounds in a reaction mixture before and after a chemical synthesis step occurs, which can be used to determine the efficiency of the synthesis for quality control purposes;

[0037] Fig.10 Reference mass spectral data for the library region are shown;

[0038] Fig.11 Compound fractions corresponding to the series of chemical synthesis steps used to construct the DNA-encoded library are shown;

[0039] Fig.12 shows a plot of sequencing read counts plotted against the number of compounds when comparing the DNA-encoded libraries disclosed herein to traditional DELs, wherein the disclosed DNA-encoded libraries achieve a higher dynamic range by having the same sequencing depth for a small number of molecules;

[0040] Fig.13An overview of the iterative DEL library process using the deep neural network architecture disclosed herein is provided;

[0041] Fig.14 Provides an illustration of a graph convolutional neural network architecture configured to receive input data and generate predicted compound ratings and binding scores according to one or more embodiments herein;

[0042] Fig.15 A star graph plotting cLogP, molecular weight (MW), hydrogen bond acceptor (HBA), hydrogen bond donor (HBD), rotatable bond (RB), topological polar surface area (TPSA), and SP3 fraction (fSP3) is shown;

[0043] Fig.16 A non-limiting example of a computing device is shown; in this case, the computing device is a device having one or more processors, memory, storage, and a network interface;

[0044] Fig.17 A non-limiting example of a web / mobile application providing system is shown; in this case, a system providing a browser-based and / or native mobile user interface;

[0045] Fig.18 A non-limiting example of an application provisioning system that may alternatively be a distributed, cloud-based architecture is shown;

[0046] Fig.19 Comparison of compound rankings based on predicted binding scores generated by the initial baseline model and an updated model with an evolved DNA-encoded library validated against true positive hits is shown;

[0047] Fig. 20 Shown is a multi-parameter score comparison between compounds in the initial DNA-encoded library and compounds identified in the first evolved DEL using the first evolved model (ML1) after the first iteration and compounds identified in the second evolved DEL using the second evolved model (ML2) after the second iteration;

[0048] Fig.21 Predicted compound binding ratings for two highly related protein domains (95% sequence similarity) within a single protein are shown, with compound scores for domain 1 plotted on the Y-axis and compound scores for domain 2 plotted on the X-axis. Each compound is represented by a blue dot and the score is derived directly from the DEL data; and

[0049] Fig. 22showing predicted compound binding evaluations for two related proteins in the same class of chromatin regulators, with many compounds predicted to specifically bind only one of the two proteins; and

[0050] Fig.23 Non-limiting examples of model architectures for predicting enrichment and adjustment values ​​for trisynthons and disynthons are shown;

[0051] Fig.24 A bar graph showing the loss and R on the test data for the full ML model compared to the other models is shown. 2 score;

[0052] Fig.25 shows a bar graph showing the area under the ROC curve for 150 external molecules tested using the full ML model and various other models; and

[0053] Fig.26 A bar graph showing the hit rates of the top 10 exomolecules out of 150 exomolecules tested using the full ML model and the flat yield model is shown. DETAILED DESCRIPTION

[0054] Disclosed herein is a platform, system and method for compound discovery using a DNA-encoded chemical library (referred to as "DEL") in an iterative process using artificial intelligence. Compound discovery can involve an integrated process that combines DEL design, experiment, calculation / ML analysis and subsequent experiments to achieve a fast feedback loop between experiment and calculation. The end-to-end integration of the entire iterative process allows higher quality data to be generated with greater scale, flexibility, specificity and speed. In addition, the ML-assisted iteration of DEL can include dynamic reconfiguration of the building block selection with each iteration, which generates richer and more useful data. The ML model can then select building blocks and synthesis schemes that have almost no overlap with the original library. This can help ensure that only the best compounds are built, even if these compounds cannot be tested in the source library. In some cases, the generation of libraries is more flexible and can allow for more diverse library synthesis schemes. For example, library design, synthesis schemes and production methods that would otherwise be uneconomical but can produce high-value structures become feasible because ML-guided synthesis decisions are more likely to succeed and allow for more focused efforts.

[0055] DNA Encoded Library

[0056] DEL is a collection of compounds, which can (but not necessarily) be stored in a single tube, wherein the compounds in the tube are each physically connected to a unique DNA sequence representing the compound (which can be used as a barcode for identifying the compound). In some cases, as used herein, the DNA sequence, structure, synthesis or other data or information associated with the compound can be referred to as a compound descriptor. In some embodiments, the compound descriptor includes data or information associated with the compound in the compound library. Data or information may include the binding affinity of the compound to the target molecule (optionally, wherein the target molecule is a drug target such as a protein or nucleic acid), the activity of the compound, the physical properties of the compound (e.g., lipophilicity), the toxicity of the compound, the stability of the compound, the permeability of the compound, the sequencing reads associated with the abundance of the compound in the experiment, information associated with the synthesis of the compound, labeling data, process quality control data, or yield associated with the synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, mixed data, structural data, or building block verification data, or a combination of two or more thereof. By creating connections between up to billions of compounds and their corresponding unique DNA sequences, a single DEL tube effectively miniaturizes a collection of compounds that would otherwise take up an entire warehouse.

[0057] DELs can be constructed by combinatorial chemistry, in which chemical building blocks (e.g., monomers) are assembled together onto the ends of DNA fragments, thereby synthesizing various coding combinations in a single tube. Each time a new building block is added, a new DNA segment is also added, and the DNA acts as a barcode for the building block. Therefore, the DNA barcode corresponding to the compound grows as new building blocks are added, so that the sequence of the DNA barcode can decode the identity and order of addition of the building blocks. When a DNA sequence from a DNA encoding library is given, the key that maps the DNA barcode to the building block can be used to determine what compound (i.e., compound structure) has been prepared.

[0058] like Figure 6 As shown, traditional HTS isolates individual compounds, typically one per well, spatially "labeling" the compounds with their well positions. Each of those compounds can be tested one at a time, typically in an automated fashion, to see if the compound binds or otherwise interacts with the target. Such compounds can be selected for further optimization. This approach is practically limited to low single-digit numbers of compounds, and small numbers of measurements of each compound. In contrast, DEL can label each compound with a DNA sequence that can later be decoded using a DNA sequencer. This allows the compounds to be mixed in a single tube without compromising the ability to identify each compound later. As Figure 6 As shown, a mixture of compounds can be filtered by exposing it to an immobilized protein and observing which DNA sequences are recovered as binders to the target. This can allow billions of compounds to be tested in parallel under hundreds of different scenarios (i.e., different proteins or different versions of proteins). This helps generate large amounts of dense data for machine learning purposes.

[0059] In some embodiments, the DEL library is configured to have a size within an optimized range. One of the advantages of a DEL library compared to high throughput screening is that it can screen hundreds of millions of compounds instead of millions of compounds. In some cases, the DEL library has at least 100 million, 200 million, 300 million, 400 million, 500 million, 600 million, 700 million, 800 million, 900 million, 1 billion, 2 billion, 3 billion, 4 billion, 10 billion uniquely encoded compounds.

[0060] However, while larger DEL libraries can theoretically allow for testing of a larger chemical space, overly large libraries can include too many non-drug compounds, poor chemical fidelity, and inherently poor signal-to-noise ratios, all of which can contaminate downstream analysis with useless and / or uninterpretable data. Several parameters that can control the DEL signal-to-noise ratio include the number of copies of each individual molecule at the beginning of the experiment, the number of selection rounds performed, and the number of sequencing reads that the sequencer can produce. In an overly large library, each individual compound may be represented by only a few copies in the tube, and after several rounds of selection, many compounds may no longer exist, while the remaining compounds may be represented by fewer copies. The sequencer may output only a fixed number of reads. By starting with more possible compounds, each individual compound receives only a small number of reads, making it difficult to distinguish the performance of one compound from that of another. Therefore, as the DEL library increases, especially beyond a certain threshold, the DEL library may become increasingly noisy. In some cases, the DEL library has no more than 100 million, 200 million, 300 million, 400 million, 500 million, 600 million, 700 million, 800 million, 900 million, 1 billion, 2 billion, 3 billion, 4 billion, 5 billion, 10 billion, up to or more than 100 billion uniquely encoded compounds. In some aspects, the methods and systems herein utilize or include multiple DELs, e.g., as input or output of a model, and thus, the size of each DEL can vary.

[0061] In addition, conventional DEL methods can use relatively small libraries to provide clearer signals, but they do so at the expense of their ability to explore chemical space extensively. In contrast, the platform, system and method disclosed herein can enable the computational design of efficient libraries to cover large-scale chemical space, although making the total compounds downstream less. For example, an initial DEL library set can be constructed and tested whether it is combined with a given target. The resulting data set can be composed of a smaller set of compounds with different binding affinities to the target and a larger set of compounds without obvious binding. The compounds that bind to the target in the initial library will inevitably be much less diverse than the initial library due to their small number. In addition, any compound that is chemically similar to these initial compounds (e.g., via molecular fingerprints and Tanimoto distance) will be difficult to expand this relatively narrow chemical space. Although smaller libraries can more easily and reliably identify useful compounds, the number of those compounds is more limited.

[0062] In some embodiments, the training methods and models described herein can be used to train machine learning (ML) algorithms to build ML models that have a preliminary understanding of the chemical space suitable for binding to the target by utilizing the higher signal-to-noise ratio of the data set, thereby solving the inherent limitations described above. A preliminary understanding of the model and the ability to train the model on the data set (which provides a more comprehensive description of the compounds in the library (e.g., DEL)) allows the ML model to explore matters beyond the chemical diversity of the original hits and design a new library that can share a smaller overlap in the synthetic plan and / or building block (e.g., compound) present in the original library, where the ML model can use the binding to the target as the main criterion. Compared to the initial positive examples and / or their derivatives, the resulting library can be more diverse in structure, while more focused on chemical space regions rich in compounds that bind to the desired target. This enables the ML model to determine which building blocks and / or library designs are most effective for a given target, thereby allowing the ML model to more effectively identify more useful compounds, such as by generating fitness scores for compounds while testing fewer total compounds. In some embodiments, one or more compounds (e.g., building blocks) and / or libraries in an ML-designed synthesis can be in an initial library. In theory, a library containing all possible compounds (e.g., building blocks) can be constructed, but the cost of constructing such a library is impractically expensive and would require 100 or 1000 times more material and 100 or 1000 times more sequencing and analysis capabilities to achieve the same signal-to-noise ratio and identify useful compounds.

[0063] The platforms, systems, and methods disclosed herein are able to exploit the superior signal of smaller libraries combined with machine learning-driven iterative design of smaller libraries to more efficiently search chemical space and can provide a clearer understanding of the chemical space of a given target. Even in the case of the largest practical initial library using conventional methods and models, the ML models trained and described herein can help create and use smaller high signal-to-noise libraries to more efficiently search for compounds, while still providing a clearer understanding of chemical space in view of the target. In some embodiments, the signal-to-noise ratio can be the percentage of compounds that are confirmed to work in the secondary assay among the total compounds tested in the secondary assay.

[0064] In some embodiments, by adding building blocks or compounds 211 to a tube containing a substrate, the Figure 2 A library (e.g., DEL) is constructed as shown. A building block 211 can be coupled to a DNA fragment 212 having a unique sequence representing the building block 211. This is performed during a single synthesis cycle 210. In some embodiments, a second cycle 220 can be performed, in which new building blocks 221 and DNA tags 222 can be added to the chain from the first cycle 210. In some embodiments, a third cycle 230 can be performed on the new building blocks to form a longer chain 231 that includes a new DNA tag 232 each time. In some embodiments, more than three cycles can be performed, and these processes can be repeated until one or more desired DELs are constructed.

[0065] The size of the library (e.g., DEL) can be further expanded using split-pool synthesis methods, such as Figure 4-Figure 5 In some embodiments, Figure 3 As shown, the substrate is divided into several tubes, but more tubes (e.g., separators) can also be used. A different building block 311, 321, 331 is added to each tube. A corresponding DNA tag that labels each building block 312, 322, 332 can also be added to each tube. In some embodiments, each tube can undergo the following steps: Figure 1 The results of each tube can be combined together to create a mixture 340 of all three resulting chemical species.

[0066] In some embodiments, the split pool DEL synthesis method may further include a second step, such as Figure 5 As shown. Figure 4The compound mixture 340 obtained in the first step shown can be divided into multiple tubes (e.g., 3 tubes). In some embodiments, more than three tubes can be used. Different building blocks 351, 361, 371 can be added to each tube. Corresponding DNA tags that label each building block 352, 362, 372 can also be added to each tube. In some embodiments, each tube undergoes the following steps: Figure 1 The second cycle described in the second step 220 of . The results can be combined together to create a mixture 380 of all three resulting chemical species. Figure 4-Figure 5 As shown, by starting with mixture 340, the resulting mixture 380 of the second step can be a combination of all possible building blocks at the first position and the second position. This results in 3*3=9 different compounds in mixture 380. In some embodiments, the process can be repeated additional times to create more complex mixtures. For example, additional steps will result in 3*3*3=27 compounds.

[0067] Library construction (e.g., DEL construction) can include one or more of several construction techniques, such as: building block combination, scaffold-directed (i.e., modification of the core scaffold), or in situ ring formation (e.g., formation of diverse heterocycles). In some embodiments, the DEL library is constructed to be greater than a threshold percentage that meets Lipinski's Rule of 5 (a thumb rule for evaluating compounds as drug candidates). Specifically, Lipinski's Rule of 5 requires: a maximum of 5 hydrogen bond donors (NH and OH bonds); a maximum of 10 hydrogen bond acceptors (nitrogen or oxygen atoms only); and a molecular weight of less than 500 Da; an octanol-water partition coefficient (log P) of no greater than 5. In some cases, the DEL library is constructed to be greater than 60%, 65%, 70%, 75%, or 80% compliant with the Rule of Five. In some embodiments, the library is screened to ensure that no more than a threshold percentage of compounds are unknown. The compounds (e.g., building blocks) used to generate the DEL or a portion thereof can be evaluated to determine the amount of the desired product, the amount of known intermediates and byproducts, and the amount of unknown products. In some embodiments, when the percentage of unknown compounds exceeds a threshold amount, the build unit will not proceed to production for DEL build. For example, if the percentage of unknown compounds exceeds 1%, 2%, 3%, 4%, 5%, 6%, 7%, 8%, 9%, 10%, 11%, 12%, 13%, 14%, 15%, 16%, 17%, 18%, 19%, or 20% or more, the build unit may fail testing.

[0068] In some embodiments, extensive in-library quality control can be performed. In a non-limiting example, spike-ins can be performed to monitor the non-attachable headpiece substrate entering each synthetic well to monitor individual conversions on the low molecular weight headpiece region of the molecule, which can be used to give better confidence in the high molecular weight region when conversions of small molecular weights may be difficult to monitor on large molecules of 30+ kDa.

[0069] DELs allow batch assays to be performed, rather than individual screening, which requires more time and resources. In a non-limiting example of a process 600 for isolating compounds of interest, a DEL 630 containing billions of compounds can be provided in a single tube or well for use in a target binding assay. For example, a protein of interest 613 (e.g., a "target") can be immobilized on a support 611, and the entire DEL 630 is incubated with the immobilized protein 613, which allows compounds 612 from the DEL 630 to bind to the protein 613 and / or the support 611 to form a resulting complex 610. The complex 610 can be washed (i.e., with a buffer) to remove the support and / or unwanted weak binders or non-specific binders. The remaining compounds can be eluted 620 to release the bound compounds 612 bound to the protein 613. The process 600 can be repeated with the eluted compounds and reapplied for another round 640. The DNA tags in the resulting mixture can be amplified (e.g., via PCR) and sequenced, giving a list of sequences that bind to the target protein (representing compound structures), as well as the frequency with which those sequences appear in the tube. Since the compounds (e.g., building blocks) and their corresponding DNA segments are predetermined, compounds that bind to the protein can be identified by the sequences detected in the binding compounds. Compounds that successfully bind to the target protein (but not to negative controls such as supports) can be suitable candidates as a starting point for drug design.

[0070] In some embodiments, a DEL experiment can seek to identify which compounds in the library are binding to a protein of interest. A DEL experiment may require one or more experimental conditions (e.g., each tube can be considered a mini experiment or "condition"). In some embodiments, the results of the conditions tested in a DEL experiment can be compound descriptors. In some embodiments, a simple DEL experiment can include two conditions: a binding assay with a protein of interest, and another binding assay without a protein used as a control. Because the condition tests the entire DEL of the compound, each condition creates its own large-scale data set for millions or even billions of compounds. In some embodiments, the data set can be a compound descriptor. In some embodiments, the data set can be used to generate a compound fitness score. This means that, compared to other methods such as high-throughput screening (HTS) that perform personalized assays on a massively parallel scale, DEL miniaturizes large-scale experiments into a single tube at a low cost. In a non-limiting example, DELs allow experiments to be conducted under multiple conditions, rather than studying a single condition, to ask many questions about a compound, such as information about the compound or properties of the compound (sometimes referred to as compound descriptors), including but not limited to: affinity (at high or low protein concentrations), specificity (towards mutant versions of a protein or closely related proteins that may be members of the same protein family as the target), and / or binding location (mutating a known binding pocket or adding a known competitive binder to the target to the mix). In some embodiments, a compound descriptor includes data or information associated with a compound in a compound library (e.g., a DEL). The data or information may include the binding affinity of the compound to a target molecule (optionally, where the target molecule is a drug target such as a protein or nucleic acid), the activity of the compound, a physical property of the compound (e.g., lipophilicity), the toxicity of the compound, the stability of the compound, the permeability of the compound, sequencing reads associated with the abundance of the compound in an experiment, information associated with the synthesis of the compound, labeling data, process quality control data, or yield associated with the synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuity data, structural data, or building block validation data, or a combination of two or more thereof.

[0071] In some embodiments, each condition can be run in its own tube before being read out on the sequencer. The number of conditions used in the experiment can be at least 2, 3, 4, 5, 6, 7, 8, 9 or 10 or more conditions. The conditions can include different protein concentrations, such as at least 2, 3, 4, 5, 6, 7, 8, 9 or 10 or more protein concentrations. The conditions can include different mutations of the target protein, such as at least 2, 3, 4, 5, 6, 7, 8, 9 or 10 or more different mutations. The conditions can include one or more internal controls, such as at least 1, 2, 3, 4, 5, 6, 7, 8, 9 or 10 or more internal controls. In some embodiments, the conditions can include an internal control for each experimental condition. In a non-limiting example, the binding assay of the wild-type protein can include an internal control, which includes: (1) a support (no protein), (2) different proteins (e.g., bovine serum albumin) that are not expected to produce specific binding, and (3) proteins in which the binding pocket is mutated.

[0072] DEL data can be a series of DNA sequences and the number of times the sequence is observed. The sequence can be decoded into a compound structure. The number of times the structure is observed (e.g., the number of observations or "hit" times of the sequence) can be related to the degree of binding of the structure to the target. In some embodiments, the compound fitness score is related to the "hit" number. However, the hit number can also be related to various factors before and after the binding event. These factors can include library production efficiency, PCR amplification, sequencing and many other steps that can affect counting and / or compound fitness scores. For example, the synthesis steps during library production can have different efficiencies, which causes the number of compound species in the library to be unequal. Then, the compound synthesized with lower efficiency can produce a relatively low sequence "hit" number expected based on the binding efficiency, because the compound amount in the library is less.

[0073] Conventional DEL dogma holds that the output of a DEL cannot be correlated with the affinity of the compound for the target. A compound may appear very abundant in a DEL selection (i.e., have many reads in the sequencer), but be only a weak binder, and vice versa. This is because DEL readout is a noisy process. Many factors complicate the readout, which may include, for example, variable chemical yields, unexpected byproducts generated during library synthesis, DNA binding (via barcodes), matrix / support binding, promiscuous binding, insufficient compound sampling, amplification bias, and sequencing noise.

[0074] The platforms, systems, and methods disclosed herein can account for noise sources to generate DEL outputs with read counts that correlate with binding affinity. In some embodiments, DELs are built from scratch to minimize noise and maximize signal, which allows machine learning models to be developed using data generated from DELs to effectively distinguish tighter binders from weaker binders.

[0075] Sequencing and selection

[0076] In some aspects, the platform, system and method disclosed herein utilize deep sequencing to carry out complete sequencing rather than partial reads. In some embodiments, the selection coverage of the molecule in the DEL is at least 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%, 99% or 100%. In some embodiments, the coverage range is selected to be from 50%, 55%, 60%, 65%, 70% or 75% to 80%, 85%, 90%, 95%, 99% or 100%. Machine learning can be used for screening compound libraries, to predict the compound libraries of new evolution, for the analysis of iterative rounds. Multiple rounds of selection can be carried out to identify binding substances.

[0077] Target molecule

[0078] Disclosed herein are platforms, systems and methods for predicting compounds interacting with target molecules.The target molecule can be a protein or nucleic acid structure, such as a mammalian target protein (e.g., in humans) or a non-mammalian target (e.g., in the case of a target molecule present in bacteria, yeast, fungi or parasites), or such as an mRNA transcript of a gene (e.g., MYC).The target molecule can be associated with a disease or its symptoms.In some embodiments, the target molecule can belong to a biological pathway associated with a disease or its symptoms, or regulate or regulate the pathway.Therefore, it is predicted that the compound that will bind to the target molecule can play a role in inhibiting the target, thereby regulating the associated pathway.In a non-limiting example, the compound can bind to the binding pocket of the extracellular ligand binding domain of a receptor tyrosine kinase, thereby acting as an inhibitor of receptor ligand binding.

[0079] In some embodiments, the target can be a target protein. The protein can be a wild type (i.e., the most common allele in a population) or a mutant allele. The mutation can be a naturally occurring, or an engineered mutation (such as mutating the binding pocket of the target protein) to assess the binding specificity compared to a wild type control condition. The mutation can be a silent mutation that has no effect on protein function, or they can result in a gain of function (e.g., enhanced ligand binding) or a loss of function (e.g., a reduction or complete loss of ligand binding).

[0080] In some embodiments, the target protein can be an enzyme that catalyzes a chemical reaction. The enzyme can have a spherical structure with an active site configured for substrate binding and catalysis, and the active site is composed of a relatively small number of amino acid residues. The enzyme can also have an allosteric binding site, and the effector molecule is combined with the allosteric binding site to change the structural conformation of the enzyme, thereby enhancing or reducing its enzymatic activity. In some embodiments, the target protein is a structural protein. The structural protein can be a fibrous protein (e.g., collagen) or a globular protein (e.g., actin or myosin). The target protein can participate in cell signaling, such as transmembrane receptor protein kinases. The target protein can be a transporter, such as a transmembrane ion channel protein. In some cases, the target protein can have a quaternary structure composed of two or more protein subunits. Therefore, experimental conditions can be performed to identify compounds that specifically bind to the structural features of a specific protein by comparing the binding to the target to mutation (such as in the protein-protein interface between the subunits in the substrate binding site or the quaternary structure).

[0081] Screening and Validation

[0082] Compounds identified by the platform, system and method disclosed herein can be verified by additional screening. Various established techniques can be used to verify individual compounds for target binding. For example, known binding assays can utilize one or more of absorbance, fluorescence, luminescence, radioactivity, NMR, crystallography, microscopy including cryo-electron microscopy, mass spectrometry or Raman scattering. For example, surface plasmon resonance (SPR) measures the reflection of polarized light, which can detect changes in the reflection angle (refractive index). Fixing or binding a ligand (compound) to a surface (which contains a fixed target protein) affects the quality or thickness of the surface, which changes the refractive index. Other methods (e.g., fluorescence resonance energy transfer (FRET)) measure changes in fluorescence intensity when a target and its ligand are bound, and changes in intensity are associated with interactions destroyed by the compound.

[0083] Machine Learning Algorithms

[0084] Disclosed herein are platforms, systems and methods for providing an iterative process, wherein laboratory experiments and machine learning reinforce each other to generate new data and understand which compounds can interact with medically relevant targets. In some embodiments, a machine learning algorithm is utilized to determine compound properties (such as binding). In some embodiments, the machine learning algorithm herein uses one or more forms of labels, including but not limited to human annotation labels and semi-supervised labels. In some embodiments, a machine learning algorithm utilizes regression modeling, wherein the relationship between a predictor variable and a dependent variable is determined and weighted.

[0085] Examples of machine learning algorithms may include support vector machines (SVM), naive Bayes classification, random forests, neural networks, deep learning, or other supervised or unsupervised learning algorithms for classification and regression. A machine learning algorithm may be trained using one or more training data sets.

[0086] In some embodiments, the platform, system and method disclosed herein provide an effective search of chemical space with a large experimentally tested data set, wherein machine learning is used to explain the results. In some embodiments, a special machine learning model is established to utilize the data. In some cases, the model architecture utilizes one or more unique features constructed by the DEL library, and explicitly represents those features. This allows the model to outperform more conventional architectures when applied to the same data. Conventional models usually only receive data associating sequences with counts as input. In contrast, disclosed herein is a model that receives data associating sequences with counts and additional features, which are such as matrix binding data, mixed data and / or building block verification data (which fraction of the reaction is estimated to proceed to the next step). In particular, the estimation of reaction efficiency by combining the fraction of the reaction that is predicted to proceed to the next step solves such a technical problem: that is, when the model assumes the overall reaction yield, the change in reaction efficiency can lead to poor prediction accuracy.

[0087] Conventional methods use only large data sets in their initial searches, and quickly narrow their search space to dozens of compounds in a short period of time (such as weeks or months later). This means that searches in chemical space are slow and inefficient because many candidate compounds with potential clinical relevance can be screened out, and the number of remaining compounds is too small to effectively identify new unrelated compounds. In contrast, the platform, system, and method disclosed herein provide cutting-edge parallel chemistry and biochemistry capabilities that allow large-scale data sets to be generated and tested faster than anything else to deepen the understanding of chemical space and find better drugs. In some embodiments, the DEL designed by machine learning is an evolved DEL generated as a subsequent generation of the initial DEL library. The DEL designed by ML can have a smaller size than the conventional DEL method while maintaining a much larger size than the conventional HTS method. For example, the initial library can have a size of 1 billion compounds, and an ML design library with an evolution of more than 1 million compounds can be generated for a library specific to a target / problem.

[0088] In addition to machine learning assisted library design, selection strategies can also be designed to maximize the signal in the ML model establishment. The improvement of the ability of machine learning to understand complex relationships allows more selection conditions to be explained in parallel. In non-restrictive cases, different mutants of proteins and closely related family members can be used. The improvement of the ability of machine learning to understand complex relationships allows more selection conditions to be explained in parallel. In non-restrictive cases, different mutants of proteins and closely related family members can be used. The software application, system and method for training and generating ML models described herein are specifically designed to process and utilize these additional data to more effectively explain the selection conditions. These capabilities together allow rapid exploration of chemical space, and more efficiently and effectively find answers to difficult drug discovery problems.

[0089] After the initial DEL search finds good starting points, ML-designed DELs can be used to deepen the understanding of what are “good” compounds for a given target, significantly shortening the timeline of the lead optimization process. ML-designed DELs can be conceived while considering one or more factors (each of which can be viewed as a descriptor associated with the compounds in the DEL), including but not limited to binding affinity, chemical diversity relative to the initial training set, physicochemical properties such as solubility and log-D, and predicted ADME (adsorption, distribution, metabolism, excretion) and predicted toxicity properties. Because ML can consider more compounds and large libraries can be efficiently built and tested by the system, the chemical space is not as narrow compared to traditional methods that only consider tens to hundreds of compounds per iteration.

[0090] Traditional HTS and DEL approaches suffer from a shrinking search space problem. After an initial medium-scale (500K-2M compound HTS) or large-scale (1B+ compound DEL) search, a few dozen compounds are selected for follow-up and further experimentation. In other words, traditional HTS and DEL immediately shrink their search space by six to nine orders of magnitude, retreating into the slow, expensive process of lead optimization that accounts for more than one-third of drug development costs. This shift from a wide-angle view of chemical space to a limited window of a few dozen compounds represents a huge inefficiency in the drug discovery process.

[0091] In contrast, the platforms, systems, and methods disclosed herein can generate millions of subsequent libraries of compounds (e.g., DELs), maintaining a wide-angle view of chemical space while still focusing on the best compounds. The generated data set strengthens the ML model with more data, giving it a deeper understanding of which compounds perform well for a given target. The data set can be a compound descriptor. In some embodiments, multiple iterations can be performed to continuously refine and improve subsequent ML-generated libraries (e.g., ML-generated DELs). The ML-generated DELs not only provide more useful data for the ML model, but also enable the model to effectively learn from its own predictions. Each library iteration can be composed of compounds for which the ML model has strong assumptions. In a non-limiting example, the ML model can treat the compounds in the ML-generated DEL as strong binders or weak binders, or be uncertain about affinity, but for each compound model will make a clear prediction (e.g., compound fitness score), which can then be tested in a gold standard experiment. In some embodiments, the compound suitability score is related to the compound's oral drug solubility, intestinal absorption, permeability, hERG toxicity, CYP inhibition, blood-brain barrier permeability, P-glycoprotein activity, plasma protein binding, and / or binding affinity.

[0092] Compared with the ML model that sees and interprets its own prediction results, a new random compound set or a compound set that is simply selected to be similar to the original library will not provide so much useful information. This is because random compounds lack the initial hypothesis provided by the ML model. Therefore, the combined laboratory and ML iteration system and method described herein allow the method of training and updating the ML model to be more efficient and effective. In some embodiments, DEL data (e.g., compound fitness score) and ML prediction accuracy can be sufficient to guarantee the synthesis of a small number of compounds for lower throughput testing. At this stage, the iterative process can have generated a high-performance ML model and generated a variety of high-quality compound structure starting points from the initial process. Therefore, the resulting compound list can have less cumbersomeness and can have an optimized ML model to help search for the next improvement. In some embodiments, the compound fitness score is related to the oral drug solubility, intestinal absorption, permeability, hERG toxicity, CYP inhibition, blood-brain barrier permeability, P-glycoprotein activity, plasma protein binding and / or binding affinity of the compound.

[0093] In some embodiments, the starting size of the initial DEL is below a threshold. The initial DEL is a non-ML iterated DEL. The starting size can be smaller than a conventional DEL and does not include all compounds that could potentially be evaluated, but is designed to optimize the overall diversity of compound structures. Then, once the first DEL has been evaluated (e.g., via binding assays and sequencing of hits), the experimental data (e.g., descriptors and / or compound fitness scores) can be input into the ML model, which is then used to generate and / or identify evolved / iterated DELs. The size of the evolved DEL can be at least 100,000, at least 200,000, at least 300,000, at least 400,000, at least 500,000, at least 600,000, at least 700,000, at least 800,000, at least 900,000, at least 1 million, at least 2 million, at least 3 million, at least 4 million, at least 5 million, at least 6 million, at least 7 million, at least 8 million, at least 9 million, or at least 10 million compounds. With each iteration, the size of the DEL can remain the same or decrease until a sufficient number of compounds with a threshold quality (e.g., binding score, affinity score, activity score, compound fitness score, etc.) are identified for subsequent low-throughput analysis.

[0094] Using an initial library (e.g., DEL) below a threshold size (e.g., below 100 billion, 10 billion, 1 billion, 500 million, 100 million, etc.) can significantly improve the signal-to-noise ratio. Larger libraries can suffer from too high noise and too small a signal, which means that it will be more difficult to find identical compounds in a larger DEL due to the noise, let alone compare those compounds to other compounds in the library. In some embodiments, the evolved libraries disclosed herein (e.g., evolved DELs) can use specialized building blocks with more interesting structures that provide increased structural diversity and therefore help increase the chemical space searched to identify hits with potential (e.g., specific target binding). Subsequent evolved DELs can then include new compounds that were not previously included in the initial DEL, but which can be structurally predicted by the ML model to result in binding and / or enhanced binding. In this way, the breadth of the chemical space search is balanced with selectivity for desired properties and the ability to reliably identify compounds amidst noise.

[0095] According to the platform, system and method disclosed herein, various machine learning models and algorithms can be utilized. Machine learning methods can include artificial neural networks, decision trees, support vector machines, regression analysis, naive Bayes, random forests, gradient boosting, XGBoost and other suitable technologies. For example, the model architecture can include XGBoost on molecular features, multilayer perceptrons on fingerprint vectors or graph convolutional neural networks (GCNN). In some embodiments, the machine learning model includes a neural network. The neural network can be a convolutional neural network (CNN). In some embodiments, the machine learning network includes a graph convolutional neural network (GCNN). GCNN is very suitable for the graph type data representation of evaluation molecules. A sequence can be converted into a structure set, which is represented by a graph with an array associated with each node and edge in the graph.

[0096] In some embodiments, the neural network architecture includes one or more sequence-oriented layers to illustrate the sequence elements of the DEL (each compound is the product of a series of reaction steps). The ability to add these sequence-oriented layers is a unique feature of the neural network, which is used to provide additional useful information to the model to improve compound selection, thereby generating evolutionary / iterative DELs. In a non-limiting example, these sequence layers can be used to display a set of compounds to the ML model, which can be synthesized during the inherent noise process of library construction. In contrast, the traditional architecture only considers a single compound at a time and therefore cannot consider this complexity. The sequence layer can also contain additional information, including but not limited to the synthesis yields of various reactions, so that the ML model can better understand the relative importance of different compounds in the set shown. The architecture of the ML model described herein is capable of incorporating one or more new and different types of data into the training data set, enabling the ML model trained on the data to provide more accurate and better compound design.

[0097] In some embodiments, DELs are determined according to a selection experiment with one or more conditions. For example, DELs can be evaluated in a protein binding assay for a target protein under one or more conditions (and one or more controls), and then the bound compounds are sequenced to determine their corresponding DNA codes for identification. Sequencing data can include the number of unique DNA sequences detected and their corresponding sequence counts or hits, which can correspond to the relative abundance of the compound in the sequencing sample (although various noise sources are discussed herein). The detected compound can include the entire structure or a portion of the structure, such as a single synthon, a disynthon, a trisynthon, a byproduct, or other polysynthon.

[0098] The platforms, systems, and methods disclosed herein can utilize a unique machine learning architecture tailored specifically for DEL analysis. During DEL synthesis, individual building blocks are assembled in a combinatorial manner to generate millions or even billions of possible compounds. However, these chemical synthesis reactions are not necessarily 100% efficient, meaning that intermediates and byproducts can be generated during DEL creation. Fig.11 As shown, as each successive synthesis step adds another building block, more possible products are generated, including unreacted intermediates and by-products. The predicted compounds generated at the end of the assembly process can be modeled. In some embodiments of the platform, system, and method disclosed herein, the model generates not only the final product compound but also the intermediate compound during the intermediate synthesis step. The conventional method of ML on DEL does not account for these intermediates or by-products. In contrast, the embodiments of the platform, system, and method disclosed herein utilize a neural network that incorporates possible intermediates or by-products in its training data. As a result, the potential intermediate reactants and side reactants or products present in DEL are explicitly described as possible factors for explaining the read counts. In contrast, the conventional method that assumes that only the final product is generated assumes that the final product is synthesized with 100% efficiency, while in fact, the synthesis efficiency of many compounds can be less than 30%. Therefore, the conventional method is to learn on "incorrect" data, and the performance of the model may be affected. Therefore, the machine learning model is given input data for all final products and intermediate compounds. The ML model architecture described herein is designed to consider the final product and any intermediates as a group, and this consideration can be further enhanced with additional data in the form of measured chemical yields. This type of data is not present in the core DEL selection data, which is the only data that can be considered by the previously described model architecture due to inherent limitations in its design. Because the ML model described herein can include this additional information and is designed to consider the complete set of synthetic compounds as well as the DEL selection data, it more accurately considers the underlying biochemical processes from the experiments. By leveraging this architecture and previously inaccessible information, it is able to significantly improve its prediction accuracy.

[0099] In some embodiments, one or more quality control steps are performed to control the reaction efficiency to be low. Measurements can be collected during one or more steps in the library construction (e.g., DEL construction) step. Measurements can include reaction efficiency, for example, mass spectrometry analysis of the reaction product to identify the relative abundance of the final product and the intermediate product or the remaining reactant. This information can be used to predict the fraction of each type of product or intermediate compound, which can be used to generate weights for machine learning models to improve learning. In some embodiments, the measurement can be a compound descriptor. In some embodiments, the measurement is used to determine the compound descriptor and / or fitness score. In some embodiments, the compound fitness score is related to oral drug solubility, intestinal absorption, permeability, hERG toxicity, CYP inhibition, blood-brain barrier permeability, P-glycoprotein activity, plasma protein binding and / or binding affinity.

[0100] In some embodiments, one DNA barcode or tag can represent all compounds (final compounds and intermediate compounds) in a given fraction. In some embodiments, one DNA barcode or tag can be a compound descriptor. As described herein, since the reaction in DEL may not be complete, the DNA tag will be attached to several molecules (final products, intermediates, by-products, etc.) produced during a series of predetermined synthesis steps. Therefore, the DNA tag can be understood as following the instructions of these synthesis steps, rather than generating complete molecules. Therefore, a unique DNA tag can be used as an indicator that the molecule attached to it is one of the molecules that can be produced in this reaction scheme.

[0101] The data collected for the DEL experiment is used to train the machine learning model. A given data set may include a descriptor (compound descriptor) for each compound in the data set, which may include a representation of the compound's molecular structure, the sequence of the DNA tag associated with the compound, and the sequencing read count of the DNA tag. The data set may include additional information (e.g., descriptors), such as chemical properties or parameters (e.g., molecular weight). The data set may include a fitness score for each of the compounds in the data set. In some embodiments, the fitness score may be a score indicating that the compound has a desired property (e.g., binding affinity or activity in a biochemical assay or some ADME measurements). The trained model can then be used to evaluate another set of compounds to identify compounds that are predicted to have a desired property (e.g., a high binding affinity to a target protein). Compounds can be identified by the fitness score of each compound. A new set of compounds predicted to have a desired property can be generated as a new or evolved DEL, which then undergoes another round of selection (e.g., binding assay and sequencing). The resulting data can be used again to further train and improve the machine learning model, which can then be used again to identify new evolved DELs. This process can be iterated multiple times until a smaller set of candidate compounds is selected for further evaluation.

[0102] In some embodiments, the machine learning model is given input data generated from the DEL binding assay. The input data can be evaluated to determine an indication of binding, such as a binding score. In some cases, the indication of binding can be used to classify a given compound as a binder or non-binder relative to a target protein (e.g., via a fitness score). In some embodiments, the binding data can be combined with a biochemical activity data set to generate a fitness score. In some embodiments, a fitness score can be generated by analyzing the binding data and / or biochemical activity data to determine which compounds are more likely to have a desired biological effect.

[0103] In some embodiments, the fitness score can be a number derived or calculated from the read counts of the DEL compound and / or the read counts of a single DNA compound. A non-limiting example of a fitness score derived or calculated from the read counts of a single DNA compound can be the ratio between the number of counts of the compound under the desired conditions (target conditions) and the number of counts in one or more controls (non-target conditions). In some embodiments, the derived or calculated fitness score can be a function of the read counts and external information (such as compound synthesis data or data measuring the baseline abundance of compounds in the original library). Many different mathematical functions combining the read counts with other data can be calculated to generate a fitness score. In some embodiments, the fitness score is related to oral drug solubility, intestinal absorption, permeability, hERG toxicity, CYP inhibition, blood-brain barrier permeability, P-glycoprotein activity, plasma protein binding and / or binding affinity. In some embodiments, a fitness score is derived from the read counts.

[0104] In some embodiments, the molecular structure of the compound must be converted into a molecular representation suitable for input into a machine learning model. Examples of molecular representations include molecular graph representations (mapping the atoms and bonds of a molecule into nodes and edges), electrostatic representations, textual representations such as SMILES or SELFIES, and geometric representations.

[0105] Although a graph representation is a 2D data structure that does not explicitly document the spatial relationships between elements, 3D information can be incorporated into the graph representation. For example, a node feature matrix can include this information as node information (e.g., the stereochemistry of the atoms represented by the nodes), and an edge feature matrix can include this information as edge information (e.g., bond type).

[0106] In some embodiments, DNA tags and sequence counts are processed from raw data into a data set suitable for machine learning analysis. In some cases, sequence read counts are corrected up or down.

[0107] In contrast, conventional ML utilizes data and labels, e.g., the molecular structure is the data (i.e., the structure determined for binding to a target), and the label is the binding score (i.e., a score indicative of binding affinity generated based at least in part on sequence counts). If a synthetic reaction step is not completely 100% efficient, then the data is one of several data among the set of molecules and / or compounds. However, the methods disclosed herein allow for machine learning analysis when it is unclear what data (molecular structure) corresponds to the label. The ML model architecture is designed to consider the final product and any intermediates as a group, and this consideration can be further enhanced with additional data in the form of measured chemical yields. This type of data is not present in the core DEL selection data, which is the only data that can be considered by the previously described model architecture due to inherent limitations in its design. Because the ML model described herein can include this additional information and is designed to consider the complete set of synthetic compounds as well as the DEL selection data, it more accurately considers the underlying biochemical processes from the experiment. By leveraging this architecture and previously inaccessible information, it is able to significantly improve its prediction accuracy.

[0108] In other words, the signature or binding score is known, but there are multiple molecules in the set represented by the DNA code, and therefore it is unclear what the actual best binder in the set is. Fig.11 , the top compound in the third synthesis step may be the best binder, but other intermediates may also undergo some binding. Although the DNA barcodes cannot distinguish between the final product and the intermediates due to incomplete synthesis, the model is given all the products corresponding to the DNA barcodes in order to more fully understand or learn the set of compounds. Therefore, in some embodiments, the model sees the complete molecule, not just the disynthon in some alternative methods. Therefore, the model is configured based on the understanding that the DNA tags or barcodes can represent any intermediate product (disynth), reaction intermediate, or full product (e.g., trisynthon).

[0109] In some embodiments, the machine learning model is used to generate predictions of new input data. For example, once the machine learning model is trained on the input data generated from the DEL experiment, it will be able to receive new input data and generate predictions of fitness, such as binding to the target protein. The prediction of fitness can be output as a fitness score. The fitness score can be a combination score or other properties. In some embodiments, the prediction includes a comprehensive score of multiple compounds or drug properties. Non-limiting examples of other properties can include oral drug solubility, human intestinal absorption, permeability, hERG toxicity, CYP inhibition (2D6, 2C9), blood-brain barrier permeability, P-glycoprotein activity, and plasma protein binding.

[0110] In some embodiments, the machine learning model learns one or more factors for each possible molecule. Once the machine learning model is trained, it can generate read count predictions by aggregating all factors, which may include one or more of the following: the matrix binding propensity of each of the possible molecule sets, the promiscuity of the possible molecule sets, and the target binding propensity of the possible molecule sets.

[0111] To score molecules, predictions can consist of the "target propensity" of the target molecule (thus eliminating all other factors of no interest).

[0112] Alternatively, in some cases, a multi-parameter score is generated that includes other properties, not just a single property such as target propensity.

[0113] Non-limiting numbered implementation scheme

[0114] 1. A method comprising: i) receiving a first input data set comprising first binding interaction information (e.g., compound descriptors) between a target molecule and a compound library (e.g., a compound set); ii) processing the first input data set using a machine learning module to generate a model representation of the binding interaction (e.g., a predicted compound descriptor), wherein the model is configured to predict a binding fitness score between the target molecule and the input compound; iii) using the model representation of the binding interaction to determine an updated compound library, wherein the updated compound library comprises one or more new compounds predicted to bind to the target molecule; iv) receiving a second input data set comprising second binding interaction information between the target molecule and the updated compound library; and v) processing the second input data set using the machine learning module to update the model representation of the binding interaction, wherein at least the prediction accuracy of the updated model representation is improved.

[0115] 2. According to the method described in embodiment 1, the compound library is a compound combinatorial library.

[0116] 3. The method according to embodiment 2, wherein the compound combinatorial library is a DNA encoded library.

[0117] 4. A method according to embodiment 2, wherein the compound combinatorial library is generated using a partitioning pool approach.

[0118] 5. A method according to embodiment 2, wherein the first input data set and the second input data set include: DNA sequencing read counts of DNA barcodes tagged to each compound in the compound combinatorial library, and at least one mapping structure corresponding to each DNA barcode.

[0119] 6. The method according to embodiment 1, wherein the target molecule is a biological molecule.

[0120] 7. A method according to embodiment 6, wherein the biomolecules include macromolecules.

[0121] 8. The method of embodiment 7, wherein the macromolecule comprises a polysaccharide, a carbohydrate, a lipid or a nucleic acid.

[0122] 9. The method of embodiment 7, wherein the macromolecule comprises a protein.

[0123] 10. A method according to embodiment 1, wherein the compound library comprises small molecule compounds.

[0124] 11. The method according to embodiment 10, wherein the molecular weight of the small molecule compound does not exceed 1000 Daltons.

[0125] 12. The method according to embodiment 10, wherein the compound library consists of 100,000 to 1,000,000,000 small molecule compounds.

[0126] 13. The method according to embodiment 10, wherein the updated compound library consists of more than 10,000 small molecule compounds.

[0127] 14. A method according to embodiment 1, wherein the model representation is configured to predict the combination of multiple compound fractions (e.g., compounds considered as whole, intermediate and / or by-product data) for each combinatorial synthetic chemistry pathway used to generate the compound library.

[0128] 15. The method of embodiment 14, wherein one or more of the plurality of compound fractions each comprises a plurality of compounds corresponding to at least one target product and at least one by-product.

[0129] 16. A method according to embodiment 14, wherein the model representation is configured to model the products and any by-products of the compound fractions generated from the synthesis steps.

[0130] 17. The method of embodiment 14, wherein each of the plurality of compound fractions is encoded with a DNA barcode corresponding to a compound (eg, full, intermediate, and / or byproduct) within the compound fraction.

[0131] 18. A method according to embodiment 14, wherein the model representation is configured to generate a binding score (e.g., a fitness score) for each compound fraction and / or the compounds in each compound fraction.

[0132] 19. A method according to embodiment 1, wherein the model representation includes a neural network.

[0133] 20. A method according to embodiment 19, wherein the neural network is a graph neural network.

[0134] 21. A method according to embodiment 1, wherein the method further comprises iteratively repeating steps (c)-(e) to improve the prediction accuracy represented by the model.

[0135] 22. A method according to embodiment 21, wherein steps (c)-(e) are iteratively repeated until at least 1, 10, 50, 100, 200, 300, 400 or 500 compounds are identified as having a predicted affinity for the target molecule above a minimum threshold.

[0136] 23. A method according to embodiment 1, wherein the method also includes weighting the binding interactions represented by the model (e.g., predicted compound descriptors) based on experimental data corresponding to the efficiency of the synthesis steps used to generate the compound library, thereby reducing the signal-to-noise ratio.

[0137] 24. A method according to embodiment 23, wherein the experimental data includes abundance information based on mass spectrometry analysis of compound fractions, wherein the compound fractions include one or more synthetic compound products generated from each synthetic step.

[0138] 25. A method according to embodiment 1, wherein the model representation is a graph convolutional neural network configured to receive a graph representation of a given compound as input data.

[0139] 26. A method according to embodiment 25, wherein the graph representation of a given compound includes a graph data structure consisting of vertices and edges.

[0140] 27. The method according to embodiment 1, wherein the method further comprises performing a binding affinity experiment between the target molecule and the compound library.

[0141] 28. A method according to embodiment 27, wherein the binding affinity experiment includes: incubating the target molecule with the compound library, and purifying the target molecule together with any bound compounds.

[0142] 29. A method according to embodiment 27, wherein the binding affinity experiment includes eluting bound compounds, wherein the bound compounds are tagged with DNA barcodes.

[0143] 30. The method of embodiment 28, wherein the method further comprises amplifying the DNA barcode.

[0144] 31. A method according to embodiment 29, wherein the method further comprises sequencing the DNA barcode to obtain read count data of the bound compound tagged with the DNA barcode.

[0145] 32. A method according to embodiment 29, wherein sequencing includes deep sequencing to obtain complete sequencing of DNA barcodes.

[0146] 33. A method according to embodiment 31, wherein the selection coverage of the compounds in the compound library is at least 80%, 85%, 90%, 95%, 99% or 100%.

[0147] 34. A method according to embodiment 27, wherein one or more iterative rounds of input data generation and use of the model representation to determine the updated compound library for binding affinity experiments.

[0148] 35. A method according to embodiment 34, wherein the binding affinity experiment includes calibrating the amount of input material to the number of cycles.

[0149] 36. The method of embodiment 34, wherein the one or more iterative rounds include at least one, two, three, four or five rounds of generating input data and using the model representation to determine an updated compound library.

[0150] 37. The method according to embodiment 1, wherein the method further comprises performing a binding affinity experiment between the target molecule and the updated compound library after step (c).

[0151] 38. A method according to embodiment 1, wherein the compound library is selected as a subset of a larger library of possible compounds, and wherein the updated compound library includes one or more new compounds identified from the larger library of possible compounds as potential binders to the target molecule.

[0152] 39. The method of embodiment 1, wherein the method further comprises using deep sequencing for full sequencing rather than partial reads. In some embodiments, the selected coverage of the molecules in the DEL is at least 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%, 99% or 100%. In some embodiments, the selected coverage ranges from 50%, 55%, 60%, 65%, 70% or 75% to 80%, 85%, 90%, 95%, 99% or 100%.

[0153] 40. 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 that can be executed by the digital processing device to perform the method described in any one of embodiments 1-39.

[0154] 41. A non-transitory computer-readable storage medium encoded with a computer program comprising instructions executable by one or more processors to perform the method of any one of embodiments 1-39.

[0155] Terms and Definitions

[0156] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0157] As used herein, the singular forms "a", "an", and "the" include plural references unless the context clearly dictates otherwise. Any reference to "or" herein is intended to include "and / or" unless stated otherwise.

[0158] As used herein, the term "about" in some instances refers to an amount that is approximately the stated amount.

[0159] As used herein, the term "about" in some cases refers to an amount that is close to 10%, 5%, or 1% of the stated amount, including increments therein.

[0160] As used herein, the term "about" referring to a percentage may, in some cases, refer to an amount that is 10%, 5%, or 1% greater or less than the stated percentage, including increments therein.

[0161] As used herein, the phrases "at least one," "one or more," and "and / or" are open-ended expressions that are both conjunctive and disjunctive in operation. For example, each of the expressions "at least one of A, B, and C," "at least one of A, B, or C," "one or more of A, B, and C," "one or more of A, B, or C," and "A, B, and / or C" means A alone, B alone, C alone, A and B together, A and C together, B and C together, or A, B, and C together.

[0162] As used herein, the term "building block" can be any chemical structure that can be combined or added to another chemical structure to build one or more compounds. In some embodiments, a compound can be used as a building block, where the compound can be combined or added to one or more building blocks to build a new compound. In some embodiments, one or more compounds are used as building blocks to build a new compound in a new DEL. In a non-limiting example, a single synthon can be a first building block, and a second single synthon can be a second building block, where the two building blocks react together to form a new disynthon compound. The disynthon can then be used as a building block together with one or more building blocks to build one or more trisynthons and / or polysynthons.

[0163] As used herein, the term "compound descriptor" refers to any data or information associated with a compound. Data or information may include, but are not limited to, the binding affinity of the compound to a target molecule (optionally, where the target molecule is a drug target such as a protein or nucleic acid), the activity of the compound, the physical properties of the compound (e.g., lipophilicity), the toxicity of the compound, the stability of the compound, the permeability of the compound, the sequencing reads associated with the abundance of the compound in the experiment, information related to the synthesis of the compound, the structure of the compound, labeling data, process quality control data, or the yield associated with the synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, mixed data, structural data, or building block verification data, or a combination of two or more thereof. In some embodiments, the compound descriptor may be in the form of a diagram, a chart, a number / value, a mark, a text, a data set, an image, etc.

[0164] Machine Learning

[0165] The machine learning model may include one or more of various machine learning models. In some embodiments, the machine learning model may include one machine learning model. In some embodiments, the machine learning model may include multiple machine learning models. In some embodiments, the machine learning model may include a neural network model. In some embodiments, the machine learning model may include a random forest model. In some embodiments, the machine learning model may include a manifold learning model. In some embodiments, the machine learning model may include a hyperparameter learning model. In some embodiments, the machine learning model may include an active learning model.

[0166] Graph, graph model, and graphical model may refer to a method of conceptualizing or organizing information into a graphical representation including nodes and edges. In some embodiments, a graph may refer to a principle of conceptualizing or organizing data, where the data may be stored in various alternative forms (such as linked lists, dictionaries, spreadsheets, arrays, permanent storage, temporary storage, etc.), and is not limited to the specific embodiments disclosed herein. In some embodiments, a machine learning model may include a graph model.

[0167] The machine learning model may include various manifold learning algorithms. In some embodiments, the machine learning model may include a manifold learning algorithm. In some embodiments, the manifold learning algorithm is principal component analysis. In some embodiments, the manifold learning algorithm is a uniform manifold approximation algorithm. In some embodiments, the manifold learning algorithm is an isomap algorithm. In some embodiments, the manifold learning algorithm is a local linear embedding algorithm. In some embodiments, the manifold learning algorithm is a modified local linear embedding algorithm. In some embodiments, the manifold learning algorithm is a Hessian eigenmapping algorithm. In some embodiments, the manifold learning algorithm is a spectral embedding algorithm. In some embodiments, the manifold learning algorithm is a local tangent space alignment algorithm. In some embodiments, the manifold learning algorithm is a multidimensional scaling algorithm. In some embodiments, the manifold learning algorithm is a t-distributed stochastic neighbor embedding algorithm (t-SNE). In some embodiments, the manifold learning algorithm is a Barnes-Hut t-SNE algorithm.

[0168] The terms reduction, dimensionality reduction, projection, component analysis, feature space reduction, latent space engineering, feature space engineering, representation engineering, or latent space embedding may refer to a method of transforming given input data having an initial number of dimensions into another form of data having fewer dimensions than the initial number of dimensions. In some embodiments, the terms may refer to the principle of reducing an input dimensional set to a smaller output dimensional set.

[0169] The term normalization may refer to a collection of methods for adjusting a data set to align the data set with a common scale. In some embodiments, the normalization method may include multiplying a portion or all of the data set by a factor. In some embodiments, the normalization method may include adding or subtracting a constant from a portion or all of the data set. In some embodiments, the normalization method may include adjusting a portion or all of the data set to a known statistical distribution. In some embodiments, the normalization method may include adjusting a portion or all of the data set to a normal distribution. In some embodiments, the normalization method may include adjusting the data set so that the signal strength of a portion or all of the data set is approximately the same.

[0170] The conversion may include one or more steps of various data conversions. In some embodiments, the conversion may include normalizing the data. In some embodiments, the conversion may include performing a mathematical operation that calculates a score based on the distance between 2 points in the data. In some embodiments, the distance may include the distance between two edges in the graph. In some embodiments, the distance may include the distance between two nodes in the graph. In some embodiments, the distance may include the distance between a node and an edge in the graph. In some embodiments, the distance may include the Euclidean distance. In some embodiments, the distance may include a non-Euclidean distance. In some embodiments, the distance may be calculated in frequency space. In some embodiments, the distance may be calculated in Fourier space. In some embodiments, the distance may be calculated in Laplace space. In some embodiments, the distance may be calculated in spectral space. In some embodiments, the mathematical operation may be a monotonic function based on the distance. In some embodiments, the mathematical operation may be a non-monotonic function based on the distance. In some embodiments, the mathematical operation may be an exponential decay function. In some embodiments, the mathematical operation may be a learning function.

[0171] In some embodiments, conversion may include converting the data in a representation to another representation. In some embodiments, conversion may include converting the data to data in another form with a smaller dimension. In some embodiments, conversion may include one or more curved paths in the linearized data. In some embodiments, conversion may be performed on the data including the data in the Euclidean space. In some embodiments, conversion may be performed on the data including the data in the graph space. In some embodiments, conversion may be performed on the data in the discrete space. In some embodiments, conversion may be performed on the data including the data in the frequency space. In some embodiments, conversion may convert the data in the discrete space to continuous space, convert the data in the continuous space to discrete space, convert the data in the graph space to continuous space, convert the data in the continuous space to graph space, convert the data in the graph space to discrete space, convert the data in the discrete space to graph space, or any combination thereof. In some embodiments, conversion may include converting the data in the discrete space to frequency domain. In some embodiments, conversion may include converting the data in the continuous space to frequency domain. In some embodiments, conversion may include converting the data in the graph space to frequency domain.

[0172] In some embodiments, the method of the present disclosure further comprises reducing the compound descriptors to a reduced descriptor space using a machine learning model. In some embodiments, the method further comprises clustering the reduced descriptor space to determine one or more groups of compound descriptors having similar characteristics.

[0173] In some embodiments, reduction may include converting given input data with any initial number of dimensions into data of another form with any number of dimensions less than the initial number of dimensions. In some embodiments, reduction may include converting input data into data of another form with less dimensions. In some embodiments, reduction may include linearizing one or more curved paths in the input data into output data. In some embodiments, reduction may be performed to data including data in Euclidean space. In some embodiments, reduction may be performed to data including data in graph space. In some embodiments, reduction may be performed to data in discrete space. In some embodiments, reduction may convert data in discrete space into continuous space, convert data in continuous space into discrete space, convert data in graph space into continuous space, convert data in continuous space into graph space, convert data in graph space into discrete space, convert data in discrete space into graph space, or any combination thereof.

[0174] The terms clustering, cluster analysis, or generative module may refer to methods for grouping samples in a dataset by some similarity measure. Samples may be grouped in set space, e.g., element 'a' is in set 'A'. Samples may be grouped in continuous space, e.g., element 'a' is a point in Euclidean space that is at distance 'l' from the centroid of the elements that make up cluster 'A'. Samples may be grouped in graph space, e.g., element 'a' is highly connected to the elements that make up cluster 'A'. These terms may refer to the principle of organizing multiple elements into groups in some mathematical space based on some similarity measure.

[0175] Clustering may include grouping any number of samples in a data set by any quantitative similarity 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 on 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 constraints. In some embodiments, clustering may include using supervised learning. In some embodiments, clustering may include using unsupervised learning.

[0176] 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 the nodes and edges of each sample.

[0177] In some embodiments, comparison may include comparison between a first group and a different second group. In some embodiments, the first group or the second group may each independently be a cluster. In some embodiments, the first group or the second group may each independently be a group of clusters. In some embodiments, comparison may include comparison between a cluster and a group of clusters. In some embodiments, comparison may include comparison between a first group of clusters and a second group of clusters different from the first group of clusters. In some embodiments, a group may be a sample. In some embodiments, a group may be a group of samples. In some embodiments, comparison may include comparison between a sample and a group of samples. In some embodiments, comparison may include comparison between a group of samples and a group of samples.

[0178] Neural Networks

[0179] 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 an autoencoder. In some embodiments, the neural network includes a language model. In some embodiments, the neural network includes a converter model.

[0180] Various types of layers may be used for a neural network. 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 locally connected 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.

[0181] In some embodiments, the neural network includes a graph model. In some embodiments, graphs, graph models, and graphical models may refer to methods of modeling data in a graphical representation including nodes and edges. In some embodiments, data may be stored in various alternative forms (such as linked lists, dictionaries, spreadsheets, arrays, permanent storage, temporary storage, etc.), and are not limited to the specific embodiments disclosed herein.

[0182] 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.

[0183] 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.

[0184] train

[0185] Various loss functions may be used to train the neural network. 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 a Laplace prior. In some embodiments, the neural network may include a Gaussian posterior. In some embodiments, the neural network may include a non-Gaussian posterior. In some embodiments, the neural network may include a Laplace posterior. 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 loss function may be formulated to optimize regression loss, an evidence-based lower bound, maximum likelihood, Kullback-Leibler divergence, applied to various distribution functions, such as Gaussian, non-Gaussian, Gaussian mixture, logistic function mixture, and the like.

[0186] Various optimizers can be used to train a neural network. In some embodiments, a neural network can be trained using an Adam optimizer. In some embodiments, a neural network can be trained using a stochastic gradient descent optimizer. In some embodiments, a neural network can be trained using an active learning algorithm. A neural network can be trained using various loss functions, and derivatives of these loss functions can be calculated to update one or more parameters of the neural network. A neural network can be trained using a hyperparameter search algorithm. In some embodiments, neural network hyperparameters are optimized using a Gaussian process.

[0187] While training a neural network, various training protocols can be used. In some embodiments, a neural network can be trained with a training / validation / test data split. In some embodiments, a neural network can be trained with a k-fold data split, where k can be any positive integer.

[0188] Training a neural network may involve providing input to an untrained neural network to generate a predicted output, comparing the predicted output to the expected output, and updating the parameters of the neural network to account for the difference between the predicted output and the expected output. Based on the calculated difference, gradients with respect to each parameter may be calculated by back propagation to update the parameters of the neural network so that the output values ​​calculated by the neural network are consistent with the examples included in the training set. This process may be iterated for a certain number of iterations, or until a certain stopping criterion is met.

[0189] Computing System

[0190] refer to Fig.16 , a block diagram depicting an exemplary machine is shown, which includes a computer system 1600 (e.g., a processing or computing system) within which a set of instructions can be executed to cause the device to perform or implement any one or more of the aspects and / or methods of static code scheduling of the present disclosure. Fig.16 The components shown 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.

[0191] The computer system 1600 may include one or more processors 1601, memory 1603, and storage 1608, which communicate with each other and other components via a bus 1640. The bus 1640 may also link a display 1632, one or more input devices 1633 (for example, which may include a keypad, keyboard, mouse, stylus, etc.), one or more output devices 1634, one or more storage devices 1635, and various tangible storage media 1636. All of these elements may be connected to the bus 1640 directly or via one or more interfaces or adapters. For example, various tangible storage media 1636 may be connected to the bus 1640 via a storage medium interface 1626. The computer system 1600 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.

[0192] Computer system 1600 includes one or more processors 1601 (e.g., central processing units (CPUs) or general purpose graphics processing units (GPGPUs)) that perform functions. Processor 1601 optionally includes a cache memory unit 1602 for temporarily storing instructions, data, or computer addresses locally. Processor 1601 is configured to assist in executing computer-readable instructions. Because processor 1601 executes non-transitory processor-executable instructions embodied in one or more tangible computer-readable storage media (such as memory 1603, storage 1608, storage device 1635, and / or storage medium 1636), computer system 1600 can be Fig.16The components depicted in the embodiment provide functionality. The computer-readable medium may store software that implements a particular embodiment, and the processor 1601 may execute the software. The memory 1603 may read the software from one or more other computer-readable media (such as mass storage devices 1635, 1636) or from one or more other sources through a suitable interface (such as a network interface 1620). The software may cause the processor 1601 to perform one or more processes described or illustrated herein or one or more steps of one or more processes. Executing such processes or steps may include defining data structures stored in the memory 1603 and modifying the data structures as directed by the software.

[0193] The memory 1603 may include various components (e.g., machine-readable media), including, but not limited to, random access memory components (e.g., RAM 1604) (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 1605), and any combination thereof. The ROM 1605 may be used to transfer data and instructions unidirectionally to the processor 1601, and the RAM 1604 may be used to transfer data and instructions bidirectionally with the processor 1601. The ROM 1605 and the RAM 1604 may include any suitable tangible computer-readable media described below. In one example, a basic input / output system 1606 (BIOS), including basic routines such as those that facilitate the transfer of information between elements within the computer system 1600 during startup, may be stored in the memory 1603.

[0194] Fixed storage 1608 is optionally bidirectionally connected to processor 1601 via storage control unit 1607. Fixed storage 1608 provides additional data storage capacity and may also include any suitable tangible computer-readable medium described herein. Storage 1608 may be used to store operating system 1609, executable files 1610, data 1611, applications 1612 (applications), etc. Storage 1608 may also include an optical drive, a solid-state storage device (e.g., a flash-based system), or a combination of any of the above. Where appropriate, the information in storage 1608 may be merged into memory 1603 as virtual memory.

[0195] In one example, storage device 1635 can be removably connected to computer system 1600 via storage device interface 1625 (e.g., via an external port connector (not shown)). In particular, storage device 1635 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 1600. In one example, the software may reside in whole or in part on the machine-readable media on storage device 1635. In another example, the software may reside in whole or in part within processor 1601.

[0196] Bus 1640 connects various subsystems. In this article, where appropriate, reference to bus may include one or more digital signal lines serving a common function. Bus 1640 may be any of several types of bus structures (including but not limited to memory bus, memory controller, peripheral bus, local bus and any combination thereof) using any of various bus architectures. As an example and not limitation, this architecture includes industrial standard architecture (ISA) bus, enhanced ISA (EISA) bus, micro channel architecture (MCA) bus, video electronics standard association local bus (VLB), peripheral component interconnect (PCI) bus, PCI-Express (PCI-X) bus, accelerated graphics port (AGP) bus, hypertransport (HTX) bus, serial advanced technology attachment (SATA) bus and any combination thereof.

[0197] The computer system 1600 may also include an input device 1633. In one example, a user of the computer system 1600 may enter commands and / or other information into the computer system 1600 via the input device 1633. Examples of the input device 1633 include, but are not limited to, an alphanumeric input device (e.g., a keyboard), a pointing device (e.g., a mouse or touchpad), a touchpad, a touch screen, a multi-touch screen, a joystick, a stylus, a game controller, an audio input device (e.g., a microphone, a voice response system, etc.), an optical scanner, a video or still image capture device (e.g., a camera), and any combination thereof. In some embodiments, the input device is a Kinect, a Leap Motion, etc. The input device 1633 may be connected to the bus 1640 via any of a variety of input interfaces 1623 (e.g., input interface 1623), including but not limited to serial, parallel, game port, USB, FIREWIRE, THUNDERBOLT, or any combination thereof.

[0198] In certain embodiments, when computer system 1600 is connected to network 1630, computer system 1600 can communicate with other devices connected to network 163 (particularly mobile devices and enterprise systems, distributed computing systems, cloud storage systems, cloud computing systems, etc.). Communications to and from computer system 1600 can be sent through network interface 1620. For example, network interface 1620 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 network 1630, and computer system 1600 can store the incoming communications in memory 1603 for processing. Computer system 1600 can similarly store outgoing communications (such as requests or responses to other devices) in the form of one or more packets in memory 1603 and transmit them from network interface 1620 to network 1630. Processor 1601 can access these communication packets stored in memory 1603 for processing.

[0199] Examples of network interface 1620 include, but are not limited to, network interface cards, modems, and any combination thereof. Examples of network 1630 or network segment 1630 include, but are not limited to, distributed computing systems, cloud computing systems, wide area networks (WANs) (e.g., the Internet, enterprise networks), local area networks (LANs) (e.g., networks associated with offices, buildings, campuses, or other relatively small geographic spaces), telephone networks, direct connections between two computing devices, peer-to-peer networks, and any combination thereof. Networks such as network 1630 may employ wired and / or wireless communication modes. In general, any network topology may be used.

[0200] Information and data can be displayed by display 1632. Examples of display 1632 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 1632 can be connected to processor 1601, memory 1603, and fixed storage 1608, as well as other devices such as input devices 1633, via bus 1640. Display 1632 is linked to bus 1640 via video interface 1622, and data transmission between display 1632 and bus 1640 can be controlled via graphics controls 1621. 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, FOVEVR, 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.

[0201] In addition to the display 1632, the computer system 1600 may also include one or more other peripheral output devices 1634, including but not limited to audio speakers, printers, storage devices, and any combination thereof. Such peripheral output devices may be connected to the bus 1640 via an output interface 1624. Examples of output interface 1624 include but are not limited to a serial port, a parallel connection, a USB port, a FIREWIRE port, a THUNDERBOLT port, and any combination thereof.

[0202] In addition or in the alternative, computer system 1600 may provide functionality as a result of 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 the present disclosure may include logic, and references to logic may include software. Further, where appropriate, references to computer-readable media may include circuitry (such as an IC) storing software for execution, circuitry embodying logic for execution, or both. The present disclosure includes any suitable combination of hardware, software, or both.

[0203] Those skilled in the art will appreciate that the various illustrative logic 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 the two. In order to clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been generally described above according to their functions.

[0204] The various illustrative logical 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.

[0205] 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 of the two. 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 into a 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.

[0206] According to the description herein, as non-limiting examples, suitable computing devices include server computers, desktop computers, laptop computers, notebook computers, sub-notebook computers, netbook computers, netpad computers, set-top computers, media streaming devices, handheld computers, Internet appliances, mobile smart phones, tablet computers, personal digital assistants, video game consoles, and vehicles. Those skilled in the art will also recognize that selected televisions, video players, and digital music players with optional computer network connections are suitable for use in the systems described herein. In various embodiments, suitable tablet computers include tablet computers with brochures, notebook computers, and convertible configurations known to those skilled in the art.

[0207] 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 the execution of applications. Those skilled in the art will recognize that, by way of non-limiting example, suitable server operating systems include FreeBSD, OpenBSD, Linux, Mac OS X Windows and Those skilled in the art will recognize that suitable personal computer operating systems include, by way of non-limiting example, Mac OS and UNIX-like operating systems such as In some embodiments, the operating system is provided by cloud computing. Those skilled in the art will also recognize that, by way of non-limiting example, suitable mobile smartphone operating systems include OS, Research In BlackBerry Windows OS, Windows OS, as well as Those skilled in the art will also recognize that suitable media streaming device operating systems include, by way of non-limiting example, Apple Google Google Amazon as well as Those skilled in the art will also recognize that suitable video game console operating systems include, by way of non-limiting example, Xbox Microsoft Xbox One, Wii as well as

[0208] Non-transitory computer readable storage medium

[0209] 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, the program including instructions that can be executed by an operating system of an optional networked computing device. In further embodiments, the computer-readable storage medium is a tangible component of the computing device. In further embodiments, the computer-readable storage medium is optionally removable from the computing device. In some embodiments, as non-limiting examples, the computer-readable storage medium includes a CD-ROM, a DVD, a flash memory device, a solid-state memory, a disk drive, a tape drive, an optical disk drive, a distributed computing system including a cloud computing system and services, and the like. In some cases, programs and instructions are permanently, substantially permanently, semi-permanently, or non-temporarily encoded on the medium.

[0210] Computer Programs

[0211] In some embodiments, the platform, system, medium and method disclosed herein include at least one computer program or its use. The computer program includes a sequence of instructions, which can be executed by one or more processors of a computing device CPU, and is written to perform a specified task. Computer-readable instructions can be implemented as program modules that perform specific tasks or implement specific abstract data types, such as functions, objects, application programming interfaces (APIs), 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.

[0212] The functions 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 mobile applications, one or more stand-alone applications, one or more network browser plug-ins, extensions, add-ons or additional components, or a combination thereof.

[0213] Web Application

[0214] In some embodiments, the computer program includes a network application. Based on the disclosure provided herein, those skilled in the art will recognize that in various embodiments, the network application utilizes one or more software frameworks and one or more database systems. In some embodiments, in a computer program 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 database systems, non-relational database systems, object-oriented database systems, associative database systems, and XML database systems. In further embodiments, suitable relational database systems include, as non-limiting examples, SQL Server, mySQL TM and Those skilled in the art will also recognize that in various embodiments, the web application is written in one or more versions of one or more languages. The web application may be written in one or more markup languages, presentation definition languages, client-side scripting languages, server-side coding languages, database query languages, or combinations thereof. In some embodiments, the web application is written to some extent in a markup language such as Hypertext Markup Language (HTML), Extensible Hypertext Markup Language (XHTML), or Extensible Markup Language (XML). In some embodiments, the web application is written to some extent in a presentation definition language such as Cascading Style Sheets (CSS). In some embodiments, the web application is written to some extent in a web application such as Asynchronous Javascript and XML (AJAX), Actionscript, Javascript, or In some embodiments, the web application is written in a client-side scripting 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 server-side coding language such as PHP or Groovy. 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 a database query language such as Lotus In some embodiments, the network application includes a media player component. In various further embodiments, the media player component utilizes one or more of a number of suitable multimedia technologies, including, by way of non-limiting example, HTML 5, Java TM and

[0215] refer to Fig.17 In a specific embodiment, the application providing system includes one or more databases 1700 accessed by a relational database management system (RDBMS) 1710. Suitable RDBMS include Firebird, MySQL, PostgreSQL, SQLite, Oracle database, Microsoft SQL server, IBM DB2, IBM Informix, SAP Sybase, SAP Sybase, Teradata, etc. In this embodiment, the application providing system also includes one or more application servers 1720 (such as Java server, .NET server, PHP server, etc.) and one or more network servers 1730 (such as Apache, IIS, GWS, etc.). The network server optionally exposes one or more network services via an app application programming interface (API) 1740. Via a network such as the Internet, the system provides a browser-based and / or mobile native user interface.

[0216] refer to Fig.18 In a particular embodiment, the application providing system may alternatively have a distributed, cloud-based architecture 1800 and include elastic load balancing, auto-scaling network server resources 1810 and application server resources 1820 and synchronously replicated databases 1830.

[0217] Standalone Application

[0218] 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 typically compiled. A compiler is a computer program that converts source code written in a programming language into binary object code such as assembly language or machine code. By way of non-limiting example, 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. Compilation is typically performed, at least in part, to create an executable program. In some embodiments, a computer program includes one or more executable compiled applications.

[0219] Web browser plug-ins

[0220] In some embodiments, a computer program includes a web browser plug-in (e.g., extension, etc.). In computing, a plug-in is one or more software components that add specific functionality to a larger software application. Manufacturers of software applications support plug-ins to enable third-party developers to create extensions to the capabilities of the application, to enable easy addition of new features, and to reduce the size of the application. Where supported, plug-ins can customize the functionality of the software application. For example, plug-ins are commonly used in web browsers to play videos, generate interactivity, scan for viruses, and display specific file types. Those skilled in the art will be familiar with several web browser plug-ins, including Player, and In some embodiments, the toolbar includes one or more web browser extensions, add-ons or add-ons. In some embodiments, the toolbar includes one or more browser bars, tool belts or desktop belts.

[0221] In view of the disclosure provided herein, one skilled in the art will recognize that there are several plug-in frameworks that can be used to implement the development of plug-ins for various programming languages, including, by way of non-limiting example, C++, Delphi, Java, and the like. TM , PHP, Python TM and VB.NET or a combination of them.

[0222] A web browser (also called an Internet browser) is a software application designed for use with a network-connected computing device for retrieving, presenting, and traversing information resources on the World Wide Web. By way of non-limiting example, suitable web browsers include Internet Chrome, Opera and KDE Konqueror. In some embodiments, the web browser is a mobile web browser. Mobile web browsers (also known as micro browsers, mini browsers, and wireless browsers) are designed for use on mobile computing devices, including, by way of non-limiting example, handheld computers, tablet computers, netbook computers, sub-notebook computers, smart phones, music players, personal digital assistants (PDAs), and handheld video game systems. By way of non-limiting example, suitable mobile web browsers include Browser, RIM Browser, Blazer, Browser, for mobile devices Internet Mobile, Basic Web, Browser, Opera Mobile and PSP TM Browser.

[0223] Software Modules

[0224] 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 or a combination thereof. In further various embodiments, the software module includes multiple files, multiple code segments, multiple programming objects, multiple program structures or a combination thereof. In various embodiments, as a non-limiting example, one or more software modules include a network application, a mobile application and an independent 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 module is hosted on one or more machines in one location. In other embodiments, the software module is hosted on one or more machines in more than one location.

[0225] database

[0226] 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 DEL information and associated experimental data collected for one or more conditions. In various embodiments, as non-limiting examples, suitable databases include relational databases, non-relational databases, object-oriented databases, object databases, entity relationship model databases, associated databases, and XML databases. Further non-limiting examples include SQL, PostgreSQL, MySQL, Oracle, DB2, and Sybase. 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.

[0227] Example

[0228] The following illustrative examples represent implementations of the software applications, systems, and methods described herein and are not meant to be limiting in any way.

[0229] Example 1

[0230] The exemplary iterative machine learning based DEL platform disclosed herein is used to identify compounds that bind to a target known to be challenging for DELs. This target binds DNA, and since DELs contain a large amount of DNA, the target will generate a lot of noise in the selection. Furthermore, in order to successfully fight this target, the compound must disrupt a protein-protein interaction (PPI). PPIs are known to be challenging for small molecules, although this particular compound is known and has been clinically tested, so this problem is very difficult but not impossible.

[0231] An initial DEL consisting of approximately 450 million compounds was evaluated in a binding assay against the target. A total of approximately 3 conditions were tested, including [Condition 1, 2, 3, Control, etc.]. The sequencing results (hits) were then input into the GCNN model to identify an evolved DEL consisting of 350,000 compounds. This step was repeated 1 time until the top binding compounds predicted to bind to the target were identified. These 135 compounds identified by the ML-enabled engine were tested and compared with the known clinical compounds shown in Table 1.

[0232] Table 1 - Affinity comparison between clinical compounds (control) and predicted compounds

[0233] Compound Affinity / Activity Molecular weight cLogP Clinical Compounds 4nM IC50 532.1 8.23 AGX-59 4.5x better Kd 321.3 2.62 AGX-87 4.3x better Kd 313.4 2.8 AGX-47 4.2x better Kd 318.3 3.05 AGX-54 1.3x better Kd 310.4 1.88 AGX-39 Equivalent Kd 359.4 1.04 AGX-46 Equivalent Kd 3544 1 28

[0234] Of the 135 compounds we predicted to be potentially effective, 21% were confirmed to bind the target, and several of these met or exceeded the performance (affinity / activity) of clinical compounds in binding assays, as shown in Table 1. Two-thirds of these confirmed compounds were also experimentally determined to disrupt the relevant PPI. A compound that performs well in an assay is useless if it lacks drug-like properties. As shown in Table 1, almost all of the identified compounds fall well within the drug-like space (appropriate molecular weight and cLogP see table above, they are even better than clinical compounds).

[0235] also, Fig.15 A star graph plotting cLogP, molecular weight (MW), hydrogen bond acceptors (HBAs), hydrogen bond donors (HBDs), rotatable bonds (RBs), topological polar surface area (TPSA), and SP3 fraction (fSP3) is provided. Good drug candidates will remain within the second to last circle on the graph. Almost all compounds fall well within this boundary, with a few being slightly heavier, around 500MW. Thus, the iterative ML-driven DEL discovery process successfully validates difficult targets with higher biochemical performance compared to known clinical compounds entering clinical trials.

[0236] ML models typically perform well on the data they were trained on, but perform poorly outside of that data set. However, the instant ML model performs well outside the bounds of the DELs on which the ML was trained. Specifically, the training set is plotted in yellow, the predicted (and also successful) molecules are plotted in red, and the newly evolved DEL library is plotted in grey. As a preliminary estimate, closely related compounds will be clustered more closely together on this plot than less closely related compounds. Fig.15 As shown, the red dots are generally outside the yellow area, indicating that our ML models can identify high-performing compounds outside of the training set. Because ML models generalize outside of the training set, they can be useful when medicinal chemists come up with a molecule that looks different than what has been established in the past. Therefore, these models can guide the downstream compound optimization process.

[0237] Example 2

[0238] Based on the method of Example 1, an evolved library was established based on the results from the initial selection. Compounds were selected for the same target and the model was updated with the resulting data. Finally, two questions were answered using the newly updated model:

[0239] 1) Given an initial set of test compounds, determine the ranking of confirmed compounds. The model generates score predictions for the compounds, which are used to determine the ranking. Fig.19As shown, the graph shows that the top of the list is now enriched with more true positives than expected. In other words, more compounds are ranked closer to the top of the list, and the model is performing better as a result of seeing the evolving DEL library.

[0240] 2) Now that the model has a better understanding of which compounds should be effective, new sets of compounds are predicted to be tested. Specifically, the question is whether compounds can be found in the new set that have better predictive properties for later-stage drug discovery (such as ADME / toxicity).

[0241] To measure this, a composite score was created for the compounds that included the following predicted drug properties: oral drug solubility, human intestinal absorption, permeability, hERG toxicity, CYP inhibition (2D6, 2C9), blood-brain barrier permeability, P-glycoprotein activity, and plasma protein binding. On this metric, a score of 1 is perfect, and compounds with lower scores tend to have a poorer set of predicted properties. The relative importance of each factor can be customized for each target as needed, and either experimentally measured values ​​or predicted values ​​can be used. Next, the values ​​are plotted for the compounds in the initial DEL, the first set of predicted compounds (ML1), and the second set of predicted compounds (ML2). As Fig. 20 As shown, compound performance improved as compounds progressed from pure DEL compounds to ML predicted compounds (ML1). They improved further when the models were shown more data (i.e., updated) and asked to predict another set of compounds (ML2). These results demonstrate the value of iterative ML-driven evolution of DEL libraries in improving model performance and helping to find better drugs faster.

[0242] Example 3

[0243] The high signal-to-noise ratio, massively parallel data sets enabled by the present disclosure allow for the evaluation of compound specificity issues in an unprecedented manner. Specifically, the platforms, systems, and methods disclosed herein allow for large-scale comparison of several different isoforms of a protein or domain, and facilitate the identification of compounds that hit one isoform but not the other. To demonstrate this capability, two experiments were performed, and the results were plotted to show how the platform can identify highly specific compounds.

[0244] Two highly related protein domains (95% sequence similarity) in a single protein were evaluated. Each compound is represented by a blue dot. Fig.21As shown, the compound scores for domain 1 are plotted on the Y-axis, and the compound scores for domain 2 are plotted on the X-axis. Compounds that are specific for domain 1 will only appear on the Y-axis, and better compounds will have higher scores. Similarly, compounds that are specific for domain 2 will only appear on the X-axis. In this example, a large number of compounds were identified that are specific for one of the two highly similar domains, but not the other.

[0245] The experiment was repeated using two proteins from the same class of chromatin regulators. Fig. 22 The results shown show that a large number of molecules are predicted to bind specifically to only one of the two related proteins. This experiment further proves that the platform is able to identify a large number of potential specific molecules for two closely related proteins, while traditional drug discovery has difficulty identifying specific compounds.

[0246] Example 4

[0247] The methods described herein demonstrate a method to train custom GNNs to develop predictive models that incorporate intermediate product data into probabilistic modeling of counts, where the trained model predicts enrichment of both full products and intermediate products.

[0248] In this case, a first data set of 700 million unique molecules was established by combining 15 proprietary DELs into an initial DEL. The initial DEL can be established by one or a combination of any number of DEL libraries (e.g., from millions to billions of unique molecules). The initial DEL library is then sequenced to produce 200 million individual sequence reads, which correspond to 90 million individual molecules. From the 90 million molecules, approximately 4 million molecules are determined to be important binders of the protein target. In order to increase the number of negative examples, 1 million molecules are added to the 4 million molecules determined to be important binders of the protein target. The 1 million negative examples include molecules that have been sequenced but only bind to controls that do not contain the target (non-target controls) or to other targets, as well as an additional 1 million molecules that are not sequenced but present in the initial library. The first data set is then split into a training data set and a test data set using Murcko scaffold splitting Landrum.

[0249] To enrich this dataset with intermediates, all possible two- and three-building block molecules corresponding to a given DNA sequence were identified, and the theoretical proportion of each intermediate was then calculated based on the estimated building block yields.

[0250] The read counts of each DNA tag are modeled as a negative binomial distribution, whose mean is driven by multiple factors. For a given DNA tag / read count pair, both the full product and possible intermediate products are considered as factors driving the total read count, such as Fig.23 A general model is shown for treating trisynths as full products and disynths as intermediate products to analyze a data set and predict target binding affinity, non-target binding affinity, and the ratio adjustment of each of the full products and intermediate products. The ability of the full ML model to predict both full products and intermediate products of a compound allows the model to learn and provide more insights about the binding affinity of the compound to the target.

[0251] In NTC or target selection experiments, the read counts of DNA tag i are given by the average parameter μ i and the negative binomial distribution with dispersion parameter α:

[0252] C 靶标,i ~NB(μ 靶标,i ,α 靶标 ) and C NTC,i ~NB(μ NTC,i ,α NTC ). The α value is obtained by negative binomial regression before training.

[0253] For μ i Value, the molecule corresponding to label i for a given target (B 靶标,i ) is due to the corresponding trisynthon (R tri,i ) and three possible disynthons A combination of enrichments:

[0254]

[0255] where the p value corresponds to the ratio of trisynthons to disynthons present in the final mixture. Similarly, we assume that NTC(B NTC,i ) is also due to a combination of enrichments in trisynthons and disynthons:

[0256]

[0257] Assume that the counts for a given DNA tag against a target are due to enrichment for binding to the target, enrichment for binding in NTC experiments, and enrichment for binding using DLS experiments (C dls,i ) estimated starting material and confounding binding counts (C 混杂,i ) combination.

[0258] Therefore, the mean parameter of the negative binomial distribution is set to

[0259]

[0260] Here, the β value is learned by negative binomial regression on the training set, and σ is a softplus function. Similarly, the count of NTC for a given DNA tag can be obtained by enrichment with NTC, DLS count (C dls,i ) and promiscuous binding count (C 混杂,i ). Therefore, μ i,NTC =σ(B NTC,i +β′ DLSpdls,i +β′ 混杂 C 混杂 +β′ 常数 ) (4)

[0261] Estimate the proportion p of product k k The first is to set these values ​​based on our internal reaction yield data set: k,lab To improve this approach and take into account possible noise in the yield measurements, we propose to adjust the estimated ratio using a learning parameter for each molecule: k,调整 In this case, p k =σ(p k,调整 +p k,lab ).

[0262] In order to understand the enrichment value (R 靶标,分子 and R NTC,分子 ), two models were used. The input for both was a molecular graph corresponding to a compound. Atoms were encoded using the Weave Atom Characterizer, and bonds between atoms were encoded using the Regularized Bond Characterizer. For the first model, we used a single message passing graph neural network followed by a sequence-to-sequence layer to output a 128-dimensional encoding. The 128-dimensional encoding was then transformed using two different fully connected networks to obtain R 靶标,i and R NTC,i The second model follows the same approach as the first model, except that the 128-dimensional encoding is also converted to p through an additional fully connected network. k,调整 Then, we use a batch size of 32 and a learning rate of 10 -3 The models are tuned for 15 epochs using the Adam optimizer on the training set.

[0263] For a given DNA barcode example i, the negative binomial loss of the target and NTC values ​​is

[0264] L 值,i = -log(P(c 值,i ∣μ 值,i ,α 值 )) (5)

[0265] where P is given by μ 值,i and α 值 The probability mass function of the parameterized negative binomial distribution. The total loss of example i is

[0266]

[0267] The model was validated using R 靶标,I The values ​​rank and / or score the binding affinity of the compounds to the target.

[0268] Then, the ML model that can handle the full product ratio data disclosed in this article was evaluated against three different models. For the first model, the only source of binding enrichment was the three synthons that were the source of the binding enrichment data. (B 靶标，i =p tri，i R tri,i , B NTC,i =p tri，i R NTC,tri，i ). In the second model, the only source of binding enrichment is the disynthon.

[0269]

[0270] For the third model, each reaction was assumed to produce a 70% yield, so all = 0.147 and all p tri,i = 0.147. For each of the models (the ML model described herein and the three models evaluated against the ML model), experiments were performed in which the product ratio was fixed, and additional experiments were performed in which adjustments to the ratio were learned as described above. Fig.24 The loss and R on the test dataset are shown. 2 Scoring. Panel A shows the negative binomial loss on the test data; Panel B shows the target protein counts versus the generated μ 靶标，i Count the R between 2 score; and Figure C shows the NTC counts versus the generated μ NTC，i Count the R between 2 score.

[0271] One evaluation metric is the performance on a test set that is a subset of the internal DEL data. When evaluating the ML model against the test data from the internal DEL dataset, the loss was calculated and μ was examined. ntc With μ 靶标 R between the value and the true count 2 score.

[0272] Of the two approaches that modeled product ratios, the trisynthon-only model produced significantly worse results. This may be evidence that the trisynthon data itself is too noisy to produce reliable estimates and is unable to capture all of the DEL products. The disynthon-only model also produced worse results than the full ML model. Testing showed that modeling the disynthon is not only a denoising step, but also an integral part of the data. The model that utilized constant yield data also performed worse than modeling both yield and product ratio values ​​to train and help model the data. The model that made a learning adjustment for yield showed improved performance, which may be due to the presence of noise in the laboratory process of performing several rounds of purification.

[0273] To estimate the generalization ability of the learned models, the performance of the full ML model and the models evaluated relative to it were tested on a dataset of 150 molecules from commercial vendors. Binding affinity was measured in-house and molecules were classified as binders or non-binders of the target. Enrichment values ​​R 靶标，分子 Used to predict binding affinity. Fig.25 The area under the receiver operating characteristic (ROC) curve of various models is shown in . On this dataset, the full model with learning yield outperforms the rest of the models.

[0274] To further investigate the impact of incorporating yield data, the full ML model and a model with flat yields were tested on 150 external molecules. Fig.26 As shown, the percentage of hits among the top 10 identified hits was examined. These results demonstrate that the full ML model with laboratory yield and learning adjustment performs best in virtual screening applications, and setting the learning yield has a greater impact on relevant cases.

[0275] Overall, the tests showed that the full ML model described in this paper outperformed models that did not consider all full products and intermediates and / or models that did not include product ratios, e.g. Figure 24-26 As shown. Using only full product (trisynthon) data produces a data set that is shown to be too noisy to support effective training, and may also not adequately describe the underlying data. Using only intermediate (disynthon) data can be an effective way to aggregate and denoise DEL data, but ignores potentially useful data required for effective compound screening.

[0276] Although preferred embodiments of the present disclosure have been shown and described herein, it will be apparent to those skilled in the art that these embodiments are provided as examples only. Without departing from the present disclosure, those skilled in the art will now conceive of many variations, changes and substitutions. It should be understood that various alternatives to the embodiments of the present disclosure described herein may be adopted in practicing the present disclosure.

Claims

1. A method comprising: a) A computer-implemented method, the computer-implemented method comprising: i. receiving a first data set, the first data set comprising: (i) a first compound descriptor for each compound in a first compound library, and (ii) a compound fitness score for each compound in the first compound library; ii. training a prediction model on the first data set; iii. inputting a second data set into the model, the second data set comprising a second compound descriptor for each compound in a second compound library; and iv. generating a compound fitness score for each compound in the second compound library from the prediction model using at least one or more compound descriptors of the first compound library and / or one or more compound descriptors of the second compound library, and b) selecting a third compound library based on information including one or more compound fitness scores of the second compound library and / or one or more compound fitness scores of the first compound library.

2. The method according to claim 1, wherein the third compound library comprises: (i) compounds from the second compound library, (ii) compounds from the first compound library, (iii) compounds including two or more compounds from the second compound library, (iv) compounds including two or more compounds from the first compound library, (v) compounds including compounds from the second compound library and compounds from the first compound library, (vi) compounds not present in the first compound library and the second compound library, (vii) compounds including compounds from the second compound library and compounds not present in the first compound library and the second compound library, (viii) compounds including compounds from the second compound library and compounds not present in the first compound library and the second compound library, or (ix) a combination of two or more of (i) to (viii).

3. The method according to claim 1 or claim 2, wherein the first compound library is a first DNA-encoded library (DEL) and / or the second compound library is a second DNA-encoded library.

4. The method according to any one of claims 1 to 3, wherein step (b) is part of the computer-implemented method.

5. The method of any one of claims 1-3, wherein step (b) is not part of the computer-implemented method.

6. A method according to any one of claims 1-3, wherein step (b) includes a first sub-step and a second sub-step, the first sub-step is part of the computer-implemented method, and the second sub-step is not part of the computer-implemented method, wherein the first step and the second step are performed sequentially, and the first sub-step is performed first or the first sub-step is performed second.

7. The method of any one of claims 1-6, wherein the information further comprises an evaluation score (sometimes referred to as an external fitness score) of the compounds in the second compound library and / or an evaluation score (sometimes referred to as an external fitness score) of the compounds in the first compound library.

8. The method of claim 7, wherein the evaluation score of the compound in the second compound library is a second fitness score generated independently of the compound fitness score generated by the computer-implemented method.

9. The method of claim 7 or claim 8, wherein the assessment score of the compound in the first compound library is a first fitness score that is different from the compound fitness score of the compound in the first compound library.

10. The method of any one of claims 1-9, wherein one or more compounds in the first compound library is a first test compound (sometimes referred to as a full product or full product compound), a building block of the first test compound, a first byproduct generated during the synthesis of the first test compound (sometimes referred to as a byproduct or byproduct), or an intermediate generated during the synthesis of the first test compound (sometimes referred to as an intermediate), or a combination of two or more thereof.

11. The method of claim 10, wherein the first test compound is a desired product of a synthesis reaction that includes two or more of the building blocks of the first test compound.

12. The method of claim 11, wherein the first side product is an undesired product of the synthesis reaction comprising the two or more building blocks of the first test compound.

13. A method according to any one of claims 1-12, wherein one or more compounds in the second compound library (or optionally, a subsequent library applicable in an iterative method) is a second test compound (sometimes referred to as a full product or full product compound), a building block of the second test compound, a second byproduct generated during the synthesis of the second test compound (sometimes referred to as a byproduct or byproduct), or an intermediate generated during the synthesis of the second test compound (sometimes referred to as an intermediate), or a combination of two or more thereof.

14. The method of claim 13, wherein the second test compound is a desired product of a synthesis reaction that includes two or more of the building blocks of the second test compound.

15. The method of claim 14, wherein the second side product is an undesired product of the synthesis reaction comprising the two or more building blocks of the second test compound.

16. A method according to any one of claims 1-15, wherein one or more compounds in the third compound library are third test compounds (sometimes referred to as full products or full product compounds), building blocks of the third test compounds, third byproducts generated during the synthesis of the third test compounds (sometimes referred to as byproducts or byproducts), or intermediates generated during the synthesis of the third test compounds (sometimes referred to as intermediates), or combinations of two or more thereof.

17. The method of claim 16, wherein the third test compound is a desired product of a synthesis reaction that includes two or more of the building blocks of the third test compound.

18. The method of claim 17, wherein the third side product is an undesired product of the synthesis reaction of the two or more building blocks comprising the third test compound.

19. The method of any one of claims 1-18, wherein the first compound descriptor comprises data or information associated with the compound in the first compound library, wherein the data or information comprises binding affinity of the compound to a target molecule (optionally, wherein the target molecule is a drug target such as a protein or nucleic acid), activity of the compound, physical properties of the compound (e.g., lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing reads associated with abundance of the compound in an experiment, compound structure, information associated with synthesis of the compound, labeling data, process quality control data, yield associated with synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuity data, structural data, or building block validation data, or a combination of two or more thereof.

20. The method of any one of claims 1-19, wherein the second compound descriptor comprises data or information associated with the compound in the second compound library, wherein the data or information comprises binding affinity of the compound to a target molecule (optionally, wherein the target molecule is a drug target such as a protein or nucleic acid), activity of the compound, physical properties of the compound (e.g., lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing reads associated with abundance of the compound in an experiment, compound structure, information associated with synthesis of the compound, labeling data, process quality control data, yield associated with synthesis of the compound, sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuity data, structural data, or building block validation data, or a combination of two or more thereof.

21. The method of claim 19 or claim 20, wherein the compound is an all-product compound, an intermediate compound, or a by-product compound.

22. The method of any one of claims 1-21, comprising testing one or more of the compounds in the first compound library in an in vitro or in vivo assay.

23. The method of any one of claims 1-22, comprising testing one or more of the compounds in the third compound library in an in vitro or in vivo assay.

24. The method according to any one of claims 1-23, wherein each compound in the third compound library comprises a nucleic acid tag or is synthesized to comprise a nucleic acid tag, the method further comprising sequencing the third compound library to generate sequencing data associated with the third compound library.

25. The method of any one of claims 1-24, wherein the information of step (b) comprises sequencing data associated with an external library of compounds (e.g., a library comprising nucleic acid tags for each compound from the first library of compounds and / or the second library of compounds).

26. The method of any one of claims 1-25, wherein the compound fitness score for each compound in the first compound library is generated from data comprising sequencing data associated with the first compound library.

27. A method according to claim 24, 25 or 26, wherein the sequencing data includes read counts, quality scores associated with the read counts, and / or scores calculated based on the sequencing read counts or read count sets from different experimental conditions from the first compound library and / or the second compound library.

28. The method of claim 27, wherein the score comprises the read count, or the read count divided by the total number of reads in a compound selection or the average number of reads in a compound selection, or a similar mathematical function that utilizes (directly or indirectly) the read count.

29. The method of any one of claims 1-28, wherein at least one compound fitness score for each compound in the first compound library is generated from data comprising a first compound descriptor (e.g., sequencing data, labeling data, product data, synthesis efficiency data, mass spectrometry data, compound fraction data, binding data, matrix binding data, promiscuity data, structural data, or building block validation data, or a combination of two or more thereof).

30. The method of any one of claims 1-29, wherein the predictive model utilizes a probabilistic framework to process the first and second data sets and outputs the compound fitness score for each compound in the second compound library.

31. The method of any one of claims 10-30, wherein the fitness score is generated at least in part from data from an all-product compound including non-target counts, target counts, and / or product ratio adjustment values.

32. The method of any one of claims 10-31, wherein the fitness score is generated at least in part from data from an intermediate product compound including non-target control counts, target counts, and / or product ratio adjustment values.

33. The method of any one of claims 1-32, comprising generating a compound fitness score for each compound in the third compound library using sequencing data associated with sequencing the third compound library.

34. The method of any one of claims 1-33, comprising assaying the third library of compounds.

35. The method of claim 34, wherein the assaying comprises binding the third library of compounds to a target.

36. A method according to claim 34 or claim 35, wherein the assay comprises sequencing the third compound library or a subset of the third compound library (e.g., wherein the subset is a subset of compounds that bind to the target).

37. The method of any one of claims 1-36, wherein the fitness score for any one of the compounds comprises a binding and / or activity score for the compound.

38. The method of any one of claims 1-37, wherein the third compound library comprises one or more compounds from the second compound library having a compound fitness score greater than a threshold score.

39. The method according to any one of claims 1-39, wherein the method comprises pre-processing the first data set and / or the second data set.

40. The method of claim 39, wherein the pre-processing step is performed before step i and / or step iii of the computer-implemented method.

41. The method of any one of claims 1-40, comprising refining the fitness score generated from the predictive model, optionally wherein the refinement is performed by the predictive model, and / or optionally wherein the refinement comprises incorporating information from an external library (e.g., a nucleic acid tag library associated with the first compound library and / or the second compound library).

42. The method of any one of claims 1-41, wherein the second compound library comprises one or more compounds different from the first compound library.

43. The method of any one of claims 1-42, wherein the second compound library comprises one or more compounds different from the first compound library.

44. The method of any one of claims 1-45, further comprising repeating steps ii-vi to update the model.

45. The method of any one of claims 1-44, wherein the overall product comprises a trisynthon and the intermediate product comprises a disynthon and / or a monosynthon.

46. ​​The method of claims 1-45, wherein steps (a)-(b) are iteratively repeated to identify a set of potential compounds having one or more desired properties.

47. The method of any one of claims 1-46, wherein steps (a)-(b) are iteratively repeated to identify a set of potential compounds having one or more desired compound fitness scores.

48. The method of any one of claims 1-47, wherein the compound suitability score is related to oral drug solubility, intestinal absorption, permeability, hERG toxicity, CYP inhibition, blood-brain barrier permeability, P-glycoprotein activity, plasma protein binding and / or binding affinity of any of the compounds.

49. The method of any one of claims 1-48, wherein the first compound descriptor input into the model comprises compound structure and / or experimental data.

50. The method of any one of claims 1-49, wherein the predictive model is a machine learning model.

51. The method of claim 50, wherein the machine learning model comprises a neural network.

52. The method of claim 51 , wherein the neural network is a graph neural network.

53. A method according to any one of claims 50-52, wherein the machine learning model includes a graph neural network and an attention layer.

54. The method of claim 53, wherein the neural network is a graph attention network.

55. The method of any one of claims 1-54, wherein the method comprises performing a validation assay on at least one compound in the third compound library.

56. The method of any one of claims 1-55, wherein the method comprises performing a low throughput analysis on at least one compound in the third compound library.

57. The method of any one of claims 1-56, wherein the method comprises inputting into a secondary system a third data set comprising compound descriptors and compound fitness scores for each compound in the third compound library in a validation assay.

58. The method of claim 57, further comprising inputting data from the validation assay into the predictive model.

59. A method according to claim 57 or claim 58, wherein the validation assay includes a proxy for binding or biochemical activity, the binding or biochemical activity including one or more of absorbance, fluorescence, luminescence, radioactivity, NMR, crystallography, microscopy including cryo-electron microscopy, mass spectrometry or Raman scattering, for example, surface plasmon resonance (SPR) measures the reflection of polarized light, which can detect changes in the reflection angle (refractive index), and fixing or binding a ligand (compound) to a surface (which contains an immobilized target protein) affects the mass or thickness of the surface, which changes the refractive index.

60. The method of any one of claims 1-59, wherein the predictive model generates a predicted compound descriptor for each compound in the first compound library and / or the second compound library.

61. The method of claim 60, wherein the compound fitness score is generated at least in part from the predicted compound descriptor for each of the compounds.

62. The method of any one of claims 1-61, wherein the first compound library is about 10,000 compounds to about 100 billion compounds or about 10,000 compounds to about 10 billion compounds; and wherein the second compound library is about 10,000 compounds to about 100 billion compounds or about 10,000 compounds to about 10 billion compounds.

63. A computer-implemented system comprising: A digital processing device, the digital processing device comprising: at least one processor; an operating system configured to execute executable instructions; a memory; and a computer program, the computer program comprising instructions that can be executed by the digital processing device to perform the method described in any one of claims 1-62.

64. A non-transitory computer-readable storage medium encoded with a computer program, the computer program comprising instructions executable by one or more processors to perform the method of any one of claims 1-62.

Images