Method, information processing system, and non-transitory computer-readable medium for determining a predicted odor

The odor prediction framework with MICP and SVM addresses the lack of confidence measures in conventional models by providing reliable odor predictions with quantifiable confidence, improving decision-making in real-world applications.

WO2026234958A1PCT designated stage Publication Date: 2026-11-12YOKOGAWA ELECTRIC CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2026/015491
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-04-16
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Conventional QSOR models lack a quantifiable measure of prediction confidence, hindering their practical use in real-world applications where decision-making relies on both accuracy and reliability of odor predictions.

Method used

An odor prediction framework employing Mondrian inductive conformal prediction (MICP) to determine a predicted odor with a quantifiable measure of confidence by using a conformity measure based on non-conformity scores and user-defined significance levels, incorporating a support vector machine (SVM) for odor class prediction and uncertainty estimation.

Benefits of technology

Provides accurate odor predictions with a quantifiable confidence measure, improving the framework's usefulness in real-world applications by addressing potential biases and class imbalances, enhancing decision-making reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2026015491_12112026_PF_FP_ABST
    Figure JP2026015491_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A method, information processing system, and non-transitory computer-readable medium for determining a predicted are disclosed. Input data specifying a plurality of features of the molecule is received. An odor class prediction for the molecule is determined by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class. A conformity measure for the odor class is determined based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules. The predicted odor of the molecule is determined based on the conformity measure for the odor class.
Need to check novelty before this filing date? Find Prior Art

Description

METHOD, INFORMATION PROCESSING SYSTEM, AND NON-TRANSITORY COMPUTER-READABLE MEDIUM FOR DETERMINING A PREDICTED ODORCROSS-REFERENCE TO RELATED APPLICATION

[0001] The present application claims priority to Japanese Patent Application No. 2025-078310 filed on May 8, 2025, the entire contents of which are incorporated herein by reference.

[0002] The present disclosure relates to a method, an information processing system, and a non-transitory computer-readable medium for determining a predicted odor.Background

[0003] Quantitative structure-odor relationship (QSOR) modeling has been used for elucidating the relationship between chemical structures and olfactory perceptions. The ability to predict odors based on physicochemical features is important in various applications, including the formulation of fragrances and environmental impact assessment.

[0004] Patent literature (PTL) 1 discloses a computer implemented method for predicting olfactory perception for an olfactory target. The method includes calculating a coefficient matrix and a perceptual distance between an indexed olfactory descriptor and a descriptor for the olfactory target, and generating a perceptual descriptor prediction for the olfactory target.

[0005] PTL 1: US 11,062,216 B2Summary

[0006] Although conventional QSOR models generate predictions well, they typically do not provide a quantifiable measure of prediction confidence. This limitation hinders the practical use of these QSOR models in real-world scenarios, where decision-making relies not only on accurate predictions but also on an understanding of the reliability of those predictions. The present disclosure seeks to address this limitation by providing an odor prediction framework that provides a predicted odor together with a quantifiable measure of prediction confidence. (Solution to Problem)

[0007] Aspects of the present disclosure that address the limitations discussed above are described in the following paragraphs.

[0008] [1] A method for determining a predicted odor, the method comprising: receiving input data specifying a plurality of features of a molecule; determining an odor class prediction for the molecule by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class; determining a conformity measure for the odor class based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules; and determining the predicted odor of the molecule based on the conformity measure for the odor class.

[0009] [2] The method according to [1], wherein determining the conformity measure for the odor class comprises: calculating a non-conformity score for the molecule based on the probability for the odor class; and determining the conformity measure for the odor class based on the non-conformity score for the molecule and the plurality of non-conformity scores corresponding to the plurality of calibration molecules.

[0010] [3] The method according to [2], wherein the conformity measure for the odor class is calculated as the proportion of non-conformity scores in the plurality of non-conformity scores corresponding to the plurality of calibration molecules that are greater than or equal to the non-conformity score for the molecule.

[0011] [4] The method according to any one of [1] to [3], wherein the conformity measure for the odor class is a p-value.

[0012] [5] The method according to [4], wherein the p-value is calculated according to:

[0013] [6] The method according to any one of [1] to [5], wherein the predicted odor is a prediction set, and determining the predicted odor comprises adding an odor label corresponding to the odor class to the prediction set when the conformity measure for the odor class is greater than a user-defined significance level.

[0014] [7] The method according to [6], wherein the odor label is indicative of an odor that is predicted to be perceived by a user with a confidence level corresponding to the user-defined significance level.

[0015] [8] A method for determining calibration data for use in determining a predicted odor, the method comprising: receiving sample data specifying, for each calibration molecule in a plurality of calibration molecules, a plurality of features; determining, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying an odor prediction model to the sample data, wherein the odor class prediction specifies a probability for an odor class; determining, for each calibration molecule in the plurality of calibration molecules, a non-conformity score based on the probability for the odor class; and outputting, for each calibration molecule in the plurality of calibration molecules, the non-conformity score as the calibration data.

[0016] [9] The method according to [8], further comprising training the odor prediction model based on training data specifying, for each training molecule in a plurality of training molecules, a plurality of features and a ground truth odor label.

[0017]

[0010] The method according to [9], wherein the plurality of calibration molecules and the plurality of training molecules are mutually exclusive.

[0018]

[0011] The method according to any one of [1] to

[0010] , wherein the odor prediction model is a trained machine learning model.

[0019]

[0012] The method according to

[0011] , wherein the trained machine learning model is a support vector machine.

[0020]

[0013] The method according to any one of [2] to

[0010] , wherein the non-conformity score is calculated as one minus the probability for the odor class.

[0021]

[0014] An information processing system comprising: a processor system; and a memory system; wherein the memory system stores instructions which, when executed by the processor system, cause the information processing system to implement a method for determining a predicted odor of a molecule, the method comprising: receiving input data specifying a plurality of features of the molecule; determining an odor class prediction for the molecule by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class; determining a conformity measure for the odor class based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules; and determining the predicted odor for the molecule based on the conformity measure for the odor class.

[0022]

[0015] An information processing system comprising: a processor system; and a memory system; wherein the memory system stores instructions which, when executed by the processor system, cause the information processing system to implement a method for determining calibration data for use in determining a predicted odor of a molecule, the method comprising: receiving sample data specifying, for each calibration molecule in a plurality of calibration molecules, a plurality of features; determining, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying an odor prediction model to the sample data, wherein the odor class prediction specifies a probability for an odor class; determining, for each calibration molecule in the plurality of calibration molecules, a non-conformity score based on the probability for the odor class; and outputting, for each calibration molecule in the plurality of calibration molecules, the non-conformity score as the calibration data.

[0023]

[0016] A non-transitory computer-readable medium storing instructions which, when executed by an information processing system, cause the information processing system to implement a method for determining a predicted odor of a molecule, the method comprising: receiving input data specifying a plurality of features of the molecule; determining an odor class prediction for the molecule by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class; determining a conformity measure for the odor class based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules; and determining the predicted odor for the molecule based on the conformity measure for the odor class.

[0024]

[0017] A non-transitory computer-readable medium storing instructions which, when executed by an information processing system, cause the information processing system to implement a method for determining calibration data for use in determining a predicted odor of a molecule, the method comprising: receiving sample data specifying, for each calibration molecule in a plurality of calibration molecules, a plurality of features; determining, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying an odor prediction model to the sample data, wherein the odor class prediction specifies a probability for an odor class; determining, for each calibration molecule in the plurality of calibration molecules, a non-conformity score based on the probability for the odor class; and outputting, for each calibration molecule in the plurality of calibration molecules, the non-conformity score as the calibration data. (Advantageous Effect)

[0025] According to the present disclosure, a method, an information processing system, and a non-transitory computer-readable medium for determining a predicted odor can be provided. More particularly, the predicted odor can be determined together with a quantifiable measure of prediction confidence.

[0026] In the accompanying drawings: FIG. 1 is a is a schematic diagram illustrating an odor prediction framework according to an embodiment; FIG. 2 is a schematic diagram illustrating an odor prediction component of the odor prediction framework; FIG. 3 is a schematic diagram illustrating a model training component of the odor prediction framework; FIG. 4 is an information processing system in accordance with an embodiment; FIG. 5 is a flow diagram illustrating a method for determining a predicted odor in accordance with an embodiment; and FIG. 6 is a flow diagram illustrating a method for determining calibration data for use in determining a predicted odor in accordance with an embodiment.DETAILED DESCRIPTION

[0027] Embodiments of the present disclosure pertain to an odor prediction framework that employs conformal prediction to estimate or predict an odor of a molecule on the basis of a predefined confidence level. Specifically, embodiments of the present disclosure pertain to an odor prediction framework that employs Mondrian inductive conformal prediction (MICP) to obtain an estimate of prediction confidence for an odor estimated or predicted for the molecule.

[0028] FIG. 1 is a schematic diagram showing an odor prediction framework 1 in accordance with an embodiment of the present disclosure. The odor prediction framework 1 comprises an odor prediction component 1A and a model training component 1B. The odor prediction component 1A employs an odor prediction model with MICP to determine a predicted odor for a molecule (hereinafter “the target molecule”). Specifically, the odor prediction component 1A determines a predicted odor for the target molecule on the basis of a predefined confidence level, thereby improving the usefulness of the predicted odor in real-world applications. In this context, the predicted odor determined by the odor prediction framework 1 is a semantic description of an odor that a human is predicted to perceive for the target molecule.

[0029] The odor prediction model with MICP employed by the odor prediction component 1A is generated by the model training component 1B. The model training component 1B employs machine learning to train the odor prediction model on the basis of a plurality of training molecules (hereinafter “the training samples”) with assigned odor descriptors, and generates calibration data for use in the MICP on the basis of a plurality of calibration molecules (hereinafter “the calibration samples”). Further details regarding the odor prediction component 1A and the model training component 1B are provided below with reference to FIGS. 2 and 3.

[0030] FIG. 2 is a schematic diagram showing the odor prediction component 1A in more detail. The odor prediction component 1A comprises an odor prediction model 10 and an uncertainty estimation model 20. As discussed above, the odor prediction model 10 is generated by the model training component 1B using machine learning. The uncertainty estimation model 20 employs MICP to determine the predicted odor for the target molecule, based on a predefined confidence level and the calibration data provided by the model training component 1B.

[0031] The odor prediction model 10 is trained to determine an odor class prediction for the target molecule. The odor prediction model 10 receives input data that represents the target molecule in terms of a plurality of features, such as physiochemical features. Such physiochemical features may be calculated for the target molecule using molecular descriptor calculation software, such as the Mordred molecular descriptor calculator [1].

[0032]

[0033] The odor prediction framework 1 according to the present disclosure is not limited to the “fruity” and “sweet” odor classes, and further examples of appropriate odor classes include “floral” and “green”. In further examples, an odor class may provide an indication of perceived odor quality, such as “excellent”, “good”, “fair” or “poor”. It will be appreciated that the odor classes described in this disclosure are non-exhaustive, and any other odor class that describes a perceived odor may be employed in the odor prediction framework 1 according to the present disclosure.

[0034] The uncertainty estimation model 20 employs MICP to determine a predicted odor for the target molecule based on the odor class prediction determined by the odor prediction model 10. Specifically, the uncertainty estimation model 20 uses MICP to determine a prediction set based on the odor class prediction determined by the odor prediction model 10 and the predefined confidence level. That is, the prediction set is a set of odor labels that are predicted for the target molecule at the predefined confidence level. The predefined confidence level is determined on the basis of a user-defined significance level, which will be discussed in further detail below.

[0035] To perform the MICP, the uncertainty estimation model 20 uses calibration data that specifies non-conformity scores corresponding to the calibration samples for each odor class. Specifically, for each calibration sample, the calibration data specifies a non-conformity score corresponding to each odor class used in the odor prediction framework 1. In this context, the non-conformity score represents, for the corresponding odor class, a measure of how different a given calibration sample is from all other calibration samples. Calculation of the non-conformity scores for the calibration samples will be described in further detail below.

[0036] The MICP performed by the uncertainty estimation model 20 is performed separately and independently for each odor class used in the odor prediction framework 1. That is, for a given odor class, the uncertainty estimation model 20 uses MICP to determine whether an odor label corresponding to a given odor class can be included in the prediction set at the predefined confidence level. By treating each odor class independently in this manner, potential prediction bias due to imbalances in the number of calibration samples for each class used to generate the calibration data can be avoided.

[0037]

[0038] In some embodiments, the non-conformity score for a given odor class may be calculated as a p-value according to: (Equation 1) where the p-value is calculated for each odor class independently. As discussed above, calculation of the p-value for each odor class separately is advantageous in that it addresses the problem of a potential odor class imbalance in the calibration data without the need for additional balancing techniques.

[0039]

[0040] FIG. 3 is a schematic diagram showing the model training component 1B in more detail. The model training component 1B includes a training module 30 and a calibration module 40.

[0041] The model training component 1B receives sample data that characterizes a plurality of sample molecules in terms of a plurality of features, such as physiochemical features. Each sample molecule in the plurality of sample molecules is associated with one or more odor descriptors (e.g., “fruity”, “sweet”, “floral”, “green”) corresponding to the odor classes used by the odor prediction framework 1. In this context, the odor descriptors for a given sample molecule in the plurality of sample molecules may be considered as representing the ground truth for the given sample molecule.

[0042] The sample molecules are divided into a training set (i.e., a plurality of training molecules or training samples) and a calibration set (i.e., a plurality of calibration molecules or calibration samples), which are provided to the training module 30 and the calibration module 40 respectively. Here, the training set and the calibration set are mutually exclusive.

[0043] The training module 30 is configured to train the odor prediction model 10 using the training set. For example, the odor prediction model 10 may be a support vector machine (SVM) that is trained on the basis of the training set. However, it will be appreciated that the present disclosure is not limited to use of a SVM for the odor prediction model 10 and, for example, other supervised machine learning models may be employed as the odor prediction model 10.

[0044]

[0045] The calibration data output by the calibration module represents a list of non-conformity scores for each odor class. The list of non-conformity scores may be sorted in increasing order, such that they represent Mondrian class lists. As discussed above, each non-conformity score in the calibration data represents a measure of how different the associated calibration sample is from all other calibration samples for a particular odor class.

[0046] FIG. 4 is a schematic diagram illustrating an example configuration of an information processing system 100 according to an embodiment. The information processing system is configured to implement the odor prediction component 1A and / or the model training component 1B of the odor prediction framework 1 explained above with reference to FIGS. 1 to 3 More particularly, the information processing system may perform one or more methods which implement the odor prediction component 1A and / or the model training component 1B of the odor prediction framework 1, as discussed below in more detail.

[0047] The information processing system 100 includes a processor system 102 and a memory system 104. The processor system 102 includes one or more processors and the memory system 104 includes one or more memory units. The information processing system 100 illustrated in FIG. 4 has been simplified for ease of understanding, but it will be understood that it may include various other components, such as one or more network interfaces, one or more communication interfaces, one or more input interfaces, and / or the like.

[0048] The processing system 102 may include one or more processors, one or more dedicated circuits, or a combination thereof. The one or more processors may include a general-purpose processor, such as a central processing unit (CPU), a dedicated processor optimized for a particular purpose, such as graphics processing unit (GPU), or any combination thereof. Examples of a dedicated circuit are a field-programmable gate array (FPGA) and an application specific integrated circuit (ASIC). The processing system 102 executes information processing to control operations performed by the information processing system 100.

[0049] The memory system 104 includes, for example, one or more semiconductor memories, one or more magnetic memories, one or more optical memories, or any combination thereof. The memory system 104 may function as main memory, auxiliary memory, cache memory, or any combination thereof. Examples of suitable semiconductor memory include Random Access Memory (RAM) and Read Only Memory (ROM). Examples of RAM include Static RAM (SRAM) or Dynamic RAM (DRAM). Examples of ROM include Electrically Erasable Programmable ROM (EEPROM). The memory system 14 stores instructions and information for use in operations performed by the information processing system 100.

[0050] The memory system 104 stores instructions which, when executed by the processor system 102, implement one or more functions of the information processing system 100. The instructions may be supplied to the information processing system 100 in a separate non-transitory computer-readable medium, such as an optical disk or a solid-state drive (SSD). Alternatively, the instructions may be received over a network (not shown), such as the Internet, a mobile communication network, an ad-hoc network, a local area network (LAN), a metropolitan area network (MAN), or any combination thereof.

[0051] FIG. 5 is a flow diagram showing a method for determining a predicted odor in accordance with an embodiment of the disclosure. The method shown in FIG. 5 may, for example, be implemented by the computer system 100 shown in Fig. 4.

[0052] First, the computer system 100 receives input data specifying a plurality of features of a molecule (i.e., a target molecule), and determines an odor class prediction for the molecule by applying an odor prediction model 10 to the input data (Step S102). As discussed above, the odor class prediction specifies one or more probabilities for one or more corresponding odor classes in the odor prediction framework 1. Next, the computer system 100 determines one or more conformity measures corresponding respectively to the one or more odor classes based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules (i.e., the calibration samples) (Steps S104 and Step S106). Specifically, for each or the one or more odor classes, the computer system 100 calculates a non-conformity score for the molecule based on the probability for the odor class (Step S104) and determines the conformity measure for the odor class based on the non-conformity score for the molecule and the plurality of non-conformity scores corresponding to the plurality of calibration molecules (Step S106). The computer system then determines the predicted odor of the molecule based on the one or more conformity measures (Step S108) and outputs the odor as output data.

[0053] As discussed above with reference to FIGS. 1 to 4, the computer system 100 may calculate the conformity measure for each odor class in Step S106 as the proportion of non-conformity scores in the plurality of non-conformity scores corresponding to the plurality of calibration molecules that are greater than or equal to the non-conformity score for the molecule. In this case, the non-conformity score for a given odor class is calculated as one minus the probability for the given odor class. More particularly, when the conformity measure for the odor class is a p-value, the computer system 100 may calculate the p-value according to Equation 1 above.

[0054] As discussed above with reference to FIGS. 1 to 4, in Step S108 the computer system 100 may determine the predicted odor as a prediction set. In this case, determining the predicted odor comprises adding an odor label corresponding to a given odor class to the prediction set when the respective conformity measure for the odor class is greater than a user-defined significance level. Here, the odor label is indicative of an odor that is predicted to be perceived by a user with a confidence level corresponding to the user-defined significance level.

[0055] FIG. 6 is a flow diagram showing a method for determining calibration data for use in determining a predicted odor in accordance with an embodiment of the disclosure. The method shown in FIG. 6 may, for example, be implemented by the computer system 100 shown in FIG. 4.

[0056] First, the computer system 100 receives sample data specifying, for each calibration molecule in a plurality of calibration molecules (i.e., the calibration samples), a plurality of features and determines, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying the odor prediction model 10 to the sample data (Step S202). As discussed above, the odor class prediction specifies one or more probabilities for one or more odor class in the odor prediction framework 1. Next, the computer system 100 determines, for each calibration molecule in the plurality of calibration molecules and for each odor class in the one or more odor classes, a non-conformity score based on the respective probability for the odor class (Step S204). Next, the computer system 100 outputs, for each calibration molecule in the plurality of calibration molecules, one or more the non-conformity scores corresponding to the one or more odor classes as the calibration data (Step S206).

[0057] As discussed above with reference to FIGS. 1 to 4, the computer system 100 may also train the odor prediction model 10 based on training data specifying, for each training molecule in a plurality of training molecules (i.e., the training samples), a plurality of features and a ground truth odor label. In this case, the plurality of calibration molecules and the plurality of training molecules are mutually exclusive. Here, the odor prediction model 10 is a trained machine learning model and, in some embodiments, may be a support vector machine.

[0058] In the embodiments of the present disclosure described above with reference to FIGS. 1 to 6, the input data characterizes a target molecule in terms of a plurality of features, such as physiochemical features, that are determined using molecular descriptor calculation software. However, embodiments of the present disclosure are not limited to the use of molecular descriptor calculation software, and the plurality of features characterizing the target molecule may be obtained via other means, such as a chemical sensor array. Use of a chemical sensor array may be useful in certain industrial applications, such as quality control in a food or beverage production line, or the like. Additionally, other forms of input data may include data obtained via spectroscopic methods (e.g., mass spectrometry), image-based molecular representations processed by deep learning, environmental sensor networks capturing volatile organic compound (VOC) signatures, or even human sensory evaluation data. These alternatives may enhance adaptability for various application areas such as environmental monitoring, fragrance design, or health diagnostics.Examples

[0059] An example of the odor prediction framework according to the present disclosure was applied to molecules (i.e., samples) retrieved from the Pyrfume project [2]. Specifically, sample data comprising 3523 samples were retrieved from the Leffingwell PMP 2001 database, and each sample in the sample data was input to the Cheminformatic RDKit package [3] using the Simplified Molecular Input Line Entry System (SMILES) to generate Mordred features in 1324 dimensions [1]. Next, a principal component analysis was performed to retain Mordred features with 95% variance, resulting in 50 dimensional features. Among the 3523 samples in the sample data, there was a total of 1392 samples with the “fruity” odor descriptor, 908 molecules with the “green” odor descriptor, 825 molecules with the “sweet” odor descriptor, and 553 molecules with the “floral” odor descriptor. Note that some samples were associated with more than one odor descriptor, so the total number of odor descriptors (3678) was greater than the total number of samples (3523). That is, one molecule can be associated with multiple odor descriptors (e.g., a molecule can be both “fruity” and “sweet”), such that the counts per odor descriptor are not mutually exclusive and can overlap. As a result, the sum of counts for all odor-descriptors may exceed the total number of unique molecules.

[0060] The odor prediction model was implemented using a support vector machine (SVM) that was optimized for each odor descriptor in the sample data. The uncertainty estimation model was implemented using the nonconformist package [4] for Mondrian inductive conformal prediction (MICP) with default parameters. A calibration set of 400 samples was randomly selected from the 3523 samples in the sample data and the remaining samples were used as the training set for training the odor prediction model 10. A separate set of 400 test samples was also selected from the 3523 samples and used to evaluate the performance of the odor prediction framework. The nonconformity score was calculated as one minus the odor class prediction obtained from the odor prediction model and was defined as the inverse probability error function.

[0061] To assess the accuracy and validity of odors predicted by the odor prediction framework, each odor descriptor in the sample data was treated as a binary classifier problem. Accordingly, based on the odor descriptors assigned in the sample data, each sample was classified as “fruity” or “non-fruity”, “green” or “non-green”, “sweet” or “non-sweet” and “floral” or “non-floral”. The following discussion considers only the “fruity” or “non-fruity” binary classifier for ease of explanation. Of the 400 test samples, 158 test samples were “fruity” (i.e., associated with the “fruity” odor descriptor) and 242 test samples were “non-fruity” (i.e., not associated with the “fruity” odor descriptor). In other words, of the 400 test samples for evaluation of the odor prediction framework, 158 test samples were associated with the “fruity” ground truth, and 242 test samples were associated with the “non-fruity” ground truth.

[0062] Table 1 shows the prediction set (i.e., the predicted odor) determined by the odor prediction framework for a first test sample, Ethyl Octanoate, which was associated with the “fruity” ground truth. More specifically, Table 1 shows the prediction set determined by the odor prediction framework for different user-defined significance levels (i.e., different confidence levels). For the first test sample, the odor prediction framework determined a p-value for “fruity” of 0.4900 and a p-value for “non-fruity” of 0.0771. These p-values indicate that the odor prediction framework 1 determined that the first test sample conforms better to the “fruity” odor class than the “non-fruity” odor class when compared to samples in the calibration set. When determining the predicted odor for the first test sample, the “fruity” and “non-fruity” odor labels were added to the prediction set if their respective p-values were greater than the user-defined significance level. Thus, for example, for a user-defined significance level of 0.20 (i.e., a confidence level of 0.80), only the “fruity” odor label was added to the prediction set. On the other hand, for a user-defined significance level of 0.01 (i.e., a confidence level of 0.99), both the “non-fruity” and “fruity” odor labels were added to the prediction set. Conversely, for a user-defined significance level of 0.50 (i.e., a confidence level of 0.50), neither the “non-fruity” nor the “fruity” odor labels were added to the prediction set (i.e. the prediction set was empty). In this manner, Table 1 demonstrates how a prediction set can be obtained on the basis of a user-defined significance level, therefore providing a quantifiable measure of prediction confidence.

[0063]

[0064] Table 2 shows the prediction set (i.e., the predicted odor) determined by the odor prediction framework for a second test sample, Diallyl Trisulfide, which was associated with the “non-fruity” ground truth. More specifically, Table 2 shows the prediction set determined by the odor prediction framework for different user-defined significance levels (i.e., different confidence levels). For the second test sample, the odor prediction framework determined a p-value for “fruity” of 0.0074 and a p-value for “non-fruity” of 0.9129. These p-values indicate the second test sample conforms better to the “non-fruity” odor class than the “fruity” odor class when compared to samples in the calibration set. When determining the predicted odor for the second test sample, the “fruity” and “non-fruity” odor labels were added to the prediction set if their respective p-values were greater than the user-defined significance level. Thus, for example, for a user-defined significance level between 0.50 and 0.01 (i.e., a confidence level of 0.50 to 0.99), only the “non-fruity” odor label was added to the prediction set. This indicates that the odor prediction framework predicts the “non-fruity” for test sample 2 across a broad range of confidence levels.

[0065]

[0066] Table 3 shows the confusion matrix for the predicted odors determined the odor prediction framework for the 400 test samples and the “fruity” or “non-fruity” binary classification. The main difference between the confusion matrix of Table 3 and the confusion matrix for a conventional machine learning model is that the addition of values for prediction sets that do not include either of the “fruity” or “non-fruity” odor labels (i.e., “Empty”) and for prediction sets which include both of the “fruity” and “non-fruity” odor labels (i.e., “Both”). Using a convention SVM model (i.e., without MICP), 134 test samples were predicted as “fruity” and 204 test samples were predicted as “non-fruity” correctly out of 158 “fruity” test samples and 242 “non-fruity” test samples. Similarly, according to the odor prediction framework of the present disclosure, 135 test samples were predicted as “fruity” and 194 test samples were predicted as “non-fruity”, with 14 uncertain samples (i.e., “Both”) at a user-defined significance level of 0.20. This demonstrates that odor prediction framework according to the present disclosure provides results that a comparable to convention SVM models, with the additional advantage of providing a quantifiable measure of prediction confidence.

[0067]

[0068] Table 4 shows the validity and efficiency results for the odor prediction framework according to the present disclosure for the “fruity”, “green”, and “floral” odor descriptors based on a user-defined significance level of 0.20, and for the “sweet” odor descriptor based on a user-defined significance level of 0.30. That data in Table 4 shows that the performance of the odor prediction framework is sufficient to make it highly usable, generating approximately 96% single-label predictions (efficiency) with an overall (global) validity of 86% for the “fruity” odor descriptor based on a user-defined significance level of 0.20. This performance also holds for the minority “floral” class despite the class imbalance in sample data (i.e., the “floral” odor descriptor is associated with the fewest number of test samples). This is a highly desirable feature for odor prediction, where often the minority class is of greatest interest. This balance in accuracy across classes is a particular feature of the MICP employed by the odor prediction framework of the present disclosure.

[0069]

[0070] Table 5 compares the performance of the odor prediction framework of the present disclosure (i.e., with MICP) in comparison to a conventional SVM mode (i.e., without MICP). For the odor prediction framework, as with Table 4, performance for the “fruity”, “green”, and “floral” odor descriptors is based on a user-defined significance level of 0.20, and performance for the “sweet” odor descriptor is based on a user-defined significance level of 0.30. For the SVM model, performance is based on the predicted probabilities, independent of significance level. Table 5 demonstrates that the odor prediction framework of the present disclosure, with MICP, can achieve the same level of true positive rate (TPR) as the conventional SVM model while also providing valid estimations of confidence level. This highlights that rejecting low-confidence samples identified by MICP can improve the prediction accuracy in general; for the fruity classification, for which the prediction sets for 14 samples were determined to include both the “fruity” and “non-fruity” odor labels. The MICP employed in the odor prediction framework also results in retrieval of more “fruity” samples, compared to conventional SVM models. The true positive rate and balanced accuracy (BA) increase at the expense of more false positives (40 for odor prediction framework, compared to 38 for the conventional SVM model, as shown in Table 3), thus decreasing the true negative rate (TNR). Furthermore, for imbalanced samples, such as the “floral” samples, the application of the MICP in the odor prediction framework according to the present disclosure clarifies how to adjust the cutoff level for label classification using a mathematically proven procedure for reclassification.

[0071]

[0072] While embodiments of the present disclosure have been described with reference to the drawings and examples, it should be noted that various modifications and revisions may be implemented by those skilled in the art based on this description and within the scope defined by the appended claims. Accordingly, such modifications and revisions are included within the scope of the present disclosure.

[0073] According to the present disclosure, a method, an information processing system, and a non-transitory computer-readable medium for determining a predicted odor can be provided. More particularly, the predicted odor can be determined together with a quantifiable measure of prediction confidence.REFERENCE DOCUMENTS LIST

[0074] [1] H. Moriwaki, et al., “Mordred: a molecular descriptor calculator”, J. Cheminform., 10, 4 (2018) [2] J. B Castro, et al., “Pyrfume: A Window to the World’s Olfactory Data”, bioRxiv 2022. [3] Rdkit: https: / / github.com / rdkit / rdkit [4] Nonconformist: https: / / github.com / donlnz / nonconformist

[0075] 1 Odor Prediction Framework 1A Odor Prediction Component 1B Model Training Component 10 Odor Prediction Model 20 Uncertainty Estimation Model 30 Training Module 40 Calibration Module 100 Information Processing System 102 Processor System 104 Memory System

Claims

1. A method for determining a predicted odor, the method comprising: receiving input data specifying a plurality of features of a molecule; determining an odor class prediction for the molecule by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class; determining a conformity measure for the odor class based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules; and determining the predicted odor of the molecule based on the conformity measure for the odor class.

2. The method according to claim 1, wherein determining the conformity measure for the odor class comprises: calculating a non-conformity score for the molecule based on the probability for the odor class; and determining the conformity measure for the odor class based on the non-conformity score for the molecule and the plurality of non-conformity scores corresponding to the plurality of calibration molecules.

3. The method according to claim 2, wherein the conformity measure for the odor class is calculated as the proportion of non-conformity scores in the plurality of non-conformity scores corresponding to the plurality of calibration molecules that are greater than or equal to the non-conformity score for the molecule.

4. The method according to claim 3, wherein the conformity measure for the odor class is a p-value.

5. The method according to claim 4, wherein the p-value is calculated according to:

6. The method according to claim 1, wherein the predicted odor is a prediction set, and determining the predicted odor comprises adding an odor label corresponding to the odor class to the prediction set when the conformity measure for the odor class is greater than a user-defined significance level.

7. The method according to claim 6, wherein the odor label is indicative of an odor that is predicted to be perceived by a user with a confidence level corresponding to the user-defined significance level.

8. A method for determining calibration data for use in determining a predicted odor, the method comprising: receiving sample data specifying, for each calibration molecule in a plurality of calibration molecules, a plurality of features; determining, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying an odor prediction model to the sample data, wherein the odor class prediction specifies a probability for an odor class; determining, for each calibration molecule in the plurality of calibration molecules, a non-conformity score based on the probability for the odor class; and outputting, for each calibration molecule in the plurality of calibration molecules, the non-conformity score as the calibration data.

9. The method according to claim 8, further comprising training the odor prediction model based on training data specifying, for each training molecule in a plurality of training molecules, a plurality of features and a ground truth odor label.

10. The method according to claim 9, wherein the plurality of calibration molecules and the plurality of training molecules are mutually exclusive.

11. The method according to any one of claims 1 to 10, wherein the odor prediction model is a trained machine learning model.

12. The method according to claim 11, wherein the trained machine learning model is a support vector machine.

13. The method according to any one of claims 2 to 10, wherein the non-conformity score is calculated as one minus the probability for the odor class.

14. An information processing system comprising: a processor system; and a memory system; wherein the memory system stores instructions which, when executed by the processor system, cause the information processing system to implement a method for determining a predicted odor of a molecule, the method comprising: receiving input data specifying a plurality of features of the molecule; determining an odor class prediction for the molecule by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class; determining a conformity measure for the odor class based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules; and determining the predicted odor for the molecule based on the conformity measure for the odor class.

15. An information processing system comprising: a processor system; and a memory system; wherein the memory system stores instructions which, when executed by the processor system, cause the information processing system to implement a method for determining calibration data for use in determining a predicted odor of a molecule, the method comprising: receiving sample data specifying, for each calibration molecule in a plurality of calibration molecules, a plurality of features; determining, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying an odor prediction model to the sample data, wherein the odor class prediction specifies a probability for an odor class; determining, for each calibration molecule in the plurality of calibration molecules, a non-conformity score based on the probability for the odor class; and outputting, for each calibration molecule in the plurality of calibration molecules, the non-conformity score as the calibration data.

16. A non-transitory computer-readable medium storing instructions which, when executed by an information processing system, cause the information processing system to implement a method for determining a predicted odor of a molecule, the method comprising: receiving input data specifying a plurality of features of the molecule; determining an odor class prediction for the molecule by applying an odor prediction model to the input data, wherein the odor class prediction specifies a probability for an odor class; determining a conformity measure for the odor class based on calibration data specifying a plurality of non-conformity scores corresponding to a plurality of calibration molecules; and determining the predicted odor for the molecule based on the conformity measure for the odor class.

17. A non-transitory computer-readable medium storing instructions which, when executed by an information processing system, cause the information processing system to implement a method for determining calibration data for use in determining a predicted odor of a molecule, the method comprising: receiving sample data specifying, for each calibration molecule in a plurality of calibration molecules, a plurality of features; determining, for each calibration molecule in the plurality of calibration molecules, an odor class prediction by applying an odor prediction model to the sample data, wherein the odor class prediction specifies a probability for an odor class; determining, for each calibration molecule in the plurality of calibration molecules, a non-conformity score based on the probability for the odor class; and outputting, for each calibration molecule in the plurality of calibration molecules, the non-conformity score as the calibration data.