Systems and methods for generating agricultural actions and compositions using generative artificial intelligence

Abstract

Systems and methods are provided for generating agricultural actions and compositions using generative artificial intelligence (AI). A trained generative AI model receives a starting effector and an outcome objective and proposes one or more effectors predicted to satisfy the objective. A reward model evaluates the proposed effectors and generates a reward value based on how well each effector meets the outcome objective. Reinforcement learning is used to update the generative AI model, improving its capacity to generate effectors aligned with desired outcomes. The system may include modules for context embedding, ranking, and experimental validation, and may operate across same-crop or cross-crop contexts. Applications include, but are not limited to, gene expression modulation, phenotype optimization, and trait development in agricultural organisms.

Description

Docket # 212601-WO-SEC-lSystems and Methods for Generating Agricultural Actions and Compositions Using Generative Artificial IntelligenceCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to and the benefit of U.S. Provisional Patent Application No. 63 / 682,758, filed on August 13, 2024, and U.S. Provisional Patent Application No. 63 / 748,693, filed on January 23, 2025. The contents of each of the foregoing applications are incorporated herein by reference in their entirety.FIELD OF THE INVENTION

[0002] This disclosure relates to the field of artificial intelligence and agricultural informatics, particularly to systems and methods for generating modifications and agricultural compositions using generative artificial intelligence (generative Al) .BACKGROUND

[0003] Global agriculture faces mounting challenges as it strives to meet rising demand amid increasingly constrained resources. Limited access to arable land, water, nutrients, and labor continues to pressure production systems across diverse environments. Simultaneously, rapid advancements in computational infrastructure, data analytics, and biological research are reshaping how complex problems are approached.SUMMARY

[0004] Provided herein is a method of creating one or more proposed effectors, the method including inputting, into a trained generative Al model, a starting effector and an outcome objective, where the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective; proposing, by the trained generative Al model, one or more proposed effectors for each starting effector and outcome objective; receiving, by a trained reward model, one or more entries, each entry including the starting effector, the outcome objective, and one of the one or more proposed effectors; generating, by the reward model, a reward value that predicts how well the one of the one or more proposed effectors meets the outcome objective; ranking the one or more proposed effectors based on the generated rewardDocket # 212601-WO-SEC-l values for each; selecting one or more proposed effectors based on their ranking; and creating the selected one or more proposed effectors.

[0005] Provided herein is a computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations including inputting a starting effector and an outcome objective into a trained generative Al model, where the model has been trained to propose one or more effectors predicted to satisfy the outcome objective; proposing one or more effectors for each starting effector and outcome objective; receiving entries including the starting effector, the outcome objective, and one of the proposed effectors; generating a reward value to predict how well the proposed effector meets the outcome objective; ranking the proposed effectors based on the reward values; selecting one or more effectors based on their ranking; and creating the selected effectors.

[0006] Provided herein is a computer system for creating one or more proposed effectors, the system including one or more servers, where one of the servers includes a starting effector and an outcome objective, and a computing device communicatively coupled to the one or more servers, the computing device including a memory and one or more processors configured to perform operations including inputting the starting effector and the outcome objective into a trained generative Al model, where the model has been trained to propose one or more effectors predicted to satisfy the outcome objective; proposing one or more effectors for each starting effector and outcome objective; receiving entries including the starting effector, the outcome objective, and one of the proposed effectors; generating a reward value to predict how well the proposed effector meets the outcome objective; ranking the proposed effectors based on the reward values; and selecting and creating the proposed effectors based on their ranking.

[0007] Provided herein is a method of using a generative Al model to alter a phenotype in a target plant, the method including inputting, into a trained generative Al model, a representation of a genotypic profile of a first parent of the target plant, a representation of a genotypic profile of a second parent of the target plant, an edit library including a list of all possible genome edits that may be introduced into the first and / or second parental genomes, and an outcome objective, where the trained generative Al model has beenDocket # 212601-WO-SEC-l trained to propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective; proposing, by the trained generative Al model, one or more combinations of genome edits; generating one or more populations of plants, each population generated by multiplex editing using one of the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants including members with varying combinations of target edits in an otherwise uniform genetic background; receiving, by a trained reward model, one or more entries, each entry including the representation of the first parental genotype and the second parental genotype, the outcome objective, and genotypic data for the members of the population; generating, by the reward model, a reward value to predict how well the genome edit or combination of genome edits meets the outcome objective; and using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.

[0008] Provided herein is a computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations including inputting genotypic profiles of a first and second parent, an edit library, and an outcome objective into a trained generative Al model, where the model has been trained to propose genome edits predicted to satisfy the outcome objective; proposing combinations of genome edits; generating populations of plants with varying combinations of target edits; receiving entries including parental genotypes, the outcome objective, and genotypic data for the plant population; generating a reward value to predict how well the genome edits meet the outcome objective; and adjusting weights in the generative Al model using reinforcement learning to improve its capacity to generate genome edits that satisfy the outcome objective.

[0009] Provided herein is a computer system for generating genome edits to alter a phenotype in a target plant, the system including one or more servers, where one of the servers includes a representation of a genotypic profile of a first parent of the target plant, a representation of a genotypic profile of a second parent of the target plant, an edit library including genome edits, and an outcome objective, and a computing device communicatively coupled to the one or more servers, the computing device including aDocket # 212601-WO-SEC-l memory and one or more processors configured to perform operations including inputting the genotypic profiles, the edit library, and the outcome objective into a trained generative Al model, where the model has been trained to propose genome edits predicted to satisfy the outcome objective; proposing combinations of genome edits; generating populations of plants with varying combinations of target edits; receiving entries including parental genotypes, the outcome objective, and genotypic data for the plant population; generating a reward value to predict how well the genome edits meet the outcome objective; and adjusting weights in the generative Al model using reinforcement learning to improve its capacity to generate genome edits that satisfy the outcome objective.

[0010] Provided herein is a method of using a generative Al model for plant breeding, the method including inputting, into a trained generative Al model, a parental genotypic population including genotypic information for a plurality of parental plants, an edit library including a list of all possible genome edits that may be introduced into parental plants of the parental genotypic population, and an outcome objective, where the trained generative Al model has been trained to select a first and second parental plant from the genotypic population and propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective; proposing, by the trained generative Al model, a first and second parental plant from the genotypic population and one or more combinations of genome edits; generating one or more edited populations of plants, each population generated by single-site editing, multiplex editing, or a combination thereof using one of the selected first and second parental plants and the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants including members with varying combinations of target edits in an otherwise uniform genetic background; receiving, by a breeding pipeline digital twin, one or more entries, each entry including the representation of the first parental genotype and the second parental genotype, the outcome objective, and genotypic data for the members of the edited population; simulating, by the breeding pipeline digital twin, breeding outcomes over the course of three or more generations; generating, by the breeding pipeline digital twin, a reward value to predict how well the genome edit or combination of genome edits meets theDocket # 212601-WO-SEC-l outcome objective; and using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.

[0011] Provided herein is a computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations including inputting a parental genotypic population, an edit library, and an outcome objective into a trained generative Al model, where the model has been trained to select parental plants and propose genome edits predicted to satisfy the outcome objective; proposing combinations of genome edits; generating edited plant populations; receiving entries including parental genotypes, the outcome objective, and genotypic data; simulating breeding outcomes; generating a reward value to predict how well the genome edits meet the outcome objective; and adjusting weights in the generative Al model using reinforcement learning to improve its capacity to generate genome edits that satisfy the outcome objective.

[0012] Provided herein is a computer system for using a generative Al model for plant breeding, the system including one or more servers, where one of the servers includes a parental genotypic population including genotypic information for a plurality of parental plants, an edit library including genome edits, and an outcome objective, and a computing device communicatively coupled to the one or more servers, the computing device including a memory and one or more processors configured to perform operations including inputting the genotypic population, the edit library, and the outcome objective into a trained generative Al model, where the model has been trained to select parental plants and propose genome edits predicted to satisfy the outcome objective; proposing combinations of genome edits; generating edited plant populations; receiving entries including parental genotypes, the outcome objective, and genotypic data; simulating breeding outcomes; generating a reward value to predict how well the genome edits meet the outcome objective; and adjusting weights in the generative Al model using reinforcement learning to improve its capacity to generate genome edits that satisfy the outcome objective.Docket # 212601-WO-SEC-l

[0013] Provided herein is a method of using a generative Al model to optimize agricultural product performance and profitability, the method including inputting, into a trained generative Al model, a set of parental genotypic representations, an edit library including genome edits applicable to the parental genotypes, and an outcome objective including both agronomic performance metrics and economic performance metrics, where the trained generative Al model has been trained to propose combinations of parental selections and genome edits predicted to satisfy the outcome objective; proposing, by the trained generative Al model, one or more tuples including a first parental genotype, a second parental genotype, a set of genome edits for the first parent, and a set of genome edits for the second parent; generating, based on the proposed tuples, one or more edited populations of plants, each population including members with varying combinations of target edits in an otherwise uniform genetic background; receiving, by a joint simulation system including a breeding pipeline digital twin and a seed-market digital twin, one or more entries, each entry including the proposed parental genotypes and genome edits, the outcome objective, and genotypic and phenotypic data for the edited plant population; simulating, by the breeding pipeline digital twin, multi-generational breeding outcomes, and simulating, by the seed-market digital twin, commercial lifecycle outcomes including regional adoption, pricing strategy, promotional spend, competitor product launches, and projected profitability; generating, by the joint simulation system, a reward value based on an additive combination of agronomic performance and economic performance; and using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to improve its capacity to generate parental selections and genome edit combinations that maximize both agronomic and economic outcomes.

[0014] Provided herein is a computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations including inputting parental genotypic representations, an edit library, and an outcome objective into a trained generative Al model, where the model has been trained to propose genome edits and parental selections predicted to satisfy the outcome objective; proposing tuples of parental genotypes and genome edits; generating edited plant populations; receiving entries including parental genotypes, genome edits, theDocket # 212601-WO-SEC-l outcome objective, and genotypic and phenotypic data; simulating breeding and commercial lifecycle outcomes; generating a reward value based on agronomic and economic performance; and adjusting weights in the generative Al model using reinforcement learning to improve its capacity to generate genome edits and parental selections that maximize agricultural product performance and profitability.

[0015] Provided herein is a computer system for optimizing agricultural product performance and profitability using a generative Al model, the system including one or more servers, where one of the servers includes parental genotypic representations, an edit library including genome edits applicable to the parental genotypes, and an outcome objective including agronomic and economic performance metrics, and a computing device communicatively coupled to the one or more servers, the computing device including a memory and one or more processors configured to perform operations including inputting the parental genotypic representations, the edit library, and the outcome objective into a trained generative Al model, where the model has been trained to propose genome edits and parental selections predicted to satisfy the outcome objective; proposing tuples of parental genotypes and genome edits; generating edited plant populations; receiving entries including parental genotypes, genome edits, the outcome objective, and genotypic and phenotypic data; simulating breeding and commercial lifecycle outcomes; generating a reward value based on agronomic and economic performance; and adjusting weights in the generative Al model using reinforcement learning to improve its capacity to generate genome edits and parental selections that maximize agricultural product performance and profitability.BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The disclosure can be more fully understood from the following detailed description and the accompanying drawings, which form a part of this application.

[0017] FIG. 1 shows a schematic representation of reinforcement learning with a laboratory feedback loop according to an embodiment. First the outcome objective is defined, along with the starting effector(s). For applications requiring additional external context from sources such as public literature, public databases, or proprietary data, an external context large language model (LLM) enabled by retrieval augmentedDocket # 212601-WO-SEC-l generation (RAG) can be implemented as optional input into the generative model. The generative model then designs proposed effectors which are modified versions of the starting effectors with the aim of meeting the set outcome objective. Proposed effectors are then moved to the laboratory for synthesis and testing to produce outcome data. The outcome data is used to train a reward model via supervised learning that enables further iterations of the method loop. Then after training, the reward model outputs are input into a reinforcement learning algorithm. The reinforcement learning outputs are then input into the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory results meet the outcome objective, modified plants, plant tissues, or plant cells are used for commercial product development.

[0018] FIG. 2 is an schematic illustrating the application of the method and system as described in Example 1 . Here, the objective is to increase the transcription level 2-fold over current levels for a variety of promoters. The starting promoter sequences include several hundred promoters from a pan-genome along with corresponding expression values. The generative model outputs proposed effectors comprising promoter sequences that are sent to the lab to be synthesized, transformed into protoplasts, and tested for expression using RNA-sequencing. The expression value outputs are used to train the reward model using supervised learning. Then, the trained reward model outputs reward values which are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory results meet the outcome objective, modified plants or plant cells are used for commercial product development.

[0019] FIG. 3 is a schematic illustrating the inputs to the external context module, the inputs to the generative model, and the outputs of the generative model for Examples 1 - 4. The inputs for the optional external context module, a Retrieval Augmented Generation (RAG)-enabled Large Language Model (LLM), are first pre-processed for uniform embeddings and can include published text, public databases, and proprietary data along with the outcome objective and starting effector(s). The output of the external context module can then be used as input, in addition to the outcome objective andDocket # 212601-WO-SEC-l starting effector(s) into the generative model, a Deep Neural Network (DNN) consisting of an optional initializing base model for generating the initial effectors and a Reinforcement Learning (RL)-updated model for generating effectors in subsequent iterations of the method loop. The output proposed effectors can be modified promoter sequences as in Examples 1-4.

[0020] FIG. 4 is a schematic illustrating the laboratory feedback process for Examples 1 - 3. The input proposed effectors comprising promoter sequences are sent to the laboratory where transgenic constructs are designed and generated. In Example 1 , leaf protoplast cells are transformed with a library of synthesized promoters upstream of barcoded reporters, and synthesized promoter sequences are directly used as the modified promoter sequence output. In Examples 2-3, plants are edited using CRISPR and whole plants are regenerated for screening. Edited plants are sequenced to determine the modified promoter sequence output. Expression profiling on protoplasts (1 ), multiple target tissues (2), or a single target tissue (3) may be performed using RNA sequencing, quantitative PCR, or similar methods to create the expression outputs. In Example 3, plant phenotypes are collected at maturity and used for the phenotype output. The outputs of modified promoter sequences, expression, and phenotype are then used for reward model training.

[0021] FIG. 5 is a schematic describing the use of laboratory data to train the reward model for reinforcement learning according to an embodiment. During training, the reward model accepts as input the outcome objective, the starting effector, and the modified promoter sequence obtained in the lab. Additionally, the training data includes a target signal calculated from the laboratory data outputs of expression and / or phenotype values. The reward model is then trained with supervised learning to output the reward value, which is calculated based on the projected difference between the outcome objective and observed results. The reward value, along with the outcome objective, starting effector, and modified promoter sequences, is used by the reinforcement learning algorithm enabled with a value function to output optimized weights to update the generative model.

[0022] FIG. 6 is a schematic illustrating a reward model and reinforcement learning algorithm steps of the methods and systems described herein according to variousDocket # 212601-WO-SEC-l embodiments. The inputs to the reward model include the outcome objective, the starting effector sequence, and the modified promoter sequence. During training, the reward model also receives as input a target signal calculated from the laboratory data including expression and / or phenotype data. Once trained, the reward model outputs the reward value, which is calculated based on the projected difference between the objective and observed result. The reinforcement learning algorithm receives the reward value as input along with the outcome objective, starting effector, and modified promoter sequence. Within the reinforcement learning algorithm, the value function provides baselines for the value of generative actions, thereby reducing the variance of the reinforcement learning training. The reinforcement learning algorithm produces optimized weights to update the generative model, which then proposes additional rounds of proposed effectors. The proposed effectors can either be used to A. generate laboratory data to further train and update the reward model training, or B. the proposed effectors can be input directly into the reward model as hypothetical promoters for additional rounds of reinforcement learning.

[0023] FIG. 7 is a schematic of an example showing how the trained reward model can be used for rapid reinforcement learning without the need for laboratory testing according to an embodiment. The previously trained reward model can use as input the outcome objective, the starting effector, and the proposed effectors output by the generative model. The reward model can then output a reward value that is calculated based on the project the difference between the objective and observed result. The reinforcement learning algorithm can then utilize the outcome objective, starting effectors and proposed effectors in addition to the reward value to update the generative model with optimized weights.

[0024] FIG. 8 is a schematic showing how the trained reward model can be used to identify and prioritize top proposed effectors from the generative model for laboratory testing according to an embodiment. Once the generative model generates proposed effectors, the trained reward model is used to project rewards for each proposed effector. Then, the top ranked proposals are prioritized for laboratory editing and phenotyping. Laboratory data and modified promoter sequences, along with the outcome objective and starting effectors, are then used to further update the trainedDocket # 212601-WO-SEC-l reward model. Outputs of the trained reward model are then used for further reinforcement learning.

[0025] FIG. 9 is a schematic illustrating the application of the method and system as described in Example 2. Here, the objective is to create a pattern of gene expression with target values in several different tissues and conditions using CRISPR-SDN3 type edits. The starting effectors include the wild-type sequences of the target genes’ promoters. The generative model outputs proposed promoter sequences (effector proposals), and then the sequences are sent to the lab where template and guide RNAs are designed and synthesized, vectors are constructed, and plants are edited. The modified promoters are sequenced, and qRT-PCR is used to measure gene expression for the target gene in tissues of interest. The expression values across tissues are used to train the reward model using supervised learning. Then, the reward model outputs are input into a reinforcement learning algorithm. The reinforcement learning model updates the weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory results meet the outcome objective, modified plants or plant cells are used for commercial product development.

[0026] FIG. 10 is a schematic illustrating the application of the method and system as described in Example 3. Here, the outcome objective is to reduce plant height through modulating expression of target genes using CRISPR-CAS9 editing. The starting effectors include the wild-type sequence of the target genes promoters. An external context module is used to embed information from published literature, public databases, and proprietary data along with the outcome objective and starting effectors. The output of the external context model informs the generative model. The generative model outputs proposed effectors, and then the sequences are sent to the lab where guide RNAs are designed and synthesized, vectors are constructed, and plants are edited. The modified promoters are sequenced, and qRT-PCR is used to measure gene expression for the target gene in the tissues of interest. Plants are grown to maturity in the greenhouse and measured for the plant height phenotype. The expression and phenotype values across tissues are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. TheDocket # 212601-WO-SEC-l reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory results meet the outcome objective, modified plants or plant cells are used for commercial product development.

[0027] FIG. 11 is a schematic illustrating the application of the method and system as described in Example 4. Here, the outcome objective is to reduce plant height through modulating expression of a target gene using CRISPR-CAS9 editing using pre-trained generative and reward models produced in Example 3. The starting effector is the wildtype promoter of the gene of interest. An external context module is used to embed information from published literature, public databases, and proprietary data along with the outcome objective and starting effector. The output of the external context model informs the generative model. The trained generative model produced in Example 3 generates proposed effectors that are then evaluated by the reward model also trained in Example 3. The best ranked proposals then move into laboratory production, outcome validation, and ultimately commercial product development.

[0028] FIG. 12 is a schematic illustrating the application of the method and system as described in Example 5. Here, the outcome objective is to increase plant yield by 10 bushels per acre. SNP representations of the parental genotypes are input into the generative model, along with an edit library consisting of all available edits for the generative model to select from. The generative model then creates proposed edit combinations of 10 or 50 edits to incorporate into the parent genotypes. The proposals are then tested in the lab and field by first generating the edits in doubled haploid populations, then growing and bulking the plants to test in the field as hybrids. Edited plants are compared to non-edited doubled haploid populations to create the outcome data in the form of general combining ability (GCA) in edited vs non-edited doubled haploid populations. The outcome data are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.Docket # 212601-WO-SEC-l

[0029] FIG. 13 is a schematic illustrating the application of the method and system as described in Example 6. Here, the outcome objective is to increase plant yield by 10 bushels per acre and lodging resistance by 10%. An edit library context model is utilized to incorporate gene functions, edit effects and target composition using a retrieval augmented generation large language model. The context module and generative model also use as input an edit library consisting of all available edits and a SNP representation of parental genotypes. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating the edits in doubled haploid populations, then growing and bulking the plants to test in the field as hybrids. Edited plants are compared to non-edited doubled haploid populations to create the outcome data in the form of general combining ability (GCA) and lodging change in edited vs non-edited doubled haploid populations. The outcome data are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.

[0030] FIG. 14 is a schematic illustrating the application of the method and system as described in Example 7. Here, the outcome objective is to increase plant yield by 10 bushels per acre and decrease plant height by 2 cm. An edit library context model is utilized to incorporate gene functions, edit effects and target composition using a retrieval augmented generation large language model. The context module and generative model also use as input an edit library consisting of all available edits and a SNP representation of parental genotypes. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating the edits in doubled haploid populations, then growing and bulking the plants to test in the field as hybrids. Further breeding to the R2 advancement stage is performed and yield and plant height are measured, and the change in these phenotypes relative to non-edited controls is used as outcome data. Then, outcome data is used to inform the breeding pipeline digital twin which is enabled with a breeding context module (FIG. 15). TheDocket # 212601-WO-SEC-l breeding pipeline digital twin outputs are then input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.

[0031] FIG. 15 is a schematic illustrating the breeding pipeline digital twin. The digital twin is trained with prior breeding data including attributes such as selection choices, genotype data, and phenotype data. Furthermore, a breeding context module with a RAG-enabled LLM structure is used to incorporate textual information from various sources, such as predictive models, breeding strategies, product concepts, business constraints, and algorithmically-optimized selection strategies.

[0032] FIG. 16 is a schematic illustrating the application of the method and system as described in Example 8. Here, the outcome objective is to increase the average expression of an insecticidal protein. The external context module uses as input published literature, published datasets, proprietary lab data, and insecticidal protein effects to provide context to the generative model and the reward module. The external context module and generative model also use as input the starting effector of the gene and regulatory sequence of the insecticidal gene of interest. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating edited cells and / or plants and then assessing protein expression to create the outcome data in the form of percent change in protein expression in edited vs. nonedited cells. The outcome data are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.

[0033] FIG. 17 is a schematic illustrating the application of the method and system as described in Example 9. Here, the outcome objective is to improve the insecticidal activity of an insecticidal protein. The external context module uses as input published literature, published datasets, proprietary lab data, and insecticidal protein effects toDocket # 212601-WO-SEC-l provide context to the generative model and the reward module. The external context module and generative model also use as input the starting effector of the gene and regulatory sequence of the insecticidal gene of interest. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating edited cells and / or plants and then assessing the insecticidal activity to create the outcome data in the form of percent change in insecticidal activity in edited vs. non-edited cells. The outcome data are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.

[0034] FIG. 18 is a schematic illustrating the application of the method and system as described in Example 10. Here, the outcome objective is to improve the insecticidal activity while decreasing phytotoxicity of an insecticidal protein. The external context module uses as input published literature, published datasets, proprietary lab data, and insecticidal protein effects to provide context to the generative model and the reward module. The external context module and generative model also use as input an edit library consisting of all available edits and the starting effector of the gene and regulatory sequence of the insecticidal gene of interest. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating edited cells and / or plants and then assessing the insecticidal activity and phytotoxicity to create the outcome data in the form of percent change in insecticidal activity and phytotoxicity in edited vs. non-edited cells. The outcome data are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.

[0035] FIG. 19 is a schematic illustrating the application of the method and system as described in Example 11 . Here, the outcome objective is to increase yield inDocket # 212601-WO-SEC-l environments managed with nitrogen application while maintaining yield in environments without management. The external context module uses as input published literature, published datasets, proprietary lab data, and management effects to provide context to the generative model and the reward module. The external context module and generative model also use as input an edit library consisting of all available edits and a SNP representation of parental genotypes. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating edited plants and then doubled haploid populations from the initial edits. Plant populations are tested for yield or a yield proxy to create outcome data in the form of yield change in edited vs. non-edited plants with supplemental nitrogen and yield change in edited vs. non-edited plants without supplemental nitrogen. The outcome data are used to train the reward model using supervised learning. Then, the reward values are input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.

[0036] FIG. 20 is a schematic illustrating the application of the method and system as described in Example 12. Here, the outcome objective is to increase plant yield by 10 bushels per acre and increase internal rate of return (IRR) by 15%. An edit library context model is utilized to incorporate gene functions, edit effects and target composition using a retrieval augmented generation large language model. The context module and generative model also use as input an edit library consisting of all available edits and a SNP representation of parental genotypes. The generative model generates proposed edit combinations that are then tested in the lab and field by first generating the edits in doubled haploid populations, then growing and bulking the plants to test in the field as hybrids. Further breeding to the R2 advancement stage is performed and yield and plant height are measured along with evaluating the IRR of released products, and the changes relative to non-edited controls are used as outcome data. Then, outcome data is used to inform the breeding pipeline digital twin which is enabled with a breeding context module (FIG. 15) and a seed market digital twin which is enabled withDocket # 212601-WO-SEC-l a market context module. The digital twin outputs are then input into a reinforcement learning algorithm. The reinforcement learning algorithm updates weights within the generative model, further informing and empowering the generative model to create improved proposed effectors. Once the laboratory and field results meet the outcome objective, modified plants are used for commercial product development.DETAILED DESCRIPTION

[0037] Described herein are methods and systems for the use of generative artificial intelligence (Al) in agriculture and agricultural research. Provided are methods and systems for improving a generative artificial intelligence model, training a generative artificial intelligence model, and for creating one or more proposed effectors using an artificial intelligence model. The methods include inputting, into a trained generative artificial intelligence (Al) model, a starting effector and an outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective, proposing by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective; receiving by a trained reward model one or more entries, each entry comprising the starting effector; the outcome objective; and one of the one or more proposed effectors, generating by the reward model a reward value to predict how well the one of the one or more proposed effectors meets the outcome objective; and using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate one or more effectors satisfying the outcome objective. In some embodiments, the methods also include selecting one or more proposed effectors based on their ranking; and creating the selected one or more proposed effectors.

[0038] As used herein, a generative model, generative artificial intelligence model, or a generative Al model are used interchangeably and refers to an artificial intelligence algorithm capable of generating effector proposals in response to a prompt. The type of Al algorithm used in the methods described herein is not particularly limited and may be any Al algorithm known in the art or described herein that is capable of generating effector proposals in response to a prompt. In certain embodiments of the methods described herein, the generative Al model is a deep neural network. In someDocket # 212601-WO-SEC-l embodiments, the deep neural network may include one or more fully connected layers, one or more convolutional neural network layers, one or more recurrent neural network layers, one or more multi-headed self-attention layers, one or more selective state space layers, or any combination thereof. In some embodiments, the generative Al model is capable of outputting one or more DNA sequences comprising an ordered set of nucleotides. In some embodiments, the generative Al model is capable of outputting one or more protein sequences comprising an ordered set of amino acids. In some embodiments, the model is capable of outputting one or more decisions of whether to include an edit from a pre-specified library into a population development process. In some examples, an external module may be used to provide context for the generative Al model. This module may have the structure of a Large Language Model. In some examples, the generative Al model includes a natural language model. This natural language model may include, but is not limited to, one or more dense layers, one or more recurrent (e.g. GRU or LSTM) layers, one or more convolutional layers, one or more pooling layers, one or more multi-headed self-attention layers, one or more selective state space layers, one or more batch or layer normalization layers, and one or more dropout layers. In some examples, the generative Al model includes a transformer model. This transformer model may include, but is not limited to, multiheaded self-attention layers, bidirectional encoder transformer (BERT) layers, and decoder-only masked self-attention layers. In some examples, the generative Al model includes a transformer large language model. This model may include, but is not limited to the BERT architecture, the generative pre-trained (GPT) architecture, and / or the Large Language Model Architecture (LLaMA).

[0039] In certain embodiments of the methods described herein, the generative Al model receives one or more pairs (e.g., 1 , 2, 3, 4, 5, 6, 7, 8, 9, 10, 25, 50, 100, 500, 1000, 10,000 or more pairs), each pair comprising a starting effector and an outcome objective. In certain embodiments, the generative Al model produces one or more (e.g., 1 , 2, 3, 4, 5, 6, 7, 8, 9, 10, 25, 50, 100, 500, 1000, 10,000 or more) proposed effectors.

[0040] In certain embodiments of the methods described herein, the generative Al model has been trained to propose one or more effectors that are predicted to satisfyDocket # 212601-WO-SEC-l the outcome objective based on the reward value from the reinforcement learning algorithm.

[0041] In certain embodiments of the methods described herein, the generative Al may receive as input public or proprietary information or combinations thereof. Public or proprietary information includes, but is not limited to, journal articles, internal articles, public data, and internal data in text and / or machine-readable formats.

[0042] As used herein, a trained generative Al model refers to a generative Al model that has been trained to propose one or more effectors that are predicted to satisfy or quantitatively approximate the outcome objective.

[0043] The generative Al model may be trained in any suitable manner such as, for example, using variations of policy gradient methodologies such as REINFORCE, TRPO, PPO, or A2C, or using variations or Deep Q-network (DQN) learning. In some examples, a single policy network parameterizing the action probabilities, ™(a | s) is trained. In some examples two output networks, ™[a | s) and v(s) are trained, wherein the former produces the action probabilities and the latter provides estimates of the value for each state. In some examples, one more weights 0, 4> of the policy and value networks are shared.

[0044] For example, as described in Example 1 , the generative Al model may be trained using one or more pairs, where each pair includes the starting effector and the outcome objective. The generative model produces one or more proposed effectors. One or more entries of the starting effector, the outcome objective, and proposed effector are input into the trained reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective. The training of the reward model is described elsewhere herein.

[0045] The method by which the starting effector and outcome objective are input into a trained generative artificial intelligence (Al) model is not particularly limited and may be input using any of the methods described herein or known in the art. In some embodiments, the starting effector, outcome objective, or both may be provided by a user to the computer via a prompt or entry. In other embodiments, the starting effector, outcome objective, or both may be retrieved from a database or context module.Docket # 212601-WO-SEC-l

[0046] In certain embodiments of the methods described herein, the starting effector is the initial starting sequence (polynucleotide or polypeptide), genotype, genetic variant, or plant phenotype. In some examples, the starting effector may include, but is not limited to, a promoter sequence, a gene sequence, a regulatory sequence, a protein sequence, a list of proposed edits, a list of parental genotypes, a representation of parental genotypes, a list of parental phenotypes, a representation of parental phenotypes, chemical structures, chemical activities, chemical formulations, biologicals, or chemicals. In some examples, the starting effector is a genetic component that affects an agronomic trait of interest.

[0047] As used herein, a polynucleotide refers to a single or a double-stranded polymer of deoxyribonucleotide or ribonucleotide bases. Polynucleotides may also include fragments and modified nucleotides. Thus, the terms “polynucleotide”, “nucleic acid sequence”, “nucleotide sequence” and “nucleic acid fragment” are used interchangeably to denote a polymer of RNA and / or DNA and / or RNA-DNA that is single- or doublestranded, optionally comprising synthetic, non-natural, or altered nucleotide bases. Nucleotides (usually found in their 5’-monophosphate form) are referred to by their single letter designation as follows: “A” for adenosine or deoxyadenosine (for RNA or DNA, respectively), “C” for cytosine or deoxycytosine, “G” for guanosine or deoxyguanosine, “U” for uridine, “T” for deoxythymidine, “R” for purines (A or G), “Y” for pyrimidines (C or T), “K” for G or T, “H” for A or C or T, “I” for inosine, and “N” for any nucleotide.

[0048] The terms “polypeptide,” “peptide” and “protein” are used interchangeably herein to refer to a polymer of amino acid residues. The terms apply to amino acid polymers in which one or more amino acid residue is an artificial chemical analogue of a corresponding naturally occurring amino acid, as well as to naturally occurring amino acid polymers.

[0049] As used herein, a gene includes a nucleic acid fragment that expresses a functional molecule such as, but not limited to, a specific protein, including regulatory sequences preceding (5’ non-coding sequences) and following (3’ non-coding sequences) the coding sequence. “Native gene” refers to a gene as found in its natural endogenous location with its own regulatory sequences.Docket # 212601-WO-SEC-l

[0050] As used herein, a promoter is a region of DNA involved in recognition and binding of RNA polymerase and other proteins to initiate transcription. The promoter sequence consists of proximal and more distal upstream elements, the latter elements often referred to as enhancers. An “enhancer” is a DNA sequence that can stimulate promoter activity, and may be an innate element of the promoter or a heterologous element inserted to enhance the level or tissue-specificity of a promoter. Promoters may be derived in their entirety from a native gene, or be composed of different elements derived from different promoters found in nature, and / or comprise synthetic DNA segments. It is understood by those skilled in the art that different promoters may direct the expression of a gene in different tissues or cell types, or at different stages of development, or in response to different environmental conditions. It is further recognized that since in most cases the exact boundaries of regulatory sequences have not been completely defined, DNA fragments of some variation may have identical promoter activity.

[0051] As used herein, regulatory elements refer to nucleotide sequences located upstream (5’ non-coding sequences), within, or downstream (3’ non-coding sequences) of a coding sequence, and which influence the transcription, RNA processing or stability, or translation of the associated coding sequence. Regulatory sequences include, but are not limited to, promoters, translation leader sequences, 5’ untranslated sequences, 3’ untranslated sequences, introns, polyadenylation target sequences, RNA processing sites, effector binding sites, and stem-loop structures.

[0052] As used herein, a coding sequence refers to a polynucleotide sequence which codes for a specific amino acid sequence.

[0053] As used herein, an agronomic trait of interest may include, but not be limited to, the following: disease resistance, drought tolerance, heat tolerance, cold tolerance, salinity tolerance, metal tolerance, herbicide tolerance, improved water use efficiency, improved nitrogen utilization, improved nitrogen fixation, pest resistance, herbivore resistance, pathogen resistance, yield improvement, health enhancement, vigor improvement, growth improvement, photosynthetic capability improvement, nutrition enhancement, altered protein content, altered oil content, increased biomass, increased shoot length, increased root length, improved root architecture, modulation of aDocket # 212601-WO-SEC-l metabolite, modulation of the proteome, increased seed weight, altered seed carbohydrate composition, altered seed oil composition, altered seed protein composition, altered seed nutrient composition, as compared to an isoline plant not comprising a modification derived from the methods or compositions herein.

[0054] As used herein, the outcome objective refers to a qualitative or quantitative specification of desired outcome, as effected through one or more effector modifications. In certain embodiments of the methods described herein, the outcome objective is the target change in phenotype on the micro (molecular) or macro (whole plant or whole plot) scale. In some examples, the outcome objective may include, but is not limited to, a modification (e.g., increase or decrease) in expression of a polynucleotide, gene, mRNA, or protein or domain or fragment thereof, altered temporal or spatial gene or protein expression in one or more cells or tissues, a change in a protein property (e.g., increased or decreased activity, stability, protein interaction), a change in plant phenotype as measured in the greenhouse or growth chamber, a change in average phenotype as measured in the field, or a change in hybrid phenotype as measured in the field. In some examples, the outcome objective may include, but is not limited to, a modification (e.g., increase or decrease) in the average expression of a polynucleotide, gene, mRNA, or protein or domain or fragment thereof, altered temporal or spatial gene or protein expression in one or more cells or tissues, a change in a protein property (e.g., increased or decreased activity, stability, protein interaction, phytoxicity to a host plant). In some examples, the outcome objective may include, but is not limited to, a defined range (e.g., increase or decrease) in the range of expression of a polynucleotide, gene, mRNA, or protein or domain or fragment thereof, altered temporal or spatial gene or protein expression in one or more cells or tissues, a change in a protein property (e.g., increased or decreased activity, stability, protein interaction).

[0055] As used herein, a “gene expression pattern” refers to any pattern of transcription of a nucleic acid molecule into a transcribed RNA molecule. Expression may be characterized by its temporal, spatial, developmental, tissue, environmental, physiological, pathological, cell cycle, and / or chemically responsive qualities as well as by quantitative or qualitative indications. The transcribed RNA molecule may beDocket # 212601-WO-SEC-l translated to produce a protein molecule or may provide an antisense or other regulatory RNA molecule, such as a dsRNA, a tRNA, an rRNA, a miRNA, and the like.

[0056] As used herein, “protein expression pattern” refers to any pattern of translation of a transcribed RNA molecule into a protein molecule. Protein expression may be characterized by its temporal, spatial, developmental, or morphological qualities as well as by quantitative or qualitative indications.

[0057] In certain embodiments, the change in plant phenotype comprises one or more of a change in yield (e.g., increased yield), plant height (e.g., decreased plant height), disease resistance (e.g., increased disease resistance), abiotic stress tolerance (e.g., improved drought resistance), flowering time, and nitrogen use efficiency. In certain embodiments, the change in the outcome objective is measured by comparing to a control such as, for example, the starting effector. The outcome objective may include a numerical target value, a range of target values, an upper bound target value, or a lower bound target value. In some examples, modified gene or mRNA expression refers to a quantitative or qualitative change (e.g., increase or decrease) in abundance as measured by any acceptable method including but not limited to quantitative PCR, semi-quantitative PCR, real-time PCR, RNA-sequencing, or northern blot. Similarly, in some examples, in some examples, modified protein expression level can refer to quantitative or qualitative change (e.g., increase or decrease) in abundance as measured by any acceptable method including but not limited to western blot, mass spectrometry, ELISA, and affinity chromatography.

[0058] In some examples, the outcome objective may be defined by a user or an agent. As used herein, an agent refers to a computational model implementing a probabilistic policy that maps from a state onto a set of actions. In some examples the state comprises an initial effector and an outcome objective. In some examples the state includes additional contextual information on an effector or the environment in which the effector would be deployed. In some examples, the action space comprises selection from the set of nucleotide monomers available at a specified position in a polynucleotide. In some examples, action space comprises selection from the set of amino acid residues available at a specified position in a protein. In some examples, the action space comprises the decision of whether to include a provided edit into theDocket # 212601-WO-SEC-l development of a new population, conditioned on a set of potential parents. In some examples, the outcome objective is input into the trained generative Al model by a user or an agent.

[0059] As used herein, “decrease in expression” “decreased expression” or the like refers to any detectable reduction in expression of a gene (e.g., transcribed mRNA) and / or the corresponding polypeptide. Similarly, “decrease in activity” “decreased activity” or the like refers to any detectable reduction in the activity (e.g., enzymatic activity) of the encoded polypeptide. Also, as used herein “decrease in stability” “decreased stability” or the like refers to any detectable increase in the turnover rate (e.g., shorter half-life) of the expressed polypeptide. As used herein, “increase in expression” “increased expression” or the like refers to any detectable gain in expression of a gene (e.g., transcribed mRNA) and / or the corresponding polypeptide. As used herein, “increase in activity” “increased activity” and the like refers to any detectable gain in activity (e.g., enzymatic activity) of the polypeptide.

[0060] In some embodiments of the methods described herein, the trained generative Al model may propose one or more proposed effectors for each starting effector and outcome objective. In some embodiments, the trained generative Al model proposes one or more proposed effectors based on weights within the generative model structure. As used herein, a proposed effector refers to the output of the generative model, representing some change to or modification of the starting effector proposed by the generative model to be bested in the laboratory and / or field.

[0061] As used herein, reward model refers to a supervised machine learning model capable of predicting a reward value for a set of inputs. The reward model may be any model that is capable of mapping from the space of effectors, effector proposals or modified effectors, and outcome objectives to a scalar real value indicating reward assignment for a given reward modification relative to an outcome objective. In some examples, a reward model is a deep neural network. In some examples the deep neural network may include one or more fully connected layers, one or more convolutional neural network layers, one or more recurrent neural network layers, one or more multiheaded self-attention layers, and / or one or more selective state space layers. In some embodiments, the reward model is a generalized linear model (GLM), a support vectorDocket # 212601-WO-SEC-l machine (SVM), a decision tree, a set of boosted decision trees, a random forest (RF), or a generalized additive model (GAM).

[0062] In some embodiments of the methods described herein, the trained reward model may receive one or more entries, for example, from the outcome objective, one or more starting effectors, one or more proposed effectors, and / or one or more observed effector edits, where observed effector edits are the outcome of a laboratory experiment, wherein the outcome of editing may or may not match the original proposals. In some embodiments of the methods described herein, each entry into the trained reward model includes the starting effector, the outcome objective, one of the one or more proposed effectors.

[0063] In one example, the reward model is trained to learn how to predict / measure how well the one or more proposed effectors meets the outcome objective. The reward model may be trained in any number of ways. For example, one of the entries may be input into the reward model and trained using supervised learning. A target signal may be used in the training to determine the reward value using outcome objective data for the starting effector and the outcome objective data for the proposed effector. In some examples, the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective data for the proposed effector. As used herein, the target signal is a value comparing the outcome objective data for the proposed effector and the starting effector. In some examples, the target signal is calculated based on the absolute value of the difference between an outcome objective and the experimental data. In some examples, the target signal is calculated from a norm, comparing the outcome objective and experimental data, wherein both have two or more targets associated with them. In some examples, the norm for the comparison of outcome objective and experimental data is differentially weighted among the two or more targets. In some examples, comparison of the outcome objective and experimental data is used to calculate a ranking of proposed effectors, which are transformed into a reward value using a monotonic mapping of rank to reward. In some examples, the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective. In some examples, the outcome objective data generated from one or more of theDocket # 212601-WO-SEC-l proposed effectors will not satisfy the outcome objective. These results may be included in those used for reinforcement learning and updating the weights in the generative model. In some examples, the reward model may be updated with public or proprietary information or combinations thereof.

[0064] As used herein, outcome objective data refers to the actual (observed) data generated in the laboratory, greenhouse, field, assay, simulated data, or combinations thereof. In certain embodiments of the methods described herein, the outcome object data includes observed data, simulated / predicted data, or a combination thereof.

[0065] Outcome objective data may include, but is not limited to, expression data in one or multiple tissues, enzyme activity, in vitro assays, protein binding, greenhouse phenotypic data for plant, field phenotypic data for a plot, combining ability of an inbred, yield performance of a hybrid, the BLUP of a genotype, an assay of insect toxicity, an assay of disease tolerance, an evaluation of lodging due to brittle snap, root lodging, or late stalk lodging, a sentiment score from breeders, a summary of plant or canopy structure, a time series of plant development data, or compositions thereof.

[0066] The method by which the outcome objective data is obtained is not particularly limited and may be obtained using any method described herein or known in the art. Examples of methods to obtain outcome objective data include, but is not limited to, transcript / transcriptome profiling, protein abundance assays, enzyme assays, laboratory assays, greenhouse experiments, or field experiments, although other experiments may also be performed Plant or plot-level phenotypic data may be obtained using a variety of instrumentation, for example, by RGB camera, stereo camera, time-of-f light camera, light detection and ranging (LiDAR), multi-spectral camera, hyperspectral camera, infrared (IR) sensor, photosynthetically active radiation (PAR) sensor, RADAR, or magnetic resonance imaging (MRI). Such instrumentation may be deployed proximally (in field or greenhouse), by unmanned aerial vehicle (UAV), or by satellite.

[0067] In some embodiments of the methods described herein, the reward model (e.g., trained reward model) generates a reward value to predict how well the one of the one or more proposed effectors meets the outcome objective. For example, the reward values can be assigned such that a higher value denotes a higher similarity of the actual or predicted outcome data to the outcome objective.Docket # 212601-WO-SEC-l

[0068] In some embodiments of the methods described herein, a reinforcement learning algorithm adjusts or optimizes one or more weights in the generative Al model based on the reward value to train or update the generative Al model’s capacity to predict / recommend effectors satisfying the outcome objective.

[0069] In certain embodiments of the methods described herein, the method further comprises repeated iterations of inputting, into a trained generative Al model, a starting effector and outcome objective, such that, for example, the step of inputting a starting effector and outcome objective into a trained generative Al model is repeated at least 1 , 2, 3, 4, 5, 6, 7, 8, 9, 10, 25, 50, 100, 500, 1000, 10,000 or more times. In certain embodiments, repeated iterations are performed with a different starting effector, a different outcome objective, or both a different starting effector and a different outcome objective.

[0070] In certain embodiments of the methods described herein, the method further comprises repeated iterations of receiving by a trained reward model one or more entries, each entry comprising the starting effector, the outcome objective, and one or more proposed effectors, such that, for example, the step of receiving by a trained reward model one or more entries, each entry comprising the starting effector, the outcome objective, and one or more proposed effectors is repeated at least 1 , 2, 3, 4, 5, 6, 7, 8, 9, 10, 25, 50, 100, 500, 1000, 10,000 or more times. In certain embodiments, when the step is repeated one or more of the proposed effectors is a different proposed effector than in the first iteration. In certain embodiments, when the step is repeated one or more of the proposed effectors is a different proposed effector than in any prior iteration. In certain embodiments, when the step is repeated with the same starting effector and one or more proposed effectors, but with a different outcome objective as in the initial iteration. In certain embodiments, when the step is repeated with the same starting effector and one or more proposed effectors, but with a different outcome objective than in any prior iteration.

[0071] In certain embodiments of the methods described herein, the method further comprises repeated iterations of receiving by a trained reward model one or more entries, each entry comprising the starting effector, the outcome objective, and one orDocket # 212601-WO-SEC-l more proposed effectors with the same starting effector and one or more proposed effectors, but with a different outcome objective as in the initial iteration.

[0072] In certain embodiments of the methods described herein, the generative Al model may design, direct or both design and direct, in whole or in part, an experiment to determine whether any of the one or more proposed effectors meets the outcome objective. The designing an experiment may include, but is not limited to, experimental design, experiment layout, locations, treatments, phenotyping methods, or laboratory experiments. The directing an experiment may include, but is not limited to, directing automated systems to perform lab experiments, perform plant growth experiments, collect data, or analyze data.

[0073] In certain embodiments of the methods described herein, an automated, high- throughput experiment is performed to determine whether any of the one or more of the proposed effectors meet the outcome objective. The automated high-throughput experiment may be designed by the user, designed by the generative Al model or by a combination thereof.

[0074] In some embodiments of the methods described herein, data may be obtained by performing an experiment or assay, from an external source, or combinations thereof. In some examples, the experiment is performed in a plant, plant tissue, plant cell, microbe, or microbial organism. In some examples, the experiment is performed under different environmental conditions. These conditions may include, but are not limited to, changes in temperature, nutrients, water, light, biotic pressures, abiotic pressures, planting density or combinations thereof.

[0075] In some embodiments of the methods described herein, the proposed effectors are physically created. This may include, but is not limited to, polynucleotide synthesis, molecular cloning, transgenic plants, edited plants, plants with multiplexed edits, inbred lines, or seed products. In some examples, one or more selected proposed effectors may be created by methods that may include, but are not limited to, nucleotide synthesis, cloning, protein synthesis, gene editing, CRISPR editing, plan transformation, chemical synthesis, or synthetic biology. In some examples, one or more of the proposed effectors are created in a plant, plant tissue, plant cell, or microbe usingDocket # 212601-WO-SEC-l genome editing technology. In some examples, a plant, plant tissue, plant cell, or microbe comprising the one or more proposed effectors are selected.

[0076] In some examples, the starting effector and / or one or more proposed effectors is a polynucleotide or polypeptide impacting yield, biomass, photosynthetic efficiency, nutrient use efficiency, heat tolerance, drought tolerance, herbicide tolerance, or disease resistance of a plant. In some examples, the starting effector and / or one or more proposed effectors is a polynucleotide or polypeptide impacting plant height, tillering capacity, root length, root mass, grain size, grain weight, plant resistance, disease resistance abiotic stress tolerance nutrient use efficiency of a plant.

[0077] In some embodiments of the methods provided herein, one or more proposed effectors may be selected based on their ranking. Proposed effectors may be ranked based on any number of factors, including but not limited to the predicted reward value as output by the reward model. In some examples the proposed effector rankings may be based in part on their propensity for effective editing, which may depend on guide site base content or potential off-target activity.

[0078] As used herein, the term “plant” includes plant protoplasts, plant cell tissue cultures from which plants can be regenerated, plant calli, plant clumps, and plant cells that are intact in plants or parts of plants such as embryos, pollen, ovules, seeds, leaves, flowers, branches, fruit, kernels, ears, cobs, husks, stalks, roots, root tips, anthers, and the like. In some examples, the plant may be grown in a growth chamber, greenhouse, or field.

[0079] In certain embodiments, the plants of the methods described herein are elite plant lines (e.g., elite maize lines). In certain embodiments, the plant cells, plant parts, or seeds are isolated from or produced by an elite plant line. As used herein, “elite line” refers to any line that has resulted from breeding and selection for superior agronomic performance that allows a producer to harvest a product of commercial significance. Numerous elite lines are available and known to those of skill in the art of plant breeding (e.g., soybean, canola, and sunflower breeding). An “elite population” is an assortment of elite individuals or lines that can be used to represent the state of the art in terms of agronomically superior genotypes of a given crop species, such as soybean.Docket # 212601-WO-SEC-l

[0080] The plant species of the methods of the present disclosure can be any plant species for which a modified outcome objective, such as those described herein, is desired, including, but not limited to, monocots and dicots. Examples of plants of interest include, but are not limited to, corn (Zea mays), Brassica spp. (e.g., Brassica napus, Brassica rapa, Brassica juncea), particularly those Brassica species useful as sources of seed oil, alfalfa (Medicago sativa), rice (Oryza sativa), rye (Secale cereale), sorghum (Sorghum bicolor, Sorghum vulgare), millet (e.g., pearl millet (Pennisetum glaucum), proso millet (Panicum miliaceum), foxtail millet (Setaria italica), finger millet (Eleusine coracana), sunflower (Helianthus annuus), safflower (Carthamus tinctorius), wheat (Triticum aestivum), soybean (Glycine max), tobacco (Nicotiana tabacum), potato (Solanum tuberosum), peanuts (Arachis hypogaea), cotton (Gossypium barbadense, Gossypium hirsutum), sweet potato (Ipomoea batatas), cassava (Manihot esculenta), coffee (Coffea spp.), coconut (Cocos nucifera), pineapple (Ananas comosus), citrus trees (Citrus spp.), cocoa (Theobroma cacao), tea (Camellia sinensis), banana (Musa spp.), avocado (Persea americana), fig (Ficus casica), guava (Psidium guajava), mango (Mangifera indica), olive (Olea europaea), papaya (Carica papaya), cashew (Anacardium occidentale), macadamia (Macadamia integrifolia), almond (Prunus amygdalus), sugar beets (Beta vulgaris), sugarcane (Saccharum spp.), oats, barley, vegetables, ornamentals, and conifers. In certain embodiments, the methods and systems described herein are applicable to both same plant and cross-plant contexts. In some examples of same-plant applications, the generative Al model may be trained and deployed using data derived from a single type of plant, such as maize, soybean, wheat, rice, cotton, or sunflower. In some examples of in cross-plant applications, the generative Al model is extended to operate across multiple types of plants, for example, among monocots, including but not limited to rice, wheat, and / or maize, or among dicots including but not limited to soybean, sunflower, and / or cotton. In some examples, the reward model may be trained using outcome data derived from a single type of plant or from multiple types of plants, e.g., monocots and dicots, only plants in the monocot family, or only plants in the dicot family, or subsets thereof. In certain embodiments, the methods and systems described herein are applicable to maize, maize inbreds, or maize hybrids.Docket # 212601-WO-SEC-l

[0081] Also provided are methods of training a generative Al model. The method of training the Al model may use any of the training methods described herein. In certain embodiments, the method comprises receiving by the generative Al model one or more pairs, wherein each pair comprises a starting effector and an outcome objective; producing by the generative Al model one or more proposed effectors; inputting one or more entries into a trained reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; and using a reinforcement learning algorithm to adjust or optimize, based on the reward value, one or more weights in the generative Al model to train or update the generative Al model’s capacity to generate effectors satisfying the outcome objective.

[0082] Further provided herein are systems for implementing a generative Al system in agriculture. In certain embodiments, the system comprises a computer readable medium having stored thereon instructions to predict the performance of a plant, when executed by a processor (or computing device), cause the processor to perform one or more of the steps of the methods described herein. In some examples, the computer system updates, trains, or improves a generative Al model. In some examples, the computer system is composed of one or more servers, and a computing device communicatively coupled to the one or more servers. In some examples, the computing device includes memory and one or more processors configured to perform operations. The operations performed may include, but are not limited to, one or more of receiving by a trained generative Al model the starting effector and the outcome objective, proposing one or more proposed effectors for each starting effector and outcome objective, receiving by a trained reward model one or more entries consisting of the starting effector, the outcome objective, and one or more proposed effectors, generating by the reward model a reward value to predict how well the one or more proposed effectors meets the outcome objective, and adjusts or optimizes, by a reinforcement learning algorithm, one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate one or more effectors satisfying the outcome objective.Docket # 212601-WO-SEC-l

[0083] In some examples, the computer system includes one or more processors configured to perform operations comprising displaying the ranked effectors on a user interface. In some examples, the system includes one or more processors configured to select one or more proposed effectors based on their ranking.

[0084] In some examples, the computer readable medium stores instructions to predict the performance of a plant, when executed by a processor (or computing device). In some examples, the computer readable medium may store instructions to predict the performance of a plant when executed by a processor (or computing device).

[0085] In some examples, the computer system for updating or improving a generative Al model consists of one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; a computing device communicatively coupled to the one or more servers, the computing device including memory, one or more processors configured to perform operations comprising receiving by the generative Al model one or more pairs, wherein each pair comprises a starting effector and an outcome objective, produce by the generative Al model one or more proposed effectors, receive by a trained reward model one or more entries, each entry comprising the starting effector, the outcome objective, and one of the one or more proposed effectors; generate by the trained reward model a reward value to predict how well the one of the one or more proposed effectors meets the outcome objective; and adjusting or optimizing, by a reinforcement learning algorithm, one or more weights in the generative Al model, thereby training the generative Al model to propose one or more effectors that are predicted to satisfy the outcome objective. In some examples, the computer readable medium has stored thereon instructions to predict the performance of a plant, when executed by a processor (or computing device), cause the processor to train the reward model.

[0086] Further provided herein is a method of using a generative artificial intelligence (Al) model to alter a phenotype in a target plant comprising inputting, into a trained generative artificial intelligence (Al) model, a genotypic representation of a first parent of the target plant, a genotypic representation of a second parent of the target plant, an edit library, the edit library comprising a list possible genome edits that may be introduced into the first and / or second parental genomes, and an outcome objective,Docket # 212601-WO-SEC-l wherein the trained generative Al model has been trained to propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective, proposing by the generative Al model one or more combinations of genome edits, generating one or more populations of plants, each of the one or more populations of plants generated by genome editing using one of the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits, receiving by a trained reward model one or more entries, each entry comprising the representation of the first parental genotype and the second parental genotype, the outcome objective, and genotypic data for the members of the population, generating by the reward model a reward value to predict how well the genome edit or combination of genome edits meets the outcome objective, and using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective. In certain embodiments, the method is repeated at least one (e.g., at least 1 , 2, 3, 4, 5, 10, 100, 1000, 10,000) cycle. In certain embodiments, the genome editing comprises multiplex editing. In certain embodiments, the produced population of plants comprises members with varying combinations of target edits in an otherwise uniform genetic background. In certain embodiments, the produced population of plants comprises members with varying combinations of homozygous target edits in an otherwise uniform genetic background.

[0087] A “genomic profile”, “genotypic profile”, or the like as used herein generally refers to a set of information about the entire genome of a given plant or group of plants (genome-wide), or it can encompass a specific subset of the genome of a given plant or group of plants, or any combination thereof in a given plant or group of plants. The genotypic profile (also referred to herein as genotype) of the plants and members of the populations described herein (e.g., inbred parents, inbred population, hybrid population) may be determined or generated using any method known in the art. In certain embodiments, the genomic profile of the plants and members of the populations described herein (e.g., inbred parents, inbred population, hybrid population) includes information regarding the presence or absence in the genome of a specific set ofDocket # 212601-WO-SEC-l mutations, single nucleotide polymorphisms (SNPs), insertion of bases, deletion of bases, genotypic markers, other sequence information, or any combination thereof. In certain embodiments of the methods described herein, the genotypic profile is determined using molecular biological assays such as, for example, PCR, DNA sequencing, restriction fragment length polymorphism identification, SNP genotyping or whole genome sequencing, such that the genotypic profile comprises an observed genotype (e.g., observed SNP markers). The plants can be genotyped individually or from a pooled DNA sample. In certain embodiments of the methods described herein, the genotypic profile is a predicted genotypic profile such that, for example, the genotypic profile comprises an imputed genotype (e.g., imputed SNP markers). In certain embodiments of the methods described herein, the genotypic profile comprises both observed genotypic information and predicted genotypic information, such that, for example, the genotype comprises both observed SNP markers and imputed SNP markers. In certain embodiments, the genotypic profile is predicted using variational autoencoders (VAEs). For example, in certain embodiments, to predict the genotypes of the plants and members of the populations described herein parental SNPs are imputed using VAEs trained for optimal reconstruction of the population, such as, for example, samples from a breeding program.

[0088] In certain embodiments of the methods described herein, the representation of genotypic information comprises a single nucleotide polymorphism (SNPs), polymorphic regions, indels, transgene presence, genome edits, or any combination thereof. The method by which the representation of genotypic information is input into the trained generative Al model is not particularly limited and may be done, for example, by using any file format that conveys genetic sequence variation at one or more locations of the genome such as those described herein or known in the art. In certain embodiments of the methods described herein, formats may include, but are not limited to, variant call format (VCF), graphical fragment assembly (GFA), variant graph (VG).

[0089] The genome editing technology for use in the methods described herein is not particularly limited and may be any genome editing technique that allows for the introduction of the genome edit or combination of genome edits to be introduced into a plant. In certain embodiments the genome editing technique uses an enzyme selectedDocket # 212601-WO-SEC-l from the group consisting of a polynucleotide-guided endonuclease, CRISPR-Cas endonucleases, base editing deaminases, zinc finger nuclease, a transcription activatorlike effector nuclease (TALEN) or engineered site-specific meganuclease. The type of genome edit of the methods described herein is not particularly limited, such that the method can be any method known in the art including, but not limited to SDN-1 type edits to a single cut site, SDN-2 type edits to produce base swaps, SDN-3 type edits by integrating provided template DNA sequence, and base editing.

[0090] The use of double-stranded break technologies such as Cas endonuclease- gRNA complexes, has been described, for example in U.S. Patent Application Publications 2015 / 0082478, and 2015 / 0059010, International Application Publications WO201 5 / 026886, W02016 / 007347, and WO2016 / 25131 , and US Patent No. 10,934,536. As used herein, a Cas endonuclease refers to a polypeptide encoded by a Cas (CRISPR-associated) gene. A Cas protein includes but is not limited to: a Cas9 protein, a Cpf1 (Cas12) protein, a C2c1 protein, a C2c2 protein, a C2c3 protein, Cas3, Cas3-HD, Cas 5, Cas7, Cas8, Casi o, or combinations or complexes of these. When complexed with a guide polynucleotide, the guide polynucleotide / Cas endonuclease complex”, (or “guide polynucleotide / Cas endonuclease system”, “ guide polynucleotide / Cas complex”, “guide polynucleotide / Cas system” and “guided Cas system” or “Polynucleotide-guided endonuclease”, “PGEN”” are capable of directing the Cas endonuclease to a DNA target site, enabling the Cas endonuclease to recognize, bind to, and nick or cleave (introduce a single or double-strand break) the DNA target site. A guided Cas system referred to herein can comprise Cas protein(s) and suitable polynucleotide component(s) of any known CRISPR systems (Horvath and Barrangou, 2010, Science 327:167-170; Makarova et al. 2015, Nature Reviews Microbiology Vol. 13:1 -15; Zetsche et al., 2015, Cell 163, 1-13; Shmakov et al., 2015, Molecular Cell 60, 1 -13).

[0091] In certain embodiments of the methods described herein the genome editing is performed by multiplex editing, such that, for example, two or more sites are targeted in the editing method. Two, three, four, five, six, seven, eight, nine, ten, or more target sites can be targeted at the same time in certain embodiments. A multiplex method is typically performed by a targeting method herein in which multiple different RNADocket # 212601-WO-SEC-l components are provided, each designed to guide a guide polynucleotide / Cas endonuclease complex to a unique DNA target site.

[0092] In some embodiments of the methods described herein, one or more combinations of genome edits are introduced by one or more guide RNAs. The terms “guide RNA" relates to an RNA molecule, also called a crRNA (CRISPR RNA), and said guide RNA can create a guide RNA / Cas endonuclease complex that can direct the Cas endonuclease to a DNA target site, enabling the Cas endonuclease to recognize, optionally bind to, and optionally nick or cleave (introduce a single or double-strand break) the DNA target site.

[0093] In certain embodiments, populations are generated by crossing two parents with known genetic backgrounds (with or without genome edits), and then crossing offspring so that the makeup of each individual plant is some combination of the two parental genomes and the edit pool. In some applications, a population may consist of many individuals and / or inbred lines with different combinations of parent alleles and / or genome edits. In certain embodiments, the population of plants of the methods described herein comprises members with varying combinations of homozygous target edits in an otherwise uniform genetic background. In certain embodiments, the outcome objective data comprises genotypic data for a plant of the population including data to determine which edits are present in the plant and the phenotype for the plant.

[0094] The target plant may be any plant described herein or known in the art for which altering a phenotype is desired. In certain embodiments, the target plant is a hybrid plant. Hybrid plants include, but are not limited to, hybrid crops made up from a cross between two or more heterotic groups or a filial cross of inbred in the same heterotic group. As used herein, “heterotic group” describes a group of germplasm that create hybrid vigor when crossed with plants from a genetically distinct heterotic group.

[0095] In some instances, the hybrid crop is maize, wheat, canola, or sorghum.

[0096] Further provided herein are methods for using a generative artificial intelligence (Al) model for plant breeding comprising inputting, into a trained generative artificial intelligence (Al) model, a parental genotypic population comprising genotypic information for a plurality of parental plants, an edit library, the edit library comprising a list of all possible genome edits that may be introduced into parental plants of theDocket # 212601-WO-SEC-l parental genotypic population, and an outcome objective, wherein the trained generative Al model has been trained to select a first and second parental plant from the genotypic population and propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective, proposing by the trained generative Al model proposes a first and second parental plant from the genotypic population and one or more combinations of genome edits, generating one or more edited populations of plants, where each of the one or more edited populations of plants is generated by single-site editing, multiplex editing, or a combination thereof using one of the selected first and second parental plant and the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits in an otherwise uniform genetic background, receiving by a breeding pipeline digital twin one or more entries, each entry comprising the representation of the first parental genotype and the second parental genotype, the outcome objective, and genotypic data for the members of the edited population, simulating breeding outcomes over the course of 3 or more (e.g., 3, 4, 5, 10, 20, 50, 100 or more) generations, generating a reward value to predict how well the genome edit (or combination of genome edits) meets the outcome objective, and using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.

[0097] In certain embodiments, a breeding pipeline digital twin is used to simulate breeding decisions and provide feedback for reinforcement learning. The breeding pipeline digital twin is a simulation environment covering in silico approximations of plant breeding pipeline processes spanning from population development through precommercial hybrid or varietal multi-environment testing. The digital twin may be structured using any method known in the art. In certain embodiments, the digital twin is structured as described in PCT / US24 / 50729.

[0098] In some embodiments of the method described herein, a breeding context module is used to provide context for the generative Al model. The breeding context module for use in the methods described herein may be any context module describedDocket # 212601-WO-SEC-l herein or known in the art. In certain embodiments, the breeding context module comprises a Large Language Model.

[0099] It must be noted that, as used in the specification and the appended claims, the singular forms "a," "an" and "the" include plural referents unless the context clearly dictates otherwise.

[0100] EMBODIMENTS

[0101] The present disclosure is further illustrated in the following embodiments. It should be understood that these embodiments are given by way of illustration only.

[0102] Embodiment 11 . A method of improving a generative artificial intelligence (Al) model, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a starting effector and an outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) proposing by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receiving by a trained reward model one or more entries, each entry comprising: i. the starting effector; ii. the outcome objective; and iii. one of the one or more proposed effectors;(d) generating by the reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective;(e) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate one or more effectors satisfying the outcome objective.2. The method of embodiment 1 , wherein the outcome objective is a defined or modified gene expression level or defined gene expression range compared to the gene expression level or range of the starting effector. In some examples, the outcome objective includes, but is not limited to, an average defined or average modified gene expression level or average defined gene expression range compared to the averageDocket # 212601-WO-SEC-l gene expression level or average range of the starting effector. In some examples, the starting effector is a regulatory element from a maize, cotton, or soybean event. See, for example, Table 1.3. The method of embodiment 1 , wherein the outcome objective is a defined or modified protein expression level or defined protein expression range compared to the protein expression level or defined protein expression range of the starting effector. In some examples, the outcome objective may include an average defined or modified protein expression level, or an average defined range of protein expression, relative to the average protein expression level or protein expression range of the starting effector.4. The method of embodiment 1 , wherein the outcome objective is a defined or modified temporal or spatial pattern of gene or protein expression, including expression level or range, compared to that of the starting effector.5. The method of embodiment 1 , wherein the outcome objective is a defined or modified protein property compared to the protein property of the starting effector, including but not limited to protein stability, protein activity, or protein phytotoxicity to the host plant, or combinations thereof.6. The method of embodiment 5, wherein the defined or modified protein property is that of one or more insecticidal proteins and the defined or modified protein property is an insecticidal protein activity. In some examples, the outcome objective may include simultaneously reducing phytoxoicity and increasing protein activity in a certain host plant. Examples of insecticidal proteins include but are not limited to those events in Table 1 .7. The method of embodiment 5, wherein the outcome objective includes simultaneously reducing protein phytotoxicity to the host plant and increasing protein activity, protein stability, protein activity, or combinations thereof, including, not limited to, for example, increasing insecticidal activity of one or more insecticidal proteins and decreasing one or more of the insecticidal proteins’ phytotoxicity to the host plant.8. The method of embodiment 1 , wherein the outcome objective is a defined or modified protein activity, including but not limited to a defined or modified protein activity range, compared to the protein activity or range of protein activity of the starting effector. In some examples, where the protein is an insecticidal protein, the activity mayDocket # 212601-WO-SEC-l be measured by how efficiently it kills or deters one or more certain pests either by itself or with another insecticidal protein for the same crop or plant.9. The method of embodiment 1 , wherein the outcome objective is a defined or modified phenotype of a plant, plant tissue, or plant cell compared to the phenotype of a plant, plant tissue, or plant cell arising from the starting effector.10. The method of embodiment 9, wherein the defined or modified phenotype of a plant or plant tissue is increased resistance to a fungal or bacterial disease or pest or combinations thereof.11 . The method of embodiment 1 , wherein the starting effector and one or more proposed effectors are the same type, for example, polynucleotides, promoter region, exons, introns, untranslated regions, regulatory motifs, or polypeptides, with the proposed effector being a modified version of the starting effector.12. The method of embodiment 1 , wherein the starting effector is a representation, for example, like a SNP genotype, while the proposed effector is a combination of genome edits derived from that starting genotype to achieve the desired outcome objective.13. The method of embodiment 1 and 11 , wherein the starting effector is a gene, promoter, regulatory element, coding sequence, or intron. In some examples, the starting effector is from an insecticidal event in a crop species. Nonlimiting examples of events are found in Table 1 .14. The method of embodiment 1 , wherein the starting effector is a genetic component that impacts an agronomic trait of interest in a crop plant.15. The method of embodiment 1 , wherein the starting effector and one or more proposed effectors are polypeptides.16. The method of embodiment 1 , wherein the starting effector is a protein domain.17. The method of embodiment 1 , wherein the one or more proposed effectors is a variant of the starting effector.18. The method of embodiment 1 , wherein the one or more proposed effectors are modified gene, promoter, regulatory element, coding, or intron sequences; gene, promoter, regulatory element, coding sequence, or intron modifications introduced via homology-directed repair (HDR)-directed CRISPR editing; gene, promoter, regulatoryDocket # 212601-WO-SEC-l element, coding sequence, or intron edits configured to modulate gene expression; genome edits applied to parental genotypes to enhance yield or other plant traits; genome edits informed by external contextual data; genome edits and parental selections optimized for breeding pipeline advancement; genome edits targeting regulatory elements of maize events to increase protein expression; genome edits to insecticidal gene sequences to improve insecticidal activity; genome edits to insecticidal gene sequences for simultaneous enhancement of insecticidal activity and reduction of phytotoxicity; genome edits designed to improve yield and / or other traits under specified agronomic management practices; or genome edits and parental selections configured to optimize hybrid performance in competitive market environments.19. The method of embodiment 1 , wherein the starting effector and one or more proposed effectors are chemical structures, chemical activities, formulations, biologicals or chemicals.20. The method of embodiment 1 , further comprising repeating step (a) with a different starting effector, a different outcome objective, or both.21 . The method of embodiment 1 , further comprising repeating step (c), wherein the one of the one or more proposed effectors is a different proposed effector than in step (a).22. The method of embodiment 1 , further comprising repeating step (c) with the same starting effector and one or more proposed effectors as in step (a) but with a different outcome objective than in step (a).23. The method of embodiment 1 , further comprising repeating step (c) with the same starting effector and proposed effector as in step (a) but different outcome objective than in step (a).24. The method of embodiment 1 , further comprising defining, by a user or an agent, the outcome objective.25. The method of embodiment 1 , further comprising inputting, by a user or an agent, the outcome objective into the trained generative Al model.26. The method of embodiment 1 , further comprising training the reward model to learn how to predict / measure how well the one or more proposed effectors meets theDocket # 212601-WO-SEC-l outcome objective by inputting into the reward model one of the entries and using supervised learning with one or more of the entries and a target signal.27. The method of embodiment 26, wherein the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective data for the proposed effector.28. The method of embodiment 26, wherein the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective.29. The method of embodiment 1 , further comprising updating the generative Al model by a. receiving by the generative Al model one or more pairs, wherein each pair comprises the starting effector and the outcome objective; b. producing by the generative Al model one or more proposed effectors; c. inputting one or more entries into a trained reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; and d. using a reinforcement learning algorithm to adjust or optimize, based on the reward value, one or more weights in the generative Al model to train or update the generative Al model’s capacity to generate effectors satisfying the outcome objective.30. The method of embodiment 1 , 27, 28, or 29, wherein the reward value is a gene or protein expression-based reward, for example, a change in protein concentration, an agronomic performance-based reward, or a model-predicted reward, for example, a reward that does not use direct lab or field measurements.31 . The method of embodiment of claims 1 , 27, 28, or 29, wherein the reward value is the negative absolute difference between target and observed gene expression, the sum of such differences across tissues with CRISPR-Cas penalties, combined scores from log2expression changes and plant height differences, predicted values from pretrained models, changes in general combining ability for yield, composite scores for yield and lodging resistance, simulated phenotype improvements from a digital twin, changes in insecticidal protein concentration and activity, dual signals for increasedDocket # 212601-WO-SEC-l activity and reduced phytotoxicity, or percent yield change under managed and unmanaged conditions, and phenotype-based performance metrics under specific environmental or agronomic contexts.32. The method of embodiment 1 , 27, 28, or 29, wherein the reward value is determined from one or more biological, phenotypic, and computational indicators, including but not limited to differential gene or protein expression metrics, expression- driven morphological changes, predictive outputs from pre-trained models, breeding performance indicators such as general combining ability, simulated outcomes from digital breeding environments, biochemical concentrations and activities of insecticidal agents, dual optimization signals balancing efficacy and safety, agronomic yield responses under varied management conditions, and / or context-specific phenotype performance metrics informed by environmental or operational variables.33. The method of claim embodiment 1 , 27, 28, or 29, comprising determining the reward value using metrics including but not limited to the negative absolute difference between target and observed gene or protein expression levels, log2changes in expression and associated phenotypic shifts like plant height, predicted reward values from pre-trained models, changes in general combining ability (GCA) for yield, simulated performance metrics from a breeding pipeline digital twin, changes in protein concentration of insecticidal proteins, changes in insecticidal activity, dual signals for increased insecticidal activity and reduced phytotoxicity, yield changes under managed and unmanaged conditions, and / or phenotype-based performance metrics under specific environmental or agronomic contexts.34. The method of embodiment 27 or 28, wherein the outcome objective data comprises actual (observed) data, simulated data, or combinations thereof.35. The method of embodiment 1 , further comprising selecting, by a user or an agent, the starting effector.36. The method of embodiment 1 , further comprising designing, by the generative Al model, an experiment to determine whether any of the one or more proposed effectors meet the outcome objective.Docket # 212601-WO-SEC-l37. The method of embodiment 1 , further comprising directing, by the generative Al model or agent, the performance of an experiment to determine whether any of the one or more of the proposed effectors meet the outcome objective.38. The method of embodiment 1 , the method further comprising performing an automated, high-throughput experiment to determine whether any of the one or more of the proposed effectors meet the outcome objective.39. The method of embodiment 1 , wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective based on the reward value from the reinforcement learning algorithm.40. The method of embodiment 1 , further comprising performing steps (a)-(e) when the observed outcome data for one or more of the proposed effectors does not meet the outcome objective.41 . The method of embodiment 1 , further comprising receiving, by the generative Al model, public or proprietary information or combinations thereof.42. The method of embodiment 1 , wherein new information public or proprietary information or combinations thereof is used to update the reward model.43. The method of embodiment 1 , wherein data is obtained by performing an experiment or assay or from an external source. In some examples, the experiment or assay includes one or more biological assays, including, but not limited to, mass spectrometry, qRT-PCR, Western blotting, insect feeding assays, protoplast assays, or transient expression systems44. The method of embodiment 1 , wherein the experiment is performed in a plot comprising a plurality of plants, a plant, plant tissue, plant cell, microbe, or microbial organism.45. The method of embodiment 1 , wherein the experiment is performed under different environmental conditions.46. The method of embodiment 1 , further comprising providing an external module to provide context for the generative Al model.47. The method of embodiment 1 , wherein the generative Al model includes a natural language model.Docket # 212601-WO-SEC-l48. The method of embodiment 1 , wherein the generative Al model includes a transformer model.49. The method of embodiment 1 , wherein the generative Al model includes a transformer large language model.50. The method of embodiment 1 , the method further comprising physically creating one or more of the proposed effectors.51 . The method of embodiment 1 , the method further comprising creating one or more of the proposed effectors in a plant, plant tissue, plant cell, or microbe using genome editing technology.52. The method of embodiment 51 , the method further comprising selecting a plant, plant tissue, plant cell, or microbe comprising the one or more proposed effectors.53. The method of embodiment 1 , wherein the starting effector or one or more proposed effectors is a polynucleotide or polypeptide impacting yield, biomass, photosynthetic efficiency, nutrient use efficiency, heat tolerance, drought tolerance, herbicide tolerance, or disease resistance of a plant.54. The method of embodiment 1 , wherein the starting or one or more proposed effectors is a polynucleotide or polypeptide impacting plant height, tillering capacity, root length, root mass, grain size, grain weight, pest resistance, disease resistance abiotic stress tolerance nutrient use efficiency of a plant or other agronomic trait.55. The method of embodiment 51 or 52, further comprising growing the plant.56. The method of embodiment 51 or 52, further comprising obtaining a plant from the plant cell or plant tissue.57. The method of embodiment 51 or 52, using the plant cell, plant tissue, or plant in a breeding program.58. The method of embodiment 5, 6, 7, 9, 10, 14, 44, 51 , 52, 53, 55, or 56, wherein the plant, plant tissue, or plant cell is a monocot or dicot.59. The method of embodiment 58, wherein the plant, plant tissue, or plant cell is a maize, maize inbred, maize hybrid, wheat, rice, canola, sorghum, cotton, sunflower, barley, oat, safflower, tobacco, flax, pearl millet, or sugarcane plant, plant tissue, or plant cell.Docket # 212601-WO-SEC-l60. The method of embodiment 1 , wherein the outcome objective is evaluated using one or more biological assays including but not limited to mass spectrometry, qRT-PCR, Western blotting, insect feeding assays, protoplast assays, or transient expression systems.61 . A computer readable medium having stored thereon instructions to predict the performance of a plant, when executed by a processor (or computing device), cause the processor to perform the steps of embodiment 1 .62. A computer system for updating or improving a generative Al model, the system comprising:(a) one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device including:(1 ) a memory; and(2) one or more processors configured to perform operations comprising:(a) receive by a trained generative Al model the starting effector and the outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) propose by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receive by a trained reward model one or more entries, each entry comprising: i. the starting effector; ii. the outcome objective; and iii. one of the one or more proposed effectors;(d) generate by the reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective; and(e) adjust or optimize, by a reinforcement learning algorithm, one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate one or more effectors satisfying the outcome objective.Embodiment 6363. A method of creating one or more proposed effectors, the method comprising:Docket # 212601-WO-SEC-l(a) inputting, into a trained generative artificial intelligence (Al) model, a starting effector and an outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) proposing by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receiving by a trained reward model one or more entries, each entry comprising: i. the starting effector; ii. the outcome objective; and ill. one of the one or more proposed effectors;(d) generating by the reward model a reward value that predicts how well the one of the one or more proposed effectors from step (c) meets the outcome objective;(e) ranking the one or more proposed effectors based on the generated reward values for each;(f) selecting one or more proposed effectors based on their ranking; and(g) creating the selected one or more proposed effectors.64. The method of embodiment 63, wherein the outcome objective is a defined or modified gene expression level or defined gene expression range compared to the gene expression level or range of the starting effector. In some examples, the outcome objective includes, but is not limited to, an average defined or average modified gene expression level or average defined gene expression range compared to the average gene expression level or average range of the starting effector.65. The method of embodiment 63, wherein the outcome objective is a defined or modified protein expression level or defined protein expression range compared to the protein expression level or defined protein expression range of the starting effector. In some examples, the outcome objective may include an average defined or modified protein expression level, or an average defined range of protein expression, relative to the average protein expression level or protein expression range of the starting effector.66. The method of embodiment 63, wherein the outcome objective is a defined or modified temporal or spatial pattern of gene or protein expression, including expression level or range, compared to that of the starting effector.Docket # 212601-WO-SEC-l67. The method of embodiment 63, wherein the outcome objective is a defined or modified protein property compared to the protein property of the starting effector, including but not limited to protein stability, protein activity, or protein phytotoxicity to the host plant, or combinations thereof.68. The method of embodiment 63, wherein the outcome objective is a defined or modified phenotype of a plant, plant tissue, or plant cell compared to the phenotype of a plant, plant tissue, or plant cell arising from the starting effector.69. The method of embodiment 63, wherein the defined or modified phenotype of a plant or plant tissue is increased resistance to a fungal or bacterial disease or pest or combinations thereof.70. The method of embodiment 63, wherein the starting effector and one or more proposed effectors are the same type, for example, polynucleotides, promoter region, exons, introns, untranslated regions, regulatory motifs, or polypeptides, with the proposed effector being a modified version of the starting effector.71 . The method of embodiment 63, wherein the starting effector and one or more proposed effectors are polynucleotides.72. The method of embodiment 63, wherein the starting effector is a gene, promoter, regulatory element, coding sequence, or intron.73. The method of embodiment 63, wherein the starting effector is a genetic component that impacts an agronomic trait of interest in a crop plant.74. The method of embodiment 63, wherein the starting effector and one or more proposed effectors are polypeptides.75. The method of embodiment 63, wherein the starting effector is a domain.76. The method of embodiment 63, wherein the one or more proposed effectors is a variant of the starting effector.77. The method of embodiment 63, wherein the starting effector and one or more proposed effectors are chemical structures, chemical activities, formulations, biologicals, or chemicals.78. The method of embodiment 63, wherein the one or more proposed effectors are modified gene, promoter, regulatory element, coding sequence, or intron sequences; gene, promoter, regulatory element, coding sequence, or intron modificationsDocket # 212601-WO-SEC-l introduced via homology-directed repair (HDR)-directed CRISPR editing; gene, promoter, regulatory element, coding sequence, or intron edits configured to modulate gene expression; genome edits applied to parental genotypes to enhance yield or other plant traits; genome edits informed by external contextual data; genome edits and parental selections optimized for breeding pipeline advancement; genome edits targeting regulatory elements of maize events to increase protein expression; genome edits to insecticidal gene sequences to improve insecticidal activity; genome edits to insecticidal gene sequences for simultaneous enhancement of insecticidal activity and reduction of phytotoxicity; genome edits designed to improve yield and / or other traits under specified agronomic management practices; or genome edits and parental selections configured to optimize hybrid performance in competitive market environments.79. The method of embodiment 63, further comprising repeating step (a) with a different starting effector, a different outcome objective, or both.80. The method of embodiment 63, further comprising repeating step (c), wherein the one of the one or more proposed effectors is a different proposed effector than in step (a).81 . The method of embodiment 63, further comprising repeating step (c) with the same starting effector and one or more proposed effectors as in step (a) but with a different outcome objective than in step (a).82. The method of embodiment 63, further comprising repeating step (c) with the same starting effector and proposed effector as in step (a) but different outcome objective than in step (a).83. The method of embodiment 63, further comprising defining, by a user or an agent, the outcome objective.84. The method of embodiment 63, further comprising inputting, by a user or an agent, the outcome objective into the trained generative Al model.85. The method of embodiment 63, further comprising training the reward model to learn how to predict / measure how well the one or more proposed effectors meets the outcome objective by inputting into the reward model one of the entries and using supervised learning with one or more of the entries and a target signal.Docket # 212601-WO-SEC-l86. The method of embodiment 85, wherein the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective data for the proposed effector.87. The method of embodiment 63 or 86, wherein the reward value measures the change in expression in a certain cell, tissue, or plant, for example, level, temporal, or spatial expression pattern.88. The method of embodiment 63 or 86, wherein the reward value measures the change in protein concentration or protein property in a cell, tissue, or plant, for example, between edited and non-edited plants.89. The method of embodiment 63 or 86, wherein the reward value measures the change in GCA for yield, or percent yield change under managed versus non-managed conditions, for example, between edited and non-edited plants.90. The method of embodiment 63 or 86, wherein the reward value measures the change in a phenotype or trait, such as plant height, tillering capacity, root length, root mass, grain size, grain weight, plant resistance, disease resistance abiotic stress tolerance nutrient use efficiency of a plant or other agronomic trait, for example, between edited and non-edited plants.91 . The method of embodiment 85 or 86, wherein the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective.92. The method of claim 63 or 86, wherein the reward value is determined from one or more biological, phenotypic, and computational indicators, including but not limited to differential gene or protein expression metrics, expression-driven morphological changes, predictive outputs from pre-trained models, breeding performance indicators such as general combining ability, simulated outcomes from digital breeding environments, biochemical concentrations and activities of insecticidal agents, dual optimization signals balancing efficacy and safety, agronomic yield responses under varied management conditions, and / or context-specific phenotype performance metrics informed by environmental or operational variables.93. The method of claim 63 or 86, comprising determining the reward value using metrics including but not limited to the difference between target and observed gene orDocket # 212601-WO-SEC-l protein expression levels, changes in expression and associated phenotypic shifts like plant height, predicted reward values from pre-trained models, changes in general combining ability (GCA) for yield, simulated performance metrics from a breeding pipeline digital twin, changes in protein concentration of insecticidal proteins, changes in insecticidal activity, dual signals for increased insecticidal activity and reduced phytotoxicity, yield changes under managed and unmanaged conditions, and / or phenotype-based performance metrics under specific environmental or agronomic contexts.94. The method of embodiment of claims 63 or 86, wherein the reward value is a gene or protein expression-based reward, an agronomic performance-based reward, or a model-predicted reward, for example, a reward that does not use direct lab or field measurements.95. The method of embodiment of claims 63 or 86, wherein the reward value is the negative absolute difference between target and observed gene expression, the sum of such differences across tissues with CRISPR-Cas penalties, combined scores from log2expression changes and plant height differences, predicted values from pre-trained models, changes in general combining ability for yield, composite scores for yield and lodging resistance, simulated phenotype improvements from a digital twin, changes in insecticidal protein concentration and activity, dual signals for increased activity and reduced phytotoxicity, or percent yield change under managed and unmanaged conditions, and phenotype-based performance metrics under specific environmental or agronomic contexts.96. The method of embodiment 63, further comprising training or updating the generative Al model by a. receiving by the generative Al model one or more pairs, wherein each pair comprises the starting effector and the outcome objective; b. producing by the generative Al model one or more proposed effectors; c. inputting one or more entries into a trained reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; andDocket # 212601-WO-SEC-l d. using a reinforcement learning algorithm to adjust or optimize, based on the reward value, one or more weights in the generative Al model to train or update the generative Al model’s capacity to predict / recommend effectors satisfying the outcome objective.97. The method of embodiment 86, wherein the outcome objective data comprises actual (observed) data, simulated data, or combinations thereof.98. The method of embodiment 63, further comprising selecting, by a user or an agent, the starting effector or one or more proposed effectors.99. The method of embodiment 63, further comprising designing, by the generative Al model, an experiment to determine whether any of the one or more proposed effectors meet the outcome objective.100. The method of embodiment 63, further comprising directing, by the generative Al model or agent, the performance of an experiment to determine whether any of the one or more of the proposed effectors meet the outcome objective.101. The method of embodiment 63, the method further comprising performing an experiment using the created one or more proposed effectors to generate data to determine whether any of the one or more of the proposed effectors meet the outcome objective.102. The method of embodiment 63, the method further comprising using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to predict / measure one or more effectors satisfying the outcome objective.103. The method of embodiment 63, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective based on the reward value from the reinforcement learning algorithm.104. The method of embodiment 63, further comprising receiving, by the generative Al model, public or proprietary information, or combinations thereof for the starting effector.105. The method of embodiment 63, wherein new information public or proprietary information or combinations thereof is used to update the reward model.106. The method of embodiment 63, wherein data is obtained by performing an experiment or assay or from an external source. In some examples, the experiment or assay includes one or more biological assays, including, but not limited to, massDocket # 212601-WO-SEC-l spectrometry, qRT-PCR, Western blotting, insect feeding assays, protoplast assays, or transient expression systems107. The method of embodiment 106, wherein the experiment is performed in a plant, plant tissue, plant cell, microbe, or microbial organism.108. The method of embodiment 106, wherein the experiment is performed under different environmental conditions.109. The method of embodiment 63, further comprising providing an external module to provide context for the generative Al model.110. The method of embodiment 63, wherein the generative Al model includes a natural language model.111. The method of embodiment 63, wherein the generative Al model includes a transformer model.112. The method of embodiment 63, wherein the generative Al model includes a transformer large language model.113. The method of embodiment 63, the method further comprising physically creating one or more of the proposed effectors.114. The method of embodiment 63, the method further comprising creating one or more of the proposed effectors in a plant, plant tissue, plant cell, or microbe using genome editing technology.115. The method of embodiment 113 or 114, the method further comprising selecting a plant, plant tissue, plant cell, or microbe comprising the created proposed effector.116. The method of embodiment 63, wherein the starting effector or one or more proposed effectors is a polynucleotide or polypeptide impacting yield, biomass, photosynthetic efficiency, nutrient use efficiency, heat tolerance, drought tolerance, herbicide tolerance, or disease resistance of a plant.117. The method of embodiment 63, wherein the starting or one or more proposed effectors is a polynucleotide or polypeptide impacting plant height, tillering capacity, root length, root mass, grain size, grain weight, plant resistance, disease resistance abiotic stress tolerance nutrient use efficiency of a plant.118. The method of embodiment 114 or 115, further comprising growing the plant.Docket # 212601-WO-SEC-l119. The method of embodiment 114 or 115, further comprising obtaining a plant from the plant cell.120. The method of embodiment 114, 115, 118, or 119, using the plant cell or plant in a breeding program.121. The method of embodiment 67, 68, 69, 73, 87, 88, 89, 90, 91 , 107, 114, 115, 118, 119, or 120, wherein the plant, plant tissue, or plant cell is a monocot or dicot.122. The method of embodiment 121 , wherein the plant, plant tissue, or plant cell is a maize, wheat, rice, canola, sorghum, cotton, sunflower, barley, oat, safflower, tobacco, flax, pearl millet, or sugarcane plant, plant tissue, or plant cell.123. The method of embodiment 63, wherein the outcome objective is evaluated using one or more biological assays selected from mass spectrometry, qRT-PCR, Western blotting, insect feeding assays, protoplast assays, or transient expression systems.124. A computer readable medium having stored thereon instructions to predict the performance of a plant, when executed by a processor (or computing device), cause the processor to perform the steps of embodiment 63.Embodiment 125125. A computer system for updating or improving a generative Al model, the system comprising:(a) one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device including:(1 ) a memory; and(2) one or more processors configured to perform operations comprising:(a) receive by a trained generative Al model the starting effector and the outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) propose by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receive by a trained reward model one or more entries, each entry comprising:Docket # 212601-WO-SEC-l1 . the starting effector;2. the outcome objective; and3. one of the one or more proposed effectors;(d) generate by the reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective; and(e) rank the one or more proposed effectors based on the generated reward values for each.126. The system of embodiment 125, the one or more processors configured to perform operations further comprising:(f) display the ranked effectors on a user interface.127. The system of embodiment 125 or 126, the one or more processors configured to perform operations further comprising: select one or more proposed effectors based on their ranking.128. A method for training a generative Al model, the method comprising the steps of: a. receiving by the generative Al model one or more pairs, wherein each pair comprises a starting effector and an outcome objective; b. producing by the generative Al model one or more proposed effectors; c. inputting one or more entries into a trained reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; and d. using a reinforcement learning algorithm to adjust or optimize, based on the reward value, one or more weights in the generative Al model to train or update the generative Al model’s capacity to generate effectors satisfying the outcome objective.129. A computer readable medium having stored thereon instructions, when executed by a processor (or computing device), to cause the processor to perform the steps of embodiment 128.Embodiment 130130. A computer system for training a generative Al model, the system comprising:Docket # 212601-WO-SEC-l(a) one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device including:(1) a memory; and(2) one or more processors configured to perform operations comprising:(a) receive by the generative Al model one or more pairs, wherein each pair comprises a starting effector and an outcome objective;(b) produce by the generative Al model one or more proposed effectors;(c) receive by a trained reward model one or more entries, each entry comprising:1 . the starting effector;2. the outcome objective; and3. one of the one or more proposed effectors;(d) generate by the trained reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective; and(e) adjust or optimize, by a reinforcement learning algorithm, one or more weights in the generative Al model, thereby training the generative Al model to propose one or more effectors that are predicted to satisfy the outcome objective.Embodiment 131131. A method for training a reward model, the method comprising the steps of: a. inputting one or more entries into a reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; and b. using supervised learning with one or more of the entries and a target signal.132. The method of embodiment 131 , wherein the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective data for the proposed effector.Docket # 212601-WO-SEC-l133. The method of embodiment 131 , wherein the target signal is used to determine the reward value using outcome objective data for the starting effector and the outcome objective.134. A computer readable medium having stored thereon instructions, when executed by a processor (or computing device), cause the processor to perform the steps of embodiment 131.Embodiment 135135. A computer system for training a reward model, the system comprising:(a) one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device including:(1) a memory; and(2) one or more processors configured to perform operations comprising: a. receive by a reward model one or more entries, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; and b. using supervised learning with one or more of the entries and a target signal.136. The system of embodiment 135, wherein the target signal is used to determine a reward value using outcome objective data for the starting effector and the outcome objective data for the proposed effector.137. The method of embodiment 135, wherein the target signal is used to determine a reward value using outcome objective data for the starting effector and the outcome objective.Embodiment 138138. A method of using a generative artificial intelligence (Al) model to alter a phenotype in a target plant, the method comprising: a. inputting, into a trained generative artificial intelligence (Al) model, a genotypic representation of a first parent of the target plant, a genotypic representation of a second parent of the target plant, an edit library, the edit library comprising a list of allDocket # 212601-WO-SEC-l possible genome edits that may be introduced into the first and / or second parental genomes, and an outcome objective, wherein the trained generative Al model has been trained to propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective; b. proposing by the trained generative Al model one or more combinations of genome edits; c. generating one or more populations of plants, each of the one or more populations of plants generated by multiplex editing using one of the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits in an otherwise uniform genetic background; d. receiving by a trained reward model one or more entries, each entry comprising: i.the representation of the first parental genotype and the second parental genotype; ii.the outcome objective; and iii. genotypic data for the members of the population; e. generating by the reward model a reward value to predict how well the genome edit or combination of genome edits from step (c) meets the outcome objective; and f. using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.139. The method of embodiment 138, further comprising providing an external module to provide context for the generative Al model.140. The method of embodiment 138, wherein the generative Al model includes a natural language model.141. The method of embodiment 138, wherein the generative Al model includes a transformer model.142. The method of embodiment 138, wherein the generative Al model includes a transformer large language model.Docket # 212601-WO-SEC-l143. The method of embodiment 138, wherein the population of plants comprises members with varying combinations of homozygous target edits in an otherwise uniform genetic background.144. The method of embodiment 138, wherein the outcome objective data comprises genotypic data for a plant of the population to determine which edits are present in the plant and the phenotype for the plant.145. The method of embodiment 138, wherein the genotypic representation comprises a SNP representation of the parent genotype.146. The method of embodiment 138, wherein the genotypic representation comprises a polymorphic representation of the parent genotype.147. The method of embodiment 146, wherein the polymorphic representation comprises single nucleotide polymorphisms (SNPs), indels, transgene presence, genome edits, or any combination thereof.148. The method of any of embodiments 138-147, wherein the target plant is a hybrid crop produced from a cross between two or more heterotic groups or a filial cross of inbred plants from the same heterotic group.149. The method of embodiment 148, wherein the hybrid crop is maize. Embodiment 150150. A method of using a generative artificial intelligence (Al) model for plant breeding, the method comprising: a. inputting, into a trained generative artificial intelligence (Al) model, a parental genotypic population comprising genotypic information for a plurality of parental plants, an edit library, the edit library comprising a list of all possible genome edits that may be introduced into parental plants of the parental genotypic population, and an outcome objective, wherein the trained generative Al model has been trained to select a first and second parental plant from the genotypic population and propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective; b. proposing by the trained generative Al model a first and second parental plant from the genotypic population and one or more combinations of genome edits;Docket # 212601-WO-SEC-l c. generating one or more edited populations of plants, each of the one or more edited populations of plants generated by single-site editing, multiplex editing, or a combination thereof using one of the selected first and second parental plant and the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits in an otherwise uniform genetic background; d. receiving by a breeding pipeline digital twin one or more entries, each entry comprising: i.the representation of the first parental genotype and the second parental genotype; ii.the outcome objective; and iii. genotypic data for the members of the edited population; e. simulating by the breeding pipeline digital twin breeding outcomes over the course of 3 or more generations; f. generating by the breeding pipeline digital twin a reward value to predict how well the genome edit or combination of genome edits from step (c) meets the outcome objective; and g. using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.151. The method of embodiment 150, wherein the population of plants comprises members with varying combinations of homozygous target edits in an otherwise uniform genetic background.152. The method of embodiment 150, further comprising providing an external module to provide context for the generative Al model.153. The method of embodiment 150, wherein the generative Al model includes a natural language model.154. The method of embodiment 150, wherein the generative Al model includes a transformer model.155. The method of embodiment 150, wherein the generative Al model includes a transformer large language model.Docket # 212601-WO-SEC-l156. The method of embodiment 150, further comprising providing a breeding context module to provide context for the generative Al model.157. The method of embodiment 150, wherein the outcome objective data comprises genotypic data for a plant of the population to determine which edits and naturally occurring variants are present in the plant and the phenotype for the plant.158. The method of embodiment 150, wherein the parental genotypic population comprises a SNP representation of the parent genotypes for a plurality of members.159. The method of embodiment 150, wherein the genotypic representation comprises a polymorphic representation of the parent genotype for a plurality of members.160. The method of embodiment 159, wherein the polymorphic representation comprises single nucleotide polymorphisms (SNPs), indels, transgene presence, genome edits, or any combination thereof.161. The method of embodiment 150, wherein the one or more combinations of genome edits are introduced by one or more guide RNAs.162. The method of any of embodiments 150-161 , wherein the target plant is a hybrid crop produced from a cross between two or more heterotic groups or a filial cross of inbred plants from the same heterotic group.163. The method of embodiment 162, wherein the hybrid crop is maize.164. A method of improving a generative artificial intelligence (Al) model, the method comprising:(a) inputting, into a trained generative Al model:(1) a starting effector comprising a plant genotype;(2) an outcome objective comprising (i) increased plant yield under a specified management practice and (ii) maintained plant yield under nonmanaged conditions;(3) an edit library comprising genome edits applicable to the genotype;(4) and a management context module comprising data on one or more management practices;(b) proposing, by the trained generative Al model, one or more combinations of genome edits predicted to satisfy the outcome objective;(c) introducing the proposed genome edits into the genotype to generate a population ofDocket # 212601-WO-SEC-l edited plants, including doubled haploid lines;(d) growing the edited and non-edited plants under two conditions: a test condition incorporating the specified management practice and a control condition excluding the management practice;(e) measuring plant performance under each condition using one or more of: yield at maturity or a proxy indicator trait correlated with the management practice;(f) training, by supervised learning, a reward model using the measured performance data, wherein the reward model is structured to embed phenotype-specific outcome objectives;(g) generating, by the reward model, reward values for each edit combination based on the percent change in yield or proxy indicator between edited and non-edited plants under both conditions;(h) adjusting, by a reinforcement learning algorithm, one or more weights in the generative Al model based on the reward values to improve its capacity to generate genome edits that optimize yield under managed and unmanaged conditions; and(i) iterating the method steps any number of times to achieve the desired outcome objective.165. The method of claim 164, wherein the plant genotype is a parental genotype, a parental genotypic representations such as a genotype selected from a breeding population, a parental line genotype from a heterotic group, a SNP-encoded representation of a maize inbred line, a founder genotype from a mapping population, a commercial event genotype, an elite line genotype, a wild-type accession genotype, a transgenic line genotype, or a mutant line genotype.166. The method of claim 164 comprising: selecting one or more proposed effectors based on their predicted reward values.167. The method of claim 164 comprising: creating, by genome editing, one or more plants, plant tissues, or plant cells comprising the selected proposed effector.168. The method of claim 164 comprising: using the created plant, plant tissue, or plant cell in a breeding program, agronomic trial, or commercial product development pipeline.Docket # 212601-WO-SEC-l169. The method of claim 164, wherein the specified management practice comprises one or more of:(a) application of a nitrogen stabilizer; (b) use of microbial inoculants; (c) intercropping;(d) variable-rate fertilizer application; (e) no-till cultivation; (f) application of biostimulants; (g) precision irrigation; (h) fungicide or insecticide treatment;(i) or any combination thereof.170. A method of using a generative artificial intelligence (Al) model to optimize agricultural product performance and profitability, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a set of parental genotypic representations, an edit library comprising genome edits applicable to the parental genotypes, and an outcome objective comprising both agronomic performance metrics and economic performance metrics, wherein the trained generative Al model has been trained to propose combinations of parental selections and genome edits predicted to satisfy the outcome objective;(b) proposing, by the trained generative Al model, one or more tuples comprising a first parental genotype, a second parental genotype, a set of genome edits for the first parent, and a set of genome edits for the second parent;(c) generating, based on the proposed tuples, one or more edited populations of plants, each population comprising members with varying combinations of target edits in an otherwise uniform genetic background;(d) receiving, by a joint simulation system comprising a breeding pipeline digital twin and a seed-market digital twin, one or more entries, each entry comprising: i. the proposed parental genotypes and genome edits; ii. the outcome objective; and iii. genotypic and phenotypic data for the edited plant population;(e) simulating, by the breeding pipeline digital twin, multi-generational breeding outcomes, and simulating, by the seed-market digital twin, commercial lifecycle outcomes including regional adoption, pricing strategy, promotional spend, competitor product launches, and projected profitability;(f) generating, by the joint simulation system, a reward value based on an additive combination of agronomic performance and economic performance;Docket # 212601-WO-SEC-l(g) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to improve its capacity to generate parental selections and genome edit combinations that maximize both agronomic and economic outcomes.171. The method of embodiment 170, wherein the parental genotypic representations comprises: a genotype selected from a breeding population, a parental line from a heterotic group, a SNP-encoded representation of a maize inbred line, a founder genotype from a mapping population.172. The method of embodiment 170, wherein the agronomic performance comprises one or more of: (a) yield advantage over competitor benchmarks; (b) yield stability across multiple environments; (c) yield under abiotic stress conditions including drought, heat, or salinity; (d) yield under biotic stress conditions including pest or disease pressure; (e) biomass accumulation; (f) plant height, tillering capacity, or root mass; (g) flowering time or synchrony; (h) lodging resistance; or (i) nutrient-use efficiency or water-use efficiency.173. The method of embodiment 170, wherein the agronomic performance comprises resistance to one or more of: fungal pathogens, bacterial pathogens, viral pathogens, insect pests, or nematodes.174. The method of embodiment 170, wherein the agronomic performance comprises photosynthetic efficiency, carbon assimilation rate, or harvest index.175. The method of embodiment 170, wherein the economic performance comprises one or more of: (a) projected internal rate of return (IRR); (b) net present value (NPV);(c) gross margin per acre; (d) cost of goods sold (COGS); (e) return on R&D investment; (f) breakeven point; or (g) projected market share gain in a target maturity zone.176. The method of embodiment 170, wherein the economic performance comprises one or more of: (a) adoption rate among growers; (b) price elasticity of demand; (c) promotional spend efficiency; (d) royalty cost per unit sold; (e) time to market; or (f) product lifecycle revenue.177. The method of embodiment 170, wherein the economic performance comprises one or more of: (a) revenue volatility under climate scenarios; (b) supply chain resilience score; or (c) regulatory approval likelihood;Docket # 212601-WO-SEC-l

[0103] While the invention has been particularly shown and described with reference to a preferred embodiment and various alternate embodiments, it will be understood by persons skilled in the relevant art that various changes in form and details can be made therein without departing from the spirit and scope of the invention. For instance, while the particular examples below may illustrate the methods and embodiments described herein using a specific plant, the principles in these examples may be applied to any plant. Therefore, it will be appreciated that the scope of this invention is encompassed by the embodiments of the inventions recited herein and in the specification rather than the specific examples that are exemplified below. All cited patents and publications referred to in this application are herein incorporated by reference in their entirety, for all purposes, to the same extent as if each were individually and specifically incorporated by reference.EXAMPLES

[0104] The following are examples of specific embodiments of some aspects of the invention. The examples are offered for illustrative purposes only, and are not intended to limit the scope of the invention in any way. Efforts have been made to ensure accuracy with respect to numbers used (e.g., amounts, temperatures, etc.), but some experimental error and deviation should, of course, be allowed for.

[0105] Example 1 : Increasing gene expression two-fold above wild-type values

[0106] The methods and systems described herein can be used to alter gene expression, for example, through editing promoter sequences of target genes (FIG. 1 ). In one example, the methods and systems can be used to increase the expression of a gene above wild-type levels (FIG. 2).

[0107] In this example, the outcome objective is to double gene expression in leaf protoplasts of soybean through introducing SNPs and small INDELs to existing promoter sequences. The starting effectors for the generative model include wild-type promoter sequences spanning 2 kb upstream through the 5’ UTR chosen from the pangenome of soybean, although promoters of different lengths, and / or promoter types including variants may also be used as starting effectors. Several hundred wild type promoter sequences are chosen for proposed modifications, and sequences are input along with the outcome objective represented as the expression level in leaf protoplast,Docket # 212601-WO-SEC-l based on adding a base-2 logarithmic unit to the currently estimated wild-type expression level given in base-2 logarithmic units. The promoters used as starting effectors are selected from genes spanning a gradient of leaf expression from low to high over the expression orders-of-magnitude present in wild-type leaf.

[0108] The initializing base policy for the generative model is based on a genetic algorithm approach (FIG. 3). Selected wild-type promoter sequences being used as starting effectors undergo 75 generations of in silica mutation, recombination, and selection, wherein the fitness function during selection is specified by the negative of the absolute difference between the target and predicted expression of each candidate sequence. A further fitness penalty may be imposed that is proportional to the number of mutations relative to the wild-type sequence to favor parsimonious modifications. For each promoter, the generative model outputs 100 proposed effectors, although any desired number of proposed effectors may be selected for output. Proposals are in the form of promoter sequences, each with a modification to the starting promoter that includes one or more base pair substitutions and / or small indels. For each modification, the outcome objective expression value is recorded alongside the modified promoter sequence, the proposed effector in this example.

[0109] In the laboratory, modified promoter sequences (proposed effectors in this example) are synthesized and inserted into a vector with a barcoded reporter library (FIG. 4). Leaf protoplasts are transformed with the library pool and reporter expression for the promoters is quantified using RNA-sequencing. The resulting expression value is recorded alongside the synthesized promoter sequence and the outcome objective expression value.

[0110] Following collection of output expression data from the laboratory, the reward model is trained using supervised learning with the outcome objective, starting effector sequence, and modified proposed effector sequence as input (FIG. 5). The reward model is initialized with a pre-trained transformer-based backbone architecture, which may begin with the same backbone weights as the predictor model used for fitness evaluation in the genetic algorithm that provided the initializing sequence. The target signal for the reward model is the negative absolute difference between the target expression given by the outcome objective and the expression value generated byDocket # 212601-WO-SEC-l laboratory experiment. This allows the reward model to provide higher (less negative) rewards to modifications that more closely align with the specified outcome objective. Once trained, the reward model is run on combinations of wild type promoter sequence (starting effector), modified promoter sequence, and outcome objective expression, and it outputs reward values based on the predicted negative of absolute error between target expression (outcome objective) and true expression in leaf protoplast.

[0111] The reinforcement learning algorithm uses the trained reward model to provide feedback to the generative model, which serves as the policy function within the reinforcement learning framework. Proximal policy optimization (PPO) - a policy gradient method - is used to update the generative model. For each wild-type promoter (starting effector) and expression objective (outcome objective), the generative model samples a new modified version of the original sequence using beam search over predicted token probabilities from the generative model, where the generative model outputs probabilities of tokens representing segments of DNA sequence when given the initial promoter sequence and objective as inputs. This process is implemented using causal language modelling with a decoder-only transformer architecture. A value function, which is instantiated as a neural network with the same base architecture as the generative model, outputs estimates of the value of each state (starting promoter and outcome objective), subtracting it from the reward to calculate the advantage for the generative action. The PPO surrogate objective is then used to update the generative model with optimized weights for generative actions, while the squared error between estimated values and rewards provides a loss for updating the weights of the value function. The surrogate objective is given below, where the expectation:

[0113] Where Atis the advantage at iteration t , ngis the policy specified by the generative model proposed with weights 6 under the effector proposals atwith starting proposal and outcome objective specified by stand a pre-specified clipping value given by e.Docket # 212601-WO-SEC-l

[0114] In subsequent iterations of the method, the generative model outputs proposed effectors based on the reinforcement learning updated model (FIG. 3). These proposed effectors are synthesized and evaluated for expression within the laboratory.

[0115] Following each return of laboratory output data, the reward model weights are updated using the available additional training data. Data from previous rounds of proposed effectors may also be leveraged during the model updating process, in order to maximize the value of collected data (FIG. 6). After each further round of reward model training, it can facilitate further PPO-based training of the generative model and value functions (FIG. 7).

[0116] In addition to providing rewards during the reinforcement learning process, once trained, the reward model may be used to predict rewards for proposed effector sequences prior to laboratory testing. Then, the top ranked proposed effectors can be prioritized for testing, and the results can be used to update the reward model training (FIG. 8). The method loop can be iterated any number of times to reach the desired outcome.

[0117] Example 2: Create target expression pattern across tissues through HDR- directed CRISPR editing

[0118] The methods and systems described herein may also be used with the outcome objective to create a pattern of gene expression with specific target values in several different tissues and / or conditions using CRISPR-SDN3 type edits. CRISPR-SDN3 edits utilize homology directed repair (HDR) to incorporate a template sequence of several hundred base pairs in place of native sequences already present in the genome (FIG.9). Here, the outcome objective includes several tissues with target expression values in each tissue.

[0119] The initial promoter sequences for each target gene serve as starting effectors. The initializing base model generates approximately 10 promoter modification proposals (proposed effectors) for each of 100 wild-type promoter sequences, although any desired number of proposed effectors may be selected for output. Each promoter modification comprises a combination of base pair substitutions and indels that can encompass up to several hundred base pairs of the original promoter. Candidate sequences are chosen based on having varying, non-zero expression in the tissues ofDocket # 212601-WO-SEC-l interest (FIG. 3). An initial base policy using a genetic algorithm provides the initial set of proposed effectors. The algorithm is specified as described in Example 1 , with two differences. First, each outcome objective covers multiple tissues and may specify varying degrees of up or down-regulation in each tissue. Second, penalties based on the CRISPR-Cas system to be used is incorporated into the fitness function used for promoter designs. See, published PCT patent application, W02024006802.

[0120] To test the proposed effectors, template and guide RNA vectors are synthesized in the lab with the goal of achieving the desired promoter sequence using homologydependent repair (HDR) with the template (FIG. 4). Vectors containing the guides, templates, and Cas enzyme are used to create edits, and edited TO plants are regenerated. Genomic DNA from edited plants is sequenced to determine the modified promoter sequences. TO plants are self-pollinated to create T1 plants, and T1 plants containing edited promoters are self-pollinated to generate homozygous edited T2 plants for expression profiling. RNA from target tissues is extracted, and gene expression is assayed using qRT-PCR, with primers designed for each target gene.

[0121] Outcome data from this experiment includes the mRNA expression in each tissue and allele sequence for each edit. However, because perfect promoter swap events occur in less than 10% of attempted CRISPR SDN3 edits, the outcome data also includes results from a combination of on-policy (target) and off-policy (non-target) edits. While the on-policy edits provide information on the results of model proposals, the off- policy edits - comprising various small indels from non-homologous end joining (NHEJ) and imperfect HDR events, off-policy events can still provide useful information on the causal impact of variation at target promoters. In both cases, the modified promoter sequence are determined by sequencing the promoters of edited plants in the laboratory.

[0122] The outcome data for each edit, both on-policy and off-policy, is then used to train a reward model using supervised learning (FIG. 5). This reward model is structured so that each tissue-based outcome objective is a separate example, with an embedding specific for each tissue type and the target expression in that tissue specified as a Iog2 expression value. As in Example 1 , the target reward value is based on the negative absolute difference between the target and assayed expression values. The output ofDocket # 212601-WO-SEC-l the reward model is used by the advantage actor-critic (A2C) reinforcement learning algorithm to train the generative model, though any variation of policy gradient reinforcement learning methods may be used. The generative model in this case may be based on a selective state space model such as MAMBA, which can be computationally more efficient than multi-head self-attention based transformers. Input to the generative model includes the original promoter sequence (starting effector) and embeddings corresponding to target tissues, along with their target expression values (outcome objective). An embedding corresponding to the desired CRISPR-Cas editing system may also be provided to allow for Cas-aware constraints. The generative model then proposes additional rounds of proposed effectors comprising additional edits to previously proposed promoter sequences (FIG. 3). During training, combinations of the starting effector sequence, proposed effector sequence, and tissue-contextualized target expression (outcome objectives), are provided to the reward model for evaluation. The sum of rewards from multiple tissue types may be additively combined to produce a composite reward value. Additionally, a deterministic Cas-specific penalty is additively included in the reward values, penalizing promoter modifications that are less feasible based on the properties of the target editing system.

[0123] Following training of the generative model, the reward model is updated with additional training data as subsequent rounds of on-policy and off-policy edits are made and evaluated in the laboratory (FIG. 6). As in Example 1 , the reward model may be used to accept as input proposed effectors along with starting effectors and the outcome objective to predict reward values for untested edits. Proposed promoters can then be ranked based on predicted reward value, and promoters with the highest reward value may be prioritized as candidates for additional rounds of laboratory editing.

[0124] Example 3: Utilize public literature and previous trait data to change phenotypes through editing expression

[0125] The methods and systems described herein may also be applied to contexts where the outcome objective is to alter an agronomic phenotype through changes to gene expression (FIG. 10). In this example, the outcome objective is to reduce plantDocket # 212601-WO-SEC-l height by 2 cm by changing the expression of one or more genes in a hormone signaling pathway using CRISPR-CAS9 editing.

[0126] To provide biological context to the experiment, an external context module utilizing retrieval augmented generation (RAG)-enabled large language model (LLM) is used upstream of the generative model (FIG. 3). The external context module uses text and data from multiple sources as input, along with the outcome objective and starting effector(s), to provide embedding of the scientific context. In this case, published texts, public databases, and proprietary data are used to enable the external context module to embed information on the relationship between the expression of the gene and plant height, such that an outcome objective provided to the external context module in terms of plant height is translated into an objective on gene expression level.

[0127] The starting effector(s) are the promoter sequence(s) spanning 2 kb upstream through the 5’ UTR for the target gene(s), for example Dwarf 8, whose expression can be modulated to reduce plant height, although promoters of different lengths, and / or promoter types including variants may also be used as starting effectors.

[0128] The initializing base policy for the generative model for this example is the previously trained generative model resulting from the completion of the experiment proposed in Example 2. The generative model generates 50-100 proposed effectors of target promoter sequences containing edits of SNPs and small insertions or deletions relative to the starting promoter sequence(s) (starting effector(s), although any number of proposed effectors can be created.

[0129] The proposed effectors and starting effectors are then sent to the laboratory where guide RNA vectors are designed and synthesized with the goal of achieving promoters with the desired edits using CRISPR-CAS9 (FIG. 4). Vectors containing guides and the CAS enzyme are used to create edits, and guides may be multiplexed to create multiple edits in individual promoters at once. TO plants are regenerated, and the promoters for each edited plant are sequenced to determine the modified promoter sequence. TO plants are self-pollinated and heterozygous T1 plants are self-pollinated to create T2 plants. T2 plants segregating for homozygous, heterozygous, and nonedited promoters are grown to maturity and plant height is measured along with expression of the target gene in the target tissue.Docket # 212601-WO-SEC-l

[0130] Next, supervised learning is used to train a reward model on inputs of the starting promoter sequence (starting effector), modified promoter sequence, and outcome objective (FIG. 5). The training dataset also utilizes a training signal that is calculated as the Iog2 change in expression values along with the difference in plant height measurements between homozygous edited and non-edited sibling plants from the laboratory experiment. The reward model outputs a reward value representing how closely the laboratory results match the outcome objective such that proposed effectors resulting in a higher similarity to the outcome objective have a higher reward value.

[0131] The output of the reward model is then used by the reinforcement learning algorithm. Any suitable policy gradient algorithm may be used, such as proximal policy optimization (PPO), advantage actor-critic (A2C), or REward Increment = Non-negative Factor x Offset Reinforcement x Characteristic Eligibility (REINFORCE). In this example, the Trust Region Policy Optimization (TRPO) reinforcement learning algorithm is used to update the weights of the generative model. Following estimation of the advantages using the value function and calculation of the policy gradient, each update of the weights is proposed based on a scaled version of the vector specified by where Hkis the Hessian matrix of the Kullback-Liebler (KL)divergence between the current weights and the kth proposal and gkis the policy gradient at that proposal, and 8 is a specific constant used to determine the size of the trust region. A backtracking line search is then used to find the maximum of the Taylor series approximate surrogate objective function gT(0 - 0k~) for current weights 0 and proposed weights 0k, under the TRPO constraint for the Taylor-series approximation of the KL-divergence, -(0 - 0k)TH (0 - 0k) < 8 This process is iterated throughout the course of reinforcement learning, until convergence, to

[0132] Following the generative model's updated training, additional proposed effectors for the same starting effectors or for new starting effectors are generated to more closely approximate the outcome objective plant height change. As in Examples 1 and 2, the new proposed effectors may either be input directly into the reward model to expedite reinforcement learning (FIG. 7), they may be sent to the laboratory to generateDocket # 212601-WO-SEC-l new data to update the reward model and the generative model via reinforcement learning (FIG. 8), or both.

[0133] Example 4: Utilizing pre-trained models to generate proposals for agricultural applications

[0134] Once trained to predict causal relationships between promoter sequence and gene expression, the generative model can be used to generate proposed effectors, the reward model can be used to rank the proposed effectors, and the top proposals can be used to generate edited plants directly for integration into breeding programs.

[0135] In this example, the outcome objective is to reduce plant height by 2 cm by changing the expression of a different gene in the same hormone signaling pathway as in Example 3 using CRISPR-CAS9 editing.

[0136] The external context module is used upstream of the generative model (FIG. 3) and receives as input additional data about the gene of interest from public and proprietary databases. The external context module embeds the information about the relationship between gene expression and plant height along with the outcome objective and starting effector sequence for use by the generative model.

[0137] The starting effector is the promoter sequence spanning 2 kb upstream through the 5’ UTR for the target gene whose expression can be modulated to reduce plant height, although promoters of different lengths or variants may also be used as starting effectors.

[0138] In this example, the generative model trained in Examples 2-3 is used to create 100 proposed effectors comprising promoter sequences containing edits of SNPs and small insertions or deletions relative to the starting promoter sequence (starting effector). The proposed effectors are then input into the reward model trained in Example 3. The reward model outputs the predicted reward value for each of the proposed effectors. These proposals can be ranked, and any number of promoters (proposed effectors) with the highest reward values move forward into laboratory production (FIG. 11 ).

[0139] The proposed effector sequences and starting effector sequences are received by the laboratory, and scientists design and synthesize guide RNA vectors with the goalDocket # 212601-WO-SEC-l of achieving promoters with the desired edits using CRISPR-CAS9 (FIG. 4). Vectors containing guides and the CAS enzyme are used to create edits. TO plants are regenerated, and the promoters for each edited plant are sequenced. Plants with edits that match the proposed effectors are grown to maturity and self-pollinated to create T1 plants. T1 plants are grown to maturity and self-pollinated. T2 plants that are homozygous for the edited promoter are then validated for plant height changes.

[0140] Edited plants with reduced height are incorporated into breeding programs for elite variety development.

[0141] Example 5: Utilizing generative Al with reinforcement learning to select edits for improved yield

[0142] One application of this method is to utilize the generative Al model to select from a pool of possible edits with the goal of influencing a phenotype such as yield in a crop plant such as maize, although this method can be used on any combination of phenotype and crop species.

[0143] In this example, the outcome objective is to increase average maize yield for a population derived from two specific parents by 10 bushels per acre by introducing some combination of genome edits. The generative model accepts as input this outcome objective, a SNP representation of the starting genotype for the parents, and an edit library consisting of a list of all possible edits that may be introduced to the parental genomes.

[0144] The initializing base policy for the generative model randomly selects 50 combinations of 10 proposed edits and 50 combinations of 50 proposed edits from the edit library for a total of 100 edit combinations for each parental genotype, although any number of proposed edits and edit combinations may be used. The initial policy may also select a small number of targeted edits and include additional random edits.

[0145] The proposed edit combinations are then sent to the lab where multiplexed edits are created and doubled haploid populations are generated. This results in a population of maize plants with varying combinations of homozygous target edits in an otherwise uniform genetic background. The doubled haploid population is then used to create a hybrid population through crossing to one or more members of the opposite heterotic group. These hybrids are planted in a representative set of locations in the targetDocket # 212601-WO-SEC-l population of environments for yield trialing. Yield values are processed through a mixed linear model analysis, and best linear unbiased predictors (BLUPs) for general combining ability (GCA) are calculated for the edited inbred lines. Each doubled haploid line in the population is genotyped to determine which of the proposed edits are present in each line. In parallel, hybrid populations from unedited parents are generated to calculate GCA of the parents. The output data from the laboratory and field experiment is the change in GCA for yield between the edited and non-edited plants. The specific editing and breeding methods depend on the crop species, for example yield in a varietal crop such as soybean is tested in inbred rather than hybrid plants.

[0146] The field data is then used to train the reward model via supervised learning. The reward model accepts as input the outcome objective, the starting genotype of the parents represented with SNPs, and the set of edits introduced into the doubled haploid plants. The target signal for the reward model is the change in GCA between edited lines and the GCA of the parents, with a positive number indicating a yield gain in edited plants and a negative number indicating a yield loss. Once trained, the reward model is run on all combinations of achieved edits, and it outputs reward values based on yield change.

[0147] As detailed above in Example 1-3, the outputs of the reward model are then used by the reinforcement learning algorithm to produce updated weights for the generative model, and the loop is iterated as many times as desired.

[0148] Example 6: Using generative Al to select edits given context from variable sources

[0149] In applications aiming to propose edits for breeding programs, it is common to have external context about the proposed edits, so an edit library context module may be implemented (FIG 13). The edit library external context module has a RAG-enabled LLM architecture as described in Example 3, but it is specifically structured to provide context for the edits available in the edit library. This context module accepts as input any known information about the edits, such as the gene associated with the edit, any known effects of the edit on any phenotype in any genetic background, and target composition information such as the type of edit or the toxicity of the edit. This information may be made available to the model by chunking scientific literature text,Docket # 212601-WO-SEC-l previous internal study data, and details on gene functions from existing knowledge bases into overlapping textual documents. The individual documents are then processed through the LLM to produce an embedding vector for each that can be stored in a vector database. When queried, the LLM can search the vector database with an appropriate similarity function (e.g. cosine similarity function) to retrieve information relevant to the target genes.- This contextual information is then combined by the LLM into a vector embedding -used to inform the generative model, as in Example 3.

[0150] In this example, the outcome objective is to both increase yield by 10 bushels per acre and increase lodging resistance by 10%. The parental genotypes are input as SNP representations of the genetic backgrounds. These SNP representations can be processed into their own embedding vectors that serve as vectorized encodings of the genetic space. A library of available edits is available, where each edit is associated with a gene identity and the expected molecular impact of the edit on the gene (e.g. knockout, increase in expression, decrease in expression). The gene identity is used to retrieve additional contextual evidence for gene function, including molecular pathway and references in the scientific literature. The SNP representation informs both the edit library context module and is used as input for the generative model and downstream steps in the method. The edit context information is encoded into an embedding vector that can be input along with the genetic encoding and objective. The initializing base policy for the generative model for this example is the previously trained generative model resulting from the completion of the experiment proposed in Example 5. The generative model generates proposals consisting of 50 combinations of 10 proposed edits and 50 combinations of 50 proposed edits from the edit library for a total of 100 edit combinations for each parental genotype.

[0151] These proposed edit combinations are then sent to the lab then the field. As in Example 5, multiplexed edits are introduced to the parental genomes, doubled haploid populations are created and genotyped for edits, and these populations are used to create hybrid populations. The hybrid populations are then tested for yield to calculate change in GCA relative to unedited parents and lodging to calculate the percent change in lodging resistance between edited and unedited parental lines.Docket # 212601-WO-SEC-l

[0152] The outcome data for each DH is then used to train the reward model using supervised learning (FIG. 5). As in Example 2, the reward model is structured so that each phenotype-based outcome objective is a separate example, with an embedding specific for each phenotype.

[0153] The reward model outputs reward values for all combinations of edits, and these reward values are used by the reinforcement learning algorithm to update weights within the generative model in a direction producing an increase in expected total reward.

[0154] Example 7: Utilizing a breeding pipeline digital twin to train a generative Al model for plant breeding

[0155] In another example, the outcome objective is to increase yield by 10 bushels per acre and decrease plant height by 2 cm in the second advancement stage of the breeding pipeline (FIG. 14). To accomplish this objective, the reward model is replaced with a breeding pipeline digital twin to output a reward value based on actual or predicted performance in the second advancement stage of the breeding pipeline (FIG. 15). This application allows for the consideration not only of population development, but also of the advancement of lines through the breeding pipeline.

[0156] The breeding pipeline digital twin is a simulation environment covering in silico approximations of plant breeding pipeline processes spanning from population development through pre-commercial hybrid or varietal multi-environment testing. Prior to utilizing the digital twin in the generative Al pipeline, it is first initialized with existing breeding data, including selection choices (eg, hybrid parent choices, advancement decisions, and breeding populations), genotype data, and phenotype data for each stage of the pipeline in previous years. Each run of the digital twin environment simulates breeding pipeline outcomes over the course of 5 or more generations. Simulated pipeline outcomes include the generation of new inbred genotypes through population development, the generation of hybrids through crossing of inbred lines, the observation of phenotypic values within yield trial experiments, and calculation of GCA BLUPs and genomic estimated breeding values (GEBVs) based on the trial data. The observed data from previous years is used to initiate the genotypes present at each simulated stage and provide a basis for the simulated mappings of genotypes and locations onto phenotypes for the generation of phenotypic data.Docket # 212601-WO-SEC-l

[0157] To provide contextual information for selection agents trained to conduct germplasm- advancement within the digital twin, a breeding context module is implemented (FIG. 15). The structure of this module is a similar RAG-enabled LLM as the edit library context module, but incorporates information about breeding strategies, predictive models, product concepts, business constraints, and algorithmically- optimized selection strategies, although any breeding context may be included depending on the experimental design. Textual information on business rules around diversity, targets for product concepts, and the results of phenotypic and advancement predictive models can be loaded into a vector database and queried based on cosine similarity to the genetic encodings of the candidate germplasm in a similar manner to the querying of gene edit information in Example 6. The digital twin allows for the processes of the breeding pipeline to influence the reward value used for reinforcement learning for the selection of edits to introduce. Whereas the reward function of Example 6 directly used observed phenotypic data, the reward function leveraging the digital twin is calculated based on the simulated increase in yield and decrease in plant height within candidate hybrids after 5 or more generations of running the breeding pipeline, instantiated with the previous year’s data.

[0158] Beyond selecting edits to incorporate into breeding lines, the generative Al model can also be used to select the initial parents that are modified. In this case, the generative model and edit context module are given a pool of possible parents represented as SNP genotypes. In this example, 50 combinations of 50 edits with two parental genotypes from contrasting heterotic groups are selected.

[0159] The edits and parents are then sent for lab and field testing. As in Examples 5 and 6, edits are produced and doubled haploid populations are created and genotyped such that edits segregate within each parental line. The DH population is used to generate a hybrid population and tested for yield and plant height. The top performing lines are advanced into the first advancement generation of hybrid testing in more environments, and the top performing lines in the first generation are advanced to the second advancement generation of hybrid testing in additional environments. The plant height and yield of the hybrids in the second advancement generation are assessed. The phenotypes in each generation are used to parameterize the digital twin for futureDocket # 212601-WO-SEC-l iterations of the pipeline. Then, the digital twin outputs a reward value representing the average phenotype after the simulated generations.

[0160] The reward value is then used in reinforcement learning to update the weights in the generative model. The updated weights are used for additional iterations of the pipeline to select new combinations of edits and parental genotypes.

[0161] Example 8: Utilizing generative Al with reinforcement learning to select edits for increased protein expression in a maize event

[0162] One application of this method is to utilize the generative Al model to propose edits with the goal of influencing a phenotype such as protein expression in a crop plant such as maize, although this method can be used on any combination of phenotype and crop species.

[0163] In this example, the outcome objective is to increase average protein expression of an insecticidal protein in an existing maize event by introducing some combination of genome edits into one or more of the event’s regulatory elements, which may include promoter(s), intron(s), and terminator(s) (FIG. 16). The outcome objective may include both minimum and maximum target protein expression values. The generative model accepts as input this outcome objective, the full sequence of the target sequence(s) of interest, and an external context module as in Example 3 that is also informed with insecticidal protein variants and effects. Non-limiting examples of maize events that may have their regulatory elements edited to increase expression of the event’s insecticidal protein are shown in Table 1.

[0164] Table 1 Non-limiting table showing examples of insecticidal events in several crop species.Docket # 212601-WO-SEC-lDocket # 212601-WO-SEC-l

[0165] The initializing base policy for the generative model randomly selects 50 proposed HDR-directed promoter swap edits and 50 combinations of 10 proposed edits for a total of 100 edit combinations for each maize event regulatory element, although any number of proposed edits and edit combinations may be used. Proposed edits may be proposed using any of the mechanisms described in Example 1-3.

[0166] The proposed edit combinations are then sent to the lab where multiplexed edits are created and tested for protein expression / concentration of the insecticidal protein in stable maize plant transformants or maize transient expression assays. This results in a population of maize plants / cells with varying combinations of homozygous target edits in an otherwise uniform genetic background. Protein concentration of the insecticidal protein is determined by a mass spectrometry-based protein identification method using extracted protein lysates from infiltrated leaf tissues (Patterson, (1998) 10(22):1 -24, Current Protocol in Molecular Biology published by John Wiley & Son Inc), or by other methods known in the art, including Western blotting or qRTPCR to determine transcript levels. Each edited maize event plant in the population is genotyped to determine which of the proposed edits are present in each plant. In parallel, maize event plants from unedited events are grown and the protein concentration of the insecticidal protein is determined for comparison to that in the edited event plants. The output data is the change in protein concentration of the insecticidal protein between the edited and nonedited plants. The specific editing and protein concentration analysis depends on the crop species.

[0167] The protein concentration data from the edited event plants is then used to train the reward model via supervised learning. The reward model accepts as input theDocket # 212601-WO-SEC-l outcome objective, the starting effector, the external context module, and the set of edits introduced into the regulatory element edited maize event plants. The target signal for the reward model is the change in protein concentration of the insecticidal protein between edited lines and the protein concentration of the unedited event plants, with a positive number indicating a protein concentration gain in edited plants and a negative number indicating a protein concentration loss. Once trained, the reward model is run on all combinations of achieved edits, and it outputs reward values based on protein concentration change.

[0168] As detailed above in Example 1-3, the outputs of the reward model are then used by the reinforcement learning algorithm to produce updated weights for the generative model, and the loop is iterated as many times as desired.

[0169] Example 9: Utilizing generative Al with reinforcement learning to select edits for improved insecticidal activity of an insecticidal gene

[0170] One application of this method is to utilize the generative Al model to propose edits with the goal of influencing a phenotype such as insecticidal activity in a crop plant such as maize, although this method can be used on any combination of phenotype and crop species.

[0171] In this example, the outcome objective is to improve insecticidal activity of an insecticidal gene by introducing some combination of genome edits to the insecticidal gene. Improvement of insecticidal gene may occur by increasing protein stability, increasing protein activity, or by decreasing phytotoxicity, for example (FIG. 17). The generative model accepts as input this outcome objective, the full sequence of the target sequence(s) of interest, and an external context module as in Example 3 that is also informed with insecticidal protein variants and effects. Non-limiting examples of events that may have their insecticidal gene sequence edited to improve insecticidal activity of the event’s insecticidal protein are shown in Table 1 .

[0172] The initializing base policy for the generative model randomly selects 50 proposed HDR-directed promoter swap edits and 50 combinations of 10 proposed edits from the edit library for a total of 100 edit combinations for each insecticidal gene sequence, although any number of proposed edits and edit combinations may be used.Docket # 212601-WO-SEC-lThe initial policy may also select a small number of targeted edits and include additional random edits.

[0173] The proposed edit combinations are then sent to the lab where multiplexed edits are created and tested for insecticidal activity in a bacterial expression system, cell-free expression system, protoplast expression system, transient expression assays in various plant tissue, or in stable plant transformants, for example. The percent change in insecticidal activity between edited and non-edited plants is then used as outcome data.

[0174] The outcome data is then used to train the reward model via supervised learning. The reward model accepts as input the outcome objective, the full sequence of the target sequence(s) of interest, the set of edits introduced into the edited gene sequence, and the external context module. The target signal for the reward model is the change in activity between the edited sequences and the activity of the parents, with a positive number indicating an activity gain in edited plants and a negative number indicating an activity loss. Once trained, the reward model is run on all combinations of achieved edits, and it outputs reward values based on change in protein activity.

[0175] As detailed above in Example 1-3, the outputs of the reward model are then used by the reinforcement learning algorithm to produce updated weights for the generative model, and the loop is iterated as many times as desired.

[0176] Example 10: Decrease phytotoxicity while increasing insecticidal activity

[0177] In this example, the outcome objective is to decrease phytotoxicity in the host plant and improve insecticidal activity of an insecticidal gene by introducing some combination of genome edits to the insecticidal gene (FIG. 18). The outcome objective may include specific target ranges for both phytotoxicity and insecticidal activity. The generative model accepts as input this outcome objective, the full sequence of the target sequence(s) of interest, and an external context module as in Example 3 that is also informed with insecticidal protein variants and effects.

[0178] The initializing base policy for the generative model randomly selects 50 combinations of 2 proposed edits and 50 combinations of 10 proposed edits from the edit library generated in Example 8 and Example 9 for a total of 100 edit combinationsDocket # 212601-WO-SEC-l for each insecticidal gene sequence, although any number of proposed edits and edit combinations may be used. The initial policy may also select a small number of targeted edits and include additional random edits.

[0179] The proposed edit combinations are then sent to the lab where multiplexed edits are created and tested for insecticidal activity in a bacterial expression system, cell-free expression system, protoplast expression system, transient expression assays in various plant tissue, or in stable plant transformants, for example. The percent change in insecticidal activity and phytotoxicity of the host plant between edited and non-edited plants is then used as outcome data.

[0180] The outcome data is then used to train the reward model via supervised learning. The reward model accepts as input the outcome objective, the full sequence of the target sequence(s) of interest, the set of edits introduced into the edited gene sequence, and the external context module. The target signals for the reward model include 1 ) the change in activity between the edited sequences and the activity of the parents, with a positive number indicating an activity gain in edited plants and a negative number indicating an activity loss, and 2) the change in phytotoxicity between the edited and non-edited plants, with a positive number indicating lower phytotoxicity and a negative number indicating increased phytotoxicity. Once trained, the reward model is run on all combinations of achieved edits, and it outputs reward values based on change in protein activity and phytotoxicity.

[0181] As detailed above in Example 1-3, the outputs of the reward model are then used by the reinforcement learning algorithm to produce updated weights for the generative model, and the loop is iterated as many times as desired.

[0182] Example 11 : Using generative Al to propose edits for yield boosts in specific management practices

[0183] Management practices and agronomic inputs are critical components for modern agriculture. In some examples, the methods and systems described herein may be used to select edits with the goal of improving yields under specific management practices, including but not limited to any combination of biological inputs, chemical inputs and / or intercropping, while maintaining yield under non-managed field conditions.Docket # 212601-WO-SEC-l

[0184] In this example, the outcome objectives are to (1 ) increase plant yield when nitrogen stabilizer is applied, and (2) maintain plant yield when nitrogen stabilizer is not applied (FIG. 19). These outcome objectives are loaded into the Edit Library Context Module described in Example 6 along with the typical inputs of gene function, edit effects, target composition, the edit library, and SNP representations of all parent genotypes. In this example, an additional management component is input into the edit context library. The management data input includes both publicly available literature on the effects of management practices on plant traits and proprietary data produced internally.

[0185] The initializing base policy for the generative model for this example is the previously trained generative model resulting from the completion of the experiment proposed in Example 6. The generative model generates proposals consisting of 50 combinations of 10 proposed edits and 50 combinations of 50 proposed edits from the edit library for a total of 100 edit combinations for each parental genotype.

[0186] These proposed edit combinations are then sent to the lab and then the field. Multiplexed edits are introduced to the parental genotypes and doubled haploid populations are created and genotyped for edits.

[0187] The doubled haploid populations, along with non-edited controls, are then grown in two controlled environment conditions. In the test condition, the soil will be supplemented with nitrogen stabilizer, while in the control conditions, soil will not be amended. Plants will then be assessed for nitrogen uptake efficiency by either testing yield at maturity or using a proxy indicator trait that can be assessed at an earlier stage of plant development. The outcome data will consist of the percent change in yield or the proxy indicator between edited and non-edited plants under no management and the percent change in yield or the proxy indicator between edited and non-edited plants grown under control conditions.

[0188] The outcome data will then be used to train the reward model using supervised learning. The external context model is also used to provide context to the reward model. As in Example 2, the reward model is structured so that each phenotype-based outcome objective is a separate example, with an embedding specific for each phenotype.Docket # 212601-WO-SEC-l

[0189] The reward model outputs reward values for all combinations of edits, and these reward values are used by the reinforcement learning algorithm to update weights within the generative model in a direction producing an increase in expected total reward.

[0190] The method loop can be iterated any number of times to reach the desired outcome objective, and any edited plants may be used for commercial product development.

[0191] Example 12: Integrating a seed-market digital twin to optimize profit under competitive pressure

[0192] The methods and systems disclosed above can be further extended so that the reward function reflects not only agronomic performance but also projected profitability in a competitive seed marketplace. In this example, the outcome objective is to generate a portfolio of maize hybrids that (i) raises on-farm yield by 8 bu / ac over the competitor commercial leader in the same maturity zone and (ii) achieve an internal rate of return (IRR) of at least 15% when launched into the US Corn Belt.

[0193] A seed market digital twin module is appended to the breeding pipeline digital twin of Example 7 (FIG. 20). The module simulates (a) demand elasticity by maturity zone and management segment, (b) competitor product launches based on historical data derived from public trial data and reported R&D pipelines, (c) price-volume tradeoffs constrained by channel strategy and grower and adoption curves, and (d) cost of goods sold (COGS) as a function of product environment, seed quality, and trait royalty structure.

[0194] Historical sales, competitor share reports, and agronomic performance trial summaries are chunked into documents and embedded in a vector database. A RAG- enabled market context module retrieves and encodes these data into an embedding that is concatenated to the genetic and phenotypic embeddings already used by the digital twin.

[0195] Each run of the joint twin simulates 20 years, including the full breeding pipeline process (population development, hybrid development, and advancement) and the commercial lifecycle trajectory of the products. Lifecycle demand and invoice data are parameterized from historical commercial sales patterns. The digital twin modelsDocket # 212601-WO-SEC-l regional adoption, pricing strategy, promotional spend, and competitor reactions through simulated non-cooperative game-theoretic dynamics. For example, a Stackelberg or Bertrand competition model may be used to produce yearly cash flows and discounted profit metrics.

[0196] Inputs to the generative model include the SNP encodings of parental lines available for editing, the contents of the edit library, the outcome objective embedding (yield and IRR), and the market context embedding describing competitor benchmarks. The initializing base policy mirrors Example 6, proposing 50 edit sets of ten edits per parent. Parent choice is also treated as an action, such that each proposal is a tuple of (parent A, parent B, edit set A, edit set B). Rewards are produced based on an additive combination of the projected IRR and the differences in average on-farm yield between the simulated hybrids advanced to commercial and the commercial leaders in the same maturity zone.

[0197] As in examples 5-7, multiplex edits are produced, DH lines are generated, hybrids are created, and multi-environment trials supply performance data to parameterize the digital twin components with up-to-date data to ensure simulation fidelity. COGS inputs - such as seed size distribution, flowering rate, and germination percentage - are collected during seed increase. Competitor trial summaries for the same years are ingested to refresh the market context database.

[0198] Using the combined pipeline and marketplace digital twin, the reward value is used in reinforcement learning to update the weights of the generative model. The updated weights are then used to generate new combinations of parents and edit sets for maximization of both agronomic performance and economic payoff.

[0199] Detailed Descriptions of methods for plant protection

[0200] Agrobacterium-Med\ated Stable Transformation of Maize

[0201] For Agrobacterium-mediated maize transformation of the edited event, the method of Cho may be employed (M. J. Cho et al., Plant Cell Rep. 33, 1767-1777 (2014)) using PMI with selection. Briefly, immature embryos (lEs) are isolated from maize and infected with an Agrobacterium suspension containing vector constructs for the expression of the edited and non-edited event. lEs and Agrobacterium are co-cultivated on solid mediumDocket # 212601-WO-SEC-l in the dark at 21 °C for 3 days and subsequently transferred to resting medium without selection agent but supplemented with antibiotic to eliminate Agrobacterium. IES are transferred to the appropriate resting medium for 10-11 days before transferring to PMI medium containing section agent with antibiotic(s). Multiple rounds of selection are performed until sufficient quantities of tissue are obtained. Regenerative green tissues are transferred to medium such as PHI-XM medium (E. Wu et al., In Vitro Cell. Dev. Biol. Plant 50, 9-18 (2014)) with selection. Shoots are transferred to tubes containing MSB rooting medium for rooting and plantlets transplanted to soil in pots in the greenhouse. Leaf discs are excised from transformed maize plants and tested for protein concentration of the insecticidal protein by quantification methods known in the art including mass spectrometry, western blotting, and qRTPCR.

[0202] Agrobacterium-mediated transient assay

[0203] The agro-infiltration method of introducing an Agrobacterium cell suspension to plant cells of intact tissues so that reproducible infection and subsequent plant derived transgene expression may be measured or studied is well known in the art (Kapila, et. al., (1997) Plant Science 122:101 -108). Briefly, excised leaf disks are agro-infiltrated with normalized bacterial cell cultures of test and control strains. After four days leaf disks are analyzed for protein expression using mass spectrometry and / or western blot analysis or for transcript levels using qRTPCR.

[0204] E. Coli Expression of edited genes

[0205] The polynucleotde sequence is cloned into various E. coli expression vectors, for example, pET28 with N-His tag. E. coli cells are grown overnight at 37° C with 40 ug / ml Kanamycin selection and then inoculated to a fresh 2xYT medium (1 :25) and further grown to an optical density of about 0.8. At that point cells are chilled in the presence of 1 mM ITPG and further grown at 16° C. for 16 hours to induce protein expression. Cell pellets of E. coli cultures are suspended in % Bper II lysis buffer containing protease inhibitor, lysozyme and endonuclease. The crude lysate is cleared by centrifugation and the top phase of clear lysate is used for insect feeding assays. The E. coli expressed proteins are purified by immobilized metal ion chromatography using Ni-NTA agarose (Qiagen, Germany) according to the manufacturer’s protocols and recombinant protein expression is confirmed on SDS protein gel.Docket # 212601-WO-SEC-l

[0206] Insect feeding assays / Diet-based assays

[0207] Insecticidal activity bioassay screens are conducted on the clarified extract from bacterial samples to evaluate the effects of its proteins on a Coleopteran species (Western corn rootworm (Diabrotica virgifera; WCRW)), a variety of Lepidoptera species (European corn borer (Ostrinia nubilalis; ECB), corn earworm (Helicoverpa zea; CEW), black cutworm (Agrotis ipsilon; BCW), fall armyworm (Spodoptera frugiperda; FAW), Soybean looper (Pseudoplusia includens; SBL) and Velvetbean caterpillar (Anticarsia gemmatalis; VBC), for example.

[0208] Insect feeding assays: Coleopteran pests

[0209] Coleopteran in-vitro feeding assays are conducted on an artificial agar-based diet (Southland Products Inc., Lake Village, AR) in 96 well format. The diet (65 pL) is mixed with a clarified extract (15 pL). Control wells are made of diet (65 pL) and 15 pL of Lysis buffer. Three to six neonate Western corn rootworm (Diabrotica virgifera; WCRW) larvae are placed into each well to feed for 72 hours at 27°C. The effects of the protein on the larvae are scored numerically as dead (3), severely stunted (2) (little or no growth but alive and equivalent to a 1st instar larvae), stunted (1 ) (growth to second instar but not equivalent to controls), or normal (0) (similar to larvae feeding on a diet with only buffer applied). Each sample is assayed on WCRW.

[0210] Insect feeding assays: Lepidopteran pests

[0211] Lepidopteran in-vitro feeding assays are conducted on an artificial agar-based diet (Southland Products Inc., Lake Village, AR) in 96 well format. The diet (75 pL) is mixed with clarified sample (25 pL). Control wells are made of diet (75 pL) and 25 pL of lysis buffer. Two to five neonate larvae are placed into each well to feed for 72 to 96 hours at 27°C. The effects of the protein on the larvae are scored numerically as dead (3), severely stunted (2) (little or no growth but alive and equivalent to a 1st instar larvae), stunted (1 ) (growth to second instar but not equivalent to controls), or normal (0) (similar to larvae feeding on diet with only buffer applied). Each sample is assayed on ECB, CEW, BCW, FAW, SBL, and VBC, for example.

[0212] Maize Protoplast Assay

[0213] The maize protoplast data is designed to rank constructs by their level of phytotoxicity. In the assay, maize leaf mesophyll protoplasts are transfected with aDocket # 212601-WO-SEC-l plasmid DNA construct containing two expression cassettes. The first expression cassette is a fluorescent protein, such as but not limited to ZS-GREEN1. The fluorescence is used to identify transfected cells to enable cell counting over a time course. The second cassette contains a different promoter driving a gene of interest, such as an insecticidal gene of interest. No components are shared between these cassettes. In one embodiment, the fluorescent protein or the gene of interest may be driven by a constitutive promoter. In another embodiment, the fluorescent protein or the gene of interest may be driven by a mesophyll or a leaf-specific promoter.

[0214] For this assay, it is believed that high levels of phytotoxicity correlate with high levels of cell death. Cell death is measured by counting fluorescing cells over a time course. Maize protoplasts are transfected, and cell counts are calculated once every 8 hours over the course of 104, 112, or 160 hours. Note that the fluorescence used to count cells is not visible until about 8 hours post-transfection.

[0215] The correlation between levels of phytotoxicity and levels of cell death is demonstrated using a panel of reference proteins having a range of known phytotoxicity from none, low to moderate, moderate, to high.

[0216] Transient expression in Bush-bean

[0217] The corresponding recombinant polynucleotide encoding the edited protein is cloned into a transient expression system under the control of the viral promoter dMMV (Dey, et. al., (1999) Plant Mol. Biol. 40:771-782). The Agrobacterium strains containing each of the constructs are infiltrated into leaves as described in Kapila, et. al., (1997) Plant Science 122:101-108, for example. Briefly, the unifoliate leaves of bush bean (common bean, Phaseolus vulgaris) are agro-infiltrated with normalized bacterial cell cultures of test and control strains. The different transiently expressed constructs comprising edited proteins observed for phytotoxicity compared to that shown by transiently expressed constructs comprising unedited gene alone under the test conditions.

[0218] Phytotoxicity scoring for all bush bean is assessed based on the following: None- no negative plant phenotype observed, low- some negative plant phenotype observed (bruising / low levels of browning), moderate- significant negative plant phenotypeDocket # 212601-WO-SEC-l observed (browning tissue on ~50% of the leaf), strong- severe negative plant phenotype (tissue death, curling of leaves)

[0219] Transient protein expression of the fusion proteins is confirmed by a mass spectrometry-based protein identification method using extracted protein lysates from infiltrated leaf tissues as described in Patterson et al. (1998) 10(22):1-24, Current Protocol in Molecular Biology published by John Wiley & Son Inc).

[0220] Agro bacterium -mediated Stable Transformation of Maize

[0221] For Agrobacterium-mediated maize transformation of engineered polypeptides, the method of Zhao will be employed (US Patent Number 5,981 ,840 and International Patent Publication Number WO 1998 / 32326, the contents of which are hereby incorporated by reference). Briefly, immature embryos are isolated from maize and the embryos contacted with an Agrobacterium Suspension, where the bacteria are capable of transferring a polynucleotide encoding the polypeptide of interest to at least one cell of at least one of the immature embryos (step 1 : the infection step). In this step the immature embryos are immersed in an Agrobacterium suspension for the initiation of inoculation. The embryos are co-cultured for a time with the Agrobacterium (step 2: the co-cultivation step). The immature embryos are cultured on solid medium with antibiotic, but without a selecting agent, for Agrobacterium elimination and for a resting phase for the infected cells. Next, inoculated embryos are cultured on medium containing a selective agent and growing transformed callus is recovered (step 4: the selection step). The immature embryos are cultured on solid medium with a selective agent resulting in the selective growth of transformed cells. The callus is then regenerated into plants (step 5: the regeneration step), and calli grown on selective medium are cultured on solid medium to regenerate the plants.

[0222] For detection of the polypeptide of interest in leaf tissue 4 lyophilized leaf punches / sample are pulverized and resuspended in 100 pL PBS containing 0.1% TWEEN™ 20 (PBST), 1% beta-mercaoptoethanol containing 1 tablet / 7 mL complete Mini proteinase inhibitor (Roche 1183615301 ). The suspension is sonicated for 2 min and then centrifuged at 4°C, 20,000 g for 15 min. To a supernatant aliquot 1 / 3 volume of 3X NuPAGE® LDS Sample Buffer (Invitrogen™ (CA, USA), 1 % B-ME containing 1 tablet / 7 mL complete Mini proteinase inhibitor is added. The reaction is heated at 80°C for 10Docket # 212601-WO-SEC-l min and then centrifuged. A supernatant sample is loaded on 4-12% Bis-Tris Midi gels with MES running buffer as per manufacturer’s (Invitrogen™) instructions and transferred onto a nitrocellulose membrane using an iBIot® apparatus (Invitrogen™). The nitrocellulose membrane is incubated in PBST containing 5% skim milk powder for 2 hours before overnight incubation in affinity-purified rabbit antibody specific to the protein of interest in PBST overnight. The membrane is rinsed three times with PBST and then incubated in PBST for 15 min and then two times 5 min before incubating for 2 hours in PBST with goat anti-rabbit-HRP for 3 hours. The detected proteins are visualized using ECL Western Blotting Reagents (GE Healthcare cat # RPN2106) and Kodak® Biomax® MR film. For detection of the polypeptide of interest in roots the roots are lyophilized and 2 mg powder per sample is resuspended in LDS, 1 % beta-mercaptoethanol containing 1 tablet / 7 mL Complete Mini proteinase inhibitor is added. The reaction is heated at 80° C for 10 min and then centrifuged at 4°C, 20,000g for 15 min. A supernatant sample is loaded on 4-12% Bis-Tris Midi gels with MES running buffer as per manufacturer’s (Invitrogen™) instructions and transferred onto a nitrocellulose membrane using an iBIot® apparatus (Invitrogen™). The nitrocellulose membrane is incubated in PBST containing 5% skim milk powder for 2 hours before overnight incubation in affinity-purified polyclonal rabbit antibody in PBST overnight. The membrane is rinsed three times with PBST and then incubated in PBST for 15 min and then two times 5 min before incubating for 2 hours in PBST with goat anti-rabbit-HRP for 3 hrs. The antibody-bound insecticidal proteins are detected using ECL™ Western Blotting Reagents (GE Healthcare cat # RPN2106) and Kodak® Biomax® MR film.

[0223] Transgenic maize plants positive for expression of the insecticidal proteins are tested for pesticidal activity using standard bioassays known in the art. Such methods include, for example, root excision bioassays and whole plant bioassays. See, e.g., US Patent Application Publication Number US 2003 / 0120054 and International Publication Number WO 2003 / 018810.

Claims

1. Docket # 212601-WO-SEC-lWe claim:1 . A method of improving a generative artificial intelligence (Al) model, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a starting effector and an outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) proposing by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receiving by a trained reward model one or more entries, each entry comprising:1 . the starting effector;2. the outcome objective; and3. one of the one or more proposed effectors;(d) generating by the reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective;(e) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate one or more effectors satisfying the outcome objective.

2. The method of claim 1 , wherein the outcome objective is(a) a defined or modified gene expression level compared to the gene expression level of the starting effector; (b) a defined or modified protein expression level compared to the protein expression level of the starting effector; (c) a defined or modified temporal or spatial gene or protein expression level or pattern compared to the temporal or spatial gene or protein expression level or pattern of the starting effector; (d) a defined or modified protein property compared to the protein property of the starting effector; (e) a defined or modified protein activity compared to the protein activity of the starting effector; (f) a defined or modified phenotype of a plant, plant tissue, or plant cell compared to the phenotype of a plant, plant tissue, or plant cell arising from the starting effector.Docket # 212601-WO-SEC-l3. The method of claim 1 , wherein (a) the starting effector and one or more proposed effectors are polynucleotides (b) the starting effector and one or more proposed effectors are polypeptides; (c) one or more proposed effectors is a variant of the starting effector; (d) the starting effector or one or more proposed effectors is a polynucleotide or polypeptide impacting yield, biomass, photosynthetic efficiency, nitrogen use efficiency, heat tolerance, drought tolerance, herbicide tolerance, or disease resistance of a plant; (e) the starting effector and one or more proposed effectors are chemical structures, chemical activities, or formulations; or (f) the starting effector and one or more proposed effectors are biologicals or chemicals.

4. The method of claim 1 , further comprising:(1) repeating step (a) with a different starting effector, a different outcome objective, or both;(2) repeating step (c), wherein the one of the one or more proposed effectors is a different proposed effector than in step (a);(3) repeating step (c) with the same starting effector and one or more proposed effectors as in step (a) but with a different outcome objective than in step (a);(4) repeating step (c) with the same starting effector and proposed effector as in step (a) but different outcome objective than in step (a); or(5) performing steps (a)-(e) when the observed outcome data for one or more of the proposed effectors does not meet the outcome objective.

5. The method of claim 4, wherein the target signal is used to determine the reward value using (a) outcome objective data for the starting effector and the outcome objective data for the proposed effector or (b) outcome objective data for the starting effector and the outcome objective.

6. The method of claim 5, wherein the outcome objective data comprises actual (observed) data, simulated data, or combinations thereof.

7. The method of claim 1 , further comprising updating the generative Al model by(a) receiving by the generative Al model one or more pairs, wherein each pair comprises the starting effector and the outcome objective;(b) producing by the generative Al model one or more proposed effectors;Docket # 212601-WO-SEC-l(c) inputting one or more entries into a trained reward model to create a reward value to predict / measure how well the proposed effector meets the outcome objective, wherein each of the one or more entries comprises a starting effector, an outcome objective, and a proposed effector; and(d) using a reinforcement learning algorithm to adjust or optimize, based on the reward value, one or more weights in the generative Al model to train or update the generative Al model’s capacity to generate effectors satisfying the outcome objective.

8. The method of claim 1 , further comprising designing, by the generative Al model, an experiment to determine whether any of the one or more proposed effectors meet the outcome objective.

9. The method of claim 1 , further comprising receiving, by the generative Al model, public or proprietary information or combinations thereof.

10. The method of claim 1 , the method further comprising creating one or more of the proposed effectors in a plant, plant tissue, plant cell, or microbe using genome editing technology.

11. A computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations comprising the steps of claim 1 .

12. A computer system for updating or improving a generative Al model, the system comprising:(a) one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device comprising:(1) a memory; and(2) one or more processors configured to perform operations comprising:(a) receive by a trained generative Al model the starting effector and the outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;Docket # 212601-WO-SEC-l(b) propose by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receive by a trained reward model one or more entries, each entry comprising:1 . the starting effector;2. the outcome objective; and3. one of the one or more proposed effectors;(d) generate by the reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective; and(e) adjust or optimize, by a reinforcement learning algorithm, one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate one or more effectors satisfying the outcome objective.

13. The method of claim 1 , wherein the plant, plant tissue, or plant cell is a monocot or dicot.

14. A method of creating one or more proposed effectors, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a starting effector and an outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) proposing by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receiving by a trained reward model one or more entries, each entry comprising:1 . the starting effector;2. the outcome objective; and3. one of the one or more proposed effectors;(d) generating by the reward model a reward value that predicts how well the one of the one or more proposed effectors from step (c) meets the outcome objective;(e) ranking the one or more proposed effectors based on the generated reward values for each;Docket # 212601-WO-SEC-l(f) selecting one or more proposed effectors based on their ranking; and(g) creating the selected one or more proposed effectors.

15. A computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations comprising the steps of claim 14.

16. A computer system for updating or improving a generative Al model, the system comprising:(a) one or more servers, wherein one of the servers comprises a starting effector and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device comprising:(1) a memory; and(2) one or more processors configured to perform operations comprising:(a) receive by a trained generative Al model the starting effector and the outcome objective, wherein the trained generative Al model has been trained to propose one or more effectors that are predicted to satisfy the outcome objective;(b) propose by the trained generative Al model one or more proposed effectors for each starting effector and outcome objective;(c) receive by a trained reward model one or more entries, each entry comprising:1 . the starting effector;2. the outcome objective; and3. one of the one or more proposed effectors;(d) generate by the reward model a reward value to predict how well the one of the one or more proposed effectors from step (c) meets the outcome objective; and(e) rank the one or more proposed effectors based on the generated reward values for each.Docket # 212601-WO-SEC-l17. A method of using a generative artificial intelligence (Al) model to alter a phenotype in a target plant, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a representation of a genotypic profile of a first parent of the target plant, a representation of a genotypic profile of a second parent of the target plant, an edit library, the edit library comprising a list of all possible genome edits that may be introduced into the first and / or second parental genomes, and an outcome objective, wherein the trained generative Al model has been trained to propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective;(b) proposing by the trained generative Al model one or more combinations of genome edits;(c) generating one or more populations of plants, each of the one or more populations of plants generated by multiplex editing using one of the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits in an otherwise uniform genetic background;(d) receiving by a trained reward model one or more entries, each entry comprising:1 . the representation of the first parental genotype and the second parental genotype;2. the outcome objective; and3. genotypic data for the members of the population;(e) generating by the reward model a reward value to predict how well the genome edit or combination of genome edits from step (c) meets the outcome objective; and(f) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.

18. The method of claim 17, further comprising providing an external module to provide context for the generative Al model.Docket # 212601-WO-SEC-l19. The method of claim 17, wherein the population of plants comprises members with varying combinations of homozygous target edits in an otherwise uniform genetic background.

20. The method of claim 17, wherein the outcome objective data comprises genotypic data for a plant of the population to determine which edits are present in the plant and the phenotype for the plant.21 . The method of claim 17, wherein the genotypic representation comprises (a) a SNP representation of the parent genotype; (b) a polymorphic representation of the parent genotype; or (c) single nucleotide polymorphisms (SNPs), indels, transgene presence, genome edits, or any combination thereof.

22. The method of claim 17, wherein the target plant is a hybrid crop produced from a cross between two or more heterotic groups or a filial cross of inbred plants from the same heterotic group.

23. The method of claim 17, wherein the hybrid crop is maize.

24. A computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations comprising the steps of claim 17.

25. A computer system for generating genome edits to alter a phenotype in a target plant, the system comprising:(a) one or more servers, wherein one of the servers comprises a representation of a genotypic profile of a first parent of the target plant, a representation of a genotypic profile of a second parent of the target plant, an edit library comprising a list of genome edits, and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device comprising:(1 ) a memory; and(2) one or more processors configured to perform operations comprising:(a) receiving, by a trained generative artificial intelligence (Al) model, the genotypic profiles of the first and second parents, the edit library, and the outcome objective, wherein the generative Al model has been trained to propose combinations of genome edits predicted to satisfy the outcome objective;Docket # 212601-WO-SEC-l(b) proposing, by the generative Al model, one or more combinations of genome edits;(c) generating one or more populations of plants via multiplex editing using the proposed combinations of genome edits to produce members with varying combinations of target edits in an otherwise uniform genetic background;(d) receiving, by a trained reward model, one or more entries, each entry comprising the genotypic profiles of the first and second parents, the outcome objective, and genotypic data for the members of the population;(e) generating, by the reward model, a reward value to predict how well the genome edits meet the outcome objective; and(f) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to improve its capacity to generate genome edits that satisfy the outcome objective.

26. A method of using a generative artificial intelligence (Al) model for plant breeding, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a parental genotypic population comprising genotypic information for a plurality of parental plants, an edit library, the edit library comprising a list of all possible genome edits that may be introduced into parental plants of the parental genotypic population, and an outcome objective, wherein the trained generative Al model has been trained to select a first and second parental plant from the genotypic population and propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective;(b) proposing by the trained generative Al model a first and second parental plant from the genotypic population and one or more combinations of genome edits;(c) generating one or more edited populations of plants, each of the one or more edited populations of plants generated by single-site editing, multiplex editing, or a combination thereof using one of the selected first and second parental plant and the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits in an otherwise uniform genetic background;Docket # 212601-WO-SEC-l(d) receiving by a breeding pipeline digital twin one or more entries, each entry comprising:1 . the representation of the first parental genotype and the second parental genotype;2. the outcome objective; and3. genotypic data for the members of the edited population;(e) simulating by the breeding pipeline digital twin breeding outcomes over the course of 3 or more generations;(f) generating by the breeding pipeline digital twin a reward value to predict how well the genome edit or combination of genome edits from step (c) meets the outcome objective; and(g) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.

27. The method of claim 26, wherein the population of plants comprises members with varying combinations of homozygous target edits in an otherwise uniform genetic background.

28. The method of claim 26, further comprising providing an external module to provide context for the generative Al model.

29. The method of claim 26, further comprising providing a breeding context module to provide context for the generative Al model.

30. The method of claim 26, wherein the outcome objective data comprises genotypic data for a plant of the population to determine which edits and naturally occurring variants are present in the plant and the phenotype for the plant.31 . The method of claim 26, wherein the parental genotypic population comprises a SNP representation of the parent genotypes for a plurality of members.

32. The method of claim 26, wherein the genotypic representation comprises a polymorphic representation of the parent genotype for a plurality of members.Docket # 212601-WO-SEC-l33. The method of claim 26, wherein the polymorphic representation comprises single nucleotide polymorphisms (SNPs), indels, transgene presence, genome edits, or any combination thereof.

34. The method of claim 26, wherein the one or more combinations of genome edits are introduced by one or more guide RNAs.

35. The method of claim 26, wherein the target plant is a hybrid crop produced from a cross between two or more heterotic groups or a filial cross of inbred plants from the same heterotic group.

36. The method of claim 26, wherein the hybrid crop is maize.

37. A computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations comprising the steps of claim 26.

38. A computer system for using a generative artificial intelligence (Al) model for plant breeding, the system comprising:(a) one or more servers, wherein one of the servers comprises a parental genotypic population comprising genotypic information for a plurality of parental plants, an edit library comprising a list of all possible genome edits that may be introduced into parental plants of the parental genotypic population, and an outcome objective; and(b) a computing device communicatively coupled to the one or more servers, the computing device comprising:(1 ) a memory; and(2) one or more processors configured to perform operations comprising:(a) receiving, by a trained generative Al model, the parental genotypic population, the edit library, and the outcome objective, wherein the trained generative Al model has been trained to select a first and second parental plant from the genotypic population and propose a combination of genome edits selected from the editing population that are predicted to satisfy the outcome objective;(b) proposing, by the trained generative Al model, a first and second parental plant from the genotypic population and one or more combinations ofDocket # 212601-WO-SEC-l genome edits;(c) generating one or more edited populations of plants, each of the one or more edited populations of plants generated by single-site editing, multiplex editing, or a combination thereof using one of the selected first and second parental plant and the one or more combinations of genome edits proposed by the trained generative Al model to produce a population of plants comprising members with varying combinations of target edits in an otherwise uniform genetic background;(d) receiving, by a breeding pipeline digital twin, one or more entries, each entry comprising:1 . the representation of the first parental genotype and the second parental genotype;2. the outcome objective; and3. genotypic data for the members of the edited population;(e) simulating, by the breeding pipeline digital twin, breeding outcomes over the course of 3 or more generations;(f) generating, by the breeding pipeline digital twin, a reward value to predict how well the genome edit or combination of genome edits meets the outcome objective; and(g) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to update or improve the generative Al model’s capacity to generate combinations of genome edits to satisfy the outcome objective.

39. A method of improving a generative artificial intelligence (Al) model, the method comprising:(a) inputting, into a trained generative Al model:(1) a starting effector comprising a genotype;(2) an outcome objective comprising (i) increased plant yield under a specified management practice and (ii) maintained plant yield under nonmanaged conditions;(3) an edit library comprising genome edits applicable to the genotype;Docket # 212601-WO-SEC-l(4) and a management context module comprising data on one or more management practices;(b) proposing, by the trained generative Al model, one or more combinations of genome edits predicted to satisfy the outcome objective;(c) introducing the proposed genome edits into the genotype to generate a population of edited plants, including doubled haploid lines;(d) growing the edited and non-edited plants under two conditions: a test condition incorporating the specified management practice and a control condition excluding the management practice;(e) measuring plant performance under each condition using one or more of: yield at maturity or a proxy indicator trait correlated with the management practice;(f) training, by supervised learning, a reward model using the measured performance data, wherein the reward model is structured to embed phenotype-specific outcome objectives;(g) generating, by the reward model, reward values for each edit combination based on the percent change in yield or proxy indicator between edited and non-edited plants under both conditions;(h) adjusting, by a reinforcement learning algorithm, one or more weights in the generative Al model based on the reward values to improve its capacity to generate genome edits that optimize yield under managed and unmanaged conditions; and(i) iterating the method steps any number of times to achieve the desired outcome objective.

40. A computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations comprising the steps of claim 39.41 . A computer system for improving a generative artificial intelligence (Al) model to optimize plant yield under managed and unmanaged conditions, the system comprising:(a) one or more servers, wherein one of the servers comprises:(1 ) a starting effector comprising a plant genotype;(2) an outcome objective comprising:Docket # 212601-WO-SEC-l- increased plant yield under a specified management practice; and- maintained plant yield under non-managed conditions;(3)an edit library comprising genome edits applicable to the genotype; and(4) a management context module comprising data on one or more management practices;(b) a computing device communicatively coupled to the one or more servers, the computing device comprising:(1 ) a memory; and(2) one or more processors configured to perform operations comprising:(a) receiving, by a trained generative Al model, the starting effector, the outcome objective, the edit library, and the management context module;(b) proposing, by the generative Al model, one or more combinations of genome edits predicted to satisfy the outcome objective;(c) introducing the proposed genome edits into the genotype to generate a population of edited plants, including doubled haploid lines;(d) growing the edited and non-edited plants under:(1 ) a test condition incorporating the specified management practice; and(2) a control condition excluding the management practice;(e) measuring plant performance under each condition using one or more of: yield at maturity or a proxy indicator trait correlated with the management practice;(f) training, by supervised learning, a reward model using the measured performance data, wherein the reward model is structured to embed phenotypespecific outcome objectives;(g) generating, by the reward model, reward values for each edit combination based on the percent change in yield or proxy indicator between edited and non-edited plants under both conditions;(h) adjusting, by a reinforcement learning algorithm, one or more weights in the generative Al model based on the reward values to improve its capacity to generate genome edits that optimize yield under managed and unmanaged conditions; andDocket # 212601-WO-SEC-l(i) iterating the operations any number of times to achieve the desired outcome objective.

42. A method of using a generative artificial intelligence (Al) model to optimize agricultural product performance and profitability, the method comprising:(a) inputting, into a trained generative artificial intelligence (Al) model, a set of parental genotypic representations, an edit library comprising genome edits applicable to the parental genotypes, and an outcome objective comprising both agronomic performance metrics and economic performance metrics, wherein the trained generative Al model has been trained to propose combinations of parental selections and genome edits predicted to satisfy the outcome objective;(b) proposing, by the trained generative Al model, one or more tuples comprising a first parental genotype, a second parental genotype, a set of genome edits for the first parent, and a set of genome edits for the second parent;(c) generating, based on the proposed tuples, one or more edited populations of plants, each population comprising members with varying combinations of target edits in an otherwise uniform genetic background;(d) receiving, by a joint simulation system comprising a breeding pipeline digital twin and a seed-market digital twin, one or more entries, each entry comprising: i. the proposed parental genotypes and genome edits; ii. the outcome objective; and iii. genotypic and phenotypic data for the edited plant population;(e) simulating, by the breeding pipeline digital twin, multi-generational breeding outcomes, and simulating, by the seed-market digital twin, commercial lifecycle outcomes;(f) generating, by the joint simulation system, a reward value based on an additive combination of agronomic performance and economic performance; and(g) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to improve its capacity to generate parental selections and genome edit combinations that maximize both agronomic and economic outcomes.Docket # 212601-WO-SEC-l43. A computer-readable medium having stored thereon instructions that, when executed by a processor, cause the processor to perform operations comprising the steps of claim 39.

44. A computer system for optimizing agricultural product performance and profitability using a generative artificial intelligence (Al) model, the system comprising:(a) one or more servers, wherein one of the servers comprises a set of parental genotypic representations, an edit library comprising genome edits applicable to the parental genotypes, and an outcome objective comprising both agronomic performance metrics and economic performance metrics;(b) a computing device communicatively coupled to the one or more servers, the computing device comprising:(1 ) a memory; and(2) one or more processors configured to perform operations comprising:(a) receiving, by a trained generative Al model, the parental genotypic representations, the edit library, and the outcome objective, wherein the generative Al model has been trained to propose combinations of parental selections and genome edits predicted to satisfy the outcome objective;(b) proposing, by the generative Al model, one or more tuples comprising a first parental genotype, a second parental genotype, a set of genome edits for the first parent, and a set of genome edits for the second parent;(c) generating, based on the proposed tuples, one or more edited populations of plants, each population comprising members with varying combinations of target edits in an otherwise uniform genetic background;(d) receiving, by a joint simulation system comprising a breeding pipeline digital twin and a seed-market digital twin, one or more entries, each entry comprising: i. the proposed parental genotypes and genome edits; ii. the outcome objective; and iii. genotypic and phenotypic data for the edited plant population;(e) simulating, by the breeding pipeline digital twin, multi-generationalDocket # 212601-WO-SEC-l breeding outcomes, and simulating, by the seed-market digital twin, commercial lifecycle outcomes;(f) generating, by the joint simulation system, a reward value based on an additive combination of agronomic performance and economic performance; and(g) using a reinforcement learning algorithm to adjust or optimize one or more weights in the generative Al model to improve its capacity to generate parental selections and genome edit combinations that maximize both agronomic and economic outcomes.

Images