Systems and methods for synergistic pesticide screening

Abstract

A computer system for predicting synergistic interactions between a pesticidal compound and a synergistic compound of a pesticidal composition is described. The system provides a trained classifier that provides a probabilistic prediction of synergistic effects on a pest between two or more compounds. The system can select features for transformation, encode features, generate one or more predictions, and combine predictions. The predictions can be evaluated by, for example, experimental testing in vitro or in a plant, and / or used to formulate and / or apply a pesticidal composition.

Description

[0001] Citation of relevant applications

[0002] This application claims priority and benefit to U.S. Provisional Patent Application No. 62 / 906341, filed September 26, 2019, and U.S. Provisional Patent Application No. 62 / 987751, filed March 10, 2020, the disclosures of which are incorporated herein by reference in their entirety. Technical Field

[0003] This disclosure relates in general to pesticide compositions, and more particularly to pesticide compositions having other active substances or formulation-related ingredients. Background Technology

[0004] Pesticides (such as fungicides, herbicides, nematicides, insecticides, fungicides, rodenticides, viricides, acaricides, algaecides, and molluscicides) are compositions used in domestic, agricultural, industrial, and commercial environments. The use of pesticides to control and / or suppress unwanted pests can, if left uncontrolled, harm plants (such as crops), animals, and people and / or other organisms. Therefore, effective pesticide compositions are needed.

[0005] There is also a desire to reduce the amount of pesticides used, whether to avoid harmful environmental effects, reduce costs, or for other reasons. For example, chemical pesticides are commonly used in agricultural environments where a variety of plant pests, such as insects, worms, nematodes, fungi, and plant pathogens (such as viruses and bacteria), are known to cause significant damage to seeds, ornamental plants, and crop plants. Such compositions are often expensive, potentially toxic (e.g., to humans, animals, and / or the environment), contribute to increased resistance in pest organisms to pesticides, are subject to regulatory restrictions, and / or have long-lasting effects after application. Farmers, consumers, and the surrounding environment generally benefit from using the minimum amount of chemical pesticides possible while continuing to control pest growth in order to maximize crop yields.

[0006] To address these issues, the use of natural or bio-derived pesticide compositions has been proposed as an alternative to some chemical pesticides. However, some natural or bio-derived pesticides have proven to be less effective or consistent in performance compared to competing chemical pesticides, leading to limited adoption.

[0007] Improvements in pesticides and pesticide compositions are generally desired to allow for the effective, economical, and environmentally safe control of unwanted pests (such as insects, plants, fungi, nematodes, mollusks, mites, rodents, viruses, and bacteria). In particular, there remains a need for pesticide compositions that reduce the amount of pesticide reagents and / or pesticide active ingredients required to achieve desired or acceptable levels of pest control during use.

[0008] Identifying improved pesticide compositions is generally challenging. Synergistic pesticide compositions in which the amount of the active pesticide ingredient is reduced through synergistic effects with some synergistic additives are extremely rare. For example, a systematic screening of approximately 120,000 two-component combinations based on a list of compounds in the references found that only 5% of two-component pairs (including fluconazole and triazole fungicides associated with certain azole agricultural fungicides) were synergistic (see Borisy et al., Systematic discovery of multicomponent therapeutics. Proc. Natl Acad. Sci. 100:7977-7982 (2003)). Screening more than 10^60 possible compositions for potential synergistic effects in a specific application is impractical using conventional experimental techniques; for example, a laboratory of 10 chemists might screen approximately 10^4–10^6 such compositions in a year.

[0009] Therefore, improved systems and methods are generally desired for screening pesticide compositions for synergistic efficacy.

[0010] The foregoing examples and related limitations of the related art are intended to be illustrative rather than exclusive. Further limitations of the related art will become apparent to those skilled in the art upon reading the specification and accompanying drawings. Summary of the Invention

[0011] The following embodiments and combinations thereof are intended to be exemplary and illustrative of systems, tools, and methods for description and illustration, and are not intended to limit the scope. In various embodiments, one or more of the problems described above have been reduced or eliminated, while other embodiments involve other improvements.

[0012] One aspect of the present invention provides a computing system including one or more processors and a memory containing instructions that cause the one or more processors to perform a method, and / or provides a non-transitory machine-readable medium storing such instructions. The method is used to generate predictions of synergistic interactions between two or more compounds against one or more pests. The method includes receiving a first representation of a pesticide compound; receiving a second representation of a synergistic compound; identifying a first chemical feature of the pesticide compound based on the first representation; identifying a second chemical feature of the synergistic compound based on the second representation; generating an encoded representation of a composition comprising the pesticide compound and the synergistic compound by encoding the first and second chemical features; and generating one or more predictions of synergistic interactions between the pesticide compound and the synergistic compound against one or more pests, the generation comprising: transforming the encoded representation based on trained parameters of a classifier trained for at least one synergistic interaction between compounds of at least one composition against at least one of one or more pests.

[0013] In some implementations, one or more predictions of synergistic interactions comprise multiple predictions, and the method further includes combining the multiple synergistic predictions into a combined synergistic effect. In some implementations, the method further includes determining at least one of the following based on the multiple predictions: confidence interval, standard deviation, and variance. In some implementations, the classifier comprises a random classifier, and generating one or more predictions comprises iteratively transforming the encoded representation based on trained parameters of the classifier, and generating a prediction for each iteration.

[0014] In some embodiments, generating a coded representation includes generating a first coded compound representation based on a first chemical characteristic of the pesticide compound and generating a second coded compound representation based on a second chemical characteristic of the synergistic compound, and wherein generating one or more predictions includes generating one or more predictions based on the first coded compound representation and the second coded compound representation.

[0015] In some implementations, generating the coded representation includes generating a coded representation that is lower in dimension than the coded representation.

[0016] In some implementations, generating the encoded representation includes converting a coded representation of at least one of the pesticide compound and the synergistic compound into a coded representation based on trained parameters of an encoder model. In some implementations, the encoder model includes an encoder portion of a variational autoencoder operable to convert the coded representation from the input space of the variational autoencoder into a latent space. In some implementations, the trained parameters of the encoder model have been trained on a training set different from the trained parameters of the classifier.

[0017] In some embodiments, the method further includes selecting a classifier from a plurality of classifiers based on one or more pests. In some embodiments, the method further includes receiving representations of one or more pests, and selecting a classifier includes selecting a classifier based on the representations of one or more pests. In some embodiments, the classifier is a first classifier among a plurality of classifiers, at least a second classifier among the plurality of classifiers has been trained for pests different from the one or more pests, and selecting a classifier from the plurality of classifiers includes selecting one of the first classifier and the second classifier based on one or more pests. In some embodiments, the classifier includes an ensemble classifier comprising a plurality of component classifiers, the plurality of component classifiers including at least a first component classifier and a second component classifier, the corresponding trained parameters of the first component classifier and the second component classifier each having been trained for at least one cooperative interaction between compounds of at least one composition against at least one of the one or more pests. In some embodiments, generating one or more predictions includes generating a first prediction based on a first component classifier and generating a second prediction based on a second component classifier.

[0018] In some embodiments, an enhanced representation of at least one of a pesticide compound and a synergistic compound is generated, the enhanced representation including enhanced chemical features, which include at least one of a first chemical feature and a second chemical feature. In some embodiments, generating the enhanced representation includes determining the enhanced chemical features based on trained parameters of a quantitative structure-activity relationship model.

[0019] In some embodiments, a third representation of a third compound is received, and compositions containing the third compound are excluded from the prediction based on determining at least one of the following: the chemical characteristics of the third compound match the exclusion rule, the availability value of the third compound is less than a threshold, the similarity measure between the third compound and the fourth compound is greater than a threshold, and the toxicity indication of the third compound matches the toxicity criterion.

[0020] In some implementations, the pesticide compound is selected from the group consisting of: fungicides, herbicides, nematicides, insecticides, bactericides, rodenticides, viricides, acaricides, and molluscicides.

[0021] In some embodiments, the method includes selecting at least one of a first chemical feature and a second chemical feature from the group consisting of: a representation of aromaticity, a representation of electronegativity, a representation of polarity, a representation of hydrophilicity / hydrophobicity, and a hybrid representation of at least one of a pesticide compound and a synergistic compound.

[0022] In some embodiments, one or more pests include at least one trained pest. In some embodiments, at least one trained pest shares the pesticide action mode with at least one of the one or more pests, without necessarily being included in the one or more pests.

[0023] In some embodiments, the trained parameters of the classifier are trained by: determining an importance metric for each of a plurality of training compositions; selecting one or more high-importance compositions from the plurality of training compositions based on the importance metric of each of the one or more high-importance compositions; and updating the trained parameters of the classifier based on the one or more high-importance compositions. In some embodiments, determining the importance metric of a given composition includes determining the importance metric of a given training composition based on the variance of one or more training predictions of synergistic interactions between the pesticide compound of the training composition and the synergistic compound of the training composition.

[0024] In some embodiments, selecting one or more highly important compositions includes selecting one or more highly important compositions based on representativeness criteria. In some embodiments, selecting one or more highly important compositions based on representativeness criteria includes identifying multiple clusters of a variety of training compositions, and selecting at least one highly important composition from each of at least two of the multiple clusters. In some embodiments, identifying multiple clusters of a variety of training compositions includes determining a graphical similarity measure between at least one graph representing at least one compound of a first training composition in the training compositions and at least one graph representing at least one compound of a second training composition in the training compositions.

[0025] In some embodiments, predictions of synergistic interactions are validated or evaluated by combining relevant pesticide compounds and synergistic compounds to generate a composition and exposing one or more pests to the composition in a test environment. In some embodiments, predictions of synergistic interactions are used to formulate a pesticide composition containing relevant pesticide compounds and synergistic compounds by formulating pesticide compounds. In some embodiments, predictions of synergistic interactions are used to manufacture a pesticide composition by mixing relevant pesticide compounds and synergistic compounds with any desired formulation component or additive to generate a pesticide composition. In some embodiments, predictions of synergistic interactions are used to treat one or more pests affecting non-target organisms by exposing non-target organisms to a pesticide composition containing pesticide compounds and synergistic compounds. In some embodiments, to treat one or more pests affecting non-target organisms, multiple predictions of synergistic interactions are identified and evaluated to select a combination of one pesticide compound from a plurality of pesticide compounds and a corresponding synergistic compound from a plurality of synergistic compounds. The non-target organism is then exposed to a composition containing a selected combination of one pesticide compound from a plurality of pesticide compounds and a corresponding synergistic compound from a plurality of synergistic compounds.

[0026] In addition to the exemplary aspects and embodiments described above, other aspects and embodiments will become apparent from the accompanying drawings and from the following detailed description. Attached Figure Description

[0027] Exemplary embodiments are illustrated in the accompanying drawings. The embodiments and drawings disclosed herein are intended to be illustrative and not restrictive.

[0028] Figure 1 An example system is schematically illustrated for predicting synergistic and / or antagonistic interactions between two or more compounds in a candidate pesticide composition on at least one pest.

[0029] Figure 2 It is used to pass Figure 1 A flowchart of an example method for predicting synergistic and / or antagonistic interactions between two or more compounds in a system that generates a candidate pesticide composition on at least one pest.

[0030] Figure 3 It is used to pass Figure 1 The flowchart shows an example method for screening candidate pesticide compositions using an example selector in a system.

[0031] Figure 4 It is used to pass Figure 1 The flowchart shows an example method for encoding candidate pesticide compositions using an example encoder in a system.

[0032] Figure 5 It is used to pass Figure 1 The flowchart illustrates an example method for generating one or more predictions of synergistic and / or antagonistic interactions between compounds in a candidate pesticide composition using a system example classifier.

[0033] Figure 6 It is used for training Figure 1 The flowchart shows the example method for the parameters of an example classifier in the system.

[0034] Figure 7 schematically shown Figure 1 Example data stream of the system's example combiner.

[0035] Figure 8 Showing suitable for providing Figure 1 An exemplary computer system.

[0036] Figure 9 An exemplary method is shown for evaluating the efficacy of a pesticide composition prepared using synergistic interactions.

[0037] Figure 10 An exemplary method for formulating pesticide compositions using predictions of synergistic interactions is shown.

[0038] Figure 11 An exemplary method for manufacturing pesticide compositions is shown, which uses predictions of synergistic interactions among multiple candidate pesticide compositions.

[0039] Figure 12 This paper demonstrates a method for using predictions of cooperative interactions to address one or more harmful organisms that affect non-target organisms.

[0040] Figure 13 This paper demonstrates a method for treating one or more pests affecting non-target organisms by predicting the synergistic interactions of multiple candidate pesticide compositions. Detailed Implementation

[0041] Specific details are set forth throughout the following description in order to provide a more thorough understanding to those skilled in the art. However, well-known elements are not shown or described in detail to avoid unnecessarily obscuring this disclosure. Therefore, the description and figures should be regarded as illustrative rather than restrictive.

[0042] Overview

[0043] The conventional approach to determining synergistic (and / or antagonistic) interactions between pesticide compounds and other compounds typically involves a series of laboratory screening and field trials. Initial plate tests in the laboratory screening phase usually reveal no synergistic interactions. Subsequent testing is typically conducted in plants and can consume considerable resources; for example, in an agricultural context, such testing can last for several growing seasons, involve numerous personnel and significant growing space and infrastructure, and may need to be repeated to mitigate systematic errors and / or respond to specific problems that arise during testing.

[0044] This disclosure provides systems and methods for screening candidate pesticide compositions of two or more compounds against synergistic interactions against one or more pests. In certain cases, the described systems and methods can efficiently and accurately predict which candidate pesticide compositions may have synergistic interactions against one or more pests. The described systems and methods can be used in addition to (e.g., before and / or simultaneously with) or even instead of conventional laboratory-based screening. Subsequent testing to predict compositions that may lack the desired synergistic interaction can be reduced or eliminated, potentially accelerating the discovery of synergistic pesticide compositions.

[0045] The systems and methods described herein predict synergistic interactions (or lack thereof) against at least one pest in compositions comprising at least one pesticide active ingredient and at least one synergistic compound. (As used herein, "synergistic compound" does not require that the compound is actually synergistic, but rather refers to the fact that the compound is evaluated for synergistic interaction with the pesticide active ingredient.) Depending on the intended use, the synergistic pesticide composition screening system can be configured to operate in a variety of different modes of operation. In some embodiments, the synergistic pesticide composition screening system generates predictions relating to the probability that a synergistic interaction is likely to be present in a candidate pesticide composition. Such predictions allow the user to select candidate pesticide compositions that may have a synergistic interaction for further testing steps (e.g., confirming the predicted synergistic interaction).

[0046] In some implementations, the synergistic pesticide composition screening system generates a prediction of the degree (if any) of synergistic interactions exhibited by candidate pesticide compositions. Such predictions allow users to select candidate pesticide compositions that are most likely to exhibit synergistic interactions, or that may exhibit at least some degree of synergistic interaction, for further testing.

[0047] In some embodiments, the synergistic pesticide composition screening system predicts a synergistic measure describing the synergistic interactions exhibited by candidate pesticide compositions. Any suitable synergistic measure can be predicted; for example, the system can predict the minimum inhibitory concentration (MIC) and / or fractional inhibitory concentration index (FICI) values ​​of the candidate pesticide composition. The system can alternatively or additionally predict any of a variety of other synergistic measures available, including, for example, those described by Greco et al., The search for synergy: a critical review from a response surface perspective, Pharmacological Reviews 47, 331–85, which are incorporated herein by reference.

[0048] In some implementations, a synergistic pesticide composition screening system predicts a measure of improved pesticide effectiveness of candidate pesticide compositions on one or more pest organisms. The predictive measure can be used to predict the amount of candidate pesticide composition required to achieve pesticide effectiveness in the field. Such predictions enable users to screen candidate pesticide compositions based on such predicted amounts. For example, the predicted amount can be combined (e.g., by multiplication) with an estimated cost per unit of the candidate pesticide composition to determine the predicted cost per unit of efficacy. Candidate pesticide compositions can be screened, ranked, presented to a user, or otherwise output based on such predicted amounts and / or the predicted cost per unit of efficacy.

[0049] One or more of the foregoing embodiments can provide an operating mode for a synergistic pesticide composition screening system. As described in more detail below, the synergistic pesticide composition screening system generates predictions based on trained parameters. In some embodiments, the trained parameters can be further trained based on the results of laboratory and / or field tests performed after the system generates predictions.

[0050] The foregoing overview generally refers to cooperative interactions. Antagonistic interactions can also be predicted or alternatively. Unless the context otherwise requires, this disclosure applies equally to cooperative and antagonistic interactions.

[0051] These and other aspects and advantages will become apparent when the following description is read in conjunction with the accompanying drawings.

[0052] definition

[0053] As used in this specification, the following definitions apply:

[0054] Candidate pesticide compositions: combinations of at least two candidate compounds, including at least one pesticide compound and at least one potentially synergistic and / or antagonistic compound (for convenience, generally referred to herein as a synergistic compound), having or not having a defined mixing ratio, and optionally containing one or more additional compounds. Candidate pesticide compositions may comprise mixtures.

[0055] Non-target organisms: Non-target organisms are organisms to which a pest has a harmful effect. Non-target organisms can include plants, animals, and any other affected organisms, and particularly include crop plants and crop animals, such as domestic farm animals. For example, non-target organisms include (but are not limited to) crop plants, such as cucumber and soybean plants; and crop animals, such as pigs and cattle.

[0056] Pests: Undesirable organisms that live in the environment and typically have a harmful effect on one or more host organisms (such as crop plants). Pests can be insects, plants, fungi, nematodes, mollusks, mites, rodents, viruses, bacteria, and / or other organisms. An example of a pest is powdery mildew, which grows on (and harms) multiple crop plants (such as soybean plants).

[0057] MIC: Minimum Inhibitory Concentration is the lowest concentration of a chemical substance that can be used to prevent the growth of harmful organisms.

[0058] FICI: Graded Inhibition Concentration Index: A measure of synergistic effect. Indicates the degree of “synergistic effect” (FICI ≤ 0.5), “antagonistic effect” (FICI > 4.0), and “no interaction” (FICI > 0.5–4.0).

[0059] Metric: A standard system used for measurement. A metric value is a distinct value within a specified measurement system. An example of a metric is the FICI, and the calculated FICI score is a metric value. Metrics do not need to be directly derived from measurements and can be predicted (e.g., metric values ​​are predicted by referring to synergistic pesticide composition screening systems, as described herein).

[0060] Synergistic interaction: The combined effect of two or more chemical compounds is greater than the sum of their individual effects at the same dosage. Compositions containing two or more compounds with synergistic interactions are said to have a synergistic effect.

[0061] Antagonistic interaction: The effect of two or more chemical compounds combined is less than the sum of their individual effects at the same dosage. Compositions containing two or more compounds with antagonistic interactions are said to have antagonistic effects.

[0062] Active ingredient: one or more chemical compounds (e.g., molecules, complexes, mixtures, etc.) that have the effect of inhibiting, stimulating, or otherwise altering the production or biological activity of at least one harmful organism. Compounds containing active ingredients are sometimes referred to as "active compounds".

[0063] Pesticides: Substances that effectively inhibit the growth and / or biological activity of one or more harmful organisms.

[0064] When used in the fields of chemistry and biochemistry, all other terms retain their normal meanings.

[0065] Overview of synergistic pesticide composition screening systems and methods

[0066] This disclosure provides a synergistic pesticide composition screening system and a method of operating thereof. In some embodiments, the synergistic pesticide composition screening system predicts the probability that two or more candidate compounds exhibit one or more synergistic (and / or antagonistic) interactions. In some embodiments, the synergistic pesticide composition screening system predicts the degree of synergistic (and / or antagonistic) interactions between candidate compounds. In some embodiments, the synergistic pesticide composition screening system predicts metrics, such as MIC and / or FICI values, describing the synergistic (and / or antagonistic) interactions of candidate pesticide compositions. The synergistic pesticide composition screening system generates predictions by transforming a numerical representation of candidate compounds based on a set of trained parameters as described in more detail herein. The predictions generated by the system can be used, for example, in industrial chemical composition screening processes to predict whether candidate pesticide compositions are likely to have synergistic (and / or antagonistic) interactions, and optionally predict the degree of such interaction (e.g., strong / weak) and / or metrics describing such interaction (e.g., MIC and / or FICI values, the amount of composition required to achieve a certain level of efficacy, etc.).

[0067] The active ingredient in a pesticide composition (and therefore the pesticide composition itself) typically has a limited lifespan. Pests can develop resistance to the mode of action of the active ingredient, making the pesticide composition less effective or ineffective over time. For example, some pests (e.g., insects, nematodes, fungi, yeasts, rust fungi) develop resistance to chemical compounds already used to manage their presence in crop fields. Because of pest resistance, commercial pesticides require new active ingredients to manage them. Synergistic pesticide composition screening systems attempt to identify previously unknown synergistic interactions between compounds through their predictions, thereby identifying candidate pesticide compositions of those compounds that are relatively more likely to be more effective against resistant organisms (relative to compositions with synergistic interactions not identified by the system). In some cases, an active ingredient that has previously become less effective or ineffective (e.g., due to increased resistance) can become effective again by combining with candidate compounds predicted by the system to have synergistic interactions with the active ingredient. Therefore, the synergistic pesticide composition screening system described herein can identify new pesticide compositions in a computationally tractable manner.

[0068] Figure 1 An example synergistic pest control composition screening system 1000 is shown, which in a first exemplary embodiment includes a computer system for predicting characteristics (e.g., presence, extent, and / or relevant metrics) of synergistic and / or antagonistic interactions between two or more compounds against at least one pest. The system 1000 and its operation are described herein.

[0069] System 1000 is a computer system providing selector 200, encoder 210, integrated classifier 300, and combiner 400. System 1000 optionally communicates with one or more data storage areas, such as databases 250, 251, and 570. Selector 200, encoder 210, integrated classifier 300, and combiner 400 may be provided by hardware and / or software and are generally referred to herein as “modules” of system 1000. At a high level, selector 200 receives digital representations of one or more candidate pesticide compositions and selects one or more selected candidate pesticide compositions (e.g., according to method 3000 described elsewhere herein). Encoder 210 receives one or more selected candidate pesticide compositions, and for each selected candidate pesticide composition, classifier 300 generates an encoded representation of the selected candidate pesticide composition for classification (e.g., according to method 4000 described elsewhere herein). Classifier 300 receives each encoded representation and generates one or more predictions for each encoded representation based on one or more sets of trained parameters (e.g., according to method 5000 described elsewhere herein). In some embodiments, including the depicted embodiments, classifier 300 includes an ensemble classifier comprising a plurality of trained classifiers 310a…310n, each of which generates a prediction. In at least some embodiments in which classifier 300 generates multiple predictions for selected candidate pesticide compositions, combiner 400 receives multiple predictions and generates a combined prediction 450 based on the multiple predictions (e.g., as referenced). Figure 7 (To be described in more detail).

[0070] System 1000 can be trained to predict any of the various interactions between compounds in a candidate pesticide composition. In some embodiments, system 1000 generates prediction 450 by predicting the predicted probability of the presence of synergistic (and / or antagonistic) interactions between compounds of the candidate pesticide composition and at least one pest, the predicted extent of such interactions, and / or a predicted metric describing such interactions. In some embodiments, system 1000 additionally or alternatively generates prediction 450 by predicting the toxicity of the candidate pesticide composition to at least one organism (e.g., at least one pest, at least one crop, etc.). In some embodiments, system 1000 generates prediction 450 by determining one or more metrics and / or other properties of predicted synergistic and / or antagonistic interactions between compounds derived from the candidate pesticide composition and / or at least one pest, such as predicted resistance mitigation by at least one pest or one or more pests, predicted effectiveness of the candidate pesticide composition, and / or predicted composition formulas (e.g., expressed as compound ratios).

[0071] Figure 2 An example method 2000 is illustrated for predicting synergistic and / or antagonistic interactions between two or more compounds for generating a candidate pesticide composition. This method is performed by a computer system (e.g., system 1000). At 2010, the computer system receives a representation of the candidate pesticide composition. Action 2010 may be performed, for example, by selector 200 of system 1000 and may include any actions described below with reference to method 3000, such as enhancing the representation of the composition and / or constituent compounds, filtering the composition, feature selection, etc. In some embodiments, action 2010 includes receiving a representation of the pesticide compound (at 2012) and receiving a representation of the synergistic compound (at 2014). In some embodiments, action 2010 includes receiving a representation of one or more pests that will evaluate the synergistic pesticide efficacy of the candidate pesticide composition. In some embodiments, action 2010 also or alternatively includes receiving mixture information, such as mixture ratios and / or mixture ratio ranges.

[0072] At 2020, the computer system generates encoded representations of candidate pesticide compositions for classification by classifier 300, based on the representation received at 2010, encoding the chemical characteristics of the pesticide compounds and synergistic compounds. Action 2020 may be performed, for example, by encoder 210 and / or classifier 300 of system 1000 (which may optionally be provided by a machine learning model), and may include any actions described below with reference to method 4000, such as compression, feature selection, and / or transcoding (e.g., the latent space defined by encoder 210 and / or classifier 300). Action 2030 includes converting each raw representation into an encoded representation of the candidate pesticide composition (which may include a global representation, such as a single eigenvector of the composition, and / or multiple representations, such as a representation of each compound in the candidate pesticide composition).

[0073] At 2030, the computer system generates a prediction of the synergistic efficacy of the candidate pesticide composition against one or more pests based on the encoded representation generated at 2020 and on the trained parameters of the classifier model. Action 2030 may be performed, for example, by the classifier 300 of system 1000 (e.g., trained according to method 6000) and may include any action described below with reference to method 5000. In at least some embodiments, action 2030 includes transforming the encoded representation based on the trained parameters of the classifier, which have been trained for at least one synergistic interaction between compounds of at least one composition against at least one of one or more pests. Action 2030 may include, for example, generating multiple predictions via a random classifier, as described in more detail elsewhere herein.

[0074] At point 2040, the computer system optionally combines multiple predictions to generate a combined prediction (e.g., prediction 450). Action 2040 can be performed, for example, by combiner 400 of system 1000, and can include combiner 400 as referred to below. Figure 7 The data flow diagram describes any actions. In some implementations, action 2040 includes generating a confidence metric (e.g., confidence interval) for the combined predictions, as described in more detail elsewhere in this document.

[0075] Select candidate pesticide compositions

[0076] In at least some implementations, the operation of system 1000 begins with selector 200. Figure 3 This is a flowchart of an example method 3000 for selecting candidate pesticide compositions using system 1000. Method 3000 may be performed wholly or partially by selector 200 of system 1000. Method 3000 selects candidate pesticide compositions for system 1000 to evaluate synergistic potential. Since many candidate pesticide compositions will generally be available, in at least some embodiments, method 3000 includes considering the removal of certain compounds and / or compositions from further evaluation.

[0077] At 3005, system 1000 (e.g., via selector 200) receives a digital representation of at least a portion of each of one or more compounds. The one or more compounds may be provided by a user, provided by another computing system, retrieved from a data storage area, and / or otherwise obtained via any suitable technology. Each digital representation includes a representation of the compound's chemical structure and / or chemical properties (which may include, for example, the compound's known effects on a class of organisms, such as pests, crop plants, etc.). The one or more compounds may include natural and / or synthetic compounds. System 1000 may also optionally receive a representation of at least one pest. In some embodiments, system 1000 also receives formulation parameters of the candidate pesticide composition, such as the component ratio and / or composition percentage of at least one compound in the candidate pesticide composition. The various representations and parameters received by system 1000 are collectively referred to herein as received representations of the candidate pesticide composition.

[0078] In some embodiments, system 1000 receives at 3005 a representation of a compound of a candidate pesticide composition, for example, in embodiments where classifier 300 and / or encoder 210 are trained for synergistic interactions between the synergistic compound and the pesticide compound. In this case, the pesticide compound may be implicitly represented by the trained classifier 300 and / or encoder 210, without necessarily requiring the receipt of an explicit representation of the pesticide compound. In some embodiments, the pesticide compound is predetermined, and its representation is made available to system 1000 at the start of time method 3000; access to the predetermined representation during method 3000 is included within the meaning of "receiving" such a representation.

[0079] Optionally, at 3010, system 1000 enhances the receiver representation with additional chemical properties to generate an enhanced representation. For example, selector 200 may obtain descriptions of atomic and molecular information (e.g., molecular structure, molecular weight, constituent atoms, bond type (e.g., single, double, triple, aromatic bonds)), atomic information (e.g., number of atoms, hybridization, aromatic ring membership, implicit and explicit valence, degree (number of bonds)), and / or other chemical properties (e.g., functional groups at specific positions, charge distribution) of various compounds from a data storage area (such as local memory, database 250, database 570, or other suitable data storage area). In some embodiments, system 1000 includes a training model for generating additional chemical properties (e.g., as part of selector 200), and enhances the receiver representation by generating such additional chemical properties based on trained parameters of the training model. For example, system 1000 may include a quantitative structure-activity relationship (QSAR) model, and 3005 may include generating one or more properties from the QSAR model and adding at least one of the one or more properties to the enhanced representation.

[0080] In some embodiments, the digital representation of at least a portion of the compounds in the candidate pesticide composition may include the identification of the composition or a class of compounds (thus allowing indirect identification of compounds). In some embodiments, if the candidate pesticide composition contains such a composition, and additional information for that composition is available to system 1000 (e.g., in an accessible data storage area), system 1000 (e.g., at selector 200) enhances the received representation by retrieving at least a portion of that additional information and adding the retrieved information to the enhanced representation. In some embodiments, such additional information includes the chemical composition and / or ratios of the composition. For example, selector 200 may add constituent compounds and optionally their associated concentrations to the enhanced representation of the candidate pesticide composition. Chemical composition information may be stored in a reference chemical database (e.g., Figure 1The system 1000 can add such constituent compounds to candidate pesticide compositions. (Databases 250 and / or 570).

[0081] In some implementations, if at least a portion of the representations received by system 1000 includes one or more identifiers that identify compounds of one or more classes as components of a candidate pesticide composition, system 1000 can generate multiple candidate pesticide compositions based on the one or more classes of compounds (e.g., by selector 200). For example, for each identified compound class, selector 200 can (e.g., based on information in a data storage area, such as database 250 and / or database 570) determine a set of compounds in that class. Selector 200 can generate multiple candidate pesticide compositions by generating multiple enhanced representations, each enhanced representation including a different one of the compounds in the identified class. (In the case of identifying multiple components in this way, each enhanced representation will include different combinations of compounds from the corresponding class; a given compound can be repeated by arranging the representations.)

[0082] In some implementations, if a candidate pesticide composition with multiple formulations is selected (e.g., this may be the case for natural compositions such as extracts), system 1000 (e.g., via selector 200) can select one or more such formulations. For example, selector 200 may generate multiple enhanced representations of the candidate pesticide compositions, each corresponding to a different formulation. Selector 200 may select one or more formulations in any suitable manner, including: selecting all available formulations, selecting each formulation that satisfies a rule (e.g., selecting the formulation with the lowest complexity based on a complexity metric, selecting the lowest environmental impact based on an environmental metric, selecting the lowest cost based on cost information associated with each formulation, etc.), selecting multiple formulations with the highest ranking based on a ranking algorithm, pseudo-randomly selecting one or more formulations, requesting a user to make a selection, and / or otherwise selecting one or more formulations in any suitable manner. In some implementations, system 1000 determines an average mixture ratio (e.g., via an arithmetic mean, pattern, or other suitable metric) based on available formulations and adds that average mixture ratio to the enhanced representation of the candidate pesticide composition.

[0083] In some embodiments, if the candidate pesticide composition contains a compound having more than one isomer, system 1000 (e.g., via selector 200) can select the isomer in any suitable manner, including any selection techniques described above with respect to the formulation. If more than one isomer is selected, system 1000 can generate multiple enhanced representations of the candidate pesticide composition, each enhanced representation corresponding to a different isomer.

[0084] In some embodiments, 3010 includes receiving a mixture ratio and / or a range of mixture ratios of one or more compounds (and / or constituents and / or compound classes, as appropriate) to be included in a candidate pesticide composition. If system 1000 receives a range of mixture ratios, system 1000 can (e.g., via selector 200) select one or more mixture ratios within the range and generate multiple enhanced representations of the candidate pesticide composition, each enhanced representation corresponding to a different one of the mixture ratios. System 1000 can generate such mixture ratios, for example, based on predetermined parameters (e.g., system 1000 can generate n mixture ratios for some parameter n, which are uniformly spaced within the range and include extreme values), user selection, and / or any other suitable selection.

[0085] In some embodiments, 3010 includes determining one or more fingerprints for each candidate compound. In such embodiments, the enhanced representation generated by system 1000 may include one or more fingerprints. In some embodiments, the fingerprint of a compound includes a combination of graphical representations of the compound in combination with additional properties of the candidate compound (e.g., the various properties described above). The graphical representation of each compound represents the structure of a compound molecule having nodes of a graph for each atom in the molecule and bonds represented as graphical edges. System 1000 may further enhance the graphical representation of each node (atom) in the compound, which has atomic properties such as the number of atoms, hybridization (whether or not the atom is part of an aromatic ring structure), implicit valence, and / or the degree of its bonds. System 1000 may additionally or alternatively enhance the graphical representation of each graphical edge (bond) with properties such as the type of bond (e.g., single bond, double bond, triple bond, aromatic bond).

[0086] In various implementations, different types of fingerprints can be used, including normalized Coulomb matrices (Rupp et al.), "binding bags" (Hansen et al.), and other fingerprinting algorithms, such as those provided by RDKit, such as atom pairs, topological torsion, extended connectivity fingerprints (ECFP), E-state fingerprints, Avalon fingerprints, ErG, Morgan, and MACCS. In some implementations, system 1000 determines multiple fingerprints (e.g., for similarity screening at 3035, as described in more detail elsewhere herein). In at least one implementation, system 1000 determines the Morgan and MACCS fingerprints for each candidate compound and adds these two fingerprints to the enhanced representation.

[0087] At point 3015, for each of one or more pests, system 1000 optionally obtains a representation of the pest. The representation may include, for example, an identifier of the pest (such as a name, index, and / or category variable) and / or a representation of at least a portion of the pest's genome. System 1000 may add representations of one or more pests (and / or information derived therefrom – for example, an index may be derived from the name of the pest received by system 1000) to an enhanced representation of the composition and / or otherwise associate representations of one or more pests (and / or information derived therefrom) with an enhanced representation of the composition. Representations of one or more pests may be predefined, received from a user, received from a data storage area and / or another computer system, and / or otherwise received by system 1000.

[0088] In some embodiments, for each of one or more non-target organisms, system 1000 alternatively or additionally receives a representation of the non-target organism. Non-target organisms may include, for example, host plants, animals, or other organisms on which pests feed, inhabit, or otherwise approach during the application of the pesticide composition. The representation may include, for example, an identifier of the non-target organism (such as a name, index, and / or category variable) and / or a representation of at least a portion of the non-target organism's genome. System 1000 may add representations of one or more non-target organisms (and / or information derived therefrom – for example, an index may be derived from the name of a non-target organism received by system 1000) to an enhanced representation of the composition and / or otherwise associate representations of one or more non-target organisms (and / or information derived therefrom) with an enhanced representation of the composition. Representations of one or more non-target organisms may be predefined, received from a user, received from a data storage area and / or another computer system, and / or otherwise received by system 1000.

[0089] In some embodiments, system 1000 performs action 3015 via selector 200. In some embodiments, system 1000 performs action 3015 at encoder 210, classifier 300, and / or via any other suitable module. Representations of one or more pests and / or one or more non-target organisms can be used to modulate the behavior of classifier 300. For example, system 1000 can select training models 320a,…320n of classifier 300 based on representations of one or more pests (e.g., selecting such models based on representations trained for at least one of one or more pests), as described in more detail below. As another example, system 1000 can modulate the behavior of classifier 300 by providing it with representations of one or more pests and / or one or more non-target organisms as input, for example, to inform the candidate pesticide composition of synergistic efficacy against pests and / or predictions of toxicity between the candidate pesticide composition and non-target organisms.

[0090] Candidate pesticide compositions received, identified, generated, or otherwise obtained at actions 3005, 3010, and / or 3015 form an initial set of candidate pesticide compositions (which may include the representation received at 3005 and / or the enhanced representation generated at 3010 and / or 3015). In some embodiments, system 1000 performs one or more filtration actions (such as optional filtration actions 3020, 3030, 3035, 3040 as described herein) to determine a final set of candidate pesticide compositions based on the initial set of candidate pesticide compositions. Actions 3010 and / or 3015 may be performed before, after, and / or simultaneously with one or more filtration actions; for example, system 1000 may enhance the compound representation as described above after performing one or more filtration actions.

[0091] At 3020, system 1000 optionally filters candidate pesticide compositions based on compound exclusion criteria (e.g., based on the representation received at 3005 and / or the enhanced representation generated at 3010). For example, system 1000 can retrieve a list of compounds and / or atoms to be excluded from candidate pesticide compositions from a data storage area (e.g., databases 250 and / or 570). As an illustrative example, an example exclusion criterion could exclude compositions containing arsenic and metals heavier than calcium. As another illustrative example, applying an example exclusion criterion could include determining a measure of chemical complexity and excluding compositions containing compounds whose chemical complexity measure exceeds a threshold. For example, such an exclusion criterion could exclude alkane (or other acyclic organic) molecules with chain lengths greater than a threshold. Such exclusion criteria could include rules (e.g., matching atoms to atomic masses greater than 40.078 or to an atomic number of 33), lists (e.g., a list of arsenic and all metals heavier than calcium), combinations thereof, and / or any other suitable criteria. Exclusion criteria can be predefined and retrieved by system 1000 from data storage areas (e.g., databases 250, 570, and / or parameter storage areas (not shown)). In some embodiments, system 1000 retrieves multiple exclusion criteria at 3020. System 1000 can apply all retrieved exclusion criteria or select a subset to apply.

[0092] In some embodiments, system 1000 filters candidate pesticide compositions at 3020 based on a chemical complexity criterion. The chemical complexity criterion may include excluding compounds based on their chemical structure. For example, system 1000 may exclude compounds with a chemical structure containing more than a threshold number of atoms (e.g., compounds with more than 50 atoms). The threshold may be predefined, user-provided, generated by system 1000 (e.g., the threshold may be set equal to a measure of the candidate compound's pass complexity, such as a measure of chemical complexity at the 10th, 20th, 30th, 40th, 50th, or another percentile in atomic number order), and / or otherwise obtained by system 1000. In some embodiments, system 1000 filters candidate pesticide compositions based on a subset of the constituent compounds of such compositions. For example, system 1000 may filter candidate pesticide compositions based on a chemical complexity criterion applied to candidate synergistic compounds, rather than necessarily based on a chemical complexity criterion applied to candidate pesticide compounds.

[0093] In some embodiments, system 1000 filters candidate pesticide compositions at 3020 based on an ingredient whitelist criterion. For example, system 1000 may exclude any candidate pesticide compositions containing compounds whose atoms are not on a predefined list of excluded atoms. For example, system 1000 may be configured to increase the probability that the selected candidate synergistic compound is inert and may exclude candidate pesticide compositions in which the candidate synergistic compound contains atoms that are not on a list of atoms with a high incidence in inert compounds. Such a list may include, for example, C, O, H, N, P, Cl, and F, because compounds containing atoms outside this list tend to be more likely to have undesirable and / or unpredictable biological activities. In some embodiments, system 1000 filters candidate pesticide compositions at 3020 based on an ingredient blacklist criterion. For example, system 1000 may exclude any candidate pesticide compositions containing compounds whose atoms are on a predefined list of excluded atoms (e.g., such a list may include As, Sc, Ti, V, Cr, and atoms such as heavy metals).

[0094] In some embodiments, system 1000 filters candidate pesticide compositions at 3020 based on chemical property criteria. For example, system 1000 may exclude candidate pesticide compositions containing compounds that have certain chemical properties, such as those identified by system 1000 as highly flammable, unstable, and / or having certain known interactions with other compounds in the same candidate pesticide composition (e.g., a mixture of potassium atom and water). System 1000 may determine the chemical properties of compounds based on, for example, an enhanced representation of the chemical compounds generated at actions 3010 and / or 3015, which may include a record of such properties. System 1000 may also, or alternatively, retrieve chemical property information from data storage areas (such as databases 250 and / or 570). For each compound in a candidate pesticide composition, chemical property information may be retrieved from a Material Safety Data Sheet (MSDS).

[0095] In some embodiments, chemical property information is retrieved for each compound in the candidate pesticide composition. In some embodiments, such information is retrieved for a subset of the compounds in the candidate pesticide composition. For example, in an embodiment where system 1000 is configured to increase the probability that a selected candidate synergistic compound is inert, system 1000 may retrieve such information for candidate pesticide compounds without necessarily retrieving such information for candidate synergistic compounds (e.g., where there is otherwise a high confidence that a candidate synergistic compound is inert). As another example, in an embodiment where system 1000 is configured to increase the probability that a selected candidate synergistic compound is inert, system 1000 may retrieve such information for candidate synergistic compounds to filter out candidate synergistic compounds with chemical properties that may lead to the candidate synergistic compound being non-inert (e.g., where there is otherwise no high confidence that a candidate synergistic compound is inert), without necessarily retrieving such information for other compounds in the candidate pesticide composition (e.g., where candidate pesticide compounds are pre-selected and / or otherwise not filtered individually).

[0096] As mentioned above, such exclusions can be limited to a subset of compounds, such as by excluding candidate pesticide compositions based on the atomic composition and / or other chemical properties of the candidate synergistic compounds, not necessarily other compounds. For example, it is assumed that compounds containing heavy metal atoms are excluded; thus, compositions having candidate synergistic compounds containing heavy metals can be excluded, but compositions that lack any heavy metal atoms may also be acceptable, even if the composition also contains candidate pesticide compounds containing heavy metal atoms.

[0097] At 3030, system 1000 optionally determines the availability of one or more compounds from one or more data storage areas (e.g., database 570). Such data storage areas may include inventory systems, such as those provided by users and / or by commercial chemical suppliers, such as Sigma-Aldrich. System 1000 may query such data storage areas to obtain the availability of one or more compounds. If a compound is identified as unavailable, and / or if its availability is less than an availability threshold, system 1000 may exclude candidate pesticide compositions containing that compound. The availability threshold may be the same or different for different compounds, and may be predetermined and / or provided by the user.

[0098] In some embodiments, at 3030, system 1000 additionally or alternatively retrieves resource metrics describing the allocation of resources per unit associated with one or more compounds. For example, system 1000 may retrieve resource metrics including the amount of time required to synthesize, transport, and / or otherwise procure a quantity of the compound, a measure of synthetic complexity (e.g., the number of atoms in the compound, which tends to correspond generally to the resources required to synthesize it), the amount of funds required to procure the compound and / or its composition, and / or any other suitable resource metrics. System 1000 may exclude candidate pesticide compositions containing compounds with relevant resource metrics exceeding a resource threshold. Resource thresholds may be predetermined, provided by a user, and / or retrieved from another computer system, for example. In some embodiments, system 1000 generates estimated compositional resource metrics based on one or more resource metrics associated with the compounds in a candidate pesticide composition and excludes candidate pesticide compositions with relevant estimated compositional resource metrics exceeding a resource threshold (which may be the same as or different from the resource threshold applied on a per-compound basis). System 1000 may generate estimated compositional resource metrics for candidate pesticide compositions based, for example, determining the sum and / or maximum value of resource metrics for the compounds in the candidate pesticide composition. System 1000 may scale, add to, or otherwise increase the estimated resource metrics, for example, based on a predetermined and / or user-provided estimate of the process overhead of preparing the candidate pesticide composition from its components. In some embodiments, System 1000 records candidate pesticide compositions excluded due to exceeding resource thresholds and / or unavailability in data storage areas (e.g., databases 250 and / or 570). System 1000 may, for example, display such candidate pesticide compositions to a user and / or generate a proposed list of future tests (e.g., sorted by resource metrics and / or availability).

[0099] At 3035, system 1000 optionally filters candidate pesticide compositions based on a measure of similarity (or dissimilarity) between each candidate pesticide composition and other candidate pesticide compositions, for example, to limit the selected candidate pesticide compositions generated by method 3000 to those having similar candidate synergistic compounds. In one embodiment, filtering can be performed using the fingerprint of each compound, as described elsewhere herein. System 1000 can encode each candidate compound based on its fingerprint (e.g., Morgan and / or MACCS fingerprint). For example, system 1000 can encode the molecular structure of each candidate compound in bitmap form based on its fingerprint; system 1000 can determine the measure of similarity between different compounds within a composition and / or between a compound in a composition and another compound (e.g., a compound previously excluded or included by system 1000) by determining a measure of similarity between the bitmaps of the comparing compounds. The measure of similarity can be determined via any suitable similarity technique, such as by determining a Jaccard index between bitmaps (and / or between any other suitable representations of the compounds).

[0100] When performing action 3035, several operating modes exist for system 1000. In some embodiments, system 1000 excludes compositions containing compounds having a similarity measure to any one of the compounds that is greater than (or, in some embodiments, less than) a threshold. In some embodiments, system 1000 excludes compositions containing compounds having a similarity measure to each of the one or more compounds that is greater than (or, in some embodiments, less than) a threshold. In some embodiments, system 1000 includes only those compositions containing compounds having a similarity measure to any one of the compounds that is greater than (or, in some embodiments, less than) a threshold. In some embodiments, system 1000 includes only those compositions containing compounds having a similarity measure to each of the one or more compounds that is greater than (or, in some embodiments, less than) a threshold. The threshold can be, for example, predetermined, provided by a user, and / or retrieved from another computer system. Operating modes can be predetermined and / or selected by a user. For example, a 60% threshold can be stored in a parameter storage area, optionally stored along with an "exclude <= threshold" option. In this case, action 3035 may include excluding all candidate pesticide compositions containing compounds that do not meet the minimum 60% similarity test using the Jaccard index. Users can configure system 1000 to include or exclude similar or dissimilar compounds and candidate pesticide compositions by applying appropriate settings.

[0101] In some implementations, system 1000 excludes candidate pesticide compositions based on a similarity measure of a subset of compounds in each candidate pesticide composition. For example, system 1000 may exclude candidate pesticide compositions based on a similarity measure of a candidate synergistic compound relative to a reference synergistic compound, without having to determine similarity measures of other compounds in the candidate pesticide composition. The reference synergistic compound may be provided by a user, pre-ordered, retrieved from another computer system, and / or otherwise obtained (e.g., a first candidate synergistic compound received by system 1000 while processing a batch of candidate pesticide compositions may serve as a reference synergistic compound). Where appropriate, limiting candidate synergistic compounds to those similar to a particular synergistic compound can limit the number of unstable or otherwise impractical compounds selected by system 1000, since compounds with chemical similarity to known stable compounds (e.g., formic acid) are often more likely to be stable than arbitrary compounds.

[0102] In some embodiments, system 1000 determines multiple similarity measures and includes and / or excludes candidate pesticide compositions based on these measures. For example, system 1000 may determine a first similarity measure of a candidate synergistic compound (e.g., relative to a reference synergistic compound) based on a first fingerprint, such as a MACCS fingerprint. System 1000 may further determine a second similarity measure of a candidate synergistic compound (e.g., relative to a reference synergistic compound) based on a second fingerprint, such as a Morgan fingerprint. If both similarity measures are above a threshold (e.g., 50%, 60%, 70%, 80%, 90%, and / or some other suitable threshold, which may be the same or different for the two fingerprints), system 1000 may, for example, include the candidate pesticide composition and otherwise exclude it.

[0103] At 3040, system 1000 optionally filters candidate compounds based on toxicity criteria and / or suitability criteria. For example, system 1000 may obtain a toxicity representation for each compound in a candidate pesticide composition, e.g., by retrieving toxicity representations from a compound's reception representation, enhancement representation, and / or from a data storage area such as databases 250 and / or 570. If a compound in a candidate pesticide composition has a corresponding toxicity representation that meets the toxicity criteria, system 1000 may exclude the candidate pesticide composition. For example, system 1000 may exclude all candidate pesticide compositions containing compounds having any known toxicity. As another example, system 1000 may exclude candidate pesticide compositions containing compounds having certain toxicity types (e.g., one or more toxicities identified by a dataset such as Tox21). As yet another example, system 1000 may exclude candidate pesticide compositions containing compounds having at least a certain threshold toxicity level (e.g., for toxicity types measured by a 5-point ratio, system 1000 may exclude candidate pesticide compositions containing compounds having a toxicity level of 2 or greater, without necessarily excluding those with a level of 1). As another example, system 1000 can exclude candidate pesticide compositions that contain compounds that are toxic to organisms on the list; for example, if toxicity to humans and certain crops is considered undesirable, the list can include humans and those crops, but can exclude other organisms (e.g., pests, against which toxicity may be desirable).

[0104] In some embodiments, action 3040 optionally includes filtering candidate pesticide compositions based on suitability criteria. For example, system 1000 may retrieve a list of compounds known to be suitable and / or known to be unsuitable from a data storage area (such as databases 250 and / or 570). System 1000 may exclude candidate pesticide compositions containing compounds listed as known to be unsuitable, and / or may exclude candidate pesticide compositions containing compounds not listed as known to be suitable. For example, system 1000 may query an EPA-provided database of compounds previously registered as pesticides and collect information about previous registrations, such as information that they are known to be effective against pests. System 1000 may exclude any candidate pesticide composition that does not contain at least one compound registered as effective against one or more pests identified at 3015, and / or is not registered as effective as a pesticide of a certain class (e.g., in the context of fungicide, compositions containing only compounds known to be effective as fungicides may generally be included).

[0105] At 3045, system 1000 optionally selects one or more features of the candidate pesticide composition and generates a reduced representation of the candidate pesticide composition. For example, system 1000 may generate an enhanced representation at 3010, which includes multiple features such as chemical properties (e.g., generated via a QSAR model), and may select certain features for generation (in which case 3045 may be a component action of 3010) and / or remove one or more such features after generation (in which case 3045 may be a component or independent action and may occur at any suitable time).

[0106] Features that have been identified from thousands of available features that contribute to the accurate identification of synergistically effective pesticide compositions for at least some embodiments of System 1000 include features related to aromaticity, electronegativity, polarity, hydrophilicity / hydrophobicity, and hybridization. In some embodiments, the features are selected from one or more groups of the following: electrochemical features (and specifically: electronegativity of each atom of the compound, partial charge of the compound, valence molecular connectivity index (e.g., Chi index), aromaticity and local dipole moment), topological features (and specifically: atomic hybridization, figure distance index (e.g., Weiner index) and number of polar bonds), conformational features (and specifically: number of single bonds, number of double bonds, number of triple bonds, number of aromatic bonds, number of aromatic rings, functional group orientation, representation of cis-trans isomers and representation of enantiomers), and surface-related and biochemical properties (and specifically: measures of partition coefficients (e.g., log P), measures of distribution coefficients (e.g., log D), measures of polar surface area, measures of molecular surface area, unsaturation index, hydrophilicity index and total hydrophobic surface area).

[0107] For example, in at least one example embodiment, a large number of features (e.g., approximately 2000 in the case of the RDKit QSAR model) can be generated via a QSAR model for each of one or more constituent compounds in a candidate pesticide composition. Such features may include, for example, scalar properties (e.g., magnetic properties), two-dimensional matrix properties (e.g., functional groups), and / or three-dimensional matrix properties (e.g., geometric / conformational properties) of the compounds.

[0108] System 1000 can select features that are expected to contribute to the predictions of classifier 300. For example, system 1000 can select features related to pesticide efficiency, and / or can remove features that have a low (or no) correlation with pesticide efficiency. For example, system 1000 can remove features from the enhanced representation (e.g., by instructing the QSAR model not to generate) and / or prevent the QSAR model from generating features such as: the count of the number of iodine atoms in the compound, the molecular weight of the compound, and / or the count of the number of atoms in the compound.

[0109] As another example, system 1000 may select chemical features with variance exceeding a threshold, and / or may remove features with variance below a threshold. (For example, in at least some embodiments, the same feature may be omitted on all compounds screened by system 1000 because they will have zero variance.) In some embodiments, one or more classification features are binarized; for example, a feature describing the number of rings possessed by compounds dominated by quantities 0 and 1 may be binarized into a feature describing whether a compound has a ring (i.e., transforming the feature such that 0 maps to FALSE / 0 and all other values ​​map to TRUE / 1). At 3050, system 1000 generates a final set of candidate pesticide compositions based on the representations of candidate pesticide compositions obtained at 3005, 3010, and / or 3015, and optionally based on candidate pesticide compositions excluded at 3020, 3030, 3035, and / or 3040. In some embodiments, system 1000 performs the actions of method 3000 asynchronously. System 1000 may query a data storage area, such as database 250, for records of candidate pesticide compositions and / or constituent compounds in asynchronous and / or other embodiments, and determine whether a record is ready to be encoded by encoder 210. System 1000 may perform such queries periodically. System 1000 may determine that each action recorded in other actions of method 3000 (excluding optional actions not provided in the embodiments) is ready to be encoded when performed for the corresponding candidate pesticide composition of the record. In some embodiments, system 1000 excludes any candidate pesticide compositions previously encoded by encoder 210 and / or predicted by classifier 300 from the final set of candidate pesticide compositions. System 1000 may optionally tag records of such candidate pesticide compositions to reflect such previous encoding and / or prediction, and may retrieve such tagging at 3050 and exclude candidate pesticide compositions accordingly.

[0110] In some implementations, system 1000 filters any candidate pesticide compositions that are used as part of the training set of the training model for classifier 300. System 1000 may store a list of previously trained compounds and / or candidate pesticide compositions in a data storage area such as database 250.

[0111] After action 3050, method 3000 is completed.

[0112] System 1000 may record representations of candidate pesticide compositions received and / or generated at actions 3005, 3010, 3015, and / or 3050 to a data storage area, such as database 250 and / or 570. The data storage area may be used by other modules of system 1000, users, and / or other computer systems. Where other modules of system 1000 receive information (also described herein as being stored in such data storage areas), receiving such information may include retrieving the information from such data storage areas.

[0113] System 1000 may additionally or alternatively record candidate pesticide compositions excluded at one or more filtration actions 3020, 3030, 3035, 3040 in a data storage area, such as databases 250 and / or 570. System 1000 can identify candidate pesticide compositions and / or specific constituent compounds excluded in such records. System 1000 may explicitly (e.g., by recording an indication that the compound is unavailable, on an exclusion list, or some other applicable reason) and / or implicitly (e.g., by recording the composition and / or compound in different data storage areas according to the reason for exclusion, such that compounds rejected due to unavailability are recorded in one data storage area, compounds rejected due to an exclusion list are recorded in another data storage area, etc.) record the reason for exclusion. In some embodiments, System 1000 queries such data storage areas and excludes candidate pesticide compositions previously excluded before, during, and / or after the application of filtration actions 3020, 3030, 3035, 3040.

[0114] Encoding candidate pesticide compositions

[0115] System 1000 encodes a representation of the candidate pesticide composition at encoder 210. Figure 4 An example method 4000 is illustrated for encoding a representation of a candidate pesticide composition that can be executed by encoder 210 and / or any suitably configured computer system. At 4010, encoder 210 receives a representation of each candidate pesticide composition, which may include a received representation and / or enhanced representation of the compounds of the candidate pesticide composition, formulation parameters of the candidate pesticide composition, a fingerprint of the compound, a graphical representation of the compound, atomic information, molecular information (e.g., atom count, bond type, and bond count), quantum mechanical information (e.g., electron charge distribution), and / or other information regarding the candidate pesticide composition and / or its constituent compounds as described herein. In at least some embodiments, encoder 210 receives a representation of each candidate pesticide composition in the final set of candidate pesticide compositions generated at action 3050 of method 3000. For the purpose of describing encoder 210, the representation of the candidate pesticide composition received by encoder 210 is referred to as the original representation.

[0116] At 4030, system 1000 (e.g., at encoder 210) converts each raw representation into an encoded representation of the candidate pesticide composition. The encoded representation of the candidate pesticide composition may include a holistic representation (e.g., a single eigenvector) or multiple representations (e.g., a representation of each compound in the candidate pesticide composition). The conversion performed by encoder 210 may include one or more of the following: compression, feature selection, and / or transcoding to generate encoded representations of candidate pesticide compositions suitable for classification by classifier 300. For example, encoder 210 may convert atomic, molecular, quantum dynamical, and / or other information about the candidate pesticide composition (including, for example, features of the constituent compounds) into a regular structured encoded representation that encodes at least a portion of that information while conforming to the structure required for input to classifier 300. For example, the structure of the encoded representation may correspond to the structure of the input layer of a classifier 300 that includes a neural network (e.g., if the classifier 300 takes a 32-variable input with numerical values, the encoder may generate a 32-variable encoded representation including numerical values, two 16-variable encoded representations including numerical values, and / or another set of encoded representations aligned with the input required by the classifier 300). The encoded representation may optionally be lower dimensional than the original representation and / or include fewer features than those provided by the original representation, as described in more detail below.

[0117] In some implementations, encoder 210 compresses the raw representation of candidate pesticide compositions. Raw representations of pesticide compositions (including their constituent compounds) are often complex and high-dimensional, involving numerous data points. For example, an augmented representation of a compound, including molecular information generated by QSAR, can provide more than 3,000 variables—a number that is difficult to train on at least some computer systems. Encoder 210 can transform such representations into lower-dimensional encoded representations of candidate pesticide compositions.

[0118] For example, at least one illustrative embodiment of encoder 210 converts a raw representation with more than 3,000 variables into an encoded representation with 32 variables. Encoder 210 can be configured to convert a raw representation into an encoded representation with any number of variables (e.g., 10, 16, 20, 25, 30, 40, 50, 64, 100, 128, etc.). Such encoding can be lossless and / or lossy. Suitable encoders, such as those described below, can provide a high degree of reconstruction fidelity (i.e., low reconstruction loss), meaning that in at least some embodiments, the lower-dimensional representation can encode all or almost all the information stored in the raw representation, albeit in coded form.

[0119] Several types of encoders may be used without departing from the scope of the invention. For example, in at least some embodiments, encoder 210 compresses the raw representation according to compression techniques such as Lempel-Ziv compression, prediction by partial matching, Huffman compression, arithmetic coding, Shannon-Fano compression, etc.

[0120] Optionally, at 4020, system 1000 (e.g., at encoder 210) performs feature selection based on the original representation. Such feature selection can supplement or replace the feature selection of action 3045 of method 3000. (Action 3045 can optionally be performed entirely or partially by encoder 210.) Encoder 210 can, for example, discard portions of the original representation and retain other portions to generate a lower-dimensional encoded representation that includes only the retained portions. Although feature selection is in the form of (typically lossy) compression, the retained portions are not necessarily compressed or otherwise encoded (although encoder 210 can optionally encode, for example, the retained portions as described herein).

[0121] In some embodiments, feature selection by encoder 210 includes extracting one or more feature descriptors based on the original representation. The feature descriptors describe characteristics of the candidate pesticide composition (e.g., characteristics of the constituent compounds of the candidate pesticide composition) and may include, for example, atomic information, molecular information (e.g., atom count, bond type, and / or bond count), quantum mechanical information (e.g., electron charge distribution), and / or other characteristics of the candidate pesticide composition (e.g., its constituent compounds). A given feature descriptor may be associated with one or more candidate pesticide compositions. Multiple feature descriptors may be associated with each other, such as when multiple feature descriptors are associated with a fingerprint (e.g., a graphical representation) of the compounds of the candidate pesticide composition.

[0122] In some embodiments, encoder 210 generates an encoded representation that includes an explicit representation of feature descriptors. For example, encoder 210 can extract an atom count from an original representation of a compound of a candidate pesticide composition and generate an encoded representation containing a value explicitly representing that atom count. For example, if the original representation of a candidate pesticide composition indicates that a first compound of the candidate pesticide composition has 10 atoms, encoder 210 can generate an encoded representation that includes the numeric scalar value 10. As another example, feature descriptors can include non-scalar (e.g., vector) values, such as where encoder 210 encodes the molecular structure of the compound in the encoded representation as a Simplified Molecular Linear Input Specification (SMILES) string. In some embodiments, encoder 210 generates an encoded representation that includes an implicit representation of feature descriptors, for example via a compressed representation, which can combine feature descriptors into a single scalar value and / or distribute information of the feature descriptors across multiple scalar values. The latent space encoded representation generated by an embodiment of encoder 210 that includes an encoder portion of a variational autoencoder is an example of such implicit feature selection.

[0123] The features selected by encoder 210 can vary across implementations. For example, atomic, molecular, quantum dynamics, and / or other features of a candidate pesticide composition (e.g., features of its constituent compounds) can be encoded differently by different encoders 210 and / or by a single encoder 210 providing different encoding schemes. Various encodings can be provided by encoder 210. System 1000 can use more than one encoding (if desired) to generate encoded representations of compounds, and / or can use different encoders 210 and / or different encodings provided by encoder 210 to generate encoded representations of different compounds. In some implementations, system 1000 provides at least two encoders—at least a first encoder for converting the raw representation of the pesticide compound; and at least a second encoder for converting the raw representation of the synergistic compound. Such first and second encoders can provide different encodings (e.g., pesticide compounds and synergistic compounds can be encoded with different numbers of values ​​and different selected features based on different trained parameters of the encoders and / or through different types of encoders).

[0124] In some embodiments, encoder 210 is configured to encode candidate pesticide compositions comprising more than two constituent compounds (e.g., comprising multiple candidate pesticide compounds, multiple candidate synergistic compounds, and / or one or more other compounds, such as adjuvants, solvents, etc.). For example, encoder 210 may generate coded representations based on three, four, or more compounds. In some embodiments, encoder 210 receives a fixed number of representations of compounds (e.g., encoder 210 may be configured to receive three compounds) and is trained against training data that includes representations of pesticide compositions having the same number of compounds. In some embodiments, encoder 210 receives a variable number of compounds depending on the number of constituent compounds in the encoded candidate pesticide composition. Encoder 210 can encode such compositions in any suitable manner; for example, encoder 210 can represent a fixed number (e.g., one, two, or more) of compounds received each time during the encoding process to generate intermediate encoded representations (e.g., 16, 32, 64, or 128 variable floating-point representations), and then combine the intermediate encoded representations via attention mechanisms, pointwise summation, and / or any other suitable method to generate a final encoded representation (e.g., in the same form as the intermediate encoded representations). Encoder 210 can optionally generate separate encoded representations for candidate synergistic compounds and candidate pesticide compounds.

[0125] In at least one exemplary embodiment, encoder 210 receives the identification of a set of feature descriptors required by classifier 300 (e.g., in the case where classifier 300 includes an ensemble classifier, it may include the identification of feature descriptors required by trained classifiers 310a,…310n) and performs feature extraction on each compound represented in the original representation of the candidate pesticide composition based on the identification of this set of feature descriptors. This set of identifications may include the identification of multiple compounds received by classifier 300 and / or, for each compound, a set of feature descriptors for that compound, and encoder 210 may perform feature extraction on each compound based on this set of feature descriptors specified for that compound. In some embodiments, encoder 210 adds mixture ratio information associated with the candidate pesticide composition (e.g., as represented by and / or associated with the original representation) to the encoded representation. For example, encoder 210 may encode representations of compounds, add these representations to the encoded representation, and add mixture ratio information to the encoded representation of the candidate pesticide composition independently of the encoding of the compounds. As another example, mixture ratio information can be encoded along with the representation of the compound, for example, by incorporating such mixture ratio information into a compressed and / or latent space representation (described below) generated by encoder 210. For example, the encoded representation of the compound can be combined (optionally along with the mixture ratio information) via concatenation, attention mechanisms, and / or any other suitable combination techniques.

[0126] In some embodiments, some information passed to classifier 300 is not encoded. For example, encoder 210 may only encode the raw representation of the candidate compound, while other information (such as formulation parameters of the candidate pesticide composition and / or the representation of one or more pests) may be passed to classifier 300 without being encoded. In some embodiments, system 1000 encodes such other information separately from the encoding of the raw representation of the compound.

[0127] In some embodiments, encoder 210 receives raw representations of compounds as input and transforms the raw representations based on a set of trained parameters of encoder 210. In some embodiments, encoder 210 independently receives and encodes the raw representations of each compound in the candidate pesticide composition, thereby generating an encoded representation for each compound. In some embodiments, system 1000 provides multiple encoders 210. System 1000 may encode a first compound (e.g., the active ingredient of the pesticide composition) of the candidate pesticide composition using a first encoder and encode a second compound (e.g., a candidate synergistic ingredient) of the candidate pesticide composition using a second encoder. The first encoder and the second encoder may be trained on the same or different training sets and include the same or different structures and / or parameters. For example, the first encoder may be trained on a training set of active ingredients of the pesticide composition, and the second encoder may be trained on a training set of synergistic (and / or antagonistic and / or non-synergistic) ingredients.

[0128] In some embodiments, encoder 210 includes at least a portion of a variational autoencoder. In at least one embodiment, encoder 210 includes an encoder portion of a variational autoencoder that has been trained together with a decoder portion but operates during encoding without the decoder portion. (The decoder portion is not necessarily part of system 1000.) Such encoder 210 transforms a (relatively sparse) raw representation x in an input space X into a (relatively dense) encoded representation z in a latent space Z characterized by the input data and a previous distribution p(z). Specifically, encoder 210 determines p(z|x) to generate a distribution over the latent space of a given compound. Encoder 210 can transform the distribution into an encoded representation in any suitable manner. In at least some embodiments, encoder 210 transforms a determined distribution into an encoded representation, for example, by determining the mean of the distribution (e.g., independently or jointly over the latent variables). Such encoder 210 can be considered to provide implicit feature compression (and in some sense, “discriminative” features of the compound) by tending to identify those features that most contribute to accurate reconstruction.

[0129] In some embodiments, encoder 210 includes an encoder of an inverse autoregressive rheovariate autoencoder. For example, encoder 210 can be trained on any suitable training dataset of chemical compositions (as described elsewhere herein) to find parameters that minimize a suitable objective function. For example, the objective function can be provided by log p(x) (and a loss function can be derived from it, for example, via inversion), which in at least some embodiments can be approximated by a lower bound based on the following:

[0130]

[0131] It can be represented in the following form:

[0132] E q [log p(x|z T )+log p(z T )-log q(z T |x)]

[0133] Where p is the true distribution trained on the inverse autoregressive hydrodynamic autoencoder, q is the approximate distribution learned by the inverse autoregressive hydrodynamic autoencoder, and z... T It is an element of the potential space and can be described as T in at least some implementations. th z i Among them, for some series of reversible transformations f i(·) , z0~q(z0|x) and z i =f i (z i-1 ,x), where x is an element from the input space.

[0134] Furthermore, in at least some implementations, log q(z) T |x) and log p(z) T It can be approximated as:

[0135]

[0136] Where ∈ is a suitable noise vector (e.g., ∈ ~ N(0, I)) and σ t，i It is the latent variable z t The variance of the i-th element.

[0137] In some implementations, encoder 210 is trained via a semi-supervised method, for example, to minimize the reconstruction loss between the input representation in the training set and the reconstructed representation generated by the decoder portion (based on the encoded representation generated by encoder 210). In some implementations, encoder 210 is pre-trained and / or trained on a dataset larger and / or more general than classifier 300. For example, classifier 300 may be trained on pesticide compositions (and / or subclasses of such compositions), while encoder 210 may be trained on a chemical dataset that is not limited to and may not even contain pesticide compositions. In some implementations, encoder 210 and classifier 300 are trained together, such that training involves updating the parameters of both encoder 210 and classifier 300 to minimize (or maximize, as appropriate) a shared objective function through shared data. For example, training data may include a subset relevant to the classifiers, and the combined loss function of encoder 210 and classifier 300 may be based on: 组合 =l 编码器 +αl 分类器 α = 1 if the given data is in the relevant subset of the classifier, otherwise α = 0. In some implementations, encoder 210 and classifier 300 are trained separately. A potential advantage of training encoder 210 and classifier 300 together is that, compared to training them separately, training together can tend to make encoder 210 tend to select features more relevant to classifier 300 with the potential cost of greater complexity and limited relevant training data.

[0138] In some embodiments, encoder 210 includes a neural network, such as a graph convolutional neural network. The neural network can receive a raw representation of the compound as input at an input layer (and / or a portion thereof, for example, encoder 210 can receive a graph representation of the compound with relevant properties), and transform the raw representation based on a set of trained parameters corresponding to the input layer and based on the activation function and nonlinearity provided by the neural network, thereby generating an intermediate representation. Encoder 210 can further transform the intermediate representation via one or more hidden layers, each hidden layer having a corresponding structure (e.g., inter-layer input / output), nonlinearity, and trained parameters, and finally generate an encoded representation at the output layer (having its own structure, nonlinearity, and trained parameter form). In at least some embodiments, the structure of the output layer corresponds to the form of the input required by classifier 300. For example, if classifier 300 receives a 32-variable input, encoder 210 can generate a 32-variable encoded representation via a 32-variable output layer. (The intermediate representation does not necessarily have, and typically will not have, the same number of variables or the same structure as the output layer).

[0139] In some embodiments, classifier 300 includes encoder 210 (i.e., encoding and classification functions may be provided by a single module). For example, in some embodiments, classifier 300 may include a graph convolutional neural network (GCNN) that receives one or more graph representations of candidate pesticide compositions (e.g., generated by selector 200) and, in an initial phase, accumulates information at the nodes and / or edges of the graphs by traversing them, thereby determining intermediate (i.e., encoded) representations of the candidate pesticide compositions to flatten those representations. In a later phase of the GCNN operation, the intermediate representations are further transformed into appropriate outputs.

[0140] For example, system 1000 can generate and provide a graphical representation of each compound of a candidate pesticide composition to a GCNN. As another example, system 1000 can generate and provide a graphical representation of a candidate pesticide composition to a GCNN, which may include a non-crosslinked subgraph representing each compound of the candidate pesticide composition. In some embodiments, system 1000 may connect such non-crosslinked subgraphs to generate a connected graph representing at least a portion of the candidate pesticide composition. In at least one embodiment, system 1000 adds edges (representing bonds) between hydrogen-bonded sites to the graphical representation of the constituent compounds of the candidate pesticide composition. System 1000 may represent bond lengths in such graphical representations; the representation of added bonds between hydrogen-bonded sites may provide lengths different from single and double bonds. For example, bond lengths may be categorized, in which case the length of a single bond may be 1, the length of a double bond may be 2, and the length of the added bond may be 3 (or in one-hot encoding, such as (1, 0, 0), (0, 1, 0), and (0, 0, 1) respectively). As another example, bond lengths can be represented continuously (e.g., based on physical length), in which case the length of an added bond can be represented as longer (i.e., weaker) than a single bond (e.g., 1 for a single bond, 0.5 for a double bond, and 2 for an added bond). In at least some experimental tests, the bond lengths of added bonds, which are significantly different from those of single bonds, have correlated with improved performance of the systems and methods described herein.

[0141] System 1000 may record coded representations of candidate pesticide compositions generated by encoder 210 to a data storage area (such as database 250 and / or 570). The coded representation may be associated with its corresponding original representation (e.g., a corresponding received representation and / or a representation identified at action 3050 of method 3000). The coded representation may also, or alternatively, be associated with the encoder (e.g., encoder 210) that generated the coded representation. Such association may include, for example, recording an identifier of the corresponding representation / encoder in a record of the coded representation, and / or recording an identifier of the coded representation in a record of the associated representation / encoder. The data storage area may be used for other modules of system 1000 (e.g., classifier 300), users, and / or other computer systems. Where other modules of system 1000 receive information (also described herein as being stored in such a data storage area), receiving such information may include retrieving the information from such a data storage area. In some implementations, if encoder 210 is modified (e.g., by updating its trained parameters through training), system 1000 can regenerate the encoded representation associated with encoder 210 by obtaining the original representation from data storage (and / or, for example, obtaining such an original representation from selector 200 based on the received representation) and converting the original representation into a new encoded representation. This can reduce the computational requirements of recoding relative to recoding all encoded representations of all encoders if system 1000 provides multiple encoders.

[0142] Synergistic effect prediction of candidate pesticide compositions

[0143] For each candidate pesticide composition, classifier 300 receives an encoded representation generated by encoder 210 and generates one or more predictions based on the encoded representation and based on one or more sets of trained parameters. Figure 5An indicative method 5000, executable by a classifier 300 and / or any suitably configured computer system, is shown for generating predictions of the synergistic efficacy of candidate pesticide compositions against one or more pests. At 5010, the classifier 300 receives a representation of each candidate pesticide composition, which may include a reception representation, an enhancement representation, and / or an encoding representation of the candidate pesticide composition (and may include such representations of the constituent compounds of the composition). At 5040, the classifier 300 translates such representations into predictions of the synergistic interactions between the constituent compounds of the candidate pesticide composition against one or more pests. The classifier 300 models complex nonlinear relationships between candidate compounds that form the basis for synergistic and / or antagonistic interactions between compounds of the candidate pesticide composition on one or more pests. For example, an active ingredient may be effective against a specific pest in a laboratory setting but cannot penetrate the cell membrane of the pest in the context of a plant or in the field due to the pest's natural defenses. A synergistic combination of two or more compounds (e.g., one or more active compounds and one or more synergistic compounds) allows the active compound to access the cellular structure of the pest, thereby making the active compound effective for use in plants and in the field. Even if anticipated by the subject matter, such interactions between compounds and harmful organisms are not easy to predict.

[0144] The classifier 300 may include any suitable classifier, such as neural networks, decision trees, logistic regression, support vector machines, stacked model classifiers, and / or any other suitable classifier. In some implementations, including Figure 1 According to the described implementation, classifier 300 includes an ensemble classifier comprising a plurality of trained classifiers 310a…310n (collectively and individually “classifier 310”), each of which generates predictions based on a corresponding set of trained parameters 320a…320n (collectively and individually “trained parameters 320”). In some implementations, classifier 310 includes a deep neural network (DNN) model with multiple computational layers. Each classifier 310 simulates interactions between compounds and also simulates interactions between one or more compounds and the natural defenses of one or more pests. System 1000 may include any number of classifiers 310. For example, system 1000 may include 8, 16, 32, 64, 128 and / or any other suitable number (not necessarily powers of two) of classifiers.

[0145] For example, classifier 300 may include multiple trained neural network classifiers (e.g., classifier 310), each of which is parameterized by a corresponding set of trained parameters 320 (e.g., classifier 310a can be parameterized by trained parameters 320a, classifier 310b can be parameterized by trained parameters 320b, etc.). Different classifiers 310 (and therefore different trained parameters 320) may be trained for different pests and / or different compounds, and may thereby simulate different interactions. For example, the trained parameters 320 of each classifier 310 may have been trained on a corresponding training dataset that includes compositions of compounds (and optionally one or more pests) that have been identified as having synergistic and / or antagonistic effects. In some embodiments of method 5000, system 1000 receives one or more representations of one or more pests (at 5020) and selects a classifier 310 trained for at least one of the one or more pests (at 5030), such as as described in more detail elsewhere herein. The selected classifier 310 is then executed to generate a prediction at 5040.

[0146] Figure 6An example method 6000 for training parameters of classifier 300 is shown. Method 6000 may optionally include parameters for training encoder 210 (e.g., by training encoder 210 together with classifier 300 and / or by training encoder 210, substantially according to the following description of method 6000). In some embodiments, action 6010 substantially corresponds to action 5010. In some embodiments, method 6010 includes selecting candidate pesticide composition representations based on co-interaction predictions (such as co-interaction predictions generated in action 5040 and / or action 6020). For example, in some embodiments, method 6000 includes training the parameters of classifier 300 via active learning, which may include, for example, determining the importance value of each of a plurality of candidate pesticide composition representations (e.g., all available candidate pesticide composition representations, candidate pesticide composition representations within a batch, candidate pesticide composition representations with corresponding co-interaction predictions having variance exceeding a threshold, or any other suitable plurality of candidate pesticide composition representations) based on co-interaction predictions generated for each such candidate pesticide composition representation (e.g., as in actions 5040 and / or 6020). In some embodiments, one or more candidate pesticide composition representations are selected at action 6010 based on the corresponding importance value of one or more candidate pesticide composition representations, and actions 6020, 6030, 6040, and 6050 are performed based on the selected candidate pesticide composition representation, thereby updating the parameters of classifier 300 based on the selected candidate pesticide composition representation.

[0147] In some embodiments, determining the importance value of a plurality of candidate pesticide composition representations includes determining an information measure for each of the plurality of candidate pesticide composition representations. The information measure may be based on the standard deviation, variance, and / or confidence interval (and in some embodiments equivalent to standard deviation, variance, and / or confidence interval) of one or more cooperative interaction predictions generated by classifier 300 for the candidate pesticide composition representation (e.g., as in actions 5040 and / or 6020). In some embodiments, such as those in which classifier 300 includes an integrated classifier, the standard deviation 7220, variance, and / or confidence interval 7220 may be referenced, and / or the variance may be determined by any other suitable determination as described elsewhere herein. In at least one embodiment, the importance measure includes determining the variance (e.g., based on standard deviation 7220). In some embodiments, such as those including a hyperplane-based classifier 300, the information measure may be based on the distance of the candidate pesticide composition representation to the nearest hyperplane. In some embodiments, other suitable importance measures may be determined additionally or alternatively.

[0148] In some embodiments, the selection of candidate pesticide composition representations further includes selecting candidate pesticide composition representations based on representativeness criteria. For example, candidate pesticide composition representations may be clustered based on a similarity metric (e.g., graphical similarity, for at least some embodiments in which a candidate pesticide composition representation includes a graphical representation of a candidate molecule and / or other constituent substituents), and one or more candidate pesticide composition representations may be selected from each of a plurality of clusters. In some embodiments, an information metric is determined for only a subset of candidate pesticide composition representations within a cluster; for example, an information metric may be determined for candidate pesticide composition representations at the center of each cluster (as defined by a clustering metric), and candidate pesticide composition representations from a plurality of clusters may be selected based on their information metrics (e.g., by selecting n candidate pesticide composition representations with the highest or lowest importance values, as appropriate; by selecting candidate pesticide composition representations with an importance metric higher than or lower than (and / or optionally equal to) a threshold, as appropriate; and / or by any other suitable selection criteria).

[0149] Appropriate representativeness criteria can facilitate dissimilarity among selected candidate pesticide composition representations and, where appropriate, can be combined with suitable information metrics, enabling the training of classifier 300 to achieve model convergence with candidate pesticide composition representations that require fewer labels than random sampling. Obtaining labeled candidate pesticide composition representations can be costly; for example, it may involve human experts performing laboratory experiments to confirm the synergistic interactions of candidate pesticide compositions. Where appropriate, this active learning approach can reduce the amount of laboratory experiments required or desired to fully train the model.

[0150] In some implementations, action 6020 substantially corresponds to action 5040. In some implementations, classifier 300 operates at action 6020 in a different mode than action 5040, such as in implementations where classifier 300 generates predictions with missing values ​​during training at action 6020 instead of action 5040.

[0151] At 6030, system 1000 receives a representation of experimental results, which includes an indication of the synergistic and / or antagonistic efficacy of the candidate pesticide composition of action 6010 against at least one training pest. In some embodiments, the at least one training pest is one or more pests that classifier 300 predicts for it. In some embodiments, the at least one training pest shares a pesticide action pattern with at least one of the one or more pests. For example, if classifier 300 predicts one or more pests for it to include lepidopteran pests (such as the codling moth), classifier 300 can be trained for experimental results that include an indication of the synergistic and / or antagonistic efficacy of the candidate pesticide composition against other pests that share a pesticide action pattern with such lepidopteran pests, such as associated lepidopteran pests (e.g., in an earlier example involving the codling moth, such associated lepidopterans could include the bollworm).

[0152] At 6040, system 1000 determines the value of an objective function (which may include, for example, a loss function) based on the representation of the prediction generated at 6020 and the experimental results received at 6030, for example, based on the difference between them. At 6050, system 1000 updates the parameters of classifier 300 based on the value of the objective function determined at 6040, for example, via backpropagation. In some implementations, different classifiers 310 have been trained on different subsets of a public training dataset. Subsets may overlap or not overlap. (Each classifier may be further validated against elements of a public training set on which it was not trained.) Subsets may be pseudo-randomly determined by identifying subranges based on some ordering of the dataset and / or by any other suitable determination criterion.

[0153] In some implementations, a subset of the public training dataset may have been determined based on the composition having tested synergistic (and / or antagonistic) interactions with pests. For example, a first classifier 310a may have been trained on a first subset of the training data, which includes compositions having known synergistic, antagonistic, or non-interacting interactions with at least a first pest. A second classifier 310b may have been trained on a second subset of the training data, which includes compositions having known synergistic, antagonistic, or non-interacting interactions with at least a second pest. Classifiers 310a and 310b may have been trained, respectively, on the interactions between the first and second pests. For example, classifier 310a may have been trained to generate predictions of synergistic effects of compositions against at least a first pest that minimize the reconstruction loss (or other suitable objective function) for the first subset of the training data, while classifier 310b may have been trained to generate predictions of synergistic effects of compositions against at least a second pest that minimize the reconstruction loss (or other suitable objective function) for the second subset of the training data. Classifier 310a is referred to herein as being trained for a first pest, and classifier 310b is trained for a second pest. In some embodiments, classifier 310 is trained for a category of pest; for example, the first classifier 310a may have been trained for fungal pests and classifier 310b may have been trained for bacterial pests.

[0154] Alternatively or additionally, subsets of the public training dataset can be determined based on the chemical properties of the compositions in the common training dataset, such as the chemical structure of the constituent compounds. Mixtures can be grouped into subsets based on, for example, the following: their broad chemical category (e.g., organic, inorganic, synthetic, and / or biological); specific chemical functional groups (e.g., having aryl, alkyl, ethyl, methyl, and / or other groups); similarity (e.g., representative compounds and their substituents, isomers, other compounds sharing a moiety with them, and other structurally related compounds); and the physical state of the composition and / or its constituent compounds (e.g., fumigant, spray, dust, etc.). For example, a first classifier 310a may have been trained on a first subset of the training data, which includes compositions containing organic pesticide active ingredients. A second classifier 310b may have been trained on a second subset of the training data, which includes compositions containing inorganic pesticide active ingredients. Classifiers 310a and 310b may have been trained on organic and inorganic pesticide active ingredients, respectively. For example, classifier 310a may have been trained to generate predictions of the synergistic effects of compositions containing organic pesticide active ingredients (e.g., against one or more pests), which minimize the reconstruction loss (or other suitable objective function) for a first subset of the training data, while classifier 310b may have been trained to generate predictions of the synergistic effects of compositions containing inorganic pesticide active ingredients (e.g., against the same or different pests as the first classifier), which minimize the reconstruction loss (or other suitable objective function) for a second subset of the training data. In some embodiments, classifier 310 is trained for categories of pests; for example, first classifier 310a may have been trained for fungal pests and classifier 310b may have been trained for bacterial pests.

[0155] System 1000 can store, receive, and / or be operable to retrieve records during operation indicating which compounds and / or pests each classifier 310 has been trained against. In some embodiments, classifier 300 selects one or more classifiers 310 from a plurality of classifiers 310 based on candidate pesticide compositions to be treated (e.g., based on received, enhanced, raw, and / or encoded representations of the candidate pesticide compositions), and generates predictions with the selected classifiers 310 based on their associated parameters 320 and the encoded representations of the candidate pesticide compositions. For example, if classifier 300 predicts the likelihood of a candidate pesticide composition synergistically resisting Varroa mites, and if classifiers 310a and 310b have been trained against Varroa mites but classifier 310c has not, then classifier 300 can select and generate predictions with classifiers 310a and 310b (based on parameters 320a and 320b) instead of using classifier 310c. As another example, if a candidate pesticide composition contains an active ingredient that has been trained on classifiers 310b and 310c for (e.g., a composition containing that compound and various synergistic compounds), but classifier 310a has not, then classifier 300 can use classifiers 310b and 310c (based on parameters 320b and 320c) to select and generate predictions, instead of using classifier 310a to select or generate predictions.

[0156] In some implementations, classifier 300 selects and retrieves trained parameters 320 from a trained parameter database 251. Each classifier 310 independently generates a prediction of synergistic (and / or antagonistic) interactions based on the corresponding trained parameters 320. Predictions may include, for example, the probability (and / or confidence interval) of such synergistic interaction, the degree of such synergistic interaction, and / or a metric describing such synergistic interaction (e.g., MIC and / or FICI values). Classifier 310 is not limited to generating predictions and may generate additional and / or alternative outputs; for example, classifier 310 may also (or alternatively) predict the toxicity and / or volatility of candidate pesticide compositions (and / or any constituent compounds), and the resistance of pests to candidate pesticide compositions (e.g., based on pest genomic data received as input and / or by training classifier 310 for pest resistance). Predictions (and / or other outputs) from each classifier 310 may be sent to combiner 400 for combination.

[0157] In some implementations, classifier 300 (e.g., at least one classifier 310) is randomized and can generate batches of different predictions based on a single encoded representation. In some implementations, classifier 300 generates more than one prediction based on a single encoded representation (e.g., in the case of an ensemble classifier, by means of a given classifier 310). For example, system 1000 may perform a loss during inference with classifier 300, for example by pseudo-randomly deactivating variables in the model of classifier 300 (e.g., at least one classifier 310) during inference. (A loss may also be performed optionally during training.) Thus, it is expected that different results may be generated with each inference iteration. System 1000 may combine multiple such predictions to determine a combined prediction and may assign confidence to the combined prediction based on the variance of the multiple predictions, for example as described in more detail elsewhere herein.

[0158] In some implementations, classifier 300 receives an encoded representation (e.g., from encoder 210), optionally determines the number N of classifiers 310 to be selected, optionally determines the number M of predictions to be generated for each classifier 310 (N and M are described below), selects N classifiers 310 if appropriate (e.g., based on the encoded representation and / or as described above), and generates M predictions for each of the N selected classifiers 310 based on the encoded representations corresponding to the selected classifiers 310 and trained parameters 320. The number N of classifiers 310 to be selected and / or the number M of predictions to be generated for each classifier 310 can be predetermined, provided by the user, determined by system 1000 (e.g., based on available computing resources), and / or otherwise obtained by classifier 300. For example, N can be 8, 16, 32, 64, 128, and / or any other suitable number (not necessarily a power of two). M can be 20, 40, 100, 200, 1000, and / or any other suitable number (not necessarily a multiple of 10). In at least one implementation, N is 32 and M is 100. The terms N and M may be implicit in the model; for example, classifier 300 may be configured to generate a prediction with each classifier 310 (i.e., N = n and M = 1). Classifier 300 may select N trained classifiers 310 based on, for example, the encoded representation described above (and select corresponding trained parameters 320 from the trained parameter database 251). Classifier 300 parameterizes classifier 310 using the selected trained parameters 320 and generates predictions based on the selected trained parameters 320.

[0159] System 1000 may record predictions generated by classifier 310 to a data storage area, such as database 250 and / or 570. Predictions may be associated with their corresponding encoded representations (e.g., with corresponding received representations, raw representations, and / or encoded representations). Predictions may also, or alternatively, be associated with the classifier 300 (and / or classifier 310) that generated the predictions. Such associations may include, for example, recording the identifier of the corresponding representation / classifier in the record of the prediction, and / or recording the identifier of the prediction in the record of the associated representation / classifier 300 / 310. The data storage area may be used for other modules of system 1000 (e.g., combiner 400), users, and / or other computer systems. Where other modules of system 1000 receive information (also described herein as being stored in such data storage areas), receiving such information may include retrieving the information from such data storage areas. In some implementations, if the predicted corresponding encoded representation and / or classifier 300 (and / or classifier 310) is modified (e.g., via updating trained parameters 320 through training), system 1000 can regenerate the prediction by obtaining the corresponding encoded representation from a data storage area (and / or, for example, obtaining such encoded representations from another module, including by regenerating them at such other modules) and converting the encoded representation into a new prediction via classifier 310. This can reduce the computational requirements for regenerating predictions relative to regenerating all predictions of all classifiers 310 and / or all encoded representations.

[0160] Prediction of combined synergistic effects

[0161] In at least some embodiments, combiner 400 combines multiple predictions generated by classifier 300 into a final prediction 450. In some embodiments, prediction 450 includes a measure of the probability of synergistic and / or antagonistic interactions between compounds of the candidate pesticide composition and / or one or more pests. For example, prediction 450 may include a mean and a confidence interval. In at least some embodiments where classifier 300 includes multiple classifiers 310, combiner 400 generates prediction 450 based on the predictions of each classifier 310.

[0162] Figure 7 An exemplary data stream characterizing the operation of combiner 400 is shown. Combiner 400 receives multiple predictions 7100 and generates a combined prediction 7300 based on the predictions 7100. In at least one depicted embodiment, combiner 400 receives multiple predictions 7100, which include multiple predictions 7110 generated by each classifier 310 of classifier 300 (these are depicted as...). Figure 7The rows of predictions 7110 in the matrix depicting predictions 7100 in the data stream. In some implementations, each classifier 310 can generate a number M predictions 7110 during M iterations. Therefore, predictions 7100 can include multiple predictions 7120 generated for each iteration (these are depicted as rows of predictions 7110 in the matrix depicting predictions 7100). Figure 7 (Columns of predictions 7120 in the matrix depicting predictions 7100 in the data stream). The number of predictions 7120 in each iteration can be the same, for example, N for each iteration, or can be different between iterations, for example in an implementation in which classifier 310a generates predictions after more or fewer iterations than another classifier 310b.

[0163] In some implementations, combiner 400 generates multiple aggregated predictions 7200 based on prediction 7100, and generates combined prediction 7300 based on the aggregated predictions 7200. Combiner 400 can generate aggregated prediction 7200 by, for example, identifying multiple subsets of predictions 7100 and generating an aggregated prediction for each such subset based on the predictions 7100 of that subset. For example, combiner 400 can identify each plurality of predictions 7110 generated by classifier 310 and / or each plurality of predictions 7120 associated with an iteration as a subset, and can generate each aggregated prediction 7200 based on the corresponding plurality of predictions 7110 and / or 7120. Generating aggregated prediction 7200 by combiner 400 may include, for example, combiningr 400 determining the mean and / or standard deviation (and / or variance) of probabilities within the selected subset. Generating combined prediction 7300 by combiner 400 may include determining the mean and / or standard deviation of the aggregated predictions 7200. For example, combiner 400 may determine the mean 7210 and optionally the standard deviation (and / or variance) 7220 of each plurality of probabilities 7110 (and / or 7120) to generate each aggregate prediction 7200. Combiner 400 may further determine the mean of the mean 7210 to generate the mean 7310 of the aggregate predictions 7310. Combiner 400 may further determine the standard deviation of the mean 7310, for example by determining it directly from the prediction 7100 based on the standard deviation (and / or variance) 7220 and / or the mean 7210, and / or in any other suitable manner. Combiner 400 may also, or alternatively, determine the confidence interval 7320 of the prediction 450, for example in an implementation in which the prediction 450 includes the probability of synergistic (and / or antagonistic) interactions. Confidence interval 450 can be determined in any suitable manner, for example, by the propagation of uncertainty, and / or by the normal distribution of the mean 7310 of the combined predictions 7300 and by determining the standard deviation and / or confidence interval 7320 based on the standard deviation (and / or variance) 7220, and, if appropriate, by a critical value and / or confidence level (which can be, for example, predefined, user-provided, and / or otherwise obtained by the combiner 400). In some embodiments, system 1000 labels (i.e., identifies the user) low-confidence predictions (i.e., candidate pesticide compositions with predicted confidence levels below a threshold) for experimental validation. Regardless of whether system 1000 performs such labeling, in some embodiments, system 1000 is configured to retrain classifier 300 (via any suitable technique) on experimental results from such low-confidence predictions.

[0164] In some implementations, combiner 400 may generate aggregated prediction 7200 based on disjoint subsets of predictions 7100, for example, as described above, where each aggregated prediction 7200 is generated from predictions 7110 of different classifiers 310. In some implementations, combiner 400 generates predictions 7100 based on overlapping subsets of predictions 7100. For example, combiner 400 may generate aggregated predictions through convolution, such as generating a first aggregated prediction based on a subset of predictions 7110 of classifiers 310 having iteration exponents of 1 to m (for some m < M), and generating a second aggregated prediction based on predictions 7110 of classifiers 310 having the same iteration exponents of 2 to m+1.

[0165] Figure 7 The data flow of an exemplary embodiment of the combiner 400 is shown. The combiner receives M predictions 7100 (parameterized by corresponding trained parameters 320) from each classifier 310. The predictions 7100 can be represented as an N×M matrix, where M is the number of iterations performed by each trained classifier 310, and each trained classifier produces (potentially different) predictions 7100 with probabilities of, for example, cooperative interactions between candidate compounds and / or harmful organisms. N is the number of classifiers 310 configured to be used by the system 1000.

[0166] In at least this exemplary embodiment, the combiner 400 determines the mean and standard deviation (and / or variance) of the predictions 7100 for each iteration 1...M. This is in Figure 7 The vector is depicted as aggregate prediction 7200, and specifically as a vector of mean 7210 and standard deviation (and / or variance) 7220. Combiner 400 determines the mean on aggregate prediction 7200, and particularly the mean 7210, to generate a combined mean 7310 that includes the average probability of synergistic (and / or antagonistic) interactions. Combiner 400 optionally determines a confidence interval 7320 for the combined mean 7310, for example by performing uncertainty determination propagation on the standard deviation (and / or variance) 7220.

[0167] Further determination based on synergy prediction

[0168] In some embodiments, system 1000 generates prediction 450 by generating prediction 7300 as described above and providing prediction 7300 as prediction 450. In some embodiments (e.g., at least some of those embodiments without combiner 400), system 1000 generates prediction 450 by providing at least one of one or more predictions generated by classifier 300 (e.g., prediction 7100) as prediction 450. In some embodiments, system 1000 generates prediction 450 by further transforming one or more of predictions 7100, 7200, and / or 7300. Such further transformations may be performed by combiner 400 and / or post-processing module (not shown) of system 1000. In some embodiments, system 1000 generates multiple predictions 450, each prediction being any of the foregoing methods. For example, system 1000 can generate a first prediction 450 by providing prediction 7300, and can generate one or more other predictions 450 based on the first prediction 450, one or more previously generated other predictions 450, and / or one or more of predictions 7100, 7200, and / or 7300. For convenience, when discussing system 1000 generating prediction 450 based on the first prediction 450, one or more previously generated other predictions 450, and / or one or more of predictions 7100, 7200, and / or 7300, such predictions (based on which prediction 450 is generated) are collectively and individually referred to as the “original prediction”.

[0169] System 1000 can determine prediction 450 in any of a variety of ways. In some embodiments, system 1000 generates a discretized prediction (such as binary yes / no or category 1 / 2 / 3 / 4 / 5) based on one or more original predictions that are above or below one or more thresholds. For example, system 1000 can receive a threshold (e.g., from a parameter storage area) and compare the threshold with the original prediction. If the threshold is greater than (or, in some embodiments, not less than) the original prediction, system 1000 can generate a discretized prediction with a TRUE value; otherwise, system 1000 can generate a discretized prediction with a FALSE value.

[0170] In some embodiments, system 1000 generates prediction 450 based on one or more original predictions, which represents the predicted probability of synergistic (and / or antagonistic) interactions between compounds in the candidate pesticide composition and / or between one or more compounds in the candidate pesticide composition and one or more pests. Alternatively or additionally, system 1000 generates prediction 450 based on one or more original predictions, representing the predicted degree of such synergistic (and / or antagonistic) interaction. This degree of prediction may include a continuous value (e.g., floating-point) measure characterizing the predicted synergistic behavior of the candidate pesticide composition. This degree of prediction may include, for example, the magnitude of such measure, i.e., synergistic interaction (e.g., log2) determined by system 1000 based on the logarithm of the measure. In some implementations, system 1000 generates prediction 450, which represents a value of a known synergistic measure, such as the fractional inhibitory concentration index (FICI) and / or any other suitable measure, such as those disclosed by: Greco, WR, Bravo, G. & Parsons, JC (199). The search for synergy: a critical review from a response surface perspective. Pharmacological Reviews 47, 331–85.

[0171] In at least one example implementation, system 1000 generates a prediction 450 representing the predicted extent of a synergistic interaction, including a synergistic measure, and maps the magnitude to the outcome based on one or more discretization criteria. For example, the discretization criteria may include configured effect level bin thresholds and corresponding outcome values ​​(e.g., obtained from a parameter storage area). System 1000 may compare the obtained effect level bin thresholds with the magnitude values ​​and thereby determine which outcome value is mapped to the magnitude value. For example, exemplary effect level bin thresholds and corresponding outcome values ​​are shown in the table below.

[0172] Lower limit of measurement Upper limit of measurement result 0 2 none 2.01 4 weak 4.01 99.99 powerful

[0173] Based on the thresholds and result values ​​depicted in the table above, if the magnitude value is between 0 and 2, System 1000 maps the predicted level of cooperative interaction to "none". Similarly, if the magnitude value is greater than 2 and less than or equal to 4, System 1000 maps the predicted level of cooperative interaction to "weak", and if the magnitude value is greater than 4, System 1000 maps the predicted level of cooperative interaction to "strong". (Optionally, one or both of the top and bottom limits, i.e., limits 0 and 99.99, may alternatively be unbounded, such that any value less than 2 or greater than 4, respectively, will be mapped to bin by System 1000).

[0174] In some embodiments, system 1000 generates prediction 450, which includes a predictive measure of the effectiveness of a candidate pesticide composition against one or more pests. System 1000 can determine the predictive measure of effectiveness by determining the amount of the candidate pesticide composition that provides effectiveness in vitro, in plants, and / or in the field (e.g., the minimum amount predicted to be necessary). Determining effectiveness in a pesticide context may include determining that the predicted composition (e.g., in a given amount) inhibits and / or controls pest populations within a threshold, example thresholds include achieving at least 90% mortality of a bedbug population under laboratory conditions. (Different thresholds may be used, such as 80%, 95%, or even 100%.) System 1000 may further combine the amounts of the candidate pesticide composition having a per-unit resource allocation (e.g., per-unit cost) (e.g., as described above, such as by multiplication) to determine the predicted cost of the efficacy measure of the candidate pesticide composition.

[0175] System 1000 may output a representation of candidate pesticide compositions for which it generates prediction 450. This representation may include any of the representations of candidate pesticide compositions described elsewhere herein, and optionally include any of predictions 450, 7100, 7200, and / or 7300, and / or other information relating to the candidate pesticide compositions (collectively and individually referred to as the “output representation”). System 1000 may, for example, filter, sort, or otherwise modify the output representation of candidate pesticide compositions based on any of predictions 450, 7100, 7200, 7300, and / or other information relating to the candidate pesticide compositions.

[0176] For example, system 1000 can filter and / or rank candidate pesticide compositions based on the cost of the aforementioned efficacy metrics. System 1000 can identify, based on the predictive metrics of the corresponding effectiveness of the candidate pesticide compositions, a group of n candidate pesticide compositions with n lowest efficacy metric costs (for some values ​​n, which may be predetermined, provided by the user, and / or otherwise obtained), a group of candidate pesticide compositions with efficacy metric costs less than (or greater than) a threshold, and / or another group of one or more candidate pesticide compositions.

[0177] As another example, system 1000 can filter and / or rank candidate pesticide compositions based on the predicted probability and / or degree of synergistic (and / or antagonistic) interactions predicted by 450. For example, system 1000 can determine that the probability (and / or degree) of such interaction for a given candidate pesticide composition is less than (or greater than, not less than, or not greater than) a threshold and can remove the candidate pesticide compositions and related information from the output representation. System 1000 can alternatively or additionally rank the candidate pesticide compositions in the output representation by such probability (e.g., from highest to lowest probability) and / or degree. Thus, the output representation can be, for example, limited to candidate pesticide compositions that are predicted to exhibit a sufficient degree of synergy to guarantee further testing. (Sufficiency here can be defined by a threshold, which can be predetermined, provided by the user, and / or otherwise obtained.)

[0178] As an illustrative example, system 1000 can eliminate candidate pesticide compositions for which the corresponding prediction 450 indicates a probability of <20% synergistic (and / or antagonistic) interaction. System 1000 can then rank the remaining candidate pesticide compositions from highest to lowest probability. Alternatively or additionally, system 1000 can rank candidate pesticide compositions for which the corresponding prediction 450 indicates a probability of >80% synergistic (and / or antagonistic) interaction higher than that of other candidate pesticide compositions. It ranks higher those results with a probability of approximately >80% synergistic outcome.

[0179] In some implementations, system 1000 retrains parameter 320 by comparing prediction 450 with the results of laboratory and / or field tests and updating parameter 320 based on such comparisons (e.g., via active learning, online learning, and / or any other suitable technique). For example, system 1000 may update parameter 320 based on the difference between prediction 450 and test results to minimize (or maximize, as appropriate) the objective function. For example, system 1000 may perform gradient descent on the objective function based on test results.

[0180] Computer System

[0181] Figure 8An exemplary computer system providing system 1000 is illustrated. Each exemplary computer 500 includes one or more processors 510a, ..., 510n (collectively and individually referred to as processor 510), such as a general-purpose CPU and / or a special-purpose processor, such as an FPGA or GPU, which are operatively connected to persistent memory 530 and / or transient memory 540. These memories store information processed by system 1000 and may store executable instructions (collectively referred to herein as “programs”) that can be executed by processor 510 for performing the methods described herein (e.g., programs 8200, 8210, 8300, 8400, with reference numerals incrementing to 8000, performing actions associated with similar elements of system 1000). Programs are described in more detail below. In some cases, such as with an FPGA, programs include configuration information for adjusting processor 510 for a specific purpose. One or more processors 510 may be operatively connected to a network and communication interface 550 adapted for deployment configuration. The permanent storage 530 of computer 500 may be one or more databases 250 for storing information collected and / or calculated by a server and read, processed, and written by processor 510 under the control of a program (e.g., 8200, 8210, 8300, 8400). Computer 500 may also, or alternatively, be operatively connected to an external database 570 via a network and communication interface 550.

[0182] Persistent memory 530 may include disks, PROMs, EEPROMs, flash memory, and similar technologies, characterized by its ability to retain its contents between power-on and power-off cycles of the computer 500. Some persistent memory 530 may take the form of a file system for the computer 500 and may be used to store control and operating programs and information defining the manner of operation of the computer 500, including background and foreground processes and scheduling of periodically executed processes. Persistent memory 530 in the form of a network-attached storage device (NAS) (a storage device accessible via a network interface) may also be used, or alternatively, without departing from the scope of this disclosure. Temporary memory 540 may include random access memory (RAM) and similar technologies, characterized by its stored contents not being retained between power-on and power-off cycles of the system.

[0183] One or more databases 250, 570 may include local file storage devices, wherein the file system includes data storage and indexing schemes, relational databases, object-oriented databases, object-relational databases, NoSQL databases and / or other database structures, such as indexed record structures. Such databases 250 and / or 570 may be stored within a single persistent storage device 530, may be stored on one or more persistent storage devices 530, and / or may be stored in persistent storage devices 530 on different computers.

[0184] For clarity, system 1000 is described using multiple logical databases. System 1000 can be deployed using one or more physical databases implemented on one or more computers 500 and / or on a virtualized computer system, and / or can be implemented using clustering techniques (e.g., such that at least a portion of the data stored in the database is physically stored on two or more computers 500). In some implementations, one or more logical and / or physical databases can be implemented on remote devices and accessed via a communication network.

[0185] System 1000 further includes several programs as described above (for example, the above modules may be provided by programs of one or more computers 500).

[0186] Prediction of pesticide compositions and experimental evaluation of formulations and their uses

[0187] Once the prediction 450 is determined, the results of the prediction can be used in any desired way. For example, in Figure 9 In one example method 9000 shown, prediction 450 for one or more pests in a test environment, such as in vivo or in plants, can be performed by formulating a composition containing a candidate pest control composition (e.g., by combining a pest control compound, a synergistic compound, and any desired formulation components such as a solvent, carrier, adjuvant, stabilizer, etc.) at 9010 and exposing one or more pests to the composition at 9020. At 9030, the efficacy of the composition as a pest control is determined (e.g., by assessing the percentage of pest mortality and / or by assessing the time taken to reach peak mortality in controlling or killing one or more pests).

[0188] As another example, in Figure 10In method 9100 shown, prediction 450 can be used to formulate a pesticide composition. At 9110, it is determined whether prediction 450 meets or exceeds a predetermined probability level of synergistic interaction, for example, to determine whether there is a candidate pesticide composition containing a pesticide compound and a synergistic compound that exhibits a high probability of synergistic interaction against one or more pests. If prediction 450 meets or exceeds the predetermined probability level of synergistic interaction, then at 9120, a pesticide composition containing a pesticide compound and a synergistic compound, as well as any desired formulation components, such as solvents, carriers, adjuvants, stabilizers, etc., is formulated.

[0189] As another example, in Figure 11 In method 9200 shown, prediction 450 can be used to manufacture a pesticide composition. At 9210, multiple predictions 450 of synergistic interactions between multiple pesticide compounds and multiple synergistic compounds are determined. Each prediction 450 corresponds to a proposed candidate pesticide composition containing at least one pesticide compound and at least one synergistic compound. At 9220, the multiple predictions are evaluated and a proposed candidate pesticide composition is selected based on the desired characteristics of the predictions 450. For example, at 9220, a proposed candidate pesticide composition having a prediction 450 that satisfies or exceeds a predetermined probability level of synergistic interaction can be selected. Alternatively, at 9220, a proposed candidate pesticide composition having a prediction 450 that is higher than that of other proposed candidate pesticide compositions can be selected. At step 9230, the candidate pesticide composition selected at 9220 is generated, for example, by mixing the pesticide compounds and synergistic compounds constituting the candidate pesticide composition with any desired formulation component (such as a solvent, carrier, adjuvant, stabilizer, etc.).

[0190] As another example, in Figure 12 In method 9300 shown, prediction 450 can be used to treat one or more pests affecting a non-target organism. At 9310, it is determined whether prediction 450 meets or exceeds a predetermined probability level of synergistic interaction, for example, to determine whether there is a candidate pesticide composition containing a pesticide compound and a synergistic compound with a high probability of resisting one or more pests exhibiting synergistic interaction. If prediction 450 meets or exceeds the predetermined probability level of synergistic interaction, then at 9320, the non-target organism can be exposed to a pesticide composition containing a candidate pesticide composition. This will result in the exposure of one or more pests affecting the non-target organism to the pesticide composition to improve or eliminate the adverse effects that one or more pests may have on the non-target organism.

[0191] As another example, in Figure 13In the method 9400 shown, prediction 450 can be used to treat one or more pests affecting non-target organisms. At 9410, multiple predictions 450 of synergistic interactions between multiple pesticide compounds and multiple synergistic compounds are determined. Each prediction 450 corresponds to a proposed candidate pesticide composition containing at least one pesticide compound and at least one synergistic compound. At 9420, the multiple predictions are evaluated and a proposed candidate pesticide composition is selected based on the desired characteristics of the predictions 450. For example, at 9420, a proposed candidate pesticide composition having a prediction 450 that satisfies or exceeds a predetermined probability level of synergistic interaction can be selected. Alternatively, at 9420, a proposed candidate pesticide composition having a prediction 450 that is higher than that of other proposed candidate pesticide compositions can be selected. At step 9430, the non-target organism is exposed to the pesticide composition containing the candidate pesticide composition selected at 9420. This will result in the exposure of one or more pests that affect non-target organisms to the pesticide composition in order to improve or eliminate the adverse effects that one or more pests may have on non-target organisms.

[0192] Example Results

[0193] System 1000 is implemented to generate predictions of the probability of the presence of synergistic interactions between pairs of compounds in a set of candidate pesticide compositions. For each prediction, System 1000 receives representations of the active pesticide compound and potential synergistic compounds. These compound representations are received as SMILES strings and augmented via QSAR to generate feature vectors. (In some tests, the augmented representations include graphical representations of the compounds.) This implementation of System 1000 takes into account the selection of features including aromaticity, electronegativity, polarity, hydrophilicity / hydrophobicity, and hybridization. System 1000 includes three classifiers 310, each trained on the synergistic efficacy of pesticide compositions when applied to different pests; no pest information is provided to classifiers 310 at inference time. The encoder is trained on a general chemistry dataset, namely Tox21. This implementation does not receive information about mixture ratios.

[0194] Laboratory experiments, including in vitro tests on each predicted pest treated with a candidate pesticide composition (containing a pesticide compound and a potential synergistic compound), were conducted to evaluate the accuracy of the predictions generated by a specific test implementation of System 1000. Accuracy was assessed by determining the change in the minimum inhibitory concentration (MIC) observed in the resistance of each candidate pesticide composition to the corresponding pest relative to a pesticide compound without a potential synergistic compound. (Specific test implementations include those based on…) Figure 3(Exemplary implementation of the integrated classifier 300 and combiner 400.)

[0195] The tests covered six pesticide active compounds and three fungal pests. Each pesticide compound was selected from a class known to have pesticide action against at least one of the three pests. They are identified hereinafter as compound AF, and the pests are identified hereinafter as pest AC.

[0196] Potential synergistic compounds are selected from the group consisting of: C4-C10 unsaturated fatty acids: 10-hydroxydecanoic acid, 12-hydroxydodecanoic acid, 2,2-diethylbutanoic acid, 2-aminobutyric acid, 2-aminohexanoic acid, 2-ethylhexanoic acid, 2-hydroxybutyric acid, 2-hydroxyoctanoic acid, 2-methyldecanoic acid, 2-methyloctanoic acid, 3-aminobutyric acid, 3-decenoic acid, 3-heptenoic acid, 3-hydroxybutyric acid, 3-hydroxyhexanoic acid, 3-hydroxyoctanoic acid, 3-methylbutyric acid, 3-methylnonanoic acid, 3-nonenoic acid, 3-octenic acid, 4-hexenoic acid, 4-methylhexanoic acid, 5-hexenoic acid, 7-octenic acid, 8-hydroxyoctanoic acid, 9-decenoic acid, decanoic acid, dodecanoic acid, heptanoic acid, nonanoic acid, octanoic acid, oleic acid, sorbic acid, trans-2-nonenoic acid, trans-2-octenic acid, trans-2-undecenoic acid, trans-3-hexenoic acid.

[0197] The test implementation of System 1000 generates predictions of the probability of the presence of synergistic interactions between compounds in each candidate pest control composition against each selected pest. As described above, the predictions of System 1000 are discretized such that probabilities less than or equal to 0.5 (i.e., 50%) are mapped to 0 (indicating no predicted synergistic effect), and probabilities greater than 0.5 are mapped to 1 (indicating predicted synergistic effect). The binarized results are presented in Table 1 under the “Prediction” column. In Table 1, the values ​​in the prediction column are the discretized predictions of System 1000. The values ​​in the “Observation” column are the results observed in the aforementioned laboratory experiments, representing the degree of synergistic effect (in this case, inverse FICI). For example, a value of 4 indicates that the observed FICI value is 1 / 4. Values ​​greater than 1 indicate synergism.

[0198] Table 1: Results of the predictive test on pairwise synergistic effects of selected pest organisms .

[0199]

[0200]

[0201]

[0202]

[0203]

[0204]

[0205]

[0206]

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214]

[0215]

[0216]

[0217]

[0218]

[0219] Overall, the results of these tests indicate that, in at least some cases, the systems and methods described in this paper are comparable to those of experienced human chemists in terms of predictive accuracy.

[0220] in conclusion

[0221] While many exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain modifications, arrangements, additions, and sub-combinations thereof. Therefore, it is intended that the appended claims and the claims introduced below be construed as including all such modifications, arrangements, additions, and sub-combinations within their true substance and scope.

Claims

1. A method for generating a prediction of a synergistic interaction between two or more compounds against one or more pests, the method performed by one or more processors and comprising: receiving a first representation of a pesticide compound; receiving a second representation of a synergistic compound; generating an encoded representation of a composition comprising the pesticide compound and the synergistic compound by encoding first chemical features of the pesticide compound and second chemical features of the synergistic compound based on the respective first and second representations; and generating one or more predictions of a synergistic interaction between the pesticide compound and the synergistic compound against one or more pests, the generating comprising: transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition against at least one training pest.

2. The method of claim 1, wherein the one or more predictions of synergistic interaction comprise a plurality of predictions, and the method further comprises: combining a plurality of synergistic interaction predictions into a combined synergistic interaction.

3. The method of claim 2, wherein the method further comprises determining, based on the plurality of predictions, at least one of: a confidence interval, a standard deviation, and a variance.

4. The method of claim 3, wherein the classifier comprises a random classifier, the generating the one or more predictions comprises transforming the encoded representation based on the trained parameters of the classifier over a plurality of iterations, and generating a prediction for each iteration.

5. The method of any one of claims 1-4, wherein the generating the encoded representation comprises generating a first encoded compound representation based on the first chemical features of the pesticide compound and a second encoded compound representation based on the second chemical features of the synergistic compound, and wherein the generating the one or more predictions comprises generating the one or more predictions based on the first and second encoded compound representations.

6. The method of any one of claims 1-4, wherein the generating the encoded representation comprises generating the encoded representation to be of lower dimensionality than at least one of the first and second representations.

7. The method of any one of claims 1-4, wherein the generating the encoded representation comprises transforming the first and second chemical features of the respective pesticide and synergistic compounds into the encoded representation based on trained parameters of an encoder model.

8. The method of claim 7, wherein the encoder model comprises an encoder portion of a variational autoencoder operable to transform the first and second chemical features from an input space of the variational autoencoder to a latent space.

9. The method of claim 7, wherein the trained parameters of the encoder model have been trained on a different training set than the trained parameters of the classifier.

10. The method of any one of claims 1-4, further comprising selecting the classifier from a plurality of classifiers based on the one or more pests.

11. The method of claim 10, further comprising receiving a representation of the one or more pests, and selecting the classifier comprises selecting the classifier based on the representation of the one or more pests.

12. The method of claim 10, wherein the classifier is a first classifier of a plurality of classifiers, at least a second classifier of the plurality of classifiers having been trained for pests different from the one or more pests, and selecting the classifier from the plurality of classifiers comprises selecting one of the first classifier and the second classifier based on the one or more pests.

13. The method of claim 10, wherein the classifier comprises an ensemble classifier comprising a plurality of constituent classifiers, the plurality of constituent classifiers comprising at least a first constituent classifier and a second constituent classifier, respective trained parameters of the first constituent classifier and the second constituent classifier each having been trained for at least one synergistic interaction between compounds of at least one composition resistant to at least one of the one or more pests.

14. The method of claim 13, wherein generating one or more predictions comprises generating a first prediction based on the first constituent classifier, and generating a second prediction based on the second constituent classifier.

15. The method of any one of claims 1-4, comprising generating an enhanced representation of at least one of the pest control compound and the synergistic compound, the enhanced representation comprising an enhanced chemical feature of the at least one of the pest control compound and the synergistic compound, the enhanced chemical feature not being included by the first representation and the second representation.

16. The method of claim 15, wherein generating the enhanced representation comprises determining the enhanced chemical feature based on trained parameters of a quantitative structure-activity relationship model.

17. The method of any one of claims 1-4, comprising receiving a third representation of a third compound, and excluding an excluded composition comprising the third compound from predictions based on determining at least one of: a chemical feature of the third compound matching an exclusion rule, an availability value corresponding to the third compound being less than a threshold value, a similarity measure between the third compound and a fourth compound being greater than a threshold value, and a toxicity indication of the third compound matching a toxicity criterion.

18. The method of any one of claims 1-4, wherein the pest control compound is selected from the group consisting of a fungicide, a herbicide, a nematicide, an insecticide, a bactericide, a rodenticide, a virucide, an acaricide, an algicide, and a molluscicide.

19. The method of any one of claims 1 to 4, comprising selecting at least one of the first chemical signature and the second chemical signature from the group consisting of: a representation of aromaticity, a representation of electronegativity, a representation of polarity, a representation of hydrophilicity / hydrophobicity, and a representation of hybridization of at least one of the pesticide compound and the synergist compound.

20. The method of any one of claims 1 to 4, wherein the one or more pests comprises the at least one training pest, such that transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition that is resistant to at least one training pest comprises transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition that is resistant to at least one of the one or more pests.

21. The method of any one of claims 1 to 4, wherein the at least one training pest shares a pesticide mode of action with at least one of the one or more pests, such that transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition that is resistant to at least one training pest comprises transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition that is resistant to at least one training pest that shares a pesticide mode of action with at least one of the one or more pests.

22. The method of any one of claims 1 to 4, wherein the trained parameters of the classifier have been trained by: determining an importance measure for each training composition of a plurality of training compositions; selecting one or more high-importance compositions from the plurality of training compositions based on the importance measure of each high-importance composition of the one or more high-importance compositions; and updating the trained parameters of the classifier based on the one or more high-importance compositions.

23. The method of claim 22, wherein determining an importance measure for a given composition comprises determining the importance measure for the given training composition based on a variance of one or more training predictions of synergistic interactions between pesticide compounds of the training composition and synergist compounds of the training composition.

24. The method of claim 22, wherein selecting one or more high-importance compositions comprises selecting the one or more high-importance compositions based on a representative criterion. ​ 25. The method of claim 24, wherein selecting the one or more high- importance compositions based on a representative standard comprises determining a plurality of clusters of the plurality of training compositions, and selecting at least one high-importance composition from each of at least two of the plurality of clusters.

26. The method of claim 25, wherein determining the plurality of clusters of the plurality of training compositions comprises determining a graph similarity measure between at least one graph representing at least one compound of a first training composition of the training compositions and at least one graph representing at least one compound of a second training composition of the training compositions.

27. A computer system, the computer system comprising: one or more processors; and a memory storing instructions that cause the one or more processors to perform operations comprising: receiving a first representation of a pesticide compound; receiving a second representation of a synergistic compound; generating an encoded representation of a composition comprising the pesticide compound and the synergistic compound by encoding a first chemical feature of the pesticide compound and a second chemical feature of the synergistic compound based on the respective first and second representations; and generating one or more predictions of a synergistic interaction between the pesticide compound and the synergistic compound against one or more pests, the generating comprising: transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition against at least one training pest.

28. The computer system of claim 27, wherein the operations further comprise performing the method of any one of claims 2-26.

29. A non-transitory machine-readable medium storing instructions that cause one or more processors to perform operations comprising: receiving a first representation of a pesticide compound; receiving a second representation of a synergistic compound; generating an encoded representation of a composition comprising the pesticide compound and the synergistic compound by encoding a first chemical feature of the pesticide compound and a second chemical feature of the synergistic compound based on the respective first and second representations; and generating one or more predictions of a synergistic interaction between the pesticide compound and the synergistic compound against one or more pests, the generating comprising: transforming the encoded representation based on trained parameters of a classifier that have been trained for at least one synergistic interaction between compounds of at least one composition against at least one training pest.

30. The non-transitory machine-readable medium of claim 29, wherein the operations further comprise performing the method of any one of claims 2-26.

31. A method of assessing a prediction of a synergistic interaction between two or more compounds against one or more pests, the method comprising: ​ ​ determining a prediction of a synergistic interaction between a pesticidal compound and a synergistic compound according to the method of any one of claims 1 to 26; combining the pesticidal compound and the synergistic compound to produce a composition; exposing the one or more pests to the composition in a test environment; and evaluating the efficacy of the composition as a pesticide.

32. A method of formulating a pesticidal composition, the method comprising: determining a prediction of a synergistic interaction between a pesticidal compound and a synergistic compound against one or more pests according to the method of any one of claims 1 to 26; determining that the prediction of the synergistic interaction meets or exceeds a predetermined probability level; and formulating the pesticidal composition containing the pesticidal compound and the synergistic compound.

33. A method of formulating a pesticidal composition, the method comprising: determining a plurality of predictions of synergistic interactions between a pesticidal compound and a synergistic compound according to the method of any one of claims 1 to 26, each prediction of the plurality of predictions corresponding to a combination of one of a plurality of pesticidal compounds and a corresponding synergistic compound of a plurality of synergistic compounds; evaluating the plurality of predictions to select a combination of one of the plurality of pesticidal compounds and the corresponding synergistic compound of the plurality of synergistic compounds having (i) a probability that meets or exceeds a predetermined probability level or (ii) a probability of a synergistic interaction that is higher than at least some of other combinations of pesticidal compounds and synergistic compounds; and mixing the selected combination of the one of the plurality of pesticidal compounds and the corresponding synergistic compound of the plurality of synergistic compounds to generate the pesticidal composition.

34. A method of treating one or more pests affecting a non-target organism, the non-target organism being a crop plant, the method comprising: determining a prediction of a synergistic interaction between a pesticidal compound and a synergistic compound against the one or more pests according to the method of any one of claims 1 to 26; determining that the prediction of the synergistic interaction meets or exceeds a predetermined probability level; and exposing the non-target organism to a pesticidal composition containing the pesticidal compound and the synergistic compound.

35. A method of treating one or more pests affecting a non-target organism, the non-target organism being a crop plant, the method comprising: determining a plurality of predictions of synergistic interactions between a pesticidal compound and a synergistic compound according to the method of any one of claims 1 to 26, each prediction of the plurality of predictions corresponding to a combination of one of a plurality of pesticidal compounds and a corresponding synergistic compound of a plurality of synergistic compounds; evaluating the plurality of predictions to select a combination of one of the plurality of pesticidal compounds and the corresponding synergistic compound of the plurality of synergistic compounds having (i) a probability that meets or exceeds a predetermined probability level or (ii) a probability of a synergistic interaction that is higher than at least some of other combinations of pesticidal compounds and synergistic compounds; and mixing the selected combination of the one of the plurality of pesticidal compounds and the corresponding synergistic compound of the plurality of synergistic compounds to generate the pesticidal composition. evaluating the plurality of predictions to select a combination of one of the plurality of pesticide compounds and a corresponding one of the plurality of synergist compounds, the plurality of predictions having (i) a probability that meets or exceeds a predetermined probability level or (ii) a probability of synergistic interaction that is higher than at least some of the other combinations of pesticide compounds and synergist compounds; and exposing the non-target organism to a pesticide composition containing the selected combination of the one of the plurality of pesticide compounds and the corresponding one of the plurality of synergist compounds.

Images