Directed evolution of molecules through iterative experiments and machine learning

Abstract

A platform, system, medium, and method for machine learning-driven iterative drug discovery process are described herein. Some embodiments include: receiving a first input dataset comprising first binding interaction information between a target molecule and a library of compounds; processing the first input dataset with a machine learning module to generate a model representation of binding interactions, the representation being configured to predict binding between the target molecule and the input compounds; determining an updated library of compounds using the model representation of binding interactions, the updated library of compounds comprising one or more new compounds predicted to bind to the target molecule; receiving a second input dataset comprising second binding interaction information between the target molecule and the updated library of compounds; and processing the second input dataset with a machine learning module to update the model representation of binding interactions, the updated model representation improving prediction accuracy.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Application No. 63 / 320,137, filed March 15, 2022, and U.S. Provisional Application No. 63 / 320,143, filed March 15, 2022, which are incorporated by reference in their entireties herein. [Background technology]

[0002] Drug discovery is an expensive and time-consuming process in which candidate small molecule compounds, potentially numbering in the millions, are screened for potential pharmaceutical applications. Candidates can be identified through several approaches, including high-throughput screening (HTS), where compounds are individually tested in large numbers in biochemical assays, as well as in silico computational screening, where target binding is assessed by modeling and / or calculating molecular interactions. HTS is limited in practice to a maximum of 2-3 million compounds, but is typically performed on smaller numbers. Computational screening, on the other hand, requires a priori knowledge of both the target's crystal structure and information on how the compound should bind to the target, and often does not accurately describe how the compound will behave in the real world. Summary of the Invention

[0003] Disclosed herein are platforms, systems, and methods for improved drug discovery using DNA-encoded chemical libraries by directed molecular evolution utilizing an iterative machine learning process. Thus, clinically useful small molecules can be identified for targets and then rapidly optimized through an iterative process until the final candidate molecules are ready for clinical trials. Advantages of the present disclosure include the ability to identify molecules for a wide variety of target classes, the ability to find molecules for targets that others struggle with (i.e., are undruggable), the ability to rapidly optimize compounds for in vivo confirmation of efficacy, and the ability to efficiently bring multiple compounds to the clinical-ready stage. Instead of optimizing compounds by slowly iterating on tens of compounds every month, the present disclosure allows for the creation of large data sets at each stage of the drug discovery cycle. By working with large experimentally analyzed data sets (as opposed to in silico), multi-parameter optimization can be efficiently performed (e.g., in a single experiment with a wide range of conditions) for complex multi-condition tasks such as ADME / Tox.

[0004] An advantage of the present disclosure is that combining DNA-encoded libraries (DELs) with machine learning provides an improved screening process. The use of DELs offers advantages over high-throughput screening (HTS) and in silico computational approaches. For example, DELs allow processing 1,000 times more compounds than competing approaches (billions of compounds versus at most millions for HTS and traditional screening). DELs can efficiently process 10-100 times more parallel small-scale experiments ("conditions") than competing approaches, whereas HTS typically tests 1-2 conditions at a time due to the intense resource requirements for setting up each additional condition. In some cases, every combination of available building blocks is created in DELs, allowing for an exhaustive search of compounds, whereas HTS is heavily biased towards what has worked in the past, limiting the search space. For example, DELs do not require crystal structures or specific binding modes of action (unlike in silico virtual / computational screening). Instead, DEL can detect compounds that bind anywhere on a protein, e.g., allosteric binders, potential binding pockets, and compounds that could be a good basis for bispecific molecules (e.g., PROTACs and other proximity inducers such as molecular adhesives).

[0005] As a result, DEL can be applied to identify compounds that bind to entirely novel targets that are currently labeled "undruggable" because HTS does not cast a broad enough net net cast to discover these targets. Thus, the DEL approach generates rich and dense datasets filled with internal controls that are well suited to machine learning (ML) but at lower cost and shorter time to screen than traditional HTS.

[0006] The present platforms, systems, and methods provide an additional advantage by integrating machine learning and generating ML-designed DELs, as opposed to traditional HTS and DEL approaches, which typically track dozens of compounds at a time.

[0007] In one aspect, provided herein is a method comprising: (a) a computer-implemented method comprising: (i) receiving a first dataset comprising a first compound descriptor for each compound of a first library of compounds and a compound suitability score for each compound of the first library of compounds; (ii) training a predictive model on the first dataset; (iii) inputting the predictive model to a second dataset comprising a second compound descriptor for each compound of a second library of compounds; and (iv) generating a compound suitability score for each compound of the second library of compounds from the predictive model utilizing at least the one or more compound descriptors of the first library of compounds and / or the one or more compound descriptors of the second library of compounds; and (b) selecting a third library of compounds according to information comprising the one or more compound suitability scores of the second library of compounds and / or the one or more compound suitability scores of the first library of compounds. In some embodiments, the third library of compounds includes (i) one compound from the second library of compounds, (ii) one compound from the first library of compounds, (iii) a compound including two or more compounds from the second library of compounds, (iv) a compound including two or more compounds from the first library of compounds, (v) a compound including a compound from the second library of compounds and a compound from the first library of compounds, (vi) a compound that is not in the first library of compounds or the second library of compounds, (vii) a compound that includes a compound from the second library of compounds and a compound that is not in the first library of compounds or the second library of compounds, (viii) a compound that includes a compound from the second library of compounds and a compound that is not in the first library of compounds or the second library of compounds, or (ix) a combination of two or more of (i) through (viii). 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 the computer-implemented method. In some embodiments, step (b) is not part of the computer-implemented method. In some embodiments, step (b) comprises a first step that is part of the computer-implemented method and a second step that is not part of the computer-implemented method, and the first and second steps are performed sequentially, the first step being performed first, or the first step being performed second.

[0008] In some embodiments, the information further comprises an evaluation score for the compounds of the second library of compounds (sometimes referred to as an external fitness score) and / or an evaluation score for the compounds of the first library of compounds (sometimes referred to as an external fitness score). In some embodiments, the evaluation score for the compounds of the second library of compounds is a second fitness score generated independently from the compound fitness score generated from the computer-implemented method. In some embodiments, the evaluation score for the compounds of the first library of compounds is a first fitness score that is different from the compound fitness score for the compounds of the first library of compounds.

[0009] In some embodiments, one or more compounds of the first library are 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 (sometimes referred to as a byproduct or side product) generated during the synthesis of the first test compound, or an intermediate (sometimes referred to as an 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 a desired product of a synthesis reaction involving two or more of the building blocks of the first test compound. In some embodiments, the first byproduct is an undesired product of a synthesis reaction involving two or more of the building blocks of the first test compound. In some embodiments, one or more compounds of the second library (or optionally a subsequent library applicable in an iterative manner) are 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 by-product (sometimes referred to as a byproduct or side product) generated during the synthesis of the second test compound, or an intermediate (sometimes referred to as an 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 a desired product of a synthesis reaction involving two or more of the building blocks of the second test compound. In some embodiments, the second by-product is an undesired product of a synthesis reaction involving two or more of the building blocks of the second test compound.In some embodiments, one or more compounds of the third library are a third test compound (sometimes referred to as a full product or full product compound), a building block of the third test compound, a third by-product (sometimes referred to as a byproduct or side product) generated during the synthesis of the third test compound, or an intermediate (sometimes referred to as an intermediate) generated during the synthesis of the third test compound, or a combination of two or more thereof. In some embodiments, the third test compound is a desired product of a synthesis reaction involving two or more of the building blocks of the third test compound. In some embodiments, the third by-product is an undesired product of a synthesis reaction involving two or more of the building blocks of the third test compound. In some embodiments, a full product comprises a tricynthon and an intermediate product comprises a di-synthon and / or a mono-synthon.

[0010] In some embodiments, the first compound descriptor comprises data or information related to a compound of the first library of compounds, the data or information comprising binding affinity of the compound to a target molecule (optionally, the target molecule is a drug target such as a protein or a nucleic acid), activity of the compound, a physical property of the compound (e.g., lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing reads related to the abundance of the compound in an experiment, compound structure, information related to the synthesis of the compound, labeling data, process quality control data, yields related to 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 comprises data or information related to a compound of the second library of compounds, the data or information comprising the binding affinity of the compound to a target molecule (optionally, the target molecule is a drug target such as a protein or a 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 related to the abundance of the compound in an experiment, compound structure, information related to the synthesis of the compound, labeling data, process quality control data, yields related to 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 compound is a full product compound, an intermediate product compound, or a by-product compound.

[0011] In some embodiments, the method includes testing one or more of the compounds of the first library of compounds in an in vitro or in vivo assay. In some embodiments, the method includes testing one or more of the compounds of the third library of compounds in an in vitro or in vivo assay.

[0012] In some embodiments, each compound in the third library of compounds includes or is synthesized to include a nucleic acid tag, and the method further includes sequencing the third library of compounds to generate sequencing data associated with the third library of compounds. In some embodiments, the information in step (b) includes sequencing data associated with an external library of compounds (e.g., a library including a nucleic acid tag from each compound in the first library and / or second library). In some embodiments, the compound suitability score for each compound in the first library of compounds is generated from data including sequencing data associated with the first library of compounds. In some embodiments, the sequencing data includes a read count, a quality score associated with the read count, and / or a score calculated from a set of sequencing read counts or read counts from different experimental conditions from the first library of compounds and / or the second library of compounds. In some embodiments, the score includes the read count or the read count divided by the total number of reads in a selection of compounds or the average number of reads in a selection of compounds, or a similar mathematical function utilizing (directly or indirectly) the read count.

[0013] In some embodiments, at least one compound suitability score for each compound in the first library of compounds is generated from data including the 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). In some embodiments, the predictive model utilizes a probabilistic framework to process the first data set and the second data set and output the compound suitability score for each compound in the second library of compounds. In some embodiments, the suitability score is generated at least in part from data from full product compounds including non-target counts, target counts, and / or product proportion adjustment values. In some embodiments, the suitability score is generated at least in part from data from intermediate product compounds including non-target control counts, target counts, and / or product proportion adjustment values. In some embodiments, the method includes utilizing sequencing data associated with sequencing the third library of compounds to generate the compound suitability score for each compound in the third library of compounds.

[0014] In some embodiments, the method includes assaying the third library of compounds. In some embodiments, the assay includes binding the third library of compounds to a target. In some embodiments, the assay includes sequencing the third library of compounds or a subset of the third library of compounds (e.g., the subset is a subset of compounds that bind to a target).

[0015] In some embodiments, the suitability score of any one of the compounds comprises a binding and / or activity score of said compound. In some embodiments, said third library comprises one or more compounds from said second library having a compound suitability score greater than a threshold score.

[0016] In some embodiments, the method includes pre-processing the first and / or second datasets. In some embodiments, the pre-processing is performed before step i and / or before step iii of the computer-implemented method. In some embodiments, the method includes refining a fitness score generated from the predictive model, optionally where the refinement is performed by the predictive model, and / or optionally where the refinement includes incorporating information from an external library (e.g., a library of nucleic acid tags associated with the first and / or second libraries of compounds).

[0017] In some embodiments, the second library comprises one or more compounds that are different from the first library. In some embodiments, the second library comprises one or more compounds that are different from the first library.

[0018] In some embodiments, the method further comprises updating the model by repeating steps ii through vi. In some embodiments, iteratively repeating steps (a) through (b) identifies a set of potential compounds having one or more desired properties. In some embodiments, iteratively repeating steps (a) through (b) identifies a set of potential compounds having one or more desired compound suitability scores.

[0019] In some embodiments, the compound suitability score relates to any one of 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.

[0020] In some embodiments, the first compound descriptor input into the predictive model comprises a compound structure and / or experimental data, hi some embodiments, the predictive model is a machine learning model.

[0021] In some embodiments, the machine learning model comprises a neural network. In some embodiments, the neural network is a graph neural network. In some embodiments, the machine learning model comprises a graph neural network and an attention layer. In some embodiments, the neural network is a graph attention network.

[0022] In some embodiments, the method includes performing a validation assay on at least one compound of the third library of compounds. In some embodiments, the method includes performing a low throughput analysis on at least one compound of the third library of compounds. In some embodiments, the method includes inputting a third dataset including compound descriptors and a compound suitability score for each compound of the third library of compounds into a secondary system in a validation assay. In some embodiments, the method further includes inputting data from the validation assay into the predictive model. In some embodiments, the validation assay includes a proxy for 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, e.g., surface plasmon resonance (SPR) measures the reflection of polarized light, thereby detecting a change in the angle of reflection (refractive index), and immobilization or binding of a ligand (compound) to a surface (including an immobilized target protein) affects the mass or thickness of the surface, thereby changing the refraction.

[0023] In some embodiments, the predictive model generates a predicted compound descriptor for each compound in the first library of compounds and / or the second library of compounds, hi some embodiments, the compound suitability score is generated at least in part from each of the predicted compound descriptors for the compounds.

[0024] In some embodiments, the first library of compounds is about 10,000 compounds to about 100 billion compounds or about 10,000 compounds to about 10 billion compounds, and the second library of compounds is about 10,000 compounds to about 100 billion compounds or about 10,000 compounds to about 10 billion compounds.

[0025] In one aspect, a computer-implemented system is provided herein that includes a digital processing device, the digital processing device including at least one processor, an operating system configured to execute executable instructions, a memory, and a computer program including instructions executable by the digital processing device to perform a method described herein.

[0026] 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]

[0027] The novel features of the present disclosure are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present disclosure will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the disclosure are utilized, and the accompanying drawings. [Figure 1]FIG. 1 depicts a flowchart of a method of iteratively screening a DNA-encoded library of compounds for target binding affinity and using the data to train a machine learning model to generate compound predictions for the updated library of compounds, according to one or more embodiments disclosed herein. [Diagram 2] FIG. 2 shows a diagram of three synthetic steps in a combinatorial chemical pathway for library generation, according to one or more embodiments disclosed herein. [Diagram 3] FIG. 3 shows a comparison between the traditional compound screening process and DNA-encoded libraries, which enable large-scale screening. [Figure 4] FIG. 4 shows a diagram of the pooling and splitting process to further expand the size of the DNA-encoded library. [Diagram 5] FIG. 5 shows a diagram of the pooling and splitting process to further expand the size of the DNA-encoded library. [Figure 6] FIG. 6 illustrates an exemplary process for a binding assay experiment to identify compounds that bind to a target molecule within a combinatorial library of DNA-encoded compounds. [Figure 7] FIG. 7 shows 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. [Figure 8] FIG. 8 shows an exemplary graph mapping the compounds in the library to known products, known intermediates, and unknown compounds that can be used for quality control. [Figure 9] FIG. 9 shows mass spectrometry data of the compounds in the reaction mixture before and after the chemical synthesis step was performed, which can be used to determine the efficiency of the synthesis for quality control purposes. [Figure 10] FIG. 10 shows the reference mass spectrometry data for the library region. [Figure 11]FIG. 11 shows the compound fractions corresponding to the sequence of chemical synthesis steps used to construct the DNA-encoded library. [Figure 12] FIG. 12 shows a graph of sequencing read counts plotted against number of compounds comparing the DNA-encoded libraries disclosed herein to conventional DELs where the disclosed DNA-encoded libraries achieve a higher dynamic range by having the same sequencing depth for a smaller number of molecules. [Figure 13] FIG. 13 provides an overview of the iterative DEL library process using the deep neural network architecture disclosed herein. [Figure 14] FIG. 14 provides a diagram of a graph convolutional neural network architecture configured to receive input data and generate predicted compound fractions and binding scores, according to one or more embodiments herein. [Figure 15] FIG. 15 shows a star chart 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). [Figure 16] FIG. 16 illustrates a non-limiting example of a computing device, in this case a device having one or more processors, memory, storage, and network interfaces. [Figure 17] FIG. 17 illustrates a non-limiting example of a web / mobile application provisioning system, in this case one that provides a browser-based and / or native mobile user interface. [Figure 18] FIG. 18 illustrates a non-limiting example of an alternative application provisioning system as a distributed cloud-based architecture. [Figure 19]FIG. 19 shows a comparison of compound rankings based on predicted binding scores generated by the initial baseline model and the updated model using the evolved DNA-encoded library validated according to true positive hits. [Figure 20] FIG. 20 shows 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 after the first iteration (ML1) and compounds identified in the second evolved DEL using the second evolved model after the second iteration (ML2). [Figure 21] Figure 21 shows an evaluation of predicted compound binding to two highly related protein domains (95% sequence similarity) within a single protein, plotting compound scores for domain 1 on the Y-axis and compound scores for domain 2 on the X-axis. Each compound is represented by a blue dot and scores are derived directly from the DEL data. [Figure 22] FIG. 22 shows an evaluation of predicted compound binding to two related proteins within the same class of chromatin regulators, where some compounds were predicted to bind specifically to only one of the two proteins. [Figure 23] FIG. 23 shows a non-limiting example of a model architecture for predicting enrichment and alignment values ​​of tricynthons and di-synthons. [Figure 24] FIG. 24 is a bar graph showing the loss and R2 score on the test data for the full ML model compared to other models. [Diagram 25] FIG. 25 is a bar graph showing the area under the ROC curve for 150 external molecules tested using the full ML model and various other models. [Figure 26] FIG. 26 shows a bar graph depicting the Hit rate in the top 10 molecules of 150 external molecules tested using the full ML model and the flat yield model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0028] Disclosed herein are platforms, systems, and methods for compound discovery using DNA-encoded chemical libraries (referred to as "DELs") in an iterative process using artificial intelligence. Compound discovery can include an integrated process that combines DEL design, experiments, computational / ML analysis, and follow-up experiments to allow for a rapid feedback loop between experiments and computations. End-to-end integration of the entire iterative process allows for the generation of higher quality data at greater scale, flexibility, specificity, and speed. In addition, ML-assisted iteration of DELs can include dynamic reconfiguration of building block selection with each iteration, generating richer and more useful data. The ML model can then select building blocks and synthesis schemes that have little or no overlap with the original library. This can help ensure that only the best compounds are built, even if the compounds were not tested in the source library. In some cases, library generation can be more flexible and allow for more diverse library synthesis schemes. For example, library designs, synthesis schemes, and manufacturing methods that would otherwise be expensive but could generate high-value structures become feasible because ML-guided synthesis decisions are more likely to succeed and allow for more focused efforts.

[0029] DNA Code Library DEL is a collection of chemical compounds that can (but need not) be stored in a single tube, with each compound in the tube physically linked to a unique DNA sequence that represents the compound (which can be used as a barcode to identify the compound). In some cases, as used herein, the DNA sequence, structure, synthesis, or other data or information related to the compound may be referred to as a compound descriptor. In some embodiments, the compound descriptor includes data or information related to the compound of the library of compounds. The data or information may include the binding affinity of the compound to a target molecule (optionally, 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, sequencing leads related to the abundance of the compound in an experiment, information related to the synthesis of the compound, labeling data, process quality control data, or yields related to 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 of these. By creating links between up to billions of compounds and their corresponding unique DNA sequences, a single DEL Tube effectively miniaturizes compound collections that would otherwise take up an entire warehouse.

[0030] DELs can be constructed by combinatorial chemistry, where chemical building blocks (e.g., monomers) are assembled together on the ends of pieces of DNA, thereby synthesizing a variety of coded combinations in a single tube. Each time a new building block is added, a new segment of DNA is also added, with the DNA acting as a barcode for said building blocks. Thus, the DNA barcode corresponding to the compound grows as new building blocks are added, thereby allowing the sequence of the DNA barcode to decode the identity and order of addition of the building blocks. Given a DNA sequence from a DNA code library, a key that maps the DNA barcode to the building blocks can be used to determine which compound was made (i.e., the compound structure).

[0031] As shown in Figure 6, traditional HTS separates individual compounds, typically one compound per well, and spatially "labels" the compounds with their well location. Each of these compounds can be tested one at a time, often in an automated fashion, to see if the compound binds to or otherwise interacts with the target. Compounds that bind to or otherwise interact with the target can be selected for further optimization. This approach is limited in practice to a low order of magnitude of compounds, and to a small number of measurements per compound. In contrast, DEL can label each compound with a DNA sequence that can later be decoded using a DNA sequencer. This allows for the mixing of compounds in a single tube without compromising the ability to later identify each compound. Compound mixtures can be filtered by exposing the mixture to immobilized proteins as shown in Figure 6 and seeing which DNA sequences are recovered as binders to the target. This can allow for testing a large number of compounds in parallel against a large number of different scenarios (i.e., different proteins or different versions of a protein). This is useful for generating large amounts of dense data for machine learning purposes.

[0032] In some embodiments, DEL library is configured with an optimized size range.One advantage of DEL library over high-throughput screening is that it can screen hundreds of millions of compounds instead of millions.In some cases, 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, billion or billion unique coded compounds.

[0033] However, while larger DEL libraries may theoretically allow for the testing of a larger chemical space, overly large libraries may contain too many non-drug-like compounds, poor chemical fidelity, and inherently poor signal-to-noise ratios that contaminate downstream analyses with useful and / or uninterpretable data. Some parameters that may govern the DEL signal-to-noise ratio include the copy number of each individual molecule at the start of the experiment, the number of selection rounds performed, and the number of sequencing reads the sequencer can generate. In overly large libraries, each individual compound may be represented by only a few copies in the tube, and many compounds may no longer be present after a few selections, while those that remain may be represented by even fewer copies. The sequencer may only output a fixed number of reads. By starting with many more possible compounds, each individual compound receives only a small number of reads, thereby making it difficult to distinguish the performance of one compound from another. As a result, DEL libraries may become increasingly noisy as they grow larger, especially beyond a certain threshold. In some cases, the DEL library has 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, up to 10 billion or 100 billion unique coding compounds. In some embodiments, the methods and systems herein utilize or comprise multiple DELs, for example, as inputs or outputs of the model, and thus each DEL can vary in size.

[0034] Furthermore, while traditional DEL approaches may provide clearer signals using relatively small libraries, in doing so they compromise their ability to broadly explore chemical space. In contrast, the platforms, systems, and methods disclosed herein can enable computational design of highly efficient libraries to cover large amounts of chemical space, despite downstream production of fewer overall compounds. For example, an initial set of DEL libraries can be constructed and tested for binding to a given target. The resulting data set can consist of a smaller set of compounds with various binding affinities to the target and a much larger set of compounds that apparently did not bind. Because of their small number, the compounds that bind to the target in the initial library are necessarily much less diverse than the initial library. Furthermore, any compounds that are chemically similar to these initial compounds (e.g., via molecular fingerprints and Tanimoto distances) will struggle to expand this relatively narrow chemical space. Smaller libraries can make it easier to reliably identify useful compounds, but the number of these compounds is more limited.

[0035] In some embodiments, the training methods and models described herein may address the inherent limitations described above by taking advantage of the higher signal-to-noise ratio of the dataset to train machine learning (ML) algorithms and build ML models with a preliminary understanding of the chemical space suitable for binding to the target. The model's preliminary understanding and ability to train the model on a dataset that provides a more holistic description of the compounds (e.g., DELs) of the library allows the ML model to explore beyond the chemical diversity of the original hits and design new libraries that may share little overlap in the building blocks (e.g., compounds) present in the synthetic plan and / or synthetic plan of the original library, where the ML model may use binding to the target as a primary criterion. The resulting library may be more structurally diverse compared to the initial positive examples and / or their derivatives, while further focusing on areas of chemical space that are enriched for compounds that bind to the desired target. This allows the ML model to determine which building blocks and / or library designs are most effective for a given target, allowing the ML model to more effectively identify more useful compounds, for example, by generating a fitness score for the compounds, while testing fewer total compounds. In some embodiments, one or more compounds (e.g., building blocks) and / or libraries in ML design synthesis may be present in the initial library. In theory, it is possible to construct a library that contains all possible compounds (e.g., building blocks), but the library would be impractically expensive to construct and would require 100 or 1000 times more material and 100 or 1000 times more sequencing and analysis capacity to achieve the same signal-to-noise and identify useful compounds.

[0036] The platform, system and method disclosed herein can leverage the superior signal of smaller libraries combined with machine learning driven iterative design of smaller libraries to search chemical space more efficiently and provide a clearer understanding of the chemical space for a given target.Even with the largest practical initial library using conventional methods and models, the ML models trained and described herein can help to create and use smaller high signal-to-noise libraries to more effectively explore compounds while still providing a clearer understanding of the chemical space given the target.In some embodiments, the signal-to-noise ratio can be the percentage of compounds that are confirmed to function in the secondary assay out of all compounds tested in the secondary assay.

[0037] In some embodiments, libraries (e.g., DELs) may be constructed as illustrated in FIG. 2 by adding building blocks or compounds 211 to tubes containing substrates. The building blocks 211 may be attached to pieces of DNA 212 with unique sequences representing the building blocks 211. This is done during a single synthesis cycle 210. In some embodiments, a second cycle 220 may be performed in which new building blocks 221 and DNA tags 222 may be added to the strands from the first cycle 210. In some embodiments, a third cycle 230 may be performed each time with new building blocks to form longer strands 231 incorporating new DNA tags 232. In some embodiments, more than two cycles may be performed and the process may be repeated until one or more desired DELs are constructed.

[0038] The size of the library (e.g., DEL) may be further expanded using a split-pool synthesis method, as illustrated in Figures 4-5. In some embodiments, the substrate is split into several tubes, as illustrated in Figure 3, although more tubes (e.g., splits) may be used. A different building block 311, 321, 331 is added to each tube. A corresponding DNA tag labeling each building block 312, 322, 332 may also be added to each tube. In some embodiments, each tube may undergo a single cycle, as described in the first step 210 of Figure 1. The results from each tube can be pooled together to create a mixture 340 of all three resulting species.

[0039] In some embodiments, the split pool DEL synthesis method may further include a second step, as shown in FIG. 5. The mixture of compounds 340 resulting from the first step shown in FIG. 4 may be split into multiple tubes (e.g., 3). In some embodiments, more than three tubes may be used. Different building blocks 351, 361, 371 may be added to each tube. Corresponding DNA tags labeling each building block 352, 362, 372 may also be added to each tube. In some embodiments, each tube undergoes a second cycle, as described in the second step 220 of FIG. 1. The results may be pooled together to create a mixture 380 of all three resulting chemical species. As shown in FIG. 4-5 by the starting point being mixture 340, the mixture 380 resulting from the second step may be a combination of all possible building blocks in both the first and second positions. This results in 3*3=9 different compounds in mixture 380. In some embodiments, the process may be repeated multiple times to generate more complex mixtures. For example, the addition process results in 3*3*3=27 compounds.

[0040] Library construction (e.g., DEL construction) may include one or more of several construction techniques, such as: building block combination, scaffold-directed (i.e., modifying a core scaffold), or in situ ring formation (e.g., forming diverse heterocycles). In some embodiments, DEL libraries are constructed to exceed a threshold percentage of conformance with Lipinski's rule 5 (a rule of thumb for evaluating compounds as drug candidates). Specifically, Lipinski's rule 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); a molecular weight of less than 500 Da; and an octanol-water partition coefficient (logP) of 5 or less. In some cases, DEL libraries are constructed to be more than 60%, 65%, 70%, 75%, or 80% Rule Five conformance. In some embodiments, libraries are screened to ensure that compounds below a threshold percentage are unknown. Compounds (e.g., building blocks) or parts thereof used to generate DELs can be evaluated to determine the amount of desired product, the amount of known intermediates and by-products, and the amount of unknowns. In some embodiments, building blocks are not passed through production for DEL construction if the percentage of unknown compounds exceeds a threshold amount. For example, building blocks can fail the test if the percentage of unknown compounds is 1%, 2%, 3%, 4%, 5%, 6%, 7%, 8%, 9%, 10%, 11%, 12%, 13%, 14%, 15%, 16%, 17%, 18%, 19%, or 20% or more.

[0041] In some embodiments, extensive intra-library quality control can be performed. In a non-limiting example, spike-in can be performed to monitor non-linked headpiece substrates into every well of the synthesis to monitor single 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 low molecular weight conversions can be difficult to monitor on large 30+kDa molecules.

[0042] DEL allows for bulk assays to be performed as opposed to individual screening, which is more time and resource intensive. In a non-limiting example of a process 600 for isolating compounds of interest, DEL 630 containing billions of compounds can be provided in a single tube or well for target binding assays. For example, a protein of interest 613 (e.g., a "target") can be immobilized on a support 611, and the entire DEL 630 can be 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 the resulting complex 610. The complex 610 can be washed (i.e., with a buffer) to remove the support and / or undesired weak or non-specific binders. The remaining compounds can be eluted 620, releasing 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 to obtain a list of sequences (representing compound structures) that bind to the target protein, and how frequently these sequences were represented 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 bound compounds. Compounds that successfully bind to the target protein (and to a negative control such as a support) may be suitable candidates as a starting point for drug design.

[0043] In some embodiments, a DEL experiment may attempt to identify which compounds in a library bind to a protein of interest. A DEL experiment may involve one or more experimental conditions (e.g., each tube can be viewed as a mini-experiment or "condition"). In some embodiments, the results of the conditions tested in a DEL experiment may be compound descriptors. In some embodiments, a simple DEL experiment may include two conditions: a binding assay with the protein of interest and another binding assay without the protein that serves as a control. Since one condition tests an entire DEL of compounds, each condition creates its own large-scale data set for millions or even billions of compounds. In some embodiments, the data set may be a composite descriptor. In some embodiments, the data set may be used to generate a composite fitness score. This means that DEL miniaturizes large-scale experiments into a single tube at low cost, as opposed to other approaches such as high-throughput screening (HTS), which perform individualized assays on a massively parallel scale. In a non-limiting example, instead of investigating a single condition, DEL allows experiments to be performed across multiple conditions to ask several questions about the compound, such as information about the compound or a property of the compound (sometimes referred to as a compound descriptor), including, but not limited to, affinity (high or low concentration of protein), specificity (for a protein that may be a member of the same protein family as the target or a mutant version of a closely related protein), and / or binding location (mutating a known binding pocket or adding a known competitive binder to the mixture against the target). In some embodiments, the compound descriptor includes data or information related to a compound in a library of compounds (e.g., DEL).The data or information may include binding affinity of the compound to a target molecule (optionally the target molecule is a drug target such as a protein or a nucleic acid), activity of the compound, a physical property of the compound (e.g., lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing reads related to the abundance of the compound in an experiment, information related to the synthesis of the compound, labeling data, process quality control data, or yields related to 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.

[0044] In some embodiments, each condition may be run in its own tube before being read out on the sequencer. The number of conditions used in an experiment may be at least 2, 3, 4, 5, 6, 7, 8, 9, or 10 or more conditions. The conditions may include different protein concentrations, such as at least 2, 3, 4, 5, 6, 7, 8, 9, or 10 or more protein concentrations. The conditions may 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 may 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 may include an internal control for each experimental condition. In a non-limiting example, a binding assay for a wild-type protein may include an internal control including (1) a support (no protein), (2) a different protein that is not expected to produce specific binding (e.g., bovine serum albumin), and (3) a protein with a mutation to the binding pocket.

[0045] The DEL data can be a set of DNA sequences and the number of times that sequence has been observed. The sequence can be decoded into a composite structure. The number of times that a structure is observed (e.g., the number of sequences or "hits" observed) can be related to how tightly that structure binds to the target. In some embodiments, the composite fitness score is related to the number of "hits". However, the number of hits can also be related to various factors before and after the binding event. These factors can include library generation efficiency, PCR amplification, sequencing, and many other steps that can affect the count and / or compound fitness score. For example, synthesis steps during library generation can have different efficiencies that result in an unequal number of compound species in the library. Compounds synthesized with lower efficiency can then result in a relatively smaller number of sequence "hits" than would be expected simply based on binding efficiency, since there was a small amount of the compound in the library.

[0046] The traditional DEL dogma holds that DEL output cannot be correlated with the affinity that a compound has for a target. A compound may appear as highly enriched in DEL selection (i.e., have many reads in the sequencer) but only be a weak binder, or vice versa. This is because DEL reads are a noisy process. Many factors complicate the reads, which may include, for example, variable chemical yields, unexpected by-products generated during library synthesis, DNA binding (due to barcodes), matrix / support binding, promiscuous binding, undersampling of compounds, amplification bias, and sequencing noise.

[0047] The platform, system and method disclosed herein can account for noise sources that result in DEL output with read counts that correlate with binding affinity. In some embodiments, DEL is built from the ground up to minimize noise and maximize signal, which allows machine learning models to be developed using data generated from DEL to effectively distinguish tighter binders from weaker binders.

[0048] Sequencing and selection In some aspects, the platforms, systems, and methods disclosed herein utilize deep sequencing for complete sequencing but not partial reads. In some embodiments, the selection coverage of molecules in the DEL is at least 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%, 99%, or 100%. In some embodiments, the range of selection coverage is 50%, 55%, 60%, 65%, 70%, or 75% to 80%, 85%, 90%, 95%, 99%, or 100%. Machine learning can be used to screen libraries of compounds and predict new evolved libraries of compounds for iterative analysis rounds. Multiple rounds of selection can be performed to identify binders.

[0049] target molecule Disclosed herein are platforms, systems, and methods for predicting compounds that interact with target molecules. The target molecule may 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 target molecules present in bacteria, yeast, fungi, or parasites), or an mRNA transcript of a gene (e.g., MYC). The target molecule may be associated with a disease or a symptom thereof. In some embodiments, the target molecule may belong to a biological pathway associated with a disease or a symptom thereof, or modulate or regulate the pathway. Thus, a compound predicted to bind to the target molecule may act to inhibit the target, thereby regulating the associated pathway. In a non-limiting example, the compound may 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.

[0050] The protein may be wild type (i.e., the most common allele in a population) or a mutant allele. The mutation may be naturally occurring or may be an engineered mutation, such as mutating the binding pocket of the target protein to evaluate the specificity of binding compared to wild type control conditions. The mutation may be a silent mutation that does not affect protein function, or may result in a gain of function (e.g., enhanced ligand binding) or loss of function (e.g., reduced or complete loss of ligand binding).

[0051] In some embodiments, the target protein may be an enzyme that catalyzes a chemical reaction. An enzyme may have a globular structure with an active site configured for substrate binding and catalysis composed of a relatively small number of amino acid residues. An enzyme may also have an allosteric binding site to which an effector molecule binds to change the structural conformation of the enzyme, thereby enhancing or decreasing its enzymatic activity. In some embodiments, the target protein is a structural protein. A structural protein may be a fibrous protein (e.g., collagen) or a globular protein (e.g., actin or myosin). A target protein may be involved in cell signaling, such as a transmembrane receptor protein kinase. A target protein may be a transport protein, such as a transmembrane ion channel protein. In some cases, a target protein may have a quaternary structure composed of two or more protein subunits. Thus, experimental conditions can be performed to identify compounds that specifically bind to structural features of a particular protein by comparing binding with target mutations, such as in the substrate binding site or the protein-protein interface between subunits in the quaternary structure.

[0052] Screening and Validation Compounds identified by the platforms, systems, and methods disclosed herein may be validated by further screening. Various established techniques can be used to validate 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) can measure the reflection of polarized light and detect changes in the angle of reflection (refractive index). Immobilization or binding of a ligand (compound) to a surface (containing an immobilized target protein) affects the mass or thickness of the surface, which changes the refraction. Other methods, such as fluorescence resonance energy transfer (FRET), measure changes in fluorescence intensity as a target and its ligand come together, and the change in intensity correlates with its interaction being disrupted by a compound.

[0053] Machine Learning Algorithms Disclosed herein are platforms, systems, and methods that provide an iterative process in which laboratory experiments and machine learning mutually reinforce each other to generate new data and learn which compounds are likely to interact with medically relevant targets. In some embodiments, machine learning algorithms are utilized to determine compound properties, such as binding. In some embodiments, the machine learning algorithms herein employ one or more forms of labels, including, but not limited to, human-annotated labels and semi-supervised labels. In some embodiments, the machine learning algorithms utilize regression modeling, where the relationship between predictor variables and dependent variables is determined and weighted.

[0054] Examples of machine learning algorithms may include support vector machines (SVMs), naive Bayes classification, random forests, neural networks, deep learning, or other supervised or unsupervised learning algorithms for classification and regression. Machine learning algorithms may be trained using one or more training datasets.

[0055] In some embodiments, the platforms, systems, and methods disclosed herein provide efficient searching of chemical space with large experimentally tested data sets, where machine learning is used to interpret the results. In some embodiments, specialized machine learning models are constructed to exploit the data. In some cases, the model architecture exploits one or more unique features of DEL library construction and explicitly represents those features. This allows the model to outperform more traditional architectures when applied to the same data. Traditional models typically only receive data relating sequences to counts as input. In contrast, models are disclosed herein that receive data relating sequences to counts, as well as additional features such as matrix binding data, randomness data, and / or building block validation data (which estimate which fractions of the reaction will proceed to the next step). In particular, the incorporation of an estimate of reaction efficiency via the reaction fractions predicted to proceed to the next step addresses the technical problem that variations in reaction efficiency can result in poor prediction accuracy when the model assumes perfect reaction yields.

[0056] Traditional approaches use only large data sets in their initial searches and rapidly collapse their search space to tens of compounds in a short period of time, such as weeks or months. This means that exploration in chemical space is slow and inefficient, as many candidate compounds with potential clinical relevance can be screened and the number of remaining compounds is too small to effectively identify new unrelated compounds. In contrast, the platforms, systems, and methods disclosed herein provide cutting-edge parallel chemistry and biochemistry capabilities that allow the generation and testing of large data sets more quickly than others to refine the understanding of chemical space and find better medicines. In some embodiments, the machine learning designed DELs are evolved DELs generated as a follow-up to the initial DEL library. The ML designed DELs can have a smaller size than traditional DEL approaches while remaining much larger than traditional HTS approaches. For example, the initial library can have a size of a billion compounds, while the evolutionary ML designed library can be generated with over a million compounds for the target / problem specific library.

[0057] In addition to machine learning-assisted library design, selection strategies can be designed to maximize the signal in ML model construction. The improved ability of machine learning to understand complex relationships allows more selection conditions to be interpreted in parallel. In a non-limiting example, different variants of a protein as well as closely related family members can be used. The improved ability of machine learning to understand complex relationships allows more selection conditions to be interpreted in parallel. In a non-limiting example, different variants of a protein as well as closely related family members can be used. The software applications, systems, and methods described herein are specifically designed to process and utilize this additional data to use for training and generating ML models and to more efficiently interpret selection conditions. Taken together, these capabilities allow for rapid exploration of chemical space to find answers to challenging drug discovery problems more efficiently and effectively.

[0058] After the initial DEL search finds a good starting point, ML designed DELs can be used to refine the understanding of the conditions for a compound to be "good" for a given target, significantly shortening the lead optimization process timeline. ML designed DELs can be devised by considering one or more factors (where each factor can be considered as a descriptor associated with the compound 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, as well as predicted ADME (adsorption, distribution, metabolism, excretion), and predicted toxicity properties. Because ML can consider many more compounds and large libraries can be efficiently constructed and tested by the system, the chemical space is less narrow compared to traditional approaches that consider only tens to hundreds of compounds per iteration.

[0059] Traditional HTS and DEL approaches suffer from a collapsing search space. After performing an initial mid-scale (500,000-2 million compound HTS) or large-scale (billion compound DEL) search, a small number of Dozen compounds are selected for follow-up and further experiments. In other words, traditional HTS and DEL immediately collapse their search space by 6-9 orders of magnitude and revert to the slow and expensive process of lead optimization that is responsible for over one-third of drug development costs. This transition from a wide-angle view of chemical space to a restricted window of a small number of Dozen compounds is a huge inefficiency in the drug discovery process.

[0060] In contrast, the platform, system, and method disclosed herein can generate millions of compound follow-up libraries (e.g., DELs) and maintain its wide-angle view of chemical space while still focusing on the best compounds. The generated datasets enrich the ML model with more data, giving a precise understanding of which compounds perform well for a given target. The datasets can be composite descriptors. 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 to the ML model, but also allow the model to efficiently learn from its own predictions. Each library iteration can consist of compounds for which the ML model has a strong hypothesis. In a non-limiting example, the ML model can consider the compounds in the ML-generated DELs to be strong or weak binders, or be uncertain about affinity, but make explicit predictions (e.g., compound suitability scores) for each compound model, which can then be tested in gold standard experiments. 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.

[0061] A new set of random compounds, or a set of compounds naively selected as similar to the original library, would provide less useful information compared to an ML model that looks at and interprets the results of its own predictions. This is because the random compounds lack the initial hypotheses provided by the ML model. Thus, the combined laboratory and ML iterative systems and methods described herein allow the methods of training and updating ML models to be much more efficient and effective. In some embodiments, the DEL data (e.g., composite fitness scores) and ML prediction accuracy may be sufficient to warrant the synthesis of a smaller number of compounds for low-throughput testing. At this stage, the iterative process may have generated high-performance ML models, and a variety of high-quality compound structure starting points generated from the initial process. Thus, the resulting list of compounds may have less bias and may have an optimized ML model to aid in the search for the next improvement. In some embodiments, the compound fitness scores relate 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.

[0062] In some embodiments, the initial DEL has a starting size below a threshold.The initial DEL is a non-ML repeat DEL.The starting size may be smaller than conventional DEL, and does not include all possible compounds that can potentially be evaluated, but is instead designed to optimize the overall diversity of compound structure.Then, once the first DEL is evaluated (e.g., via binding assay and hit sequencing), experimental data (e.g., descriptors and / or composite fitness scores) can be input into the ML model, which is then used to generate and / or identify evolutionary / repeated DEL. The evolved DELs may have a size of 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. At each iteration, the DELs may remain the same or decrease in size until a sufficient number of compounds having a threshold quality (e.g., binding score, affinity score, activity score, compound suitability score, etc.) are identified for follow-up low-throughput analysis.

[0063] The use of an initial library (e.g., DEL) below a threshold size (e.g., less than 100 billion, less than 10 billion, less than 1 billion, less than 500 million, less than 100 million, etc.) can significantly improve the signal-to-noise ratio. A larger library has the drawback of being too noisy when there is too little signal, meaning that the same compound in a larger DEL is difficult to find due to noise and is much less likely to be found compared to other compounds in the library. In some embodiments, the evolved library (e.g., evolved DEL) disclosed herein can use specialized building blocks with more interesting structures that provide increased structural diversity and thus help increase the chemical space searched to identify hits with potential (e.g., specific target binding). Follow-up evolved DELs can then include new compounds that were not previously included in the initial DEL but that may be structurally predicted by the ML model to provide binding and / or enhanced binding. In this way, a wide breadth of chemical space exploration is balanced with selectivity for the desired properties and the ability to reliably identify compounds in the noise.

[0064] Various machine learning models and algorithms may be utilized in accordance with the platforms, systems, and methods disclosed herein. Machine learning methods may include artificial neural networks, decision trees, support vector machines, regression analysis, naive Bayes, random forests, gradient boosting, XGBoost, and other suitable techniques. For example, the model architecture may include XGBoost for molecular features, multi-layer perceptrons for fingerprint vectors, or graph convolutional neural networks (GCNN). In some embodiments, the machine learning model includes a neural network. The neural network may be a convolutional neural network (CNN). In some embodiments, the machine learning network includes a graph convolutional neural network (GCNN). GCNN is well suited to evaluate graph-type data representations of molecules. The sequence may be converted into a set of structures represented by a graph with an array associated with each node and edge in the graph.

[0065] In some embodiments, the neural network architecture includes one or more sequence-oriented layers that account for the sequence elements of the DELs (each compound is the product of a series of reaction steps). The ability to add these sequence-oriented layers is a unique feature of neural networks that has been exploited to provide additional useful information to the model for improved compound selection to generate evolved / iterative DELs. In a non-limiting example, these sequence layers can be used to show the ML model a set of compounds that may have been synthesized during the inherently noisy process of library construction. In contrast, traditional architectures only consider a single compound at a time and therefore cannot account for this complexity. The sequence layers can also incorporate additional information, including but not limited to the synthetic yields of various reactions, allowing the ML model to better learn the relative importance of different compounds in the shown set. The architecture of the ML model described herein allows for the incorporation of one or more new and different types of data into the training dataset, allowing the ML model trained on the data to provide more accurate and better composite designs.

[0066] In some embodiments, DEL is assayed according to a selection experiment with one or more conditions. For example, DEL can be evaluated in a protein binding assay against 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 code for identification. Sequencing data can include the unique DNA sequences detected and their corresponding number of sequence counts or hits, which can correspond to the relative abundance of the compound in the sequenced sample (notwithstanding the various noise sources discussed herein). The detected compounds can include whole structures or parts of structures, such as monosynthons, di-synthons, tricynthons, by-products or other polysynthons.

[0067] The platform, system, and method disclosed herein may utilize a unique machine learning architecture that is specifically tailored 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, which means that intermediates and by-products may be generated during DEL production. As shown in FIG. 11, 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 does not only consider the final product compounds, but also intermediate compounds generated during the intermediate synthesis steps. Traditional approaches to ML for DEL do not take these intermediates or by-products into account. In contrast, embodiments of the platform, system, and method disclosed herein utilize a neural network that incorporates possible intermediates or by-products into its training data. As a result, potential intermediates and side reactants or products present in the DEL are explicitly accounted for as possible factors explaining the lead count. In contrast, traditional approaches that assume only the final product is produced assume that the final product is synthesized with 100% efficiency when, in fact, many compounds may have been synthesized with less than 30%. Traditional approaches are therefore trained on "wrong" data, which may impair model performance. Thus, 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 may further enhance its consideration with additional data in the form of measured chemical yields. This type of data is not present in the core DEL selection data, and this is the only data that the aforementioned model architecture can consider due to inherent limitations in its design.The ML model described herein can incorporate this additional information and is designed to consider the full set of synthetic compounds along with the DEL selection data, thus more accurately accounting for the underlying biochemical processes from the experiments. By leveraging its architecture and previously inaccessible information, its predictive accuracy can be dramatically improved.

[0068] In some embodiments, one or more quality control steps are performed to control for reaction inefficiencies. Measurements may be collected during one or more of the library construction (e.g., DEL construction) steps. Measurements may include mass spectrometry of reaction products to identify reaction efficiency, e.g., intermediate products or relative abundance of final products versus remaining reactants. Such information may be used to predict the proportion of each type of product or intermediate compound, which may be utilized to generate weights for machine learning models to improve learning. In some embodiments, the measurements may be composite descriptors. In some embodiments, the measurements are used to determine compound descriptors and / or fitness scores. In some embodiments, the compound fitness scores relate 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.

[0069] In some embodiments, one DNA barcode or tag can represent all of the compounds (final and intermediate) in a given fraction. In some embodiments, one DNA barcode or tag can be a compound descriptor. As described herein, because the reactions in DEL may not be complete, DNA tags are added to several molecules (final products, including intermediate products, by-products, etc.) that are generated during a given sequence of synthetic steps. Thus, the DNA tag can be understood as an indication that these synthetic steps were followed, rather than a complete molecule being generated. As a result, a unique DNA tag serves as an indication that the molecule added to it is one of the molecules that may be generated in such a reaction scheme.

[0070] The data collected for the DEL experiment is used to train a machine learning model. A given dataset can include a descriptor (compound descriptor) for each compound in the dataset, which can include a representation of the compound's molecular structure, the sequence of the compound's associated DNA tag, and sequence read counts for the DNA tag. The dataset can include additional information (e.g., descriptors), such as chemical properties or parameters (e.g., molecular weight). In some embodiments, the fitness score can be a score that indicates a compound that has the desired property (e.g., binding affinity or activity in a biochemical assay or some ADME measurement). The trained model can then be used to evaluate another set of compounds to identify compounds that are predicted to have the desired property (e.g., 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 the desired property can be generated as a new or evolved DEL, which is then subjected to 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, and then used again to identify new evolved DELs. This process can be repeated iteratively any number of times until a smaller set of candidate compounds is selected for further evaluation.

[0071] In some embodiments, the machine learning model is provided with input data generated from the DEL binding assay. The input data can be evaluated to determine an index of binding, such as a binding score. In some cases, the index of binding can be used to classify a given compound as a binder or a non-binder to the target protein (e.g., via a compatibility score). In some embodiments, the binding data can be combined with a biochemical activity dataset to generate a compatibility score. In some embodiments, the compatibility score can be generated by analyzing the binding data and / or the biochemical activity data to determine which compounds are likely to have a desired biological effect.

[0072] In some embodiments, the fitness score can be a number derived or calculated from the read counts of the DEL compounds 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 of the counts of that compound in a desired condition (target condition) to the 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 the compound in the original library. Many different mathematical functions that combine the read counts with other data can be calculated to obtain 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, the fitness score is derived from the read counts.

[0073] In some embodiments, compounds have molecular structures that must be converted into molecular representations suitable for input into machine learning models. Examples of molecular representations include molecular graph representations (mapping the atoms and bonds of a molecule to nodes and edges), electrostatic representations, textual representations such as SMILES or SELFIES, and geometric representations.

[0074] A graph representation is a 2D data structure that does not explicitly enumerate spatial relationships between elements, but 3D information can be incorporated into the graph representation: for example, a node feature matrix may contain this information as node information (e.g., the stereochemistry of the atoms represented by the node) and an edge feature matrix may contain this information as edge information (e.g., bond types).

[0075] In some embodiments, DNA tags and sequence counts are processed from raw data into a dataset suitable for machine learning analysis. In some cases, sequence read counts are corrected upwards or downwards.

[0076] In contrast, traditional ML utilizes data and labels, e.g., molecular structures are the data (i.e., structures assayed for binding to the target) and labels are binding scores (i.e., scores indicating binding affinity generated at least in part based on sequence counts). If a synthesis reaction step does not have perfect 100% efficiency, the data is one of several of a set of molecules and / or compounds. However, the methods disclosed herein allow for machine learning analysis when it is unclear which 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 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 the aforementioned model architecture can consider due to inherent limitations in its design. The ML model described herein can incorporate this additional information and is designed to consider the complete set of synthetic compounds along with the DEL selection data, so it more accurately considers the underlying biochemical processes from the experiments. By leveraging its architecture and previously inaccessible information, its prediction accuracy can be dramatically improved.

[0077] In other words, the label or binding score is known, but there are multiple molecules in the set represented by the DNA code, and as a result, it is unknown what the actual best binder in the set is. With reference to FIG. 11, the top compound in the third synthesis step may be the best binder, but other intermediate products may also experience some binding. Although the DNA barcode cannot distinguish between the final product and the intermediate product due to incomplete synthesis, the model is given all the products corresponding to the DNA barcode in order to more fully understand or learn from the set of compounds. Thus, in some embodiments, the model looks at the complete molecule, not just the disynthon, as in some alternative methods. As a result, the model is constructed based on the understanding that the DNA tag or barcode may represent either an intermediate product (disynthon), a reaction intermediate, or a complete product (e.g., tricynthon).

[0078] In some embodiments, the machine learning model is used to generate predictions for new input data. For example, once the machine learning model has been trained on input data generated from DEL experiments, it will be able to receive new input data and generate predictions of fitness, e.g., binding to a target protein. The fitness predictions can be output as a fitness score. The fitness score can be a binding score or other property. In some embodiments, the predictions include a composite score of multiple compound 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.

[0079] In some embodiments, the machine learning model learns one or more factors per possible molecule. Once trained, the machine learning model can then generate a read number prediction by aggregating all the factors, which may include one or more of the matrix binding propensity of each of the set of possible molecules, the promiscuity propensity of the set of possible molecules, and the target binding propensity of the set of possible molecules.

[0080] To score a molecule, a prediction can be made of the "target propensity" of the target molecule (thus removing all other factors that are not of interest).

[0081] Alternatively, in some cases, a multi-parameter score is generated that includes other characteristics instead of just a single characteristic, such as target propensity. Non-Limiting Numbered Embodiments 1. A method comprising: i) receiving a first input dataset comprising first binding interaction information (e.g., compound descriptors) between a target molecule and a library (e.g., set) of compounds; ii) processing the first input dataset using a machine learning module to generate a model representation of the binding interactions (e.g., predicted compound descriptors), where the model is configured to predict a binding compatibility score between the target molecule and the input compounds; iii) determining an updated library of compounds using the model representation of the binding interactions, where the updated library of compounds comprises one or more new compounds predicted to bind to the target molecule; iv) receiving a second input dataset comprising second binding interaction information between the target molecule and the updated library of compounds; and v) processing the second input dataset using the machine learning module to update the model representation of the binding interactions, where at least the predictive accuracy of the updated model representation is improved. 2. The method of embodiment 1, wherein the library of compounds is a combinatorial library of compounds. 3. The method of embodiment 2, wherein the combinatorial library of compounds is a DNA-encoded library. 4. The method of embodiment 2, wherein the combinatorial library of compounds is generated using a split-pool method. 5. The method of embodiment 2, wherein the first input dataset and the second input dataset comprise DNA sequencing read counts of DNA barcodes tagged to each compound in the combinatorial library of compounds, and at least one mapped structure corresponding to each DNA barcode. 6. The method of embodiment 1, wherein the target molecule is a biomolecule. 7. The method of embodiment 6, wherein the biomolecule comprises a macromolecule. 8. The method of embodiment 7, wherein the macromolecule comprises a polysaccharide, a carbohydrate, a lipid, or a nucleic acid. 9. The method of embodiment 7, wherein the macromolecule comprises a protein. 10. The method of embodiment 1, wherein the library of compounds comprises small molecule compounds. 11. The method of embodiment 10, wherein the small molecule compound has a molecular weight of 1000 Daltons or less. 12. The method of embodiment 10, wherein the library of compounds consists of 100,000 to 1,000,000,000 small molecule compounds. 13. The method of embodiment 10, wherein the updated library of compounds consists of more than 10,000 small molecule compounds. 14. The method of embodiment 1, wherein the model representation is configured to predict binding for a plurality of compound fractions (e.g., compounds viewed as a set of complete product, intermediate, and / or by-product data) for each combinatorial synthetic chemical pathway used in generating the library of compounds. 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. 16. The method of embodiment 14, wherein the model representation is configured to model the products and any by-products of the compound fraction generated from the synthesis process. 17. The method of embodiment 14, wherein each of the plurality of compound fractions is encoded with a DNA barcode corresponding to a compound (e.g., a complete product, an intermediate product, and / or a by-product) within the compound fraction. 18. The method of embodiment 14, wherein the model representation is configured to generate a binding score (e.g., a compatibility score) for each compound fraction and / or for the compounds in each compound fraction. 19. The method of embodiment 1, wherein the model representation comprises a neural network. 20. The method of embodiment 19, wherein the neural network is a graph neural network. 21. The method of embodiment 1, further comprising iteratively repeating steps (c)-(e) to improve the predictive accuracy of the model representation. 22. The method of embodiment 21, wherein steps (c)-(e) are repeated iteratively until at least 1, 10, 50, 100, 200, 300, 400, or 500 compounds are identified as having a predicted affinity for the target molecule that exceeds a minimum threshold. 23. The method of embodiment 1, further comprising weighting the binding interactions of the model representation (e.g., predicted compound descriptors) based on experimental data corresponding to the efficiency of the synthetic steps used to generate the library of compounds, thereby reducing signal-to-noise. 24. The method of embodiment 23, wherein the experimental data comprises abundance information based on mass spectrometry of compound fractions comprising one or more synthetic compound products generated from each synthetic step. 25. The method of embodiment 1, wherein the model representation is a graph convolutional neural network configured to receive as input data a graph representation of a given compound. 26. The method of embodiment 25, wherein the graph representation of the given compound comprises a graph data structure composed of vertices and edges. 27. The method of embodiment 1, further comprising conducting binding affinity experiments between the target molecule and a library of compounds. 28. The method of embodiment 27, wherein the binding affinity experiment comprises incubating the target molecule with a library of compounds and purifying the target molecule together with any binding compounds. 29. The method of embodiment 27, wherein the binding affinity experiment comprises eluting the bound compounds, and the bound compounds are tagged with DNA barcodes. 30. The method of embodiment 28, further comprising amplifying the DNA barcode. 31. The method of embodiment 29, further comprising sequencing the DNA barcodes to obtain read count data for the binding compounds tagged with the DNA barcodes. 32. The method of embodiment 29, wherein the sequencing comprises deep sequencing to obtain a complete sequence of the DNA barcode. 33. The method of embodiment 31, wherein the selection coverage of the compounds in the library of compounds is at least 80%, 85%, 90%, 95%, 99%, or 100%. 34. The method of embodiment 27, wherein binding affinity experiments are performed for one or more iterative rounds of input data generation and the model representation is used to determine an updated library of compounds. 35. The method of embodiment 34, wherein the binding affinity experiment comprises calibrating the amount of input material to the round number. 36. The method of embodiment 34, wherein the one or more iterative rounds comprise at least one, two, three, four, or five rounds of input data generation and using the model representation to determine an updated library of compounds. 37. The method of embodiment 1, further comprising the step of performing binding affinity experiments between the target molecule and the updated library of compounds after step (c). 38. The method of embodiment 1, wherein the library of compounds is selected as a subset of a larger library of potential compounds, and the updated library of compounds comprises one or more novel compounds identified from the larger library of potential compounds as potential binders of the target molecule. 39. The method of embodiment 1, wherein the method further comprises using deep sequencing for complete sequencing and not partial reads. In some embodiments, the selection coverage of molecules in the DEL is at least 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%, 99%, or 100%. In some embodiments, the selection range is in the range of 50%, 55%, 60%, 65%, 70%, or 75% to 80%, 85%, 90%, 95%, 99%, or 100%. 40. A computer-implemented system comprising a digital processing device including at least one processor, an operating system configured to execute executable instructions, a memory, and a computer program including instructions executable by the digital processing device to perform any one of the methods of embodiments 1 to 39. 41. A non-transitory computer-readable storage medium coded with a computer program comprising instructions executable by one or more processors to perform any one of the methods of embodiments 1 to 39.

[0082] Terms and Definitions 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 this disclosure belongs.

[0083] 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 specifically stated otherwise.

[0084] As used herein, the term "about" refers to an amount that is approximately the specified amount, as the case may be.

[0085] As used herein, the term "about" refers to an amount near the stated amount, such as 10%, 5%, or 1%, as the case may be, including increments therein.

[0086] As used herein, the term "about" refers to an amount greater or less than the recited percentage, as the case may be, with respect to percentages, including increments therein, such as by 10%, 5%, or 1%.

[0087] As used herein, the phrases "at least one," "one or more," and "and / or" are open-ended expressions that are conjunctive and disjunctive in operation. For example, each of the phrases "at least one of A, B, and C," "at least one of A, B, and C," "one or more of A, B, and C," "one or more of A, B, and 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.

[0088] As used herein, the term "building block" may 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 may function as a building block, where a compound may 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 monosynthon may be the first building block, a second monosynthon may be the second building block, and the two building blocks react with each other to form a new di-synthon compound. The di-synthon may then be used as a building block with one or more building blocks to build one or more tricynthons and / or polysynthons.

[0089] As used herein, the term "compound descriptor" refers to any data or information related to a compound. The data or information may include, but is not limited to, the binding affinity of the compound to a target molecule (optionally, the target molecule is a drug target such as a protein or a 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, sequencing reads related to the abundance of the compound in an experiment, information related to the synthesis of the compound, the structure of the compound, labeling data, process quality control data, or yields related to 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 composite descriptor may be in the form of a graph, a chart, a number / value, a label, text, a data set, an image, etc.

[0090] Machine Learning The machine learning model may include one or more of a variety of 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.

[0091] Graphs, graph models, and graphical models can refer to ways of conceptualizing or organizing information into graphical representations that include nodes and edges. In some embodiments, a graph can refer to a principle of conceptualizing or organizing data, and data may be stored in various alternative forms, such as linked lists, dictionaries, spreadsheets, arrays, persistent storage, transient storage, and the like, and is not limited to the specific embodiments disclosed herein. In some embodiments, the machine learning model can include a graph model.

[0092] The machine learning model can include various manifold learning algorithms. In some embodiments, the machine learning model can include a manifold learning algorithm. In some embodiments, the manifold learning algorithm is a 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 locally linear embedding algorithm. In some embodiments, the manifold learning algorithm is a modified locally 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 Bames-Hut t-SNE algorithm.

[0093] 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 methods of transforming given input data having an initial dimensionality into another form of data having fewer dimensions than the initial dimensionality. In some embodiments, these terms may refer to principles that reduce a set of input dimensions to a smaller set of output dimensions.

[0094] The term normalization can refer to a collection of methods for adjusting datasets to align them to a common scale. In some embodiments, the normalization method can include multiplying a portion or the entire dataset by a factor. In some embodiments, the normalization method can include adding or subtracting a constant from a portion or the entire dataset. In some embodiments, the normalization method can include adjusting a portion or the entire dataset to a known statistical distribution. In some embodiments, the normalization method can include adjusting a portion or the entire dataset to a normal distribution. In some embodiments, the normalization method can include adjusting a dataset such that the signal intensities of the portion or the entire dataset are approximately the same.

[0095] The transformation may include one or more steps of various transformations of the data. In some embodiments, the transformation may include normalizing the data. In some embodiments, the transformation may include performing a mathematical operation to calculate a score based on a distance between two points in the data. In some embodiments, the distance may include a distance between two edges in a graph. In some embodiments, the distance may include a distance between two nodes in a graph. In some embodiments, the distance may include a distance between a node and an edge in a graph. In some embodiments, the distance may include a 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 Laplacian 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 learned function.

[0096] In some embodiments, the transformation may include converting data in one representation to another representation. In some embodiments, the transformation may include converting the data to another form of data having a smaller dimension. In some embodiments, the transformation may include linearizing one or more curved paths in the data. In some embodiments, the transformation may be performed on data including data in Euclidean space. In some embodiments, the transformation may be performed on data including data in graph space. In some embodiments, the transformation may be performed on data in discrete space. In some embodiments, the transformation may be performed on data including data in frequency space. In some embodiments, the transformation may convert data in discrete space to continuous space, continuous space to discrete space, graph space to continuous space, continuous space to graph space, graph space to discrete space, discrete space to graph space, or any combination thereof. In some embodiments, the transformation may include converting data in discrete space to frequency domain. In some embodiments, the transformation may include converting data in continuous space to frequency domain. In some embodiments, the transformation may include converting data in graph space to frequency domain.

[0097] In some embodiments, the methods of the present disclosure further include using a machine learning model to reduce the compound descriptors to a reduced descriptor space. In some embodiments, the methods further include clustering the reduced descriptor space to determine one or more groups of compound descriptors that have similar characteristics.

[0098] In some embodiments, the reduction can include converting given input data having any initial number of dimensions to another form of data having any number of dimensions less than the initial number of dimensions. In some embodiments, the reduction can include converting the input data to another form of data with fewer dimensions. In some embodiments, the reduction can include linearizing one or more curved paths in the input data into the output data. In some embodiments, the reduction can be performed on data including data in Euclidean space. In some embodiments, the reduction can be performed on data including data in graph space. In some embodiments, the reduction can be performed on data in discrete space. In some embodiments, the reduction can convert data from discrete space to continuous space, continuous space to discrete space, graph space to continuous space, continuous space to graph space, graph space to discrete space, discrete space to graph space, or any combination thereof.

[0099] The terms clustering, cluster analysis, or generative module can refer to a method of grouping samples in a dataset by some measure of similarity. Samples can be grouped in a set space, e.g., element "a" is in set "A". Samples can be grouped in a continuous space, e.g., element "a" is a point in Euclidean space a distance T away from the centroid of the elements that comprise cluster "A". Samples can be grouped in a graph space, e.g., element "a" is highly connected to the elements that comprise cluster "A". These terms can refer to the principle of organizing multiple elements into groups in some mathematical space based on some measure of similarity.

[0100] Clustering may include grouping any number of samples in a dataset by any quantitative measure of similarity. 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 computing one or more N-1 dimensional surfaces in an N dimensional space that divides the dataset into clusters. In some embodiments, clustering may include distribution-based clustering. In some embodiments, clustering may include fitting multiple prior distributions across 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 computing probability values ​​for 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.

[0101] In some embodiments, the clustering may include grouping samples based on similarity. In some embodiments, the clustering may include grouping samples based on quantitative similarity. In some embodiments, the clustering may include grouping samples based on one or more features of each sample. In some embodiments, the clustering may include grouping samples based on one or more labels of each sample. In some embodiments, the clustering may include grouping samples based on Euclidean coordinates. In some embodiments, the clustering may include grouping samples based on node and edge features of each sample.

[0102] In some embodiments, the comparison may include comparing a first group to a second group that is different. In some embodiments, the first or second group may each independently be a cluster. In some embodiments, the first or second group may each independently be a group of clusters. In some embodiments, the comparison may include comparing a cluster to a group of clusters. In some embodiments, the comparison may include comparing a first group of clusters to a second group of clusters that is different from the first group. In some embodiments, a group may be a sample. In some embodiments, a group may be a group of samples. In some embodiments, the comparison may include comparing a sample to a group of samples. In some embodiments, the comparison may include comparing a group of samples to a group of samples.

[0103] Neural Networks In some embodiments, the systems and methods of the present disclosure may include or may include using 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 including 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 transformer model.

[0104] Various types of layers may be used as 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 max pooling layers.

[0105] In some embodiments, the neural network includes a graph model. In some embodiments, graph, graph model, and graphical model can refer to a way of modeling data in a graphical representation that includes nodes and edges. In some embodiments, the data may be stored in various alternative forms, such as linked lists, dictionaries, spreadsheets, arrays, permanent storage, transient storage, etc., and is not limited to the specific embodiments disclosed herein.

[0106] 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.

[0107] The neural network may include a variety of 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. A variety of activation functions may be used with the neural network without departing from the inventive concepts disclosed herein.

[0108] training A variety of loss functions can 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 distribution. 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 Laplacian 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 Laplacian 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, loss functions may be formulated to optimize regression loss, evidence-based lower bounds, maximum likelihood, Kullback-Leiber divergence, and are applied with a variety of distribution functions, such as Gaussian, non-Gaussian, mixtures of Gaussians, mixtures of logistic functions, etc.

[0109] A variety of optimizers may be used to train the neural network. In some embodiments, the neural network may be trained using an Adam optimizer. In some embodiments, the neural network may be trained using a stochastic gradient descent optimizer. In some embodiments, the neural network may be trained using an active learning algorithm. The neural network may be trained using a variety of loss functions whose derivatives may be calculated to update one or more parameters of the neural network. The neural network may be trained using a hyperparameter search algorithm. In some embodiments, the neural network hyperparameters are optimized using a Gaussian process.

[0110] Various training protocols can be used during training of the neural network. In some embodiments, the neural network can be trained using a training / validation / test data split. In some embodiments, the neural network can be trained using a k-fold data split, where k is any positive integer.

[0111] Training a neural network may include providing inputs to an untrained neural network to generate a predicted output, comparing the predicted output to an 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 for each parameter may be calculated by backpropagation 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 repeated for a certain number of iterations or until some stopping criterion is met.

[0112] Computing Systems Referring to Figure 16, a block diagram is shown illustrating an example machine including a computer system 1600 (e.g., a processing or computing system) capable of executing a set of instructions to cause a device to perform or to perform any one or more of the aspects and / or methods for static code scheduling of the present disclosure. The components of Figure 16 are merely examples and do not limit the scope of use or functionality of any hardware, software, embedded logic components, or combinations of two or more such components to implement a particular embodiment.

[0113] The computer system 1600 may include one or more processors 1601, memory 1603, and storage 1608 that communicate with each other and with other components via a bus 1640. The bus 1640 may also link a display 1632, one or more input devices 1633 (which may include, for example, 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 interface to the bus 1640 directly or through one or more interfaces or adapters. For example, the various tangible storage media 1636 may interface with the bus 1640 through a storage media interface 1626. The computer system 1600 may have any suitable physical form, including, but not limited to, one or more integrated circuits (ICs), a printed circuit board (PCB), a mobile handheld device (such as a cell phone or PDA), a laptop or notebook computer, a distributed computer system, a computing grid, or a server.

[0114] The computer system 1600 includes one or more processors 1601 (e.g., a central processing unit (CPU) or a general purpose graphics processing unit (GPGPU)) that perform functions. The processor 1601 optionally includes a cache memory unit 1602 for temporary local storage of instructions, data, or computer addresses. The processor 1601 is configured to assist in the execution of computer-readable instructions. The computer system 1600 can provide functionality for the components illustrated in FIG. 16 as a result of the processor 1601 executing 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. The computer-readable medium can store software implementing certain embodiments, and the processor 1601 can execute the software. Memory 1603 may read software from one or more other computer-readable media (e.g., mass storage devices 1635, 1636) or from one or more other sources through a suitable interface, such as network interface 1620. The software can cause processor 1601 to perform one or more processes or one or more steps of one or more processes described or illustrated herein. Performing such processes or steps may include defining data structures stored in memory 1603 and modifying the data structures as directed by the software.

[0115] 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 operate to communicate data and instructions unidirectionally to the processor 1601, and the RAM 1604 may operate to communicate data and instructions bidirectionally with the processor 1601. The ROM 1605 and the RAM 1604 may include any suitable tangible computer-readable media as described below. In one example, a basic input / output system 1606 (BIOS), containing the basic routines that help to transfer information between elements within the computer system 1600, such as during start-up, may be stored in the memory 1603.

[0116] Persistent storage 1608 is bidirectionally connected to processor 1601, optionally via storage control unit 1607. Persistent storage 1608 provides additional data storage capacity and may include any suitable tangible computer-readable media as described herein. Storage 1608 may be used to store operating system 1609, executable files 1610, data 1611, applications 1612 (application programs), and the like. Storage 1608 may also include optical disk drives, solid-state memory devices (e.g., flash-based systems), or any combination of the above. Information in storage 1608 may be incorporated as virtual memory in memory 1603, where appropriate.

[0117] In one example, storage device 1635 may removably interface with 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 may provide nonvolatile and / or volatile storage of machine-readable instructions, data structures, program modules, and / or other data for computer system 1600. In one example, software may reside, completely or partially, within the machine-readable media on storage device 1635. In another example, software may reside, completely or partially, within processor 1601.

[0118] The bus 1640 connects a wide variety of subsystems. As used herein, reference to a bus may encompass, where appropriate, one or more digital signal lines performing a common function. The bus 1640 may be any of several types of bus structures, including, but not limited to, a memory bus, a memory controller, a peripheral bus, a local bus, and any combination thereof, using any of a variety of bus architectures. By way of example and not limitation, such architectures include an Industry Standard Architecture (ISA) bus, an Enhanced ISA (EISA) bus, a Micro Channel Architecture (MCA) bus, a Video Electronics Standards Association local bus (VLB), a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, an Accelerated Graphics Port (AGP) bus, a HyperTransport (HTX) bus, a Serial Advanced Technology Attachment (SATA) bus, and any combination thereof.

[0119] The computer system 1600 may also include input devices 1633. In one example, a user of the computer system 1600 may input commands and / or other information into the computer system 1600 via the input devices 1633. Examples of the input devices 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 gamepad, 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, Leap Motion, or the like. The input devices 1633 may be interfaced 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 of the above.

[0120] In certain embodiments, when computer system 1600 is connected to network 1630, computer system 1600 may communicate with other devices connected to network 1630, specifically mobile devices and enterprise systems, distributed computing systems, cloud storage systems, cloud computing systems, and the like. Communications to and from computer system 1600 may be sent through network interface 1620. For example, network interface 1620 may 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 may store the incoming communications in memory 1603 for processing. Computer system 1600 may similarly store outgoing communications (such as requests or responses to other devices) in the form of one or more packets in memory 1603 and communicate them from network interface 1620 to network 1630. Processor 1601 may access these communication packets stored in memory 1603 for processing.

[0121] Examples of network interface 1620 include, but are not limited to, a network interface card, a modem, and any combination thereof. Examples of network 1630 or network segment 1630 include, but are not limited to, a distributed computing system, a cloud computing system, a wide area network (WAN) (e.g., the Internet, a corporate network), a local area network (LAN) (e.g., a network associated with an office, a building, a campus, or other relatively small geographic space), a telephone network, a direct connection between two computing devices, a peer-to-peer network, and any combination thereof. A network such as network 1630 may employ wired and / or wireless communication modes. In general, any network topology may be used.

[0122] Information and data can be displayed through a display 1632. Examples of the display 1632 include, but are not limited to, a cathode ray tube (CRT), a liquid crystal display (LCD), a thin film transistor liquid crystal display (TFT-LCD), an organic liquid crystal display (OLED) such as a passive matrix OLED (PMOLED) or active matrix OLED (AMOLED) display, a plasma display, and any combination thereof. The display 1632 can interface with the processor 1601, memory 1603, and fixed storage 1608, as well as other devices such as input device(s) 1633, via a bus 1640. The display 1632 is linked to the bus 1640 via a video interface 1622, and the transfer of data between the display 1632 and the bus 1640 can be controlled via a graphics control 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 examples, HTC Vive, Oculus Rift, Samsung Gear VR, Microsoft HoloLens, Razer OSVR, FOVE VR, Zeiss VR One, Avegant Glyph, Freefly VR headsets, etc. In still further embodiments, the display is a combination of devices such as those disclosed herein.

[0123] In addition to the display 1632, the computer system 1600 may 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.

[0124] Additionally or alternatively, computer system 1600 may provide functionality as a result of logic hardwired or otherwise embodied in circuitry, which may operate in place of or in conjunction with software to perform one or more processes or one or more steps of one or more processes described or illustrated herein. References to software in this disclosure may encompass logic, and references to logic may encompass software. Furthermore, references to computer-readable media may encompass, where appropriate, circuitry (such as an IC) that stores software for execution, circuitry that embodies logic for execution, or both. This disclosure encompasses any appropriate combination of hardware, software, or both.

[0125] Those skilled in the art will appreciate that the various illustrative logical blocks, modules, circuits, and algorithmic operations described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability of hardware and software, the various illustrative components, blocks, modules, circuits, and operations have been described above generally in terms of their functionality.

[0126] The various example logic blocks, modules, and circuits described in connection with the embodiments disclosed herein may be implemented or performed using 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 alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0127] The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by one or more processors, or in a combination of the two. The software modules may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, 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 the processor such that the processor can read information from, and write information to, the storage medium. Alternatively, the storage medium may be integrated into the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. Alternatively, the processor and the storage medium may reside as discrete components in a user terminal.

[0128] In accordance with the description herein, suitable computing devices include, by way of non-limiting examples, 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 connectivity are suitable for use in the systems described herein. Suitable tablet computers, in various embodiments, include those having booklet, slate, and convertible configurations known to those skilled in the art.

[0129] In some embodiments, the computing device includes an operating system configured to execute executable instructions. An operating system is software, including, for example, programs and data, that manages the device's hardware and provides services for the execution of applications. Those skilled in the art will recognize that suitable server operating systems include, by way of non-limiting examples, FreeBSD, OpenBSD, NetBSD, Linux, Apple Mac OS X Server, Oracle Solaris, Windows Server, and Novell NetWare. Those skilled in the art will recognize that suitable personal computer operating systems include, by way of non-limiting examples, UNIX-like operating systems, such as Microsoft Windows, Apple Mac OS X, UNIX, and GNU / Linux. In some embodiments, the operating system is provided by cloud computing. Those skilled in the art will also recognize that suitable mobile smartphone operating systems include, by way of non-limiting examples, Nokia® Symbian® OS, Apple® iOS®, Research In Motion® BlackBerry OS®, Google® Android®, Microsoft® Windows Phone® OS, Microsoft® Windows Mobile® OS, Linux®, and Palm® WebOS®.Those skilled in the art will also recognize that suitable media streaming device operating systems include, by way of non-limiting examples, Apple TV®, Roku®, Boxee®, Google TV®, Google Chromecast®, Amazon Fire®, and Samsung® HomeSync®. Those skilled in the art will also recognize that suitable video game console operating systems include, by way of non-limiting examples, Sony® PS3®, Sony® PS4®, Microsoft® Xbox360®, Microsoft Xbox One, Nintendo® Wii®, Nintendo® Wii U®, and Ouya®.

[0130] Computer-readable recording medium In some embodiments, the platforms, systems, media, and methods disclosed herein include one or more non-transitory computer-readable storage media, optionally encoded with a program including instructions executable by an operating system of a networked computing device. In further embodiments, the computer-readable storage medium is a tangible component of the computing device. In still further embodiments, the computer-readable storage medium is optionally removable from the computing device. In some embodiments, the computer-readable storage medium includes, by way of non-limiting examples, CD-ROMs, DVDs, flash memory devices, solid-state memories, magnetic disk drives, magnetic tape drives, optical disk drives, distributed computing systems including cloud computing systems and services, and the like. In some cases, the programs and instructions are encoded on the medium permanently, substantially permanently, semi-permanently, or non-transitoryly.

[0131] Computer Programs In some embodiments, the platform, system, medium, and method disclosed herein include at least one computer program, or the use thereof. A computer program includes a sequence of instructions that are written to perform a specified task and are executable by one or more processors of a CPU of a computing device. The computer-readable instructions may be implemented as program modules, such as functions, objects, application programming interfaces (APIs), computing data structures, etc., that perform a particular task or implement a particular abstract data type. In light of the disclosure provided herein, those skilled in the art will recognize that computer programs may be written in various versions of various languages.

[0132] The functionality of the computer readable instructions may be combined or distributed as desired in various environments. In some embodiments, the computer program includes one sequence of instructions. In some embodiments, the computer program includes multiple sequences of instructions. 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, in part or in whole, one or more web applications, one or more mobile applications, one or more standalone applications, one or more web browser plug-ins, extensions, add-ins, or add-ons, or combinations thereof.

[0133] Web Applications In some embodiments, the computer program comprises a web application. In light of the disclosure provided herein, one skilled in the art will recognize that web applications, in various embodiments, utilize one or more software frameworks and one or more database systems. In some embodiments, the web application is created on a software framework such as Microsoft® .NET or Ruby on Rails (RoR). In some embodiments, the web application utilizes one or more database systems, including, by way of non-limiting examples, relational, non-relational, object-oriented, associative, and XML database systems. In further embodiments, suitable relational database systems include, by way of non-limiting examples, Microsoft® SQL Server, mySQL™, and Oracle®. One skilled in the art will also recognize that web applications, in various embodiments, are written in one or more versions of one or more languages. Web applications 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 client-side scripting language such as Asynchronous Javascript and XML (AJAX), Flash® Actionscript, Javascript, or Silverlight®.In some embodiments, the web application is written to some extent in a server-side coding language, such as Active Server Pages (ASP), ColdFusion®, Perl, Java®, JavaServer Pages (JSP), Hypertext Preprocessor (PHP), Python®, Ruby, Tel, Smalltalk, WebDNA®, or Groovy. In some embodiments, the web application is written to some extent in a database query language, such as Structured Query Language (SQL). In some embodiments, the web application integrates with an enterprise server product, such as IBM® Lotus Domino®. In some embodiments, the web application includes a media player element. In various further embodiments, the media player element utilizes one or more of many suitable multimedia technologies, including, by way of non-limiting examples, Adobe® Flash®, HTML5, Apple® QuickTime®, Microsoft® Silverlight®, Java®, and Unity®.

[0134] 17, in a particular embodiment, the application provisioning system comprises one or more databases 1700 accessed by a relational database management system (RDBMS) 1710. Suitable RDBMSs include Firebird, MySQL, PostgreSQL, SQLite, Oracle Database, Microsoft SQL Server, IBM DB2, IBM Informix, SAP Sybase, Teradata, and the like. In this embodiment, the application provisioning system further comprises one or more application servers 1720 (e.g., Java servers, .NET servers, PHP servers, and the like) and one or more web servers 1730 (Apache, IIS, GWS, and the like). The web servers optionally expose one or more web services via app application programming interfaces (APIs) 1740. Over a network such as the Internet, the system provides a browser-based and / or mobile native user interface.

[0135] Referring to FIG. 18, in a particular embodiment, the application provisioning system alternatively has a distributed cloud-based architecture 1800 with elastic load balancing auto-scaling web server resources 1810 and application server resources 1820 and a synchronously replicated database 1830.

[0136] Standalone Applications In some embodiments, the computer program includes a standalone application, which is a program that runs as an independent computer process, rather than an add-on to an existing process, e.g., a plug-in. Those skilled in the art will recognize that standalone applications are often 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. Suitable compiled programming languages ​​include, by way of non-limiting examples, C, C++, Objective-C, COBOL, Delphi, Eiffel, Java, Lisp, Python, Visual Basic, and VB.NET, or combinations thereof. Compilation is often performed, at least in part, to create an executable program. In some embodiments, the computer program includes one or more executable compiled applications.

[0137] Web browser plugins In some embodiments, the computer program includes a web browser plug-in (e.g., an extension, etc.). In computing, a plug-in is one or more software components that add specific functionality to a larger software application. Makers of software applications support plug-ins to allow third party developers to create the ability to extend the application, to support easy addition of new features, and to reduce the size of the application. When supported, plug-ins allow customization of the functionality of the software application. For example, plug-ins are commonly used in web browsers to play video, generate interactivity, scan for viruses, and display specific file types. Those skilled in the art are familiar with several web browser plug-ins, including Adobe® Flash® Player, Microsoft® Silverlight®, and Apple® QuickTime®. In some embodiments, the toolbar includes one or more web browser extensions, add-ins, or add-ons. In some embodiments, the toolbar comprises one or more explorer bars, tool bands, or desk bands.

[0138] Given the disclosure provided herein, one of skill in the art will recognize that several plug-in frameworks are available that allow for the development of plug-ins in a variety of programming languages, including, by way of non-limiting examples, C++, Delphi, Java, PHP, Python, and VB.NET, or combinations thereof.

[0139] A web browser (also referred to as an Internet browser) is a software application designed for use with networked computing devices to search, present, and traverse information resources on the World Wide Web. Suitable web browsers include, by way of non-limiting example, Microsoft® Internet Explorer®, Mozilla® Firefox®, Google® Chrome, Apple® Safari®, Opera Software® Opera®, and KDE Konqueror. In some embodiments, the web browser is a mobile web browser. Mobile web browsers (also referred to as microbrowsers, minibrowsers, 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, smartphones, music players, personal digital assistants (PDAs), and handheld video game systems. Suitable mobile web browsers include, by way of non-limiting example, Google® Android® browser, RIM BlackBerry® browser, Apple® Safari®, Palm® Blazer, Palm® WebOS® browser, Mozilla® Firefox® for mobile devices, Microsoft® Internet Explorer® Mobile, Amazon® Kindle® Basic Web, Nokia® browser, Opera Software® Opera® Mobile, and Sony® PSP™ browser.

[0140] In some embodiments, the platforms, systems, media, and methods disclosed herein include software, server, and / or database modules, or the use thereof. In view of the disclosure provided herein, the software modules are created by techniques known to those skilled in the art using machines, software, and languages ​​known to those skilled in the art. The software modules disclosed herein are implemented in numerous ways. In various embodiments, the software modules comprise files, sections of code, programming objects, programming structures, or combinations thereof. In further various embodiments, the software modules comprise multiple files, multiple sections of code, multiple programming objects, multiple programming structures, or combinations thereof. In various embodiments, one or more software modules comprise, non-limiting examples being web applications, mobile applications, and standalone applications. In some embodiments, the software modules are in one computer program or application. In other embodiments, the software modules are in multiple computer programs or applications. In some embodiments, the software modules are hosted on one machine. In other embodiments, the software modules are hosted on multiple machines. In further embodiments, the software modules are hosted on a distributed computing platform, such as a cloud computing platform. In some embodiments, the software modules are hosted on one or more machines in one location. In other embodiments, the software modules are hosted on one or more machines in one or more locations.

[0141] Database In some embodiments, the platform, system, medium, and method disclosed herein includes one or more databases, or the use thereof. Given the disclosure provided herein, one skilled in the art will recognize that many databases are suitable for storing and retrieving the DEL information and related experimental data collected for one or more conditions. In various embodiments, suitable databases include, as non-limiting examples, relational databases, non-relational databases, object-oriented databases, object databases, entity-relationship model databases, associative databases, and XML databases. Further non-limiting examples include SQL, PostgreSQL, MySQL, Oracle, DB2, and Sybase. In some embodiments, the database is Internet-based. In further embodiments, the database is web-based. In yet further embodiments, the database is cloud computing-based. In certain embodiments, the database is a distributed database. In other embodiments, the database is based on one or more local computer storage devices. EXAMPLES

[0142] The following illustrative examples are representative of embodiments of the software applications, systems, and methods described herein, and are not intended to be limiting in any way.

[0143] Example 1 Using the exemplary iterative machine learning-based DEL platform disclosed herein, we identified a compound that binds to a target that is known to be challenging for DEL. This target binds to DNA, and DEL contains a lot of DNA, so the target generates a lot of noise in the selection. Furthermore, to be successful against this target, the compound must disrupt protein-protein interactions (PPIs). PPIs are known to be difficult for small molecules, but this particular compound is known and has reached clinical trials, so this problem is very difficult, but not impossible.

[0144] An initial DEL consisting of approximately 450 million compounds was evaluated in a binding assay against the target. A total of about three conditions were tested, including [conditions 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 process was repeated once 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 to known clinical compounds shown in Table 1.

[0145] [Table 1]

[0146] Of the 135 compounds that we predicted might work, 21% were confirmed to bind to the target, some of which matched or exceeded the performance (affinity / activity) of the clinical compounds in the binding assay, as shown in Table 1. Two-thirds of these confirmed compounds were also experimentally determined to disrupt the relevant PPI. Compounds that perform well in the assay are not useful if they lack drug-like properties. As shown in Table 1, nearly all of the identified compounds fall well within the drug-like space (see table above for appropriate molecular weights and cLogP, which are even better than the clinical compounds).

[0147] Additionally, Figure 15 provides a star chart 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). Good candidate drugs will stay within the next to last circle on the plot. Almost all compounds fall well within this boundary, with a few slightly heavier, touching roughly 500 MW. Thus, the iterative ML-driven DEL discovery process was successfully validated against challenging targets with higher biochemical performance compared to known clinical compounds that progressed to clinical trials.

[0148] ML models often perform well on the data they were trained on, but not on other data sets. However, our ML model performed well outside the boundaries of the DELs on which the ML was trained. Specifically, we plotted the training set in yellow, the predicted (and successful) molecules in red, and the new evolved DEL library in grey. To a first approximation, on this plot, closely related compounds group themselves more closely than less closely related compounds. As shown in Figure 15, the red dots are generally outside the yellow region, indicating that our ML model can identify high-performing compounds outside the training set. Because ML models generalize outside the training set, they can be useful when medicinal chemists propose molecules that appear different from those built in the past. These models can thus guide downstream compound optimization processes.

[0149] Example 2 Based on the method of Example 1, an evolutionary library was constructed based on the results from the first selection. Compounds were selected against the same target, and the model was updated with the obtained data. Finally, the newly updated model was used to answer the following two questions:

[0150] 1) Given an initial set of compounds to be tested, determine a rank order for the confirmed compounds. The model generated a score prediction for the compounds, which was used to determine the ranking. As shown in FIG. 19, the plot 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 because it has seen the evolved DEL library.

[0151] 2) predict a set of new compounds to test that the model has a better understanding of which compounds should work. Specifically, it was asked whether compounds could be found within the new set of compounds that have better predictive properties for later stages of drug discovery such as ADME / toxicity.

[0152] To measure this, a composite score was created for the compound 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. In this metric, a score of 1 is perfect, and compounds with lower scores tend to have a worse set of predicted properties. The relative importance of each factor can be customized for each target as needed, and either experimental measurements or predicted values ​​can be used. We then plotted the values ​​for the compound in the initial DEL, the first set of predicted compounds (ML1), and the second set of predicted compounds (ML2). As shown in Figure 20, the compound's properties improved as the compound progressed from a pure DEL compound to an ML predicted compound (ML1). They improved further as the model was presented with more data (i.e., updated) and asked to predict yet another set of compounds (ML2). These results demonstrate the value of iterative ML-driven evolutionary DEL libraries in improving model performance and helping to find better drugs more quickly.

[0153] Example 3 The high signal-to-noise massively parallel data sets enabled by the present disclosure allow compound specificity questions to be evaluated in an unprecedented manner. Specifically, the platform, system, and method disclosed herein allow for comparison at the scale of several different isoforms of a protein or domain, helping to identify compounds that hit one isoform but not the other. To demonstrate this capability, two experiments were performed and the results plotted to show how this platform can identify highly specific compounds.

[0154] Two highly related (95% sequence similarity) protein domains within a single protein were evaluated. Each compound is represented by a blue dot. Compound scores for domain 1 were plotted on the Y-axis and compound scores for domain 2 were plotted on the X-axis, as shown in FIG. 21. Compounds specific to domain 1 appear only on the Y-axis, with better compounds having larger scores. Similarly, compounds specific to domain 2 appear only on the X-axis. In this example, a large number of compounds were identified that were specific to each of the two highly similar domains but not the other.

[0155] This experiment was repeated using two proteins within the same class of chromatin regulators. The results shown in Figure 22 show a large number of molecules that are predicted to specifically bind to only one of the two related proteins. This experiment further demonstrates that while traditional drug discovery struggles to identify specific compounds, the present platform can identify a large number of potentially specific molecules for two closely related proteins.

[0156] Example 4 The methods described herein demonstrate how to train a custom GNN to develop a predictive model that incorporates intermediate product data in the probabilistic modeling of counts, where the trained model predicts enrichment of the full product and enrichment of the intermediate products.

[0157] In this case, the first dataset was constructed by creating an initial DEL from a combination of 15 unique DELs totaling 700 million unique molecules. The initial DELs can be constructed from one or a combination of any number of DEL libraries (e.g., from one million to several billion unique molecules). The initial DEL library is then sequenced, generating 200 million individual sequence reads corresponding to 90 million individual molecules. From the 90 million molecules, approximately 4 million molecules were determined to be significant binders to the protein target. To increase the number of negative examples, 1 million molecules were added to the 4 million molecules determined to be significant binders to the protein target. The 1 million negative examples include controls that were sequenced but did not contain the target (non-target controls) or molecules that bound only to other targets, as well as another 1 million molecules that were not sequenced but were present in the initial library. The first dataset was then split into a training dataset and a test dataset using the Murcko scaffold split Landrum.

[0158] To enrich this data set with intermediate products, all possible two- and three-building block molecules corresponding to a given DNA sequence were determined, and the theoretical proportion of each intermediate product was then calculated from the estimated building block yields.

[0159] The read count of each DNA tag is modeled as following a negative binomial distribution whose average 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 that drive the total read count, as illustrated in Figure 23, which treats the tricynthon as a full product and the di-synthon as an intermediate product, analyzes the data set, and outlines the model that is utilized to predict the target binding affinity, non-target binding affinity, and ratio adjustment for each of the full product and intermediate product. The ability of the full ML model to predict both the full product and intermediate product of a compound allows the model to learn and provide further insight into the binding affinity of the compound to the target.

[0160] The read counts for DNA tag i in either the NTC or target selection experiment are calculated using the mean parameter μ i and is modeled by a negative binomial distribution with dispersion parameter α:

[0161]

number

[0162] The α value is obtained by negative binomial regression before training.

[0163] μ i For a given value, the binding enrichment (B target,i ) is the corresponding tricynthon (R tri,i ) and three possible synthons (R di1 , i , R di2 , i , R dij , i ) by combining the following:

[0164]

number

[0165] where the p value corresponds to the proportion of tricynthons and disynthons present in the final mixture. Similarly, NTC(B NTC,i ) is also due to a combination of enrichment in the tricynthons and di-synthons:

[0166]

number

[0167] The counts of a given DNA tag relative to the target are calculated by the binding enrichment to the target, the binding enrichment in the NTC experiment, and the binding enrichment relative to the starting material estimated using DLS experiments (C dls,i ), and promiscuous join count (C promiscuity,i ) was assumed to be a combination of

[0168] Therefore, the mean parameters of the negative binomial distribution were set as follows:

[0169]

number

[0170] where β values ​​are learned by negative binomial regression on the training set and σ is a soft positive function. Similarly, the counts of a given DNA tag relative to the NTC are expressed as the binding enrichment to the NTC, DLS counts (C dls,i ), and promiscuous join count (C promiscuity,i ) is explained by:

[0171]

number

[0172] Two different ways to estimate the fraction pk of product k. The first is to set these values ​​based on our internal reaction yield data set: p k,lab To refine this approach and take into account possible noise in the yield measurements, we propose to adjust the estimated ratios using parameters learned for each molecule: k,adjust In this case, p k =σ(p k,adjust +p k,lab ).

[0173] Concentration value, R target,molecule , and R NTC,moleculeWe used two models to train the model. The input to both is a molecular graph corresponding to a compound. We used the Weave atom featurizer to code the atoms, and a standard bond featurizer to code the bonds between atoms. 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. Two separate fully connected networks were then used to transform the 128-dimensional code and run it in R. target,i and R NTC,i The second model follows the same method as the first model, but the 128-dimensional encoding is further amplified through an additional fully connected network. k,adjust These models were then run on 10 ―3 The Adam optimizer with a learning rate of is used to train on the training set.

[0174] For a given DNA barcode example i, the negative binomial loss for the target and NTC values ​​was:

[0175]

number

[0176] In the formula, P is μ value,i and α value is the probability mass function of the negative binomial distribution parameterized by. The complete loss for example i was:

[0177]

number

[0178] The model is currently in the validation phase. target,I The values ​​are used to rank and / or score the binding affinities of compounds for the target.

[0179] ML models capable of processing the full product ratio data as disclosed herein were then evaluated against three different models: For the first model, the only source of binding enrichment was the tricynthon as the source of binding enrichment data.

[0180]

number

[0181] In the second model, the only source of binding enrichment is the disynthon.

[0182]

number

[0183] For the third model, we assumed that each reaction yielded 70% yield, so that all p dij,i = 0.147, and all p tri,i = 0.147. For each of the ML models described herein and the three models evaluated against ML, experiments were performed where the product rates were fixed and adjustments to the rates were learned as above. Figure 24 shows the loss and R on the test dataset. 2 Graph A shows the negative binomial loss on the test data, and graph B shows the relationship between the target protein counts and the generated μ target,i Count between R 2 Graph C shows the NTC count and the generated μ NTC,i Count between R 2 Show the score.

[0184] One evaluation metric was performance on a test set, which was a subset of the internal DEL data. When evaluating an ML model on test data from the internal DEL dataset, we compute the loss, μ ntc and μ target R between the value and the real number 2 I checked the score.

[0185] Across the two approaches to modeling product fractions, the tricynthon-only model performs noticeably worse. This is likely evidence that the tricynthon data itself is too noisy to generate reliable estimates and does not capture all of the DEL products. The di-synthon-only model also performed worse than the full ML model. Testing showed that modeling di-synthons is not just a noise removal step, but also an integral part of the data. The model utilizing constant yield data also performed worse than modeling yield and product fraction values ​​to help train and model the data. The model with the learned adjustment for yield shows improved performance, likely due to noise in the laboratory process where several purifications are performed.

[0186] To estimate the generalization ability of the trained model, the performance of the full ML model and the models evaluated against it were tested on a dataset of 150 molecules from a commercial vendor. Binding affinity was measured internally and molecules were classified as binders or non-binders to the target. The enrichment value, R target,molecule is used to predict binding affinity. The receiver operating characteristic (ROC) area under the curve for various models is shown in Figure 25. On this dataset, the full model with the trained yield outperforms the remaining models.

[0187] To further explore the effect of incorporating yield data, the full ML model and the model with flat yield were tested for 150 foreign molecules. The hit rate percentage of the top 10 identified hits was examined as shown in Figure 26. These results demonstrate that the full ML model with laboratory yield and learned adjustments performs best in virtual screening applications, and that setting the learned yield has more impact in relevant individual cases.

[0188] Overall, the tests showed that the full ML models described herein outperformed models that did not consider all of the full and intermediate products and / or did not include product fractions, as shown in Figures 24-26. Using only full product (tricynthon) data produces datasets that have been shown to be too noisy to support effective training and may also inappropriately describe the underlying data. Using only intermediate (dicynthon) data may be an effective approach to aggregating and denoising DEL data, but ignores potentially useful data that is needed to effectively screen compounds.

[0189] While preferred embodiments of the present disclosure have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, modifications, and substitutions will occur to those skilled in the art without departing from the present disclosure. It will be understood that various alternatives to the embodiments of the present disclosure described herein may be employed in implementing the present disclosure.

Claims

1. 1. A method comprising: a) a computer-implemented method comprising: i) receiving a first dataset comprising: (i) a first compound descriptor for each compound in a first library of compounds; and (ii) a compound suitability score for each compound in said first library of compounds; ii) training a predictive model on the first dataset; iii) inputting the predictive model to a second dataset comprising second compound descriptors for each compound of a second library of compounds; iv) generating a compound suitability score for each compound of the second library of compounds from the predictive model utilizing at least one or more compound descriptors of the first library of compounds and / or one or more compound descriptors of the second library of compounds; A computer-implemented method comprising: b) selecting a third library of compounds according to information comprising one or more compound suitability scores of the second library of compounds and / or one or more compound suitability scores of the first library of compounds; A method comprising:

2. 2. The method of claim 1 , wherein the third library of compounds comprises: (i) one compound from the second library of compounds; (ii) one compound from the first library of compounds; (iii) a compound comprising two or more compounds from the second library of compounds; (iv) a compound comprising two or more compounds from the first library of compounds; (v) a compound comprising a compound from the second library of compounds and a compound from the first library of compounds; (vi) a compound that is not in the first library of compounds or the second library of compounds; (vii) a compound that is not in the second library of compounds and a compound that is not in the first library of compounds or the second library of compounds; (viii) a compound that is not in the second library of compounds and a compound that is not in the first library of compounds or the second library of compounds; or (ix) a combination of two or more of (i) through (viii).

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

4. The method of 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 to 3, wherein step (b) is not part of the computer-implemented method.

6. 4. The method of claim 1, wherein step (b) comprises a first sub-step that is part of the computer-implemented method and a second sub-step that is not part of the computer-implemented method, and the first and second steps are performed consecutively, and the first sub-step is performed first or the first sub-step is performed second.

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

8. 8. The method of claim 7, wherein the evaluation scores for the compounds of the second library of compounds are second suitability scores generated independently from the compound suitability scores generated from the computer-implemented method.

9. 9. The method of claim 7 or claim 8, wherein the evaluation score of a compound of the first library of compounds is a first suitability score that is different from the compound suitability score of a compound of the first library of compounds.

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

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

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

13. 13. The method of any one of claims 1 to 12, wherein one or more compounds of the second library (or optionally a subsequent library applicable in an iterative manner) are 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 by-product (sometimes referred to as a byproduct or side product) generated during the synthesis of the second test compound, or an intermediate (sometimes referred to as an intermediate) generated during the synthesis of the second test compound, or a combination of two or more thereof.

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

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

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

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

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

19. 19. The method of any one of claims 1 to 18, wherein the first compound descriptor comprises data or information related to a compound of the first library of compounds, the data or information comprising binding affinity of the compound to a target molecule (optionally the target molecule is a drug target such as a protein or a nucleic acid), activity of the compound, a physical property of the compound (e.g. lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing leads related to the abundance of the compound in an experiment, compound structure, information related to the synthesis of the compound, labeling data, process quality control data, yields related to 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.

20. 20. The method of any one of claims 1 to 19, wherein the second compound descriptor comprises data or information related to a compound of the second library of compounds, the data or information comprising binding affinity of the compound to a target molecule (optionally the target molecule is a drug target such as a protein or a nucleic acid), activity of the compound, a physical property of the compound (e.g. lipophilicity), toxicity of the compound, stability of the compound, permeability of the compound, sequencing leads related to the abundance of the compound in an experiment, compound structure, information related to the synthesis of the compound, labeling data, process quality control data, yields related to 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.

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

22. 22. A method according to any one of claims 1 to 21, comprising testing one or more of the compounds of said first library of compounds in an in vitro or in vivo assay.

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

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

25. 25. The method of any one of claims 1 to 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 from each compound in the first library and / or second library).

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

27. The method of claim 24, 25, or 26, wherein the sequencing data comprises read counts, a quality score associated with the read counts, and / or a score calculated from sequencing read counts or sets of read counts from different experimental conditions from the first library of compounds and / or the second library of compounds.

28. 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 selection of compounds or the average number of reads in a selection of compounds, or a similar mathematical function utilizing (directly or indirectly) the read count.

29. 29. The method of any one of claims 1 to 28, wherein at least one compound suitability score for each compound of the first library of compounds is generated from data comprising the first compound descriptors (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. 30. The method of any one of claims 1 to 29, wherein the predictive model utilizes a probabilistic framework to process the first data set and the second data set and output the compound suitability scores for each compound in the second library of compounds.

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

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

33. 33. The method of any one of claims 1 to 32, comprising utilizing sequencing data associated with sequencing of the third library of compounds to generate said compound suitability scores for each compound in the third library of compounds.

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

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

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

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

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

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

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

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

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

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

44. A method according to any preceding claim, further comprising updating the model by repeating steps ii to vi.

45. 45. The method of any one of claims 1 to 44, wherein the complete product comprises a tricynthon and the intermediate products comprise a di-synthon and / or a monosynthon.

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

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

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

49. 49. The method of any one of claims 1 to 48, wherein the first compound descriptors input into the predictive model comprise compound structures and / or experimental data.

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

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

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

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

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

55. The method of any one of claims 1 to 54, wherein the method comprises carrying out a validation assay for at least one compound of the third library of compounds.

56. 56. The method of any one of claims 1 to 55, wherein the method comprises carrying out a low throughput analysis on at least one compound of the third library of compounds.

57. 57. The method of any one of claims 1 to 56, wherein the method comprises inputting a third dataset comprising compound descriptors and compound suitability scores for each compound of said third library of compounds into a secondary system in a validation assay.

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

59. 59. The method of claim 57 or claim 58, wherein the validation assay comprises a proxy for binding or biochemical activity comprising one or more of absorbance, fluorescence, luminescence, radioactivity, NMR, crystallography, microscopy including cryo-electron microscopy, mass spectrometry, or Raman scattering, e.g., surface plasmon resonance (SPR) measures the reflection of polarized light, thereby detecting a change in the angle of reflection (refractive index), and immobilization or binding of a ligand (compound) to a surface (including an immobilized target protein) affects the mass or thickness of the surface, thereby changing the refraction.

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

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

62. 62. The method of any one of claims 1 to 61, wherein the first library of compounds is from about 10,000 compounds to about 100 billion compounds or from about 10,000 compounds to about 10 billion compounds and the second library of compounds is from about 10,000 compounds to about 100 billion compounds or from 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 comprising instructions executable by said digital processing device to perform the method of any one of claims 1 to 62.

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

Images