Artificial intelligence for resolving domain-specific issues

Abstract

Aspects herein describe a new approach to providing an interpretable solution for a domain-specific problem using artificial intelligence. Problems of a particular domain, such as biological problems, are identified by a computing unit. The computing unit determines an initial subset of features including various features for observable characteristics of the problem. The computing unit optimizes the subset of initial features to improve the accuracy of the subset in generating a solution to the problem. The computing unit trains one or more predictive artificial intelligence models based on the optimized subset to output a solution to the problem. The computing unit then uses the trained predictive model to output solutions to the one or more additional domain-specific problems. The computing unit may additionally provide an interpretable solution for a problem of a particular field by outputting a representation of steps performed by the predictive model to output the solution.

Description

Cross-reference to related applications

[0001] This application claims priority to the following applications: U.S. Provisional Application Serial No. 63 / 527,690, filed July 19, 2023, entitled “ARTIFICIAL INTELLIGENCE METHODS FOR SOLVING BIOLOGICAL PROBLEMS”; U.S. Provisional Application Serial No. 63 / 603,160, filed November 28, 2023, entitled “ARTIFICIAL INTELLIGENCE METHODS FOR SOLVING BIOLOGICAL PROBLEMS”; and U.S. Provisional Patent Application Serial No. 63 / 573,981, filed April 3, 2024, entitled “ARTIFICIAL INTELLIGENCE METHODS FOR SOLVING BIOLOGICAL PROBLEMS”, each of which is incorporated herein by reference in its entirety for all purposes. field

[0002] The aspects described in this article involve computers, software, and artificial intelligence. More specifically, these aspects relate to the creation of interpretable artificial intelligence models for solving problems in specific domains. background

[0003] The world of artificial intelligence and machine learning can be broadly divided into two opposing approaches: (1) the automatic creation of general AI models intended for use in any field, and (2) the manual creation of domain-specific AI models based on human expertise.

[0004] One drawback of conventional methods for automatically creating general-purpose models is their lack of interpretability. This means that no human expert can readily interpret, through observation, translation, and / or other means, the mechanisms by which machine learning models make decisions, processes, steps, etc., to generate outputs (e.g., solutions to domain-specific problems as described in this paper). General-purpose AI models generated through automated methods attempt to learn complex domain-specific patterns from simple, general input features (e.g., unstructured observations corresponding to a domain-specific problem, known properties corresponding to one or more objects in the domain-specific problem, environmental conditions, etc., and / or other features of the domain-specific problem). Such models traditionally involve complex functional forms (e.g., large language models, deep neural networks, etc.), which typically require significant resources (time, money, etc.) to generate a sufficiently large training set of simple features to train the model to the desired accuracy. Furthermore, the internal workings of such models are unpredictable. That is, the processes, decisions, relevance, and / or other factors by which the model outputs solutions to domain-specific problems are "black boxes," meaning that human experts cannot determine the factors leading to a particular solution. Lacking interpretable internal mechanisms, these general-purpose models cannot be validated, improved, or relied upon by human domain experts, and their sole output is a final prediction that cannot be validated or used for any deeper level of domain analysis or reasoning. Another major drawback of these general-purpose models is that training them from the aforementioned simple, general features requires vast amounts of data because their design has no specific connection to the domains in which they are applied. In at least some examples, obtaining the necessary domain-specific data may be too expensive to make this approach feasible.

[0005] The conventional approach of manually creating domain-specific AI models (e.g., based on human expertise) also has many drawbacks. For example, the development cost of such models is high, given the combination of domain and AI expertise required to build them—one model at a time. Due to this manual, one-off process, and because of the specificity of their design and construction, such models cannot be reused, even for other problems in the same domain. Conventional techniques for manually generating such models require domain experts to invest significant time in building and selecting interpretable features suitable for solving specific domain-specific problems.

[0006] Examples of the aforementioned shortcomings in conventional approaches to using artificial intelligence to solve domain-specific problems can be readily seen in fields such as biology. The automated creation of general-purpose AI models, in the form of deep neural networks or transformer models, has been applied to biological problems, but with mixed results and a lack of interpretability. For example, one such general-purpose AI model, the AlphaFold protein folding algorithm, is based on training deep neural networks on a very large number of known protein conformations. Due to the enormous data requirements of this type of general-purpose model, its training was only possible because of the massive financial investments already spent on experimentally determining these known protein conformations and the public availability of that data. Furthermore, the AlphaFold model does not produce a reproducible solution to the protein folding problem; it merely provides an uninterpretable prediction of the final folded conformation of the amino acid sequence. While AlphaFold's prediction of this conformation may have the expected accuracy, the model does not output any form of explanation about how the resulting protein folds into that conformation. Due to this lack of interpretability, there is no way to verify the output except by performing experiments that the AI ​​model aims to avoid its costs. Biological proteins are only evolutionarily feasible if they fold correctly into the correct, biologically active conformation via incremental and dynamic processes during their biosynthesis. Traditional models such as AlphaFold cannot generate an explanation of the dynamic processes required to produce the model's output predictions.

[0007] Conventional techniques for manually generating AI models to solve specific problems in biology are also lacking. For example, some conventional methods for generating AI models to solve biological problems involve constructing “molecular fingerprints” or “reaction fingerprints.” These fingerprints are carriers of molecular features (e.g., molecular substructures within molecules involved in biological processes, reactants, etc.) manually created by experts, for example, in biochemistry, attempting to identify substructures within molecules that might be relevant to solving the biological problem. Because the relevance of a given molecule or molecular substructure depends on the biological problem being solved, there are many competing, manually defined criteria for such fingerprints. Therefore, to solve any particular biological problem, the manual effort of a domain expert needs to be repeated for each such problem to determine the fingerprint relevant to solving the problem. This can lead to significant costs in terms of time, money, and / or other resources required to solve biological problems using manually generated AI models for specific domain problems.

[0008] Therefore, there is a strong need for methods to generate and utilize artificial intelligence models to solve domain-specific problems that overcome the shortcomings of the two conventional approaches described in this paper. To overcome the deficiencies associated with the automatic generation of AI models based on simple features, there is a need for methods to generate AI models based on features corresponding to the optimal level of detail for the given domain and the domain-specific problem being solved. To further understand the need for such methods, consider the example from the field of biology again.

[0009] Biology is an extremely complex discipline, partly because there are already many levels of description available for studying it. For example, one can look at ecology, taxonomy, genomics, proteomics, molecular, atomic physics, and / or other levels. The highest levels of description tend to have the most structure for prediction, but these levels are the least directly related to the actual complexity of the problem, which must be solved through quantum physics. On the other hand, starting from the lowest levels of actual quantum physics leads to cumbersome calculations and misses the structure and constraints imposed by higher-level descriptions that could simplify the problem. For example, a chemical reaction depends on the binding energies of the various atoms involved in the reaction. This sounds like a quantum physics problem, but these binding energies are not constants for two coupled atoms; they depend on the surrounding molecular structure. Therefore, while the atomic physics level is the ultimate arbiter of the outcome, it is too low to be used for computationally feasible and accurate predictions.

[0010] Given the abundance of existing levels of description in biology, conventional methods have yet to identify the optimal level of description for training predictive AI models. Therefore, there is a strong need for guideline-based approaches to train AI models at the optimal level of description for solving domain-specific (e.g., biological) problems using features corresponding to that optimal level of description. There is also a need for models that provide interpretable white-box predictions that can be understood, verified, and used for purposes beyond mere prediction, in order to leverage the results of training predictive models at the optimal level of description. Overview

[0011] The following is a simplified overview of the various aspects described herein. This overview is not a comprehensive review, nor is it intended to identify the most important or critical elements or to describe the scope of the claims. The overview below presents some concepts in a simplified form as an introductory prelude to the more detailed description provided below.

[0012] To address the aforementioned shortcomings in conventional approaches to using artificial intelligence (AI) to solve domain-specific problems, and to provide additional benefits that will be realized upon reading this disclosure, elements of the illustrative aspects described herein address new improvements in applying AI approaches to solving domain-specific problems.

[0013] As described herein, methods for solving domain-specific problems may include identifying the optimal level of description for that domain. For example, human domain expert analysis, observational information, and / or other data from the domain may be used in a one-off process to define the optimal level of description that an AI model should consider when solving a problem corresponding to that domain. This optimal level of description may be used to generate candidate features relevant to solving the domain-specific problem (e.g., through computer systems and / or predictive AI models). The optimal level of description may include descriptive information indicating the differences between candidate features corresponding to potential predicted values. The size of the descriptive information may be less than a threshold storage capacity. The descriptive information may include (e.g., through predictive AI models, human experts, and / or other sources) all information identified as necessary to predict a solution to the domain-specific problem. As described herein, these methods applied to, for example, the biological field may include reducing the problem from standard descriptive levels such as ecology, taxonomy, genomics, proteomics, molecular, quantum physics, or others to a previously undefined descriptive level of molecular interaction mechanisms from which data corresponding to logical or non-operational biological pathways can be constructed. These molecular mechanisms can be generated by a predictive AI model trained at the optimal level of description, using a combination algorithm that combines features or sub-features of a domain-specific problem at the optimal level of description.

[0014] As described herein, methods for providing AI models that generate interpretable solutions may include one or more of the following: (1) generating domain-specific candidate features at the optimal level of description for the domain (e.g., via a combination algorithm, etc.); (2) selecting specific candidate features that positively or negatively influence the outcome of the domain-specific problem being solved; (3) eliminating redundancy among similar candidate features that have similar predictive effects on the outcome of the domain-specific problem; (4) compressing similar candidate features that have similar effects on the outcome of the domain-specific problem into interpretable composite features that correspond to the desired level of description of the domain; (5) collecting the selected, non-redundant, and / or compressed candidate features as an optimized feature subset of interpretable features for the predictive model; and / or (6) training the predictive model based on the optimized feature subset using a simple functional form that generates interpretable solutions for the domain-specific problem.

[0015] This disclosure provides several improvements to conventional AI applications for solving domain-specific problems. For example, as described herein, reducing the domain-specific problem to an optimal descriptive level and identifying the best candidate features to incorporate into the feature set improves the scope and level of detail of the auxiliary data collected when applying an AI model to solve the domain-specific problem. Using an optimal descriptive level includes interactions that are more complex than those used in automated methods for training conventional AI models. Furthermore, the use of simplified forms of AI models (e.g., shallow neural networks, etc.) provides transparency into the steps by which an AI model trained using the methods described herein generates a solution to the domain-specific problem. This transparency allows for deeper insights and extrapolations from the process of solving the domain-specific problem compared to conventional methods of manually generating AI models to solve domain-specific problems.

[0016] As set forth more fully in the claims, various aspects of the illustrative embodiments are substantially shown in the following drawings and / or described in conjunction with at least one of the following drawings.

[0017] These and other advantages, aspects, and novel features of this disclosure, as well as the details of the embodiments shown therein, will be more fully understood from the following description and accompanying drawings. Brief description of the attached diagram

[0018] A more complete understanding of the invention and its advantages can be obtained by referring to the following description and considering the accompanying drawings, in which like reference numerals denote like features, and wherein: Figure 1 An illustrative network environment that can be utilized according to various embodiments is described; Figure 2 Example user equipment that can be used in a network environment based on the illustrative aspects described herein, such as Figure 1 Client devices or servers in a network environment.

[0019] Figure 3 A general view of the interaction between the user device and the computing device during the iteration of the method according to the illustrative aspects described herein is shown; Figure 4 The steps for providing artificial intelligence for solving domain-specific problems are illustrated in the illustrative aspects described herein; and Figure 5 The steps for optimizing the feature set according to the illustrative aspects described herein are shown. Detailed description

[0020] I. Description of the basic algorithm A. Methods for the automated generation of interpretable, domain-specific AI models To address the challenge of applying conventional artificial intelligence to domain-specific problems, it is crucial to identify the optimal level of description for that problem. A middle ground is needed between the highest and lowest levels of description: the highest level has a lot of structure but lacks sufficient detail for good predictions, while the lowest level is so detailed and unstructured that it is computationally infeasible. The optimal level of detail can be identified based on an estimate of the minimum information required to solve the domain-specific problem (e.g., by human experts, machines, AI, etc.) and the estimated computational resources required (e.g., computation time, memory, etc.). In this way, the optimal level of detail can be identified such that it corresponds to descriptive information that corresponds to the domain-specific problem and indicates the distinctions between candidate features corresponding to potential predictive values. The size of the descriptive information can be below a threshold storage capacity to conserve computational resources. In some examples, the descriptive information may include all information identified as necessary to address the domain-specific problem (e.g., by a predetermined set of rules for solving the domain-specific problem, by human experts, and / or by other means).

[0021] This disclosure provides illustrative examples of applying the disclosed methods for automatically generating interpretable domain-specific AI models to the field of biology. It should be understood that these methods are applicable to any domain, and the biological examples are not limiting. As described herein, the molecular level provides an approximate optimal balance between high-level structure and low-level detail for biological problems. However, conventional approaches utilizing the molecular description level do not take into account the following concepts: (1) what matters is not a single molecule, but the interactions between multiple molecules that constitute steps in the overall biochemical interaction pathway of biological operation; and (2) relevant information about any single molecule involved in any such interaction in a biochemical pathway cannot be captured by any fixed notion (e.g., conventional “fingerprint” methods), but must be specifically determined for each such interaction step. To improve upon conventional approaches, this disclosure identifies the optimal analytical level for the best combination of feasibility and accuracy of predictive modeling in biology as the molecular mechanism level. These molecular mechanisms may include minimal information about any step in a biochemical pathway that is necessary to determine its logical operation or non-operation, and, where it is operational, the resulting output will be an expected or undesirable input that may be one or more subsequent steps in the pathway. For example, a given molecular mechanism may include molecular substructures involved in a step of a biochemical pathway, geometric constraints between one or more molecular substructures corresponding to a single molecule (e.g., multiple molecules) involved in a step of a biochemical pathway, catalysts affecting the rate of a step of a biochemical pathway, environmental conditions affecting a step of a biochemical pathway, and / or other mechanisms.

[0022] According to this definition, a molecular mechanism can include a combination of molecular substructures within one or more molecules involved in an interaction, collectively and directly contributing positively or negatively to the outcome or rate of a potential molecular interaction (e.g., reaction or binding) or to downstream effects of such an interaction. These substructures can be unique for each such interaction. If a positive or negative impact of this combination of substructures on the outcome or rate is required, the molecular mechanism can further include the geometric relationships between multiple substructures within a single molecule. The molecular mechanism can include any cofactors or catalysts that have a positive or negative impact on the outcome or rate. The molecular mechanism can also include environmental information required for the interaction to occur (e.g., temperature, pH level, etc.). The substructures should be large enough to limit the energy of the critical atoms and bonds directly involved in the interaction, but small enough to contain no more than the number of atoms and bonds required to provide sufficient accuracy. Therefore, the first step of this approach is to break down the original biological problem into one or more subtasks to automatically discover and quantify relevant molecular mechanisms and create predictive models based on these mechanisms. For example, if the original problem is expressed as an amino acid sequence, a given amino acid sequence can be translated into an actual protein or polypeptide molecule, and then the specific molecular mechanisms by which that protein or polypeptide interacts with other relevant molecules can be automatically extracted.

[0023] This method allows for a degree of flexibility in constructing molecular mechanisms, which can include overdetermined and / or underdetermined specifications of various aspects of the complete molecular mechanism. For example, overdetermining may be useful to provide a predictive model with a better ability to generalize potentially limited training data that is beyond the available scope, while underdetermining may be useful to accommodate any constraints in the available training data that prevent the determination of the complete mechanism. Some examples of this flexibility will now be presented to illustrate the range of possibilities, which can be extended to other situations by those skilled in the art.

[0024] One example where overdetermining might be useful is the substitution of similar atoms within molecular substructures involved in an interaction (e.g., in the same column of the periodic table). Instead of creating a large number of similar features that implement those substitutions, which might each have similar effects on the outcome and reduce the predictive model's ability to generalize well beyond the training data, overdetermining could be beneficial. Another example where overdetermining might be useful is the specification of cofactors or catalysts. In some cases, multiple cofactors or catalysts can have similar effects on both positive and negative outcomes of an interaction. In this case, similar to atomic substitution, the feature specification could include multiple choices of cofactors or catalysts, along with potential quantitative information about the number of cofactors or catalysts required to positively or negatively influence the interaction.

[0025] One example where underdeterminacy might be useful is when multiple substructures of the same molecule are actively or passively involved in influencing related molecular interactions, as might happen, for example, between larger molecules. In this case, the geometric relationships of these multiple substructures can be important for the outcome or rate. While such geometric constraints could be included in the feature description itself for a fully deterministic description of the complete molecular mechanism, an alternative is to include geometric information as additional quantitative features in the predictive model, where the primary feature represents the underdetermined molecular mechanism and includes at most one of the multiple substructures from each molecule involved in the potential interactions in each feature. This could be useful if a simple functional form of the predictive model might better utilize geometric information than building features based on a fully deterministic molecular mechanism.

[0026] Another example where underdetermination can be useful is when the training data includes only information on one aspect of an interaction. In this case, the original training data can be augmented with a previously constructed database of molecular mechanisms (described below), which enables the identification of the other aspect of each interaction. For example, when analyzing human diagnostic data that includes only information about biomarkers, where the biomarker, when translated into molecular terms, constitutes only one aspect of the interaction in which the biomarker affects human biology, the training data can be augmented with a database of known molecular mechanisms in human biological pathways, to which the biomarker can provide one aspect of the mechanism, thus connecting with known biological pathways. If such a database of molecular mechanisms is unavailable, or if there is a possibility that another aspect of the molecular mechanism is unknown and its discovery may be part of an unsolved biological problem, underdetermination techniques can be applied by using features that include only one aspect of the molecular mechanism. In this case, the method of this application can still provide profound insights by constructing underdetermined features that include only one aspect of the interaction, thus providing clear clues for further research. To generalize the example of analyzing human diagnostic data, if relevant disease pathways in humans are not yet known, the method of this application can automatically identify potential active sites on molecules in diagnostic assays, providing clues for identifying unknown molecules in disease pathways.

[0027] While the description of features in this paper is flexible, it is important to reiterate the differences between all variations of the method described herein and existing fingerprint-based methods, which also involve molecular substructures. First, as previously stated, existing fingerprints are not automatically identified by artificial intelligence, but rather by input from human experts. This results in generic fingerprints that may be irrelevant to a specific biological problem, or highly specialized fingerprints constructed for a specific problem through arduous experimental work. In contrast, the method of this application identifies, entirely automatically, the most relevant molecular substructures tailored to each possible biological problem from relevant data via artificial intelligence. Second, conventional fingerprinting methods also do not include the complete molecular mechanism, which constitutes the fundamental, irreducible element of a biological pathway, as a feature. Instead, features in conventional fingerprints may include only substructures from a single molecule at a time, relying on complex functional forms of predictive models built upon such features to automatically establish connections between the various molecular substructures involved in the complete mechanism. This reliance on complex functional forms leads to inaccuracies when attempting to identify the specific and complex properties of biological patterns present in the data provided as model inputs. In contrast, the method of this application treats the complete molecular mechanism as the fundamental unit of feature construction.

[0028] The first step in identifying relevant molecular mechanisms may involve obtaining sufficient data, either through existing datasets or through laboratory experiments, to assess which combinations of substructures and their geometric relationships among one or more molecules involved in the relevant interactions, and which cofactors or catalysts are reliably configured to excite or inhibit those interactions on a sufficiently large and diverse subset of data, or to differentially influence their downstream effects, thus constituting useful and meaningful features for predictive models. The mechanisms thus identified do not need to be perfectly predictive, only that they have a sufficiently large and reliable impact on the outcomes, rates, or downstream effects, so that the subsequent predictive models built in the second step on the feature set including those mechanisms can make the expected accurate predictions. This identification can be performed entirely automatically and is computationally feasible by imposing reasonable constraints on the potential substructures and geometries considered. For example, if one of the molecules is a protein, in some examples, the protein backbone may be irrelevant to the outcome because the backbone is the same length along the entire protein or any other protein; therefore, a large number of potential substructures including the protein backbone can be ignored. As another example, based on practical considerations of determining the size of the substructures necessary for the energies of the atoms and bonds directly involved in the interaction, each individual related substructure of the interacting molecule can have a lower and upper limit for its size.

[0029] When available data includes databases of known molecular interactions (e.g., reactions or non-covalent bonds), this method can identify the exact atoms and bonds directly involved in the interaction. For example, in analyzing known chemical reactions, retained substructures between reagents and products can be automatically identified through molecular matching to first identify specific bonds that were not retained but recombined during the reaction by elimination. Then, secondly, the molecular mechanism for the reaction can be constructed as a combination of appropriately sized molecular substructures surrounding each altered bond in the reagent.

[0030] Where the training data used for the predictive model comprises only a set of molecules (e.g., after translating any biomarkers at different levels of biological description into molecular terms) and experimental results regarding the effects of these molecules, the method operates by automatically screening all possible one-sided features of the aforementioned properties within the constraints described above and evaluating them for inclusion in the model feature set based on various criteria, which may include: (1) the largest differential impact on the outcome, rate, or downstream effect of the interaction; (2) compatibility with any known information about the nature of the interaction involved; and (3) compatibility with any known auxiliary data about the complete molecular mechanisms of the related interactions in which the possible features may be involved. If auxiliary data are applied, the resulting features will be complete molecular mechanisms; otherwise, they will be partial and require further investigation.

[0031] By employing novel methods, as described herein, for identifying the optimal level of description for a domain (e.g., molecular mechanisms in the biological field), new approaches can be developed for automatically generating interpretable, domain-specific AI models to solve specific problems. For example, predictive AI models for solving biological problems can be generated based on using molecular mechanisms as interpretable model features.

[0032] The method of this application offers advantages over conventional methods: (1) Due to the complex yet meaningful and domain-specific nature of the features provided to the predictive models generated as described herein, and the simplicity of the functional forms employed by these predictive models, the method described herein produces “white-box” models that go beyond mere prediction, providing in-depth and verifiable insights into how outcomes can be connected to other parts of the overall network of biological pathways, or even redesigning biology to produce better results; (2) Because in-depth, domain-specific knowledge about biology is embedded in the construction of the features themselves, the method of this application provides efficiency improvements over conventional methods that are based on the hope or assumption that biologically relevant patterns may emerge from attempts to fit general, complex functional forms for which there is no particular understanding of biology. For example, a disease diagnostic test based on the method described herein will provide more than just opaque, unverifiable yes / no answers; it will further provide connections to the biological pathways involved in the diseases leading to the yes / no diagnosis, thus providing clues for the development of potential therapeutic interventions.

[0033] Generative predictive models can include (e.g., via combinatorial algorithms) generating candidate features specific to each domain-specific problem to be solved. For example, candidate molecular mechanisms for a specific biological problem can be generated as described herein. Such combinatorial generation can produce a large number of candidate features. Generative predictive models can further include selecting candidate features based on predicted values ​​for candidate features of a specific biological problem, as described herein.

[0034] Generating predictive models may include performing automatic feature redundancy elimination. Depending on the size and nature of the training data, the screening methods described above can produce multiple similar features with similar differential effects. For example, based on available data, molecular substructures differing by only one atom may be indistinguishable in their effects. Redundant features reduce the ability of the resulting predictive model to generalize beyond the training data; therefore, redundancy elimination should be selected to best avoid the types of generalization errors expected for the specific biological problem being addressed. In many cases, the concern is with false positives, either because the selected features are too small and do not include enough structures to judge a broader dataset, or because the selected features are too large to be causally related themselves, but happen to be statistically correlated with the outcome. In such cases, if two similar features with similar differential effects exist, the feature including a larger number of atoms but still within reasonable size constraints will be selected, which may help determine the energy of the relatively small number of atoms or bonds actually involved in the interaction, while the other feature will be discarded because it is likely to generate fewer false positives. Depending on the nature of the specific biological problem, those skilled in the art can add additional automatic redundancy elimination criteria.

[0035] Generative predictive models can include compressing similar features with similar predictive effects into a single, complex feature at its optimal descriptive level. Otherwise, such similar features could introduce statistical confusion for any predictive models built upon them. Regarding molecular mechanisms, one compression approach could be to combine biochemically similar molecular mechanisms into a single, more complex specification of a molecular mechanism, allowing for some variation of its specification. For example, two identical molecular mechanisms can be combined into a single molecular mechanism with multiple choices at that atomic position, except that one atom in the same column of the periodic table is replaced by another. As another example, two identical molecular mechanisms can be combined into a single molecular mechanism with a series of permissible geometric constraints between those substructures, in addition to geometric constraints between multiple substructures.

[0036] As mentioned above, the design of the feature set incorporates domain-specific knowledge about biology, which improves the accuracy of predictions. Based on these features, relatively simple predictive models can be constructed. However, an important characteristic for the preferred predictive modeling method is its ability to make predictions with abrupt shifts or even discontinuities in the latent features. This is fundamental to the nature of biological problems. For example, the binding of a reasonably sized small molecule to an enzyme may require multiple binding sites for success, and the dependence of successful binding on the geometry between those binding sites on the enzyme can vary, shifting from highly excitatory to highly inhibitory as the distance between identical binding sites on the enzyme changes.

[0037] Examples of predictive modeling methods that meet this criterion include shallow neural networks whose final prediction layer has an activation function that has a sharp transition at the zero or first derivative. For example, a final sigmoid activation can be used for classification tasks, or a final rectified linear activation function (ReLU activation) can be used for numerical prediction tasks. Deep neural networks are generally not needed because most of the work is already done by the feature set. Another example of predictive modeling methods that meet this criterion is decision trees for classification tasks, or regression trees, or similar methods for numerical prediction tasks. For classification tasks, another example is a ranking function for the resulting classes. A ranking function can be constructed by predicting class ranking differences from individual features, and then finding the best-fit overall ranking by aggregating these individual ranking differences.

[0038] It will be apparent to those skilled in the art that biological problems can be solved by breaking them down into sub-tasks that can be addressed by the AI ​​method of this application. For example, to develop therapeutic interventions for diseases whose biological pathways are unknown, one could begin with a sub-task of generating predictive diagnostic tests based on patient assays. By acquiring a limited amount of training data on such assays and validating diagnoses, the AI ​​method of this application can generate predictive diagnostic tests based on interpretable features, namely underdetermined molecular mechanisms, which provide automatically determined substructures of molecules involved in assays linked to unknown disease pathways. Then, as a next sub-task, additional training data including known human metabolic pathways can be used to determine which pathways are actually activated by interactions involving molecular substructures identified in the first sub-task. As a final sub-task, human metabolic pathways identified by previous sub-tasks can be evaluated for interventions with candidate drugs.

[0039] B. Illustrative applications of artificial intelligence for solving domain-specific problems In the following description of various embodiments, illustrative steps for implementing the AI ​​methods described above will be described in detail. It should be understood that one or more additional or alternative steps may be performed using the methods described herein without departing from the scope of this disclosure. Although the steps described herein are generally discussed in relation to solving biological problems, it should be understood that the description is merely illustrative and that the methods described herein may be performed to apply artificial intelligence to solve problems in other specific domains without departing from the scope of this disclosure. We will take the problem of learning a diagnostic model from biomarker assays performed on a group of human patients who have been assigned various diagnostic categories by expert physicians as an example. Categories can be simple, such as yes / no, that the patient has a disease, or more precise and fine-grained, providing different diagnostic categories for different stages, variants, or physiological expressions of the disease. Another objective of applying the AI ​​methods described herein can be to use the assay data to go beyond simply black-box diagnostic predictions, to provide a fine-grained causal explanation for each diagnostic prediction based on the human biological pathways found to be activated in each patient.

[0040] A biomarker assay dataset may include a table that provides two data points for each combination of biomarker and patient case ID: (1) a measure of the amount of the biomarker present in the patient case, such as by means of a biochemical assay, such as various types of genome sequencing, chromatography or electrophoresis; and (2) the diagnostic category assigned to the patient case.

[0041] For the purpose of applying the AI ​​method of this application, computer systems, computing units, etc., can translate biomarker information from its raw biological description at the molecular level, and the method will then automatically extract the molecular mechanism from the molecular level. For example, DNA or RNA sequences can be converted into DNA or RNA molecules where the DNA or RNA molecules themselves are directly involved in the relevant interactions, or into their translation products, i.e., amino acid sequences, which can themselves be further converted into peptide molecules, such as proteins and enzymes. In a preferred embodiment, the final quantitative information for each molecule obtained in the assay will be converted into some type of molar unit. Therefore, the input to the translated assay dataset will be considered as the assay molecule and the patient ID, and for each such pair, the output will be the molar concentration of the assay molecule for that patient and the patient's diagnostic category.

[0042] To provide the benefit of identifying complete molecular mechanisms, the assay dataset can be augmented with a database of known human biological pathways, including biochemical reactions and / or binding interactions, and any enzymes or catalysts required for these interactions to proceed at biologically acceptable rates. Each step in the pathway can have metadata containing information about its biological function in the human body.

[0043] To obtain the molecular mechanisms corresponding to these pathways, each biochemical reaction can be analyzed to determine the exact bonds that are rearranged to cause the reaction. This can be determined automatically by eliminating identical molecular substructures in the products and reagents of the reaction, leaving only a small number of atoms and bonds where the rearrangement occurs. The corresponding molecular mechanism can then be considered as a neighborhood of the correct size of the altered atoms and bonds in the reagents and products, along with the desired combination of enzymes or catalysts. The size of the neighborhood can be chosen to be large enough to determine the energy of the altered bonds accurately, but small enough not to avoid including foreign atoms and bonds that are not necessary for sufficiently accurate determination. Thus, a raw database of human biological pathways can be transformed into a database of molecular mechanisms that provide information for determining the operation of each step of the pathway in the raw database, as well as any metadata about the specific biological function performed by that particular step.

[0044] A large number of candidate molecular mechanisms can be generated from assay data. This step begins by generating random combinations of molecular substructures of the assay molecule, or even all possible combinations of molecular substructures, where each molecular substructure has a reasonable size and the geometric relationships between the potential multiple molecular substructures are reasonably constrained. From these combinations of molecular substructures, combinations are selected that sufficiently match an aspect of a molecular mechanism existing in a human pathway database, whose metadata indicates its potential relevance to the diagnostic problem being addressed. For example, if the disease being analyzed is cancer, matching molecular mechanisms from the database can be limited to include only known functions associated with cancer progression. The final candidate molecular mechanisms can include a first aspect and a second aspect, the first aspect comprising a combination of molecular substructures and geometric constraints in the assay molecule that sufficiently matches an aspect of a suitable molecular mechanism from the human pathway database, and the second aspect being given by another aspect of a molecular mechanism operating in the human body from the human pathway database.

[0045] Candidate molecular mechanisms can be selected from the set mentioned above. For example, a candidate molecular mechanism (one that has a sufficiently different effect on the diagnostic category) can be considered as a function of the amount of one aspect of the molecular mechanism contained in the assay molecule. The amount of any candidate molecular mechanism can be measured in atomic molars by multiplying the molar amount of the assay molecule by the number of atoms of the mechanism present in each molecule and adding the resulting amount to all assay molecules that include one aspect of that mechanism. If the number of diagnostic categories is greater than two (yes / no), it may be useful to select a candidate mechanism that has a sufficiently different effect on only one pair of diagnostic categories.

[0046] The next step could be to perform redundancy elimination. The previous step may generate a large number of similar candidate mechanisms, differing by a single atom, which have similar differential diagnostic effects and will therefore act as confounding variables in any predictive models built on such features. The strategy for eliminating this redundancy and selecting a single representative mechanism from among many similar mechanisms is to designate the smaller mechanism as redundant if there are a sufficient number of relevant patient cases, such that a sufficiently large portion of the total number of smaller mechanisms is contained within an instance of a larger candidate mechanism that has also been selected in the previous step, has similar differential diagnostic effects, and is linked to the same complete mechanism in the human pathway database.

[0047] Feature compression can be performed to further reduce potential confounding variables. Several types of feature compression can be performed. Feature compression can involve combining similar candidate features (e.g., candidate molecular mechanisms) to prevent confounding variables (e.g., variables that might cause variable-based predictive AI models trained on variables to confuse their respective effects due to their similarity in construction and / or predictive effects). In an example relevant to the biological field, if two mechanisms are identical except for the substitution of similar atoms (e.g., in the same column of the periodic table), they can be combined into a single feature that has multiple options for the atom at that position. If multiple such atoms appear in a mechanism, this can lead to implicit generalization because all combinations of substitutional atoms may not appear in the dataset. In the example of addition or substitution, if two mechanisms are identical except for geometric constraints between multiple molecular substructures, they can be combined into a single feature where the geometric constraints take the form of a range rather than an explicit distance. This should only be done if such a range would not implicitly include candidate features that would otherwise have been rejected (e.g., because they do not have sufficient differential effects).

[0048] Feature scoring can be performed. For example, each feature in the feature set can be scored and ranked according to its predictive power (positive or negative). In some examples, feature scores can also be based on performing feature compression, redundancy elimination, and / or other optimization actions described herein. To produce the final feature set for the predictive model, an absolute or threshold score cutoff can be applied, below which features are rejected.

[0049] Based on the feature sets constructed as described in this paper, interpretable, predictive, and diagnostic models can be trained. For example, models can be trained based on biologically meaningful feature sets constructed from previous steps. The models themselves can be simple and transparent. Some options include decision trees, random forests, shallow neural networks, or ranking functions.

[0050] Predictive models can be applied to biological problems to produce interpretable diagnostic results. For example, given biomarker assay results for a new set of patients, a predictive model can be applied, and in addition to the predicted diagnosis, the model can be examined to see which biologically significant traits contribute most to the results. Through construction, each trait can be linked to a specific biological function within a complex of human pathways.

[0051] II. Illustrative System Architecture In the following description of various embodiments, reference is made to the accompanying drawings, which form a part thereof, and various embodiments in which the aspects described herein can be practiced are illustrated. It should be understood that other embodiments may be utilized and structural and functional modifications may be made without departing from the scope of this document. Aspects are capable of having other embodiments and can be practiced or performed in various ways. Furthermore, it should be understood that the wording and terminology used herein are for descriptive purposes and should not be considered limiting. Rather, the phrases and terms used herein should be given their broadest interpretation and meaning. The use of “comprising” and “including” and variations thereof means to encompass items listed thereafter and their equivalents, as well as additional items and their equivalents. The use of the terms “installation,” “connection,” “coupling,” “positioning,” “joining,” and similar terms is intended to include direct and indirect installation, connection, coupling, positioning, and joining. While the various illustrative embodiments described herein generally provide examples of applying predictive AI models to biological problems using this system architecture, it should be understood that this architecture can be used to generate and / or apply predictive AI models to solve any specific domain problem described herein without departing from the scope of this disclosure.

[0052] Figure 1 An example of a network architecture and data processing device that can be used to implement one or more illustrative aspects is shown. Various network nodes 103, 105, 107, and 109 can be interconnected via a wide area network (WAN) 101 such as the Internet. Other networks, including private intranets, corporate networks, LANs, wireless networks, personal networks (PANs), etc., may also be used or optionally used. Network 101 is for illustrative purposes and can be replaced by fewer or additional computer networks. The local area network (LAN) may have one or more of any known LAN topologies and may use one or more of various different protocols, such as Ethernet. Devices 103, 105, 107, 109, and other devices (not shown) may be connected to one or more of the network via twisted-pair cables, coaxial cables, fiber optic cables, radio waves, or other communication media.

[0053] As used herein and described in the accompanying figures, the term "network" refers not only to a system in which remote storage devices are coupled together via one or more communication paths, but also to individual devices that can be coupled to such a system with storage capabilities from time to time. Therefore, the term "network" includes not only "physical networks" but also "content networks," which consist of data residing in all physical networks (belonging to a single entity).

[0054] The components may include computing unit 103, web server 105, and client devices 107, 109. Computing unit 103 may be a general-purpose or special-purpose computer or computer farm. Computing unit 103 may provide overall access, control, and management of database and control software for performing one or more illustrative aspects described herein. Computing unit 103 may be connected to web server 105 through which users interact with requested data and obtain data. Alternatively, computing unit 103 may act as the web server itself and be directly connected to the Internet. Computing unit 103 may be connected to web server 105 via network 101 (e.g., the Internet) via a direct or indirect connection or via some other network. Computing unit 103 may have a significant ability to run multiple instances of the described methods in parallel. Computing unit 103 may also have significant bandwidth for data communication between multiple instances of the described methods. Users can interact with computing unit 103 using remote devices 107 and 109, for example, by using a web browser to connect to computing unit 103 via one or more externally exposed websites hosted by web server 105. Devices 107 and 109 can work in conjunction with computing unit 103 to access data stored therein, or can be used for other purposes. For example, as is known in the art, a user can access web server 105 from device 107 by using an internet browser or by executing a software application that communicates with web server 105 and / or computing unit 103 via a computer network such as the internet.

[0055] Servers and applications can be combined on the same physical machine and retain separate virtual or logical addresses, or they can reside on separate physical machines. Figure 1 Only one example of a network architecture that can be used is shown, and those skilled in the art will understand that the specific network architecture and data processing devices used can be varied and are secondary to the functionality they provide, as further described herein. For example, the services provided by web server 105 and computing unit 103 can be combined on a single server.

[0056] Each component 103, 105, 107, 109 can be any type of known computer, server, or data processing device. The computing unit 103 may, for example, include a processor 111 that controls the overall operation of the computing unit 103. Unit 103 may also include RAM 113, ROM 115, a network interface 117, an input / output interface 119 (e.g., keyboard, mouse, monitor, printer, etc.), and memory 121. I / O 119 may include various interface units and drivers for reading, writing, displaying, and / or printing data or files. Memory 121 may also store operating system software 123 for controlling the overall operation of the data processing device 103, control logic 125 for instructing the computing unit 103 to perform the aspects described herein, and other application software 127 that provides auxiliary functions, support functions, and / or other functions that may be used in conjunction with or without the aspects described herein. The control logic may also be referred to herein as computing unit software 125. The functionality of computing unit software can refer to a combination of operations or decisions made automatically based on rules encoded into control logic, operations or decisions made manually by the user providing input to the system, and / or automatic processing based on user input (e.g., queries, data updates, etc.).

[0057] Memory 121 may also store data for performing one or more aspects described herein, including a first database 129 and a second database 131. In some embodiments, the first database may include the second database (e.g., as a separate table, report, etc.). That is, depending on the system design, information may be stored in a single database or may be separated into different logical, virtual, or physical databases. Devices 105, 107, and 109 may have similar or different architectures as described with respect to device 103. Those skilled in the art will recognize that the functionality of data processing device 103 (or devices 105, 107, and 109) as described herein can be distributed across multiple data processing devices, for example, to distribute processing loads across multiple computers, to isolate transactions based on geographic location, user access level, quality of service (QoS), etc.

[0058] One or more aspects may be embodied in computer-usable or readable data and / or computer-executable instructions, such as in one or more program modules executed by one or more computers or other devices as described herein. Typically, program modules include routines, programs, objects, components, data structures, etc., which perform specific tasks or implement specific abstract data types when executed by a processor in a computer or other device. Modules may be written in source code programming languages ​​and subsequently compiled for execution, or they may be written in scripting languages ​​such as (but not limited to) HTML or XML. Computer-executable instructions may be stored on computer-readable media such as hard disks, optical disks, removable storage media, solid-state storage, RAM, etc. As those skilled in the art will recognize, in various embodiments, the functionality of program modules may be combined or allocated as needed. Additionally, functionality may be embodied, in whole or in part, in firmware or hardware equivalents such as integrated circuits, field-programmable gate arrays (FPGAs), etc. Specific data structures may be used to more efficiently implement one or more aspects described herein, and such data structures are contemplated within the scope of the computer-executable instructions and computer-usable data described herein.

[0059] As described above, the computing unit 103 can perform the methods described herein. Figure 2 Device 107 is shown. Figure 1 The example user equipment 200 (shown herein) allows a user to access and communicate with computing unit 103. User equipment 200 and computing unit 103 may be part of the same device or separate devices. User equipment 200 may include various components and modules, including processor 217, random access memory (RAM) 215, read-only memory (ROM) 213, and memories 201 and 203, which may include one or more data sets, such as databases, client or server software 205, output adapter 211, input interface 209, and communication interface 207. Processor 217 may include a graphics processing unit (GPU), or a separate GPU may be included in output adapter 211. Memory 201 may be configured to store electronic data, including any electronic information disclosed herein. Another memory (such as memory 203) may be configured to store different or overlapping data. In one embodiment, memories 201 and 203 may be a single non-transitory computer-readable medium. Each memory 201, 203 may or may not include a database for storing data, or may include data stored in RAM memory, which can be accessed by client / server software as needed. Data associated with the described method can be transmitted between user equipment 200 and computing unit 103 or server via transceiver or network interface (such as communication interface 207).

[0060] One or more legally compliant computer-readable media (such as media 201 or 203) may be configured to contain sensor and / or server software (graphically shown as software 205). In one or more arrangements, sensor software 205 may be configured to identify field-specific observational data and to facilitate or guide communication between two devices, including remote device 109 and / or communication devices, and other devices. Users can control the devices via input interface 209 using various types of input devices, including keyboard 223 and mouse 225. Other types of input devices may include microphones (e.g., for voice communication over a network), joysticks, motion sensing devices, touchscreen 219, and / or combinations thereof. In one or more arrangements, music or other audio, such as speech, may be included as part of the user experience of using the devices. Further collection of field-specific observational data may be facilitated by cameras, GPS, accelerometers, chemical detectors, microscopes, or any other such input structures that can help collect observational data. In this case, audio may be output via speaker 221. In some examples, the observational data may include field-specific information. For example, the observational data may include biological information, such as molecular-level information, atomic-level information, and / or other information corresponding to one or more biological phenomena.

[0061] In some embodiments, one or more actions proposed as a query response by the described method can be executed by actuator 230. These actuators 230 may include any structure or individual device that outputs instructions to a user to perform actions or performs some or all of the actions specified by the described method itself. Such actuators 230 may include, but are not limited to, various machines such as diagnostic devices, surgical devices and / or other devices, appliances such as alarm clocks or washing machines, and robots or artificial intelligence entities such as automated personal assistants. Such actuators 230 may physically be user equipment 200 or computing units (such as…) Figure 1 Part of 103 shown in the diagram. Actuator 230 can also be connected via a network (such as...) Figure 1 Actuator 230 (as shown in Figure 101) interacts with user device 200 or computing unit 103. Actuator 230 can provide query responses (e.g., instructions, recommended actions, etc.), pattern mappings (e.g., information collected by performing the methods described herein and / or representations of that information), and / or other outputs of the methods described herein. Actuator 230 can also be used to perform specific tasks to solve domain-specific (e.g., biological) problems that can be input into the described methods. In a real-world setting, actuator 230 can perform any of a multitude of tasks, including but not limited to providing diagnostic test results, inducing chemical reactions, facilitating experiments, and / or other functions.

[0062] Software 205, computer-executable instructions, and other data used by processor 217 and other components of user device 200 may be stored in memories 201, 203, RAM 215, ROM 213, or combinations thereof. Other types of memory, including volatile and non-volatile memory, may also be used. Software 205 may be stored in RAM 215, ROM 213, and / or memories 201 and 203 to provide instructions to processor 217 such that, when executed, processor 217, device 200, and / or other components thereof perform the functions and methods described herein. In one example, instructions for generating a user interface for interfacing with server 105 or user device 107 may be stored in RAM 215, ROM 213, and / or databases 201 and 203. Software 205 may include application and operating system software and may include code segments, instructions, applets, precompiled code, compiled code, computer programs, program modules, engines, program logic, and combinations thereof. Computer-executable instructions and data can also be stored on some physical form of computer-readable storage medium (referred to herein as "computer memory"), including, for example, electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, DVD or other optical disc storage, cassette tape, magnetic tape, magnetic storage, etc. Software 205 can operate to receive and process observation data so that it can be used by the described method. Software 205 may also exist to relay actions instructed by the method to the user. In some cases, such software 205 can generate a text or speech list of instructions. Furthermore, software 205 can allow an actuator to perform actions specified by the method.

[0063] Figure 3 A general view of possible communication between user equipment 301 and computing unit 302 performing the methods described herein is shown. In some examples, computing unit 302 may be a component of user equipment 301 and / or otherwise incorporated into user equipment 301. In some examples, computing unit 302 may be configured to provide user equipment 301 with one or more instructions (e.g., query responses) to cause user equipment 301 (e.g., via actuator 230) to perform one or more actions. In the illustrated embodiment, user equipment 301 includes input structures and actuators. However, as mentioned above, such user equipment 301 does not need to include all those elements. Figure 3As shown, user equipment 301 can relay information to computing unit 302. For example, user equipment 301 can send, transmit, and / or otherwise relay information to computing unit 302, such as domain-specific queries (e.g., requests to solve problems in the field of biology (e.g., providing diagnostics, analyzing test results, suggesting molecular compounds for performing tasks, and / or any other biological problems) and / or to solve any other domain-specific queries), preliminary data 304 (e.g., information captured by sensors such as cameras, thermometers, barometers, potentiometers, microphones, and / or any other type of sensor, information provided by computing devices (e.g., mobile phones, laptops, servers, etc.) and / or other devices (e.g., microscopes, diagnostic equipment, chemical detectors, and / or other devices), manually uploaded datasets (e.g., laboratory experiment results, pre-existing datasets corresponding to biological problems, and / or other preliminary data), and / or other information).

[0064] In some examples, a user can input a domain-specific query 303 into a user device for relay to a computing unit 302. Additionally and / or alternatively, in some examples, the domain-specific query 303 can be identified based on preliminary data 304. The described preliminary data 304 and domain-specific query 303 can be identified through... Figure 1 Any of the various communication methods exemplified in the above description is transmitted to computing unit 302. Using preliminary data 304 and / or a domain-specific query 303, computing unit 302 executes the described interpretable artificial intelligence methods to solve a domain-specific problem while providing an interpretable approach to problem-solving. For example, computing unit 302 may determine one or more steps to take, choices to make, incidental strategies to implement, and / or other actions to be performed by a robot and / or a human in response to domain-specific query 303. In some examples, computing unit 302 may provide these responses as query responses 305 to user device 301. In some examples, computing unit 302 may perform these actions itself (e.g., by providing query responses to actuators). In some examples, computing unit 302 may determine query responses 305 based on providing preliminary data 304 to a predictive artificial intelligence model trained using the methods described herein. The computing unit 302 can generate decision representations 306 (e.g., lists, summaries, pattern maps, and / or other information corresponding to the query response 305, such as data used by the model, relevances made by the model, stored and / or otherwise used, decisions made by the model to generate a solution, and / or other information corresponding to the query response 305) and / or other representations of the steps taken by the predictive model to determine the query response 305 by utilizing a predictive artificial intelligence model trained using the interpretable artificial intelligence methods described herein.

[0065] The computing unit 302 can execute interpretable artificial intelligence methods for solving domain-specific problems as described herein. Figure 4 An overview of 400 steps in an interpretable artificial intelligence approach is provided. (Reference) Figure 4 At step 402, an interpretable artificial intelligence approach can begin by receiving preliminary data. For example, computing unit 302 may receive preliminary data from user device 301 corresponding to one or more domain-specific problems (e.g., biological problems). As described herein, preliminary data may include information captured by sensors (such as cameras, thermometers, barometers, potentiometers, microphones, and / or any other type of sensor), information provided by computing devices (e.g., mobile phones, laptops, servers, etc.) and / or other devices (e.g., microscopes, diagnostic equipment, chemical detectors, and / or other devices), domain-specific datasets (e.g., laboratory experiment results, pre-existing datasets corresponding to biological problems), and / or other information relevant to solving the domain-specific problem. In some examples, preliminary data may correspond to a real-world environment (e.g., geographical location, laboratory setup, and / or other real-world environments). In some examples, preliminary data may correspond to a virtual or simulated world (e.g., computer programs, virtual reality worlds, and / or other simulated worlds).

[0066] At step 404, computing unit 302 can identify a domain-specific problem. In some examples, computing unit 302 can identify the domain-specific problem based on preliminary data. For example, computing unit 302 can perform a preliminary analysis of the preliminary data to identify problems related to the preliminary data. In some examples, identifying a domain-specific problem can include identifying the optimal level of detail and / or description for that domain, as described herein. For example, computing unit 302 can identify the optimal level of description based on a balance between the estimated minimum information required to solve the domain-specific problem and the estimated amount of computation time required to solve the domain-specific problem. Computing unit 302 can estimate the minimum information and / or amount of computation time required based on historical information about the relevant domain-specific problem, user input, and / or other parameters.

[0067] In biological examples, for instance, based on parsing, reading, and / or otherwise processing preliminary data including incomplete gene sequences, computing unit 302 can identify biological problems such as completing gene sequences. In some examples, computing unit 302 can identify domain-specific problems based on receiving domain-specific queries. For example, computing unit 302 may receive from 301 messages, computer-executable instructions, and / or other queries (e.g., as part of or separate from the preliminary data in step 402) that guide computing unit 302 in solving a specific problem. For example, computing unit 302 may receive from 301 domain-specific queries corresponding to the fields of biology, organic chemistry, and / or others.

[0068] At step 406, computing unit 302 may transform the preliminary data. In transforming the preliminary data, computing unit 302 may automatically (e.g., without manual input) transform the preliminary data by converting, decomposing, and / or otherwise transforming the preliminary data corresponding to one or more different descriptive levels (e.g., ecology, taxonomy, genomics, proteomics, molecular, atomic physics, and / or other descriptive levels) into a single uniform descriptive level. For example, computing unit 302 may identify multiple categories of the preliminary data based on the preliminary data. The preliminary data for each category may correspond to a descriptive level as described herein. Based on the identification of multiple categories of preliminary data, computing unit 302 may reduce the preliminary data to information for a single category.

[0069] To provide benefits attributable to reducing the complexity of the forms of artificial intelligence required to solve domain-specific problems, as described herein, computing unit 302 can transform preliminary data corresponding to a biological problem into molecular information. For example, based on preliminary data corresponding to an atomic physical description level, computing unit 302 can process (e.g., analyze, cluster, classify, and / or otherwise process) the preliminary data to identify one or more molecular interactions between molecules, such as those containing atoms identified at the atomic physical description level. Furthermore or alternatively, computing unit 302 can replace similar atoms within the molecular substructures involved in the molecular interactions (e.g., in the same column of the periodic table). Thus, in these and other examples, computing unit 302 can reformat, summarize, and / or otherwise transform atomic physical information into molecular information. In some examples, when transforming the preliminary data, computing unit 302 can additionally or alternatively convert the preliminary data from a first unit of measurement to a second unit of measurement. For example, computing unit 302 can convert one or more portions of the preliminary data into molar units as part of converting the preliminary data into, for example, molecular information. In some examples, when transforming preliminary data, the calculation unit 302 can transform the preliminary data by indicating an operation or non-operation corresponding to a step of the biochemical pathway and / or by indicating one or more results of performing the steps of the biochemical pathway.

[0070] At step 408, based on and / or by transforming the preliminary data, computation unit 302 can generate multiple candidate features. Candidate features may include information corresponding to interactions between observable characteristics of the domain, information corresponding to the preliminary data and / or to a specific domain problem, and / or other subsets of information that can be used to train and / or utilize predictive AI models. For example, candidate features may include information corresponding to interactions between objects, devices, organisms, molecules, environmental conditions, and / or other observable characteristics of the domain. In generating multiple candidate features, computation unit 302 may perform one or more randomization operations on the transformed preliminary data, and subsequently reduce, optimize, and / or otherwise revise the randomized candidate features to a set of candidate features corresponding to the specific domain problem to be solved.

[0071] In some instances, the domain-specific problem can be a biological problem, as described herein. In these examples, when generating multiple candidate features as described above, computing unit 302 can generate random combinations of molecular substructures corresponding to one or more molecules involved in molecular interactions identified by molecular information (e.g., preliminary data transformed into molecular information), random combinations of molecular aspects involved in steps of a biochemical pathway corresponding to a biological problem, random combinations of environmental conditions (e.g., temperature, pressure, etc.) corresponding to steps in a biochemical pathway, and / or random combinations of all of these. When generating random combinations, computing unit 302 can use a combinatorial algorithm configured to combine aspects of molecules, environmental conditions, interactions between molecules, and / or other factors into information that determines the operation of at least one step in a biochemical pathway related to the biological problem being solved. For example, computing unit 302 can identify multiple molecular substructures based on molecular interactions and subsequently generate random groups of molecular substructures. In some examples, the size of a single molecular substructure generated by computing unit 302 can be based on one or more predetermined parameters included in a domain-specific query. When generating candidate features, computing unit 302 can generate multiple candidate molecular mechanisms based on random combinations of molecular substructures. Candidate molecular mechanisms may include information indicating features involved in one or more steps of a biochemical pathway. For example, a candidate molecular mechanism may include one or more aspects of the molecules involved in the steps of a biochemical pathway (e.g., molecular substructure, geometric relationships between molecules, molecular weight, molecular polarity, bonding properties, physical state, water solubility, melting temperature, boiling point, electrical conductivity, and / or other aspects of the molecules). In some examples, a candidate molecular mechanism may also or alternatively include combinations of molecular substructures corresponding to one or more actions to be taken (e.g., producing a vaccine, generating a chemical reaction, and / or other actions) (e.g., combinations generated using combinatorial algorithms, etc.), combinations of molecular substructures corresponding to downstream effects of one or more actions to be taken (e.g., a representation of molecular substructures included in an organism responding to one or more actions), a representation of geometric relationships between molecular substructures, catalysts / cofactors contributing to the downstream effects of one or more actions, and / or other information. For example, based on a biological problem of treating a disease, computational unit 302 may generate a candidate molecular mechanism representing actions to create a treatment for the disease and the effects of that treatment on an organism carrying the disease. It should be understood that these examples are merely illustrative, and without departing from the scope of this disclosure, computational unit 302 can generate additional or alternative molecular mechanisms (and / or other candidate features) based on any combination of the transformed preliminary data. Although biological problems are described, it should be understood that additional or alternative problems in one or more different fields can also be solved using the methods described herein.For example, problems in the field of organic chemistry can be solved based on the steps of a chemical synthesis cycle rather than the steps of a biological pathway as described in this article.

[0072] At step 410, computing unit 302 may identify one or more parameters for a solution to a domain-specific problem. For example, computing unit 302 may identify a set of selection parameters for determining a subset of candidate features for generating a solution to the domain-specific problem. In some examples, this set of selection parameters may include criteria for determining which subset of candidate features is relevant to the domain-specific problem. For example, if the problem is a treatment for a disease, the set of selection parameters may include criteria indicating that candidate features corresponding to known treatments for the disease should be eliminated (e.g., to avoid generating redundant solutions for biological problems). Additionally or alternatively, in some examples, one or more parameters may include instructions for executing one or more optimization functions to optimize features used to train and / or leverage predictive AI models to generate a solution to the domain-specific problem. In some examples, when identifying one or more parameters, computing unit 302 may identify one or more parameters within a domain-specific query. For example, in addition to identifying information about the domain-specific problem, the domain-specific query may also include one or more parameters. In some examples, when recognizing one or more parameters, the computing unit 302 may (e.g., to the user device 301, to a display associated with the computing unit 302, and / or to other devices) output a prompt requesting the user to input one or more parameters.

[0073] In some examples, computing unit 302 may identify one or more parameters that instruct computing unit 302 to select a subset of multiple candidate features in order to optimize the features used to solve a domain-specific problem, save resources, and / or achieve other benefits. In these examples, computing unit 302 may proceed to step 412. It should be understood that in some examples, computing unit 302 may use all the candidate features generated at step 408 to solve the biological problem. In these examples, computing unit 302 may proceed to step 414 without performing the actions described at step 412.

[0074] At step 412, computation unit 302 may select an initial feature subset. Computation unit 302 may select an initial feature subset from multiple candidate features as part of a process for selecting an optimized (e.g., interpretable) feature set, which includes features that can be used by a predictive AI model to output decisions, processes, steps, etc., used by a machine learning model to identify solutions to a domain-specific problem. For example, computation unit 302 may select a subset of multiple candidate molecular mechanisms from multiple candidate features that include multiple candidate molecular mechanisms. Computation unit 302 may select the initial feature subset based on the relevance of each of the selected candidate features to the overall domain-specific problem.

[0075] In some examples, computational unit 302 may select an initial subset of features based on one or more downstream effects of a single candidate feature. For example, if the domain-specific problem is a biological problem, computational unit 302 may identify multiple biological effects, each corresponding to at least one candidate molecular mechanism. In some examples, these biological effects may be the effects described in step 408. For example, a biological effect may be and / or include an effect that involves taking one or more actions involving a combination of molecular substructures corresponding to a given candidate molecular mechanism. Biological effects may include the results of utilizing combinations of molecular substructures (e.g., diagnosing a disease, preparing a treatment for a disease, inducing a chemical reaction, and / or performing other uses of combinations of molecular substructures).

[0076] In some examples, computational unit 302 may select an initial subset of features by selecting candidate features (e.g., molecular mechanisms) corresponding to the largest differential effect in performing one or more actions based on the interaction of candidate features. For example, computational unit 302 may identify multiple biological effects. Each biological effect may be and / or include the result of an interaction corresponding to at least one candidate molecular mechanism. In some examples, based on the identified biological effects, computational unit 302 may generate multiple effect scores. For example, computational unit 302 may use one or more algorithms, ranking systems, etc., identified by one or more parameters identified at step 410 to generate an effect score for each candidate molecular mechanism. The effect score may indicate the degree and / or severity of the differential effect on the result, rate, or downstream effect of the interaction corresponding to the candidate molecular mechanism. In some examples, based on the generated effect scores, computational unit 302 may compare each effect score with a threshold effect score. For example, computational unit 302 may compare the effect scores with a threshold effect score identified by one or more parameters to determine which effect scores meet (e.g., conform to or exceed) the threshold effect score. In these examples, computation unit 302 can select an initial feature subset by selecting a subset of molecular mechanisms, wherein each molecular mechanism in the subset corresponds to an effect score that satisfies a threshold effect score.

[0077] At step 414, based on the selected initial feature subset, computation unit 302 can identify whether the initial feature subset needs to be optimized. For example, computation unit 302 can identify whether one or more optimization actions need to be performed to reduce multiple candidate features to an optimized feature subset based on a specific domain query and / or based on one or more parameters. In some examples, computation unit 302 can determine whether the initial feature subset needs optimization as a binary yes / no choice. In some examples, when determining whether the initial feature subset needs optimization, computation unit 302 can identify the degree of optimization required. For example, computation unit 302 can identify multiple optimization actions to be performed during optimization (e.g., redundancy elimination, feature compression, feature scoring, and / or, as described herein, for example, regarding...). Figure 5 (Other optimization actions described). Based on the identification that optimization is needed, calculation unit 302 can proceed to step 416 and optimize the feature subset. Based on the identification that optimization is not needed, calculation unit 302 can proceed to step 418 without performing the function described in step 416.

[0078] In step 416, based on the identification that the initial feature subset needs optimization, computation unit 302 can optimize the initial feature subset. For example, computation unit 302 can generate an optimized feature subset by performing one or more optimization actions to eliminate at least one candidate feature from the initial feature subset. In generating the optimized feature subset, computation unit 302 can perform actions described herein, for example, regarding... Figure 5 One or more and / or other optimization actions described further.

[0079] At step 418, computation unit 302 can train a predictive AI model. For example, computation unit 302 can train the predictive AI model based on an optimized subset of features and / or based on an initial subset of features. When training the predictive AI model, computation unit 302 can train a predictive AI model that includes simplified functional forms relative to some conventional AI models. For example, instead of generating an AI model that requires a large training dataset (e.g., a large language model, deep neural networks, etc.), computation unit 302 can train a predictive AI model capable of making predictions using abrupt changes and / or discontinuities in the feature set used as training data, without exceeding a predetermined amount of training data. Computation unit 302 can train a predictive AI model, for example, in the form of a shallow neural network. In these examples, the final sigmoid activation can be used by the predictive AI model for classification tasks (e.g., transforming initial data and / or other classification tasks) and / or ReLU activation can be used for numerical prediction tasks (e.g., generating scores and / or other prediction tasks). Furthermore or alternatively, computation unit 302 can train a predictive AI model, for example, in the form of one or more tree generation algorithms. For example, a predictive AI model may include algorithms for generating decision trees for classification tasks and / or regression trees for numerical prediction tasks. Additionally or alternatively, computation unit 302 may train a predictive AI model, for example, including one or more ranking functions for predicting classification ranking differences based on individual characteristics. It should be understood that while the described methods indicate a preference for training predictive AI models that include simplified functional forms, the methods described herein can be applied to other types of machine learning models.

[0080] In some examples, when training a predictive AI model, computation unit 302 may train the predictive AI model based on an initial feature set and / or an optimized feature set as described herein. For example, computation unit 302 may provide the initial feature set and / or the optimized feature set as input to the training set of the predictive AI model. Training the predictive AI model may configure it to output a representation of a query response and / or a decision made by the predictive model (e.g., a pattern map, an overview, a list of steps, and / or other representations of the decision as described herein) for use with domain-specific queries. For example, training the predictive AI model may configure it to output a representation of a query response and a decision corresponding to a solution to a biological problem based on an input of a set of molecular mechanisms. In some examples, in order to configure and / or otherwise train a predictive AI model, computing unit 302 may apply natural language processing, natural language understanding, supervised machine learning techniques (e.g., regression, classification, neural networks, support vector machines, random forest models, Naive Bayes models and / or other supervised techniques), unsupervised machine learning techniques (e.g., principal component analysis, hierarchical clustering, K-means clustering and / or other unsupervised techniques) and / or other techniques to enable the predictive AI model to process an initial feature set and / or optimize the feature set.

[0081] In some examples, when configuring and / or otherwise training the predictive AI model, computation unit 302 may configure the predictive AI model to perform one or more of the functions described herein in steps 402-416. For example, computation unit 302 may cause the predictive AI model to store portions of an initial feature set and / or an optimized feature set with one or more correlations between these features and the functions described in steps 402-416 used to generate the initial feature set and / or the optimized feature set, and / or otherwise cause the predictive AI model to perform one or more of the functions described in steps 402-416 based on inputs of a domain-specific query, preliminary data, and / or a feature set (e.g., molecular mechanisms). It should be understood that although steps 402-416 describe the generation of a single initial feature set and / or optimized feature set, in some examples, the functions described in steps 402-416 may be repeated before and / or after training the predictive AI model as described herein to generate one or more additional feature sets. In these examples, additional feature sets can be used to update, refine, and / or otherwise further configure the representations of the predictive AI model's output query responses and / or decisions, as described herein. Therefore, computation unit 302 can create iterative loops configured to continuously or periodically update, refine, and / or otherwise configure the predictive AI model to improve the accuracy of the query responses output by the predictive AI model and to improve the efficiency of the processes used to generate representations of query responses and decisions (e.g., pattern mappings, etc.). For example, computation unit 302 can configure the predictive AI model to perform iterative update loops to output representations of query responses and / or decisions for additional domain-specific queries based on the execution of one or more of the functions described in steps 402-416, and to refine the predictive AI model based on the results of these outputs. For example, computing unit 302 can configure the predictive AI model to modify and / or otherwise update one or more algorithms used to generate representations of query responses and / or decisions as described herein, based on identified patterns in domain-specific queries, user feedback on one or more query responses, and / or other results that represent the output query responses and / or decisions. In this way, improvements to the conventional methods described herein can be further enhanced through iterative update loops.

[0082] At step 420, based on the trained predictive AI model, computing unit 302 can identify a domain-specific problem. For example, computing unit 302 can identify additional and / or new domain-specific problems based on further input from user device 301. For example, computing unit 302 can identify a domain-specific problem based on a domain-specific query requesting the use of the predictive AI model to generate a query response. In these examples, computing unit 302 can perform the functions described herein in steps 402-404 to identify the domain-specific problem and / or receive preliminary data corresponding to the domain-specific problem.

[0083] At step 422, computing unit 302 may generate a feature set. For example, computing unit 302 may generate an initial feature set and / or an optimized feature set corresponding to a specific domain problem identified at step 420. When generating the feature set, computing unit 302 may input preliminary data and / or a specific domain query corresponding to the specific domain problem into the predictive AI model. In these examples, computing unit 302 may cause the predictive AI model to perform one or more of the functions described in steps 406-416 to generate the feature set.

[0084] At step 424, based on the generated feature set, computation unit 302 can output a solution. For example, computation unit 302 can output a representation of the query response and / or the decision made by the predictive AI model to output the query response. When outputting the representation of the query response and / or the decision, computation unit 302 can input the feature set from step 422 into the predictive AI model to cause the predictive AI model to output the representation of the query response and / or the decision. When outputting the query response, the predictive AI model can, for example, process the feature set. For example, the predictive AI model can identify one or more solutions to a biological problem based on comparing, combining, and / or otherwise identifying the interactions between each feature in the feature set. For example, based on an input feature set including multiple molecular mechanisms, computation unit 302 can cause the predictive AI model to generate simulations of some or all of a set of possible effects that perform actions associated with each molecular mechanism. Based on comparing, combining, and / or otherwise identifying the interactions between each feature in the feature set, the predictive AI model can identify which interactions between which features produce solutions corresponding to domain-specific queries of the feature set, as further described herein. In these examples, the predictive AI model can generate query responses that include one or more instructions for solving a domain-specific query. For example, computation unit 302 can enable the predictive AI model to output query responses that include formulas, genome sequences, a range of treatments, and / or other methods for solving biological problems.

[0085] In some examples, the output solution may include and / or lead to the application of one or more decisions made by the predictive AI model to one or more additional domain-specific problems. For example, a solution whose output includes a decision representation of a solution to a domain-specific problem made by the predictive AI model may enable one or more algorithms, machine or human experts to train additional predictive AI models using one or more decisions made by the predictive AI model to output solutions to problems adjacent to the domain-specific problem, and / or otherwise utilize one or more decisions made by the predictive AI model to solve additional domain-specific problems. For example, solutions to problems adjacent to one or more biological problems solved by the predictive AI model may be generated and output based on defining a decision-making process for the additional predictive model and using the decision-making process to output solutions to biological problems, which include one or more parameters (e.g., environmental conditions, involved molecules, etc.) that match at least one parameter in one or more biological problems for which the predictive AI model has previously output solutions. Furthermore or alternatively, one or more decisions may be used to revise the predictive AI model through iterative loops. For example, computation unit 302 may enable the predictive AI model to receive feedback information based on the representation of the aforementioned decisions. Feedback information may include corrections to the query response, modified parameters, and / or additional feedback from one or more machine and / or human experts in the domain corresponding to a specific domain problem. In these examples, the iterative loop may include updating the predictive AI model based on the feedback information (e.g., by performing one or more configuration / training steps as described in step 420). The iterative loop may further include repeatedly receiving feedback information and updating the predictive AI model for one or more biological questions until all biological questions in the query response have been resolved.

[0086] When outputting a representation of a decision, computational unit 302 can enable the predictive AI model to generate pattern maps, summaries, lists of steps, etc., while generating a query response. These summaries include information and / or representations of information collected by performing the methods described herein. For example, the predictive AI model can generate a “white-box” summary of the steps performed by the model in determining a query response and / or the interactions compared. For example, the predictive AI model can generate a pattern map that includes information indicating one or more biological patterns identified by the model as molecular mechanisms compared by the model to identify the optimal set of mechanisms for solving biological problems.

[0087] In some examples, when outputting a representation of the query response and / or decision, computing unit 302 may send (e.g., transmit) the representation of the query response and / or decision to 301. In some examples, when sending the representation of the query response and / or decision to 301, computing unit 302 may send one or more instructions that cause the actuator to perform one or more functions to achieve a solution to a problem in a specific domain (e.g., biology).

[0088] It should be understood that each time computation unit 302 receives a new domain-specific query, steps 420-424 can be repeated to provide a representation of the query response and / or decision using the predictive AI model. In these examples, computation unit 302 may further refine and / or otherwise update the predictive AI model based on performing steps 420-424 for each new domain-specific query (e.g., as part of an iterative loop as described herein).

[0089] As described in step 416, the method described herein may include optimizing the feature set (e.g., an initial subset of the feature set). Figure 5 The steps for optimizing the feature set according to the illustrative aspects described herein are illustrated. (Reference) Figure 5 At step 502, the calculation unit 302 may identify optimization parameters. For example, the calculation unit 302 may execute, or has previously executed, the functions described at steps 410 and / or 414 to identify one or more optimization actions to be performed (e.g., based on configuration settings, user instructions, etc.) as part of the output query response for a domain-specific query.

[0090] At step 504, computation unit 302 may identify whether redundancy elimination should be performed. For example, computation unit 302 may identify whether the optimization parameters include instructions for eliminating any redundant features (e.g., molecular mechanisms and / or other features) from the feature set, such redundant features correspond, for example, to an indication that the probability of a feature being a false positive exceeds a threshold probability (e.g., a predetermined score, value, etc., corresponding to the probability of a feature being a false positive). A false positive may be, for example, one of two or more candidate features that have similar predictive effects and / or other similarities to solutions to a particular domain problem. Based on the identification that redundancy elimination should be performed, computation unit 302 may proceed to step 506 and eliminate the redundant feature. Based on the identification that redundancy elimination is not required, computation unit 302 may proceed to step 508.

[0091] At step 506, computation unit 302 may perform redundancy elimination. For example, computation unit 302 may identify whether the feature set (e.g., a subset of the initial feature set) includes any redundant features. In identifying whether the feature set includes redundant features, computation unit 302 may identify whether measurable differences between features meet or exceed a threshold tolerance level (e.g., threshold differences among multiple observable features, threshold differences among multiple unique features, threshold similarity between the predictive effects of two or more features, and / or other tolerance levels). In some examples, computation unit 302 may eliminate candidate features based on comparison indicators (e.g., scores, labels, rankings, etc.) of the probability that a candidate feature is a false positive (e.g., the probability that a candidate feature has a predictive effect and / or value similar to solving a problem in a specific domain). For example, computation unit 302 may identify multiple biological effects. Each biological effect may correspond to at least one molecular mechanism included in the feature set (e.g., the initial feature set). Computation unit 302 may compare each of these biological effects to identify similar effects within tolerance levels. For example, computing unit 302 can compare two biological effects, such as two diagnostic effects corresponding to molecular interactions, and identify that the two effects produce the same downstream effects in, for example, an organism. Furthermore or alternatively, computing unit 302 can identify, for example, molecular mechanisms corresponding to two biological effects that differ in composition by only a single atom. Therefore, computing unit 302 can identify molecular mechanisms corresponding to two biological effects that meet or exceed a threshold tolerance level, which indicates the likelihood that the corresponding candidate feature is a false positive. Computing unit 302 can eliminate redundant features from the feature set. For example, in the example of the two biological effects described above, computing unit 302 can eliminate one of the two features corresponding to the molecular mechanism. In some examples, when performing redundancy elimination, computing unit 302 can implement one or more additional or alternative comparison mechanisms to identify and eliminate redundant features.

[0092] In some examples, computation unit 302 may eliminate candidate features based on comparison indicators that further indicate an instance of a molecular mechanism corresponding to a first candidate feature has been incorporated into an instance of a second molecular mechanism corresponding to a different candidate feature. For example, computation unit 302 may eliminate the first candidate feature based on a comparison indicator indicating that an instance of a molecular mechanism corresponding to a first candidate feature has been incorporated into an instance of a second molecular mechanism exceeding a frequency threshold (e.g., a limit on the number of times a molecular mechanism can be incorporated into an instance of another molecular mechanism before being marked as redundant, capacity, etc.). Furthermore, or alternatively, in some examples, the elimination of the first candidate feature may be based on the size of the corresponding molecular mechanism. For example, computation unit 302 may eliminate the first candidate feature if a molecular mechanism corresponding to the first candidate feature has been incorporated into an instance of a second molecular mechanism by an amount exceeding a frequency threshold, and the second molecular mechanism is larger in size than the molecular mechanism corresponding to the first candidate feature.

[0093] At step 508, the calculation unit 302 can identify whether feature compression should be performed. For example, the calculation unit 302 can identify whether the optimization parameters include instructions for compressing features that exceed a threshold similarity level. Based on the identification that feature compression should be performed, the calculation unit 302 can proceed to step 510 and compress one or more features. Based on the identification that feature compression does not need to be performed, the calculation unit 302 can proceed to step 512.

[0094] At step 510, computational unit 302 may perform feature compression. Feature compression may include combining candidate features. In some examples, computational unit 302 may combine candidate features by and / or based on combining multiple geometric constraints into a series of geometric constraints. In some examples, when performing feature compression, computational unit 302 may combine features based on the similarity of one or more features (e.g., features corresponding to molecular mechanisms, etc.). For example, computational unit 302 may combine multiple molecular substructures sharing a threshold percentage of biochemical traits (e.g., a threshold determined by a set of parameters provided as input, a threshold determined by human experts, and / or other thresholds) into a molecular substructure with multiple specification choices. Specification choices may include rules, licenses, etc., that govern modifications that can be made to the molecular substructure to produce solutions to domain-specific problems. For example, specification choices may include allowing the substitution of a first atom with a second atom from the same column of the periodic table, or allowing the substitution of a first catalyst with a second catalyst sharing a similar trait (e.g., the effect on the rate of reaction and / or other traits). The calculation unit 302 can determine the similarity of two or more given features based on a comparison of similarity scores for each feature. The similarity score can be and / or include values ​​corresponding to similarities in the construction of the two or more features (e.g., information used to construct the features, the process used to construct the features, etc.), similarities in the predictive performance of the features, and / or other similarities. In some examples, the calculation unit 302 can compress features based on comparing the similarity score to a threshold similarity level and / or a threshold similarity score. For example, the calculation unit 302 can compress two or more features (e.g., molecular mechanisms, etc.) exceeding a threshold similarity level into a single feature. The threshold similarity level can be or includes predetermined benchmarks, scores, tests, etc., included in a domain-specific query and / or one or more parameters.

[0095] When identifying whether a feature exceeds a threshold similarity level, computational unit 302 can compare one or more observable properties of the feature. For example, computational unit 302 can identify multiple atoms corresponding to a feature set (e.g., an initial feature set). For example, computational unit 302 can identify which atoms correspond to each molecule and / or molecular substructure involved in a given molecular mechanism. Based on comparing the atoms of each molecule and / or molecular substructure, computational unit 302 can identify, for example, that two molecular mechanisms are identical (e.g., in construction, function, and / or effect), except that similar atoms are substituted in one molecular mechanism (e.g., atoms occupying the same column in the periodic table). In these examples, computational unit 302 can compress two molecular mechanisms into a single molecular mechanism. For example, computational unit 302 can generate a new molecular mechanism that includes indicators for whether atoms corresponding to each atom of the molecular mechanism can substitute for similar atoms. By performing feature compression, computational unit 302 can reduce or eliminate the presence of confounding variables in the optimized feature set (i.e., variables that, due to their similarity in construction and / or predicted effect, may cause predictive AI models trained on variables to confuse their corresponding effects).

[0096] At step 512, the calculation unit 302 can identify whether feature scoring should be performed. For example, the calculation unit 302 can identify whether the optimization parameters include instructions for scoring features based on, for example, predictive ability and eliminating features from the feature set based on the feature scores. Based on the identification that feature scoring should be performed, the calculation unit 302 can proceed to step 514 and score one or more features. Based on the identification that feature scoring does not need to be performed, the calculation unit 302 can terminate the execution of method 500.

[0097] At step 514, the computing unit 302 may score one or more features. In some examples, the computing unit 302 may score features based on one or more predetermined parameters, algorithms, etc., to indicate the predictive power of the features (i.e., the likelihood that the features will predict the correct solution to a domain-specific query). For example, the computing unit 302 may generate a prediction score for each feature based on an algorithm included in the optimization parameters, the complexity of the feature, the cost of performing one or more actions corresponding to the feature, and / or other variables indicating the likelihood that the feature will predict a solution to a domain-specific problem. Additionally or alternatively, in some examples, the computing unit 302 may generate a prediction score for a feature based on whether the computing unit 302 has previously performed redundancy elimination, feature compression, and / or any other optimization actions. For example, if the feature is a compressed feature, the computing unit 302 may increase the prediction score of the feature, thereby indicating an increased likelihood that the feature will be useful in predicting solutions to domain-specific (e.g., biology) problems.

[0098] At step 516, the calculation unit 302 may compare a feature with a threshold. For example, the calculation unit 302 may compare a predicted score corresponding to each feature included in the feature set with a threshold predicted score. In comparing the predicted score with the threshold predicted score, the calculation unit 302 may determine whether each predicted score meets or exceeds the threshold predicted score.

[0099] At step 518, the calculation unit 302 may select features from the feature set to select, retain, and / or otherwise generate a subset of candidate features. For example, the calculation unit 302 may eliminate features from the feature set based on the result of comparing the predicted score of each feature with a threshold predicted score. For example, the calculation unit 302 may eliminate any features corresponding to predicted scores that do not meet or exceed the threshold predicted score.

[0100] It should be understood that the optimization actions described in this article are merely illustrative, and regarding... Figure 5 The optimization actions described may be performed in different orders without departing from the scope of this disclosure. For example, feature scoring and selection of candidate feature subsets may be performed first, candidate feature elimination may be performed second, and feature compression may be performed third. Furthermore, it should be understood that one or more additional or alternative optimization actions may be performed to generate an optimized feature set without departing from the scope of this disclosure.

[0101] III. Illustrative Use Case Scenarios The components of the described interpretable artificial intelligence approach will now be explained in an example of a computational unit analyzing biomarker assays performed on a cohort of human patients, as further described in Part I of this document. In this example, the computational unit may receive domain-specific queries requesting diagnostic predictions for different stages, variants, or physiological expressions of a disease.

[0102] The device configuration may include a personal computer as a user device (e.g., belonging to a doctor, researcher, etc.) and a cluster of remote servers maintained by an entity (e.g., a hospital, service provider, etc.) as a computing unit. These two can communicate via wireless and wired internet. To initiate the process, the computing unit may receive biomarker measurements as preliminary data and queries from the user device.

[0103] The computational unit can then use the functions described herein to transform a biomarker assay (e.g., the quantity of a certain measure of a biomarker present in each patient case, such as by means of a biochemical assay (e.g., various types of genome sequencing, chromatography, or electrophoresis); and the diagnostic category assigned to that patient case) into multiple candidate molecular mechanisms. For example, each biochemical reaction and / or binding interaction detailed in the biomarker assay can be analyzed to determine the molecular substructures involved in the reaction or interaction. The computational unit can select an initial subset of features that includes a portion of multiple candidate molecular mechanisms. The computational unit can then refine the initial subset of features into an optimized subset of features using one or more optimization actions as described herein. Based on the generated optimized subset of features, the computational unit can train one or more predictive AI models to output a representation of a query response and / or decision that can be used by physicians and / or researchers to study, treat, or cure the disease corresponding to the biomarker assay. Subsequently, the computing unit may receive biomarker assay results from a new set of patients and use one or more predictive AI models to generate representations of query responses and / or decisions that contribute to understanding and / or solving biological problems (i.e., treating diseases). In some examples, the decision representations may be used and / or otherwise applied to additional predictive models to solve additional (e.g., adjacent) domain-specific problems. Furthermore, or alternatively, in some examples, the computing unit may output instructions (e.g., audio instructions, written instructions, actuator commands, etc.) to the user device to guide the user device and / or the user in solving biological problems.

[0104] In the following text, various features will be highlighted in a set of numbered clauses or paragraphs. These features should not be construed as limiting the invention or inventive concept, but are provided merely as a highlighting of some of the features described herein, without implying a particular order of importance or relevance of these features.

[0105] The following paragraphs (M1) through (M15) describe examples of methods that can be implemented according to this disclosure.

[0106] (M1) A method comprising: receiving preliminary data corresponding to one or more biological problems at a computing device; generating a plurality of candidate features based on a transformation of the preliminary data, each feature including a molecular mechanism, wherein a given molecular mechanism includes: determining information on operations corresponding to a step in at least one biochemical pathway of the biological problem, wherein the determining information may include one or more aspects of a molecule involved in a step of the at least one biochemical pathway, or one or more environmental conditions corresponding to a step of the at least one biochemical pathway; reducing the plurality of candidate features to an optimized feature set by performing one or more feature reduction operations, wherein one or more feature reduction operations include: selecting a subset of candidate features based on one or more prediction scores indicating the probability that a given candidate feature predicts a solution to the biological problem, and based on a comparison of the probability that a first candidate feature among the plurality of candidate features is a false positive. Indicators are used to eliminate the first candidate feature, wherein a comparison indicator indicates that the probability of the first candidate feature being a false positive exceeds a threshold, or based on one or more similarity scores, a combination of two or more candidate features that have predictive power for one or more solutions to one or more biological problems and correspond to matching similarity scores in one or more similarity scores; a prediction model is trained based on an optimized feature set, wherein the training of the prediction model configures the prediction model to output a solution to a biological problem and a representation of the decision made by the prediction model to output a solution to the biological problem; one or more solutions to one or more biological problems are output using the prediction model; a representation of the decision made by the prediction model to output one or more solutions is output using the prediction model; and one or more decisions made by the prediction model are applied to at least one additional prediction model based on the representation of the decision made by the prediction model to output one or more solutions.

[0107] (M2) The method as described in paragraph (M1), wherein applying one or more decisions comprises: revising the predictive model through an iterative loop and based on one or more solutions to one or more biological problems, wherein the iterative loop comprises: receiving feedback information corresponding to one or more human biological experts based on the output of a representation of the decision made by the predictive model; updating the predictive model based on the feedback information; and repeatedly receiving feedback information and updating during the output of one or more solutions to one or more biological problems until a solution has been output for each of the one or more biological problems.

[0108] (M3) The method as described in any one of paragraphs (M1) to (M2), wherein applying one or more solutions to one or more decisions made by the prediction model to at least one additional prediction model comprises: defining a decision process for at least one additional prediction model based on one or more decisions made by the prediction model; and using the output of at least one additional prediction model to provide solutions to problems adjacent to one or more biological problems, wherein problems adjacent to one or more biological problems include one or more parameters that match at least one parameter of one or more biological problems.

[0109] (M4) The method as described in any one of paragraphs (M1) to (M3), wherein the given molecular mechanism further comprises one or more of the following: a molecular substructure involved in a step of at least one biochemical pathway, a geometric constraint between one or more molecular substructures corresponding to a single molecule among a plurality of molecules involved in a step of at least one biochemical pathway, a catalyst affecting the rate of a step of at least one biochemical pathway, or environmental conditions affecting a step of at least one biochemical pathway.

[0110] (M5) The method as described in any one of paragraphs (M1) to (M4), wherein the selection of a subset of candidate features is based on one or more differential effects of each molecular mechanism corresponding to a plurality of candidate features on at least one outcome of solving one or more biological problems.

[0111] (M6) The method as described in any one of paragraphs (M1) to (M5), wherein the elimination of the first candidate feature is based on a comparison indicator that further indicates that an instance of a first molecular mechanism corresponding to the first candidate feature is incorporated into an instance of a second molecular mechanism corresponding to at least one second candidate feature and having a size greater than that of the first molecular mechanism.

[0112] (M7) The method as described in any one of paragraphs (M1) to (M6), wherein the combined candidate features include one or more of the following: combining multiple geometric constraints into a series of geometric constraints, or combining multiple molecular substructures of biochemical traits sharing a threshold percentage into a molecular substructure having multiple specification choices, wherein the multiple specification choices include one or more of the following: allowing the substitution of a first atom with a second atom in the same column of the periodic table, or allowing the substitution of a first catalyst with a second catalyst of similar traits sharing a threshold percentage.

[0113] (M8) The method as described in any one of paragraphs (M1) to (M7), wherein the prediction model comprises: a ranking function, a decision tree, a random forest of decision trees, or a shallow neural network.

[0114] (M9) The method as described in any one of paragraphs (M1) to (M8), wherein the transformation of the preliminary data includes at least one of the following: an instruction for operation or non-operation corresponding to a step of at least one biochemical pathway, or an instruction for one or more results of performing a step of at least one biochemical pathway.

[0115] (M10) A method for providing interpretable artificial intelligence, comprising: receiving a domain-specific dataset corresponding to a domain-specific problem at a computing device; identifying an optimal level of description for the domain corresponding to the domain-specific problem based on an estimate of the minimum information required to solve the domain-specific problem and an estimate of the computational time required to solve the domain-specific problem; generating multiple candidate features based on the domain-specific dataset, wherein generating multiple candidate features includes: transforming one or more portions of the domain-specific dataset to correspond to the optimal level of description; and generating multiple combinations of features corresponding to the optimal level of description and included in the domain-specific dataset based on the transformation and using a combinatorial algorithm; reducing the multiple candidate features to an optimized feature set by performing one or more feature reduction operations, wherein one or more feature reduction operations are performed to reduce the multiple candidate features to an optimized feature set. The subtraction operation includes one or more of the following: selecting a subset of candidate features based on one or more prediction scores indicating the likelihood of predicting a solution to a domain-specific problem for a given candidate feature; eliminating one or more redundant features; or compressing multiple candidate features to combine features exceeding a threshold similarity score; training a prediction model based on an optimized feature set, wherein training the prediction model configures the prediction model to output a solution to a domain-specific problem and a decision representation of the output solution to the domain-specific problem made by the prediction model; using the prediction model to output a solution to a domain-specific problem and a representation of the decision made by the prediction model for the output solution; and applying one or more decisions made by the prediction model to at least one additional domain-specific problem based on the representation of the decision made by the prediction model for the output solution.

[0116] (M11) The method as described in paragraph (M10), wherein applying one or more decisions comprises: revising the predictive model through an iterative loop and based on a solution to a domain-specific problem, wherein the iterative loop comprises: receiving feedback information corresponding to one or more human experts in the domain corresponding to the domain-specific problem, based on the output of a representation of the decision made by the predictive model; updating the predictive model based on the feedback information; and for one or more additional domain-specific problems, repeating the output of the solution, receiving feedback information, and updating for each additional domain-specific problem.

[0117] (M12) The method as described in any one of paragraphs (M10) to (M11), wherein the optimal level of description corresponds to descriptive information having the following attributes: the descriptive information indicates the differences between candidate features corresponding to potential predicted values; the size of the descriptive information is less than a threshold storage capacity; and the descriptive information includes all information identified as necessary for predicting a solution to a problem in a particular domain.

[0118] (M13) The method as described in any one of paragraphs (M10) to (M12), wherein the prediction model comprises: a ranking function, a decision tree, a random forest of decision trees, or a shallow neural network.

[0119] (M14) The method as described in any one of paragraphs (M10) to (M13), wherein eliminating one or more redundant features comprises: eliminating the first candidate feature based on the probability that the first candidate feature among a plurality of candidate features is a false positive exceeding a threshold.

[0120] (M15) The method as described in any one of paragraphs (M10) to (M14), wherein compression comprises combining two or more candidate features with a similarity score exceeding a threshold based on comparing one or more similarity scores corresponding to corresponding features among a plurality of candidate features.

[0121] The following paragraphs (A1) through (A6) describe examples of computing systems that can be implemented according to this disclosure.

[0122] (A1) A computing system comprising: one or more processors; a memory storing computer-executable instructions that, when executed by the processor, cause the computer system to: receive a domain-specific dataset corresponding to a domain-specific problem at a computing device; identify an optimal level of description for a domain corresponding to the domain-specific problem based on an estimate of the minimum information required to solve the domain-specific problem and an estimate of the computation time required to solve the domain-specific problem; generate a plurality of candidate features based on the domain-specific dataset, wherein generating the plurality of candidate features includes: transforming one or more portions of the domain-specific dataset to correspond to the optimal level of description; and generating a plurality of combinations of features corresponding to the optimal level of description and included in the domain-specific dataset based on the transformation and using a combinatorial algorithm; and reducing the plurality of candidate features to an optimal level by performing one or more feature reduction operations. The feature set, wherein one or more feature reduction operations include one or more of the following: selecting a subset of candidate features based on one or more prediction scores indicating the likelihood of predicting a solution to a domain-specific problem for a given candidate feature; eliminating one or more redundant features; or compressing multiple candidate features to combine features exceeding a threshold similarity score; training a prediction model based on the optimized feature set, wherein training the prediction model configures the prediction model to output a solution to a domain-specific problem and a decision representation of the output solution to the domain-specific problem made by the prediction model; using the prediction model to output a solution to a domain-specific problem and a representation of the decision made by the prediction model for the output solution; and applying one or more decisions made by the prediction model to at least one additional domain-specific problem based on the representation of the decision made by the prediction model for the output solution.

[0123] (A2) The computing system as described in paragraph (A1), wherein, when executed, the instruction configures the computing system to apply one or more decisions by revising a predictive model through an iterative loop and based on a solution to a domain-specific problem, wherein the iterative loop includes: receiving feedback information corresponding to one or more human experts in the domain corresponding to the domain-specific problem, based on the output of a representation of the decision made by the predictive model; updating the predictive model based on the feedback information; and for one or more additional domain-specific problems, repeatedly outputting the solution, receiving feedback information, and updating for each additional domain-specific problem.

[0124] (A3) A computational system as described in any one of paragraphs (A1) to (A2), wherein the optimal level of description corresponds to descriptive information having the following attributes: the descriptive information indicates the differences between candidate features corresponding to potential predicted values; the size of the descriptive information is less than a threshold storage capacity; and the descriptive information includes all information identified as necessary for predicting a solution to a problem in a particular domain.

[0125] (A4) A computational system as described in any one of paragraphs (A1) to (A3), wherein the prediction model comprises: a ranking function, a decision tree, a random forest of decision trees, or a shallow neural network.

[0126] (A5) A computing system as described in any one of paragraphs (A1) to (A4), wherein, when executed, the instruction configures the computing system to compress a plurality of candidate features by combining two or more candidate features with a similarity score exceeding a threshold based on comparing one or more similarity scores corresponding to corresponding features among the plurality of candidate features.

[0127] (A6) A computing system comprising: one or more processors; and a memory storing computer-executable instructions that, when executed by the processor, cause the computing system to perform the method as described in any one of paragraphs (M1) to (M15).

[0128] The following paragraph (S1) describes an example of a system of devices that can be implemented according to this disclosure.

[0129] (S1) A system comprising: a computing device configured to perform the method as described in any one of paragraphs (M1) to (M15); and a user device configured to send domain-specific queries and preliminary data.

[0130] The following paragraph (CRM1) describes examples of computer-readable media that can be implemented according to this disclosure.

[0131] (CRM1) A non-transitory computer-readable medium containing one or more stored instructions that, when executed by a computing system including at least one processor, a communication interface and a memory, cause the computing system to perform the method as described in any one of paragraphs (M1) to (M15).

[0132] Although the subject matter has been described in language specific to structural features and / or methodological behavior, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or behaviors described above. Rather, the specific features and actions described above are disclosed as exemplary forms for implementing the claims.

Claims

1. A method for providing explainable artificial intelligence, comprising: Receive preliminary data corresponding to one or more biological questions at a computing device; Based on the transformation of the preliminary data, multiple candidate features are generated, each feature including a molecular mechanism, wherein a given molecular mechanism includes: --To determine information on the operation of at least one step in a biochemical pathway corresponding to one or more of the biological problems. --The information to be determined can include one or more of the following: ---One or more aspects of the molecules involved in the steps of the at least one biochemical pathway, or ---One or more environmental conditions corresponding to the steps of the at least one biochemical pathway; The plurality of candidate features are reduced to an optimized feature set by performing one or more feature reduction operations, wherein the one or more feature reduction operations include one or more of the following: --A subset of candidate features is selected based on one or more prediction scores that indicate the likelihood of predicting a solution to a biological problem given the candidate features. --Eliminate the first candidate feature based on a comparison indicator of the probability that the first candidate feature is a false positive among the plurality of candidate features, wherein the comparison indicator indicates that the probability that the first candidate feature is a false positive exceeds a threshold, or --Based on one or more similarity scores, combine two or more candidate features that have predictive power for one or more solutions to the one or more biological problems and correspond to the matching similarity scores among the one or more similarity scores; A prediction model is trained based on the optimized feature set, wherein training the prediction model configures the prediction model to output a solution to a biological problem and a representation of the decision made by the prediction model to output the solution to the biological problem; The predictive model is used to output one or more solutions to the one or more biological problems. The predictive model outputs a representation of the decision made by the predictive model for one or more solutions; and Based on the representation of the decision of one or more solutions made by the prediction model, the decision made by the prediction model is applied to at least one additional prediction model.

2. The method according to claim 1, wherein, Applying one or more of the aforementioned decisions includes: The predictive model is revised through iterative cycles and based on one or more solutions to the one or more biological questions, wherein the iterative cycles include: --Based on the output of a representation of the decision made by the prediction model, receive feedback information corresponding to one or more human biological experts; --Update the prediction model based on the feedback information; and --During the output of one or more solutions for the one or more biological problems, feedback information and updates are repeatedly received until a solution has been output for each of the one or more biological problems.

3. The method according to claim 1, wherein, Applying one or more solutions to the at least one additional prediction model to one or more decisions made by the prediction model includes: Define a decision process for the at least one additional prediction model based on the one or more decisions made by the prediction model; and The solution to the problem adjacent to the one or more biological problems is output using at least one additional predictive model. The problem adjacent to the one or more biological problems includes one or more parameters that match at least one parameter of the one or more biological problems.

4. The method according to claim 1, wherein, The given molecular mechanism also includes one or more of the following: The molecular substructures involved in the steps of the at least one biochemical pathway. Geometric constraints between one or more molecular substructures corresponding to individual molecules among a plurality of molecules involved in the steps of the at least one biochemical pathway. A catalyst that affects the rate of the step in the at least one biochemical pathway, or Environmental conditions that affect the steps of the at least one biochemical pathway.

5. The method according to claim 1, wherein, The selection of a subset of the candidate features is based on one or more differential effects of each molecular mechanism corresponding to the plurality of candidate features on at least one outcome of solving the one or more biological problems.

6. The method according to claim 1, wherein, Eliminating the first candidate feature is based on the comparison indicator, which further indicates that an instance of a first molecular mechanism corresponding to the first candidate feature is incorporated into an instance of a second molecular mechanism corresponding to at least one second candidate feature and having a size greater than that of the first molecular mechanism.

7. The method according to claim 1, wherein, Combined candidate features include one or more of the following: Combining multiple geometric constraints into a series of geometric constraints, or Multiple molecular substructures sharing a common threshold percentage of biochemical traits are combined into a molecular substructure having multiple specification choices, wherein the multiple specification choices include one or more of the following: --It is permissible to substitute the first atom with the second atom in the same column of the periodic table, or --Allowing the replacement of the first catalyst with a second catalyst that shares a similarity threshold percentage.

8. The method according to claim 1, wherein, The prediction model includes: Sorting functions Decision tree Random forest of decision trees, or Shallow neural networks.

9. The method according to claim 1, wherein, The transformation of the preliminary data includes at least one of the following: Instructions for operation or non-operation corresponding to the steps of the at least one biochemical pathway, or An indication of one or more results of the steps of the at least one biochemical pathway.

10. A method for providing explainable artificial intelligence, comprising: Receive a domain-specific dataset corresponding to a domain-specific problem at a computing device; Based on the estimated minimum information required to solve the problem in the specific domain and the estimated computation time required to solve the problem in the specific domain, the optimal level of description for the domain corresponding to the problem in the specific domain is identified; Multiple candidate features are generated based on the specific domain dataset. Generating the plurality of candidate features includes: --Transform one or more portions of the domain-specific dataset to correspond to the optimal level of description; and --Based on the transformation, multiple combinations of features corresponding to the optimal description level and included in the specific domain dataset are generated using a combination algorithm; The plurality of candidate features are reduced to an optimized feature set by performing one or more feature reduction operations, wherein the one or more feature reduction operations include one or more of the following: --A subset of candidate features is selected based on one or more prediction scores that indicate the likelihood of predicting a solution to a problem in the specific domain given the candidate features. --Eliminate one or more redundant features, or --Compress the multiple candidate features to combine features that exceed the threshold similarity score; A prediction model is trained based on the optimized feature set, wherein training the prediction model configures the prediction model to output a solution to a domain-specific problem and a representation of the decision made by the prediction model to output a solution to the domain-specific problem; The predictive model outputs a solution to the specific domain problem and a representation of the decision made by the predictive model to output the solution; and Based on the representation of the decision made by the predictive model that outputs the solution, one or more decisions made by the predictive model are applied to at least one additional domain-specific problem.

11. The method according to claim 10, wherein, Applying one or more of the aforementioned decisions includes: The predictive model is refined through iterative loops and based on solutions to the specific domain problem, wherein the iterative loops include: --Based on the representation of the decision made by the prediction model, receive feedback information corresponding to one or more human experts in the domain corresponding to the specific domain problem; --Update the prediction model based on the feedback information; and --For one or more additional domain-specific problems, repeatedly output solutions, receive feedback, and update for each additional domain-specific problem.

12. The method according to claim 10, wherein, The optimal level of description corresponds to descriptive information with the following properties: The descriptive information indicates the differences between candidate features corresponding to potential predicted values; The size of the descriptive information is less than the threshold storage capacity; and The descriptive information includes all the information identified as necessary for predicting solutions to the problem in the specific domain.

13. The method according to claim 10, wherein, The prediction model includes: Sorting functions Decision tree Random forest of decision trees, or Shallow neural networks.

14. The method of claim 10, wherein, Eliminating one or more redundant features includes: eliminating the first candidate feature based on a comparison indicator that the probability of the first candidate feature being a false positive exceeds a threshold.

15. The method according to claim 10, wherein, The compression includes combining two or more candidate features that exceed the threshold similarity score based on comparing one or more similarity scores corresponding to corresponding features among the plurality of candidate features.

16. A computing system, comprising: One or more processors; as well as A memory storing computer-executable instructions that, when executed by the one or more processors, cause the computing system to: Receive domain-specific datasets corresponding to specific domain problems; Based on the estimated minimum information required to solve the problem in the specific domain and the estimated computation time required to solve the problem in the specific domain, the optimal level of description for the domain corresponding to the problem in the specific domain is identified; Multiple candidate features are generated based on the specific domain dataset. Generating the plurality of candidate features includes: --Transform one or more portions of the domain-specific dataset to correspond to the optimal level of description; and --Based on the transformation, multiple combinations of features corresponding to the optimal description level and included in the specific domain dataset are generated using a combination algorithm; The plurality of candidate features are reduced to an optimized feature set by performing one or more feature reduction operations, wherein the one or more feature reduction operations include one or more of the following: --A subset of candidate features is selected based on one or more prediction scores that indicate the likelihood of predicting a solution to a problem in the specific domain given the candidate features. --Eliminate one or more redundant features, or --Compress the multiple candidate features to combine features that exceed the threshold similarity score; A prediction model is trained based on the optimized feature set, wherein training the prediction model configures the prediction model to output a solution to a domain-specific problem and a representation of the decision made by the prediction model to output a solution to the domain-specific problem; The predictive model outputs a solution to the specific domain problem and a representation of the decision made by the predictive model to output the solution; and Based on the representation of the decision made by the predictive model that outputs the solution, one or more decisions made by the predictive model are applied to at least one additional domain-specific problem.

17. The computing system according to claim 16, wherein, When executed, the instruction configures the computing system to apply one or more decisions in the following manner: The predictive model is refined through iterative loops and based on solutions to the specific domain problem, wherein the iterative loops include: --Based on the representation of the decision made by the prediction model, receive feedback information corresponding to one or more human experts in the domain corresponding to the specific domain problem; --Update the prediction model based on the feedback information; and --For one or more additional domain-specific problems, repeatedly output solutions, receive feedback, and update for each additional domain-specific problem.

18. The computing system according to claim 16, wherein, The optimal level of description corresponds to descriptive information with the following properties: The descriptive information indicates the differences between candidate features corresponding to potential predicted values; The size of the descriptive information is less than the threshold storage capacity; and The descriptive information includes all the information identified as necessary for predicting solutions to the problem in the specific domain.

19. The computing system according to claim 16, wherein, The prediction model includes: Sorting functions Decision tree Random forest of decision trees, or Shallow neural networks.

20. The computing system according to claim 16, wherein, When executed, the instruction configures the computing system to compress the plurality of candidate features by combining two or more candidate features that exceed the threshold similarity score based on comparing one or more similarity scores corresponding to corresponding features among the plurality of candidate features.

Images