Method and system for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition

Abstract

The method (100) for predicting physicochemical and / or odor characteristic values ​​of a chemical structure or composition comprises: defining a representation of a chemical structure or composition (105); Executing (110) an end-to-end trained ensemble neural network or multi-branch neural network model based on said defined representation to predict physicochemical and / or odor attribute values; - providing said physicochemical and / or odor characteristic values ​​(115). The method further comprises: - providing example data to an end-to-end ensemble neural network or multi-branch neural network device (120), wherein the end-to-end ensemble neural network or multi-branch neural network device: several neural network sub-devices configured to make independent predictions; a layer outputting at least one value of said distribution of independent predictions, - providing (120), wherein the layer comprises a sampling device configured to output random values; - operating (125) the end-to-end ensemble neural network or multi-branch neural network device; - obtaining (130) the trained ensemble neural network or multibranch neural network model.

Description

[Technical Field]

[0001] The present invention is directed to a method for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition, a system for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition, and a method for efficiently assembling a chemical structure or composition.

[0002] The invention has particular application in the fragrance and perfume industry.

[0003] Background technology In scientific experiments, measurements 305 stored in a database, such as that shown in FIG. 3, vary depending on the environment of the experiment. To achieve stable conditions, an ideal environment, such as the International Space Station, is typically required. Even with near-perfect environmental conditions, technicians' equipment and sample preparation can vary slightly from one experiment to another. Integrating experimental data from different sources can be challenging because experimental variability varies depending on the measurement conditions used. Statistical methods have been developed to homogenize experimental data and reduce these variability with moderate to good success. In the field of machine learning, such variability can exist between training and test sets, as well as between known and future data.

[0004] Another well-known problem with machine learning models is the number of hyperparameters in the model, which can significantly affect the model's ability to overfit to the training data 310, as shown in Figure 3. The amount of available data may be insufficient to train the number of hyperparameters in a neural network. Therefore, it is best to consider the minimum necessary to extract a meaningful digital representation for the data. The latter can be achieved by prioritizing layers with fewer parameters. One example of such an attempt is replacing densely connected layers with compact convolutional layers. Convolutional layers are particularly specialized for extracting local features from data, such as images. However, even by using compact and efficient convolutional layers, networks with billions of parameters can still be generated. Trending megamodels have shown that such models can only be trained with very large datasets, such as those available for images, literature, and music. Conversely, to avoid expensive experiments and eliminate animal testing, scientific fields such as chemistry and biology often have limited datasets with hundreds to thousands of data points. Therefore, in such fields, the goal should be to reduce the number of parameters to match the size of the data.

[0005] One way to compensate for the large size of a network is through data augmentation. Indeed, performance improvements with increasing augmentation rates indicate that more data is needed for the chosen network size, suggesting that the network size may be reduced (Tetko, IV, Karpov, P., Van Deursen, R. et al. State-augmented NLP transformer models for direct and single-step retrosynthesis. Nat Commun 11, 5575 (2020)). At the same time, augmentation can be used to identify whether the network is critically parameterized, i.e., the point at which augmentation has no or little impact on the model's performance. Not all models support data augmentation. For example, graph neural networks (GNNs) are invariant to representation shuffling. Therefore, GNNs are incompatible with existing data augmentation methods used for natural language processing or images.

[0006] The third issue is that the model training procedure 315 defines a critical aspect of modeling. Often, there are too many variables that can influence the model decision. This can explain why hyperparameterization optimization strategies may be required to improve the model in terms of performance or efficiency. Apart from the selected model, the issue of data partitioning between the training set and the test set also plays an important role. Several methods, from full leave-one-out, random partitioning, to K-fold cross-validation, can be used to simulate and estimate model quality for unseen data. Ultimately, model predictions are merely educated guesses depending on the training conditions, model size, optimization parameters, and data partitioning used. Once complete, it is impossible to know whether the best model was actually trained. However, it is assumed that the calculated model is the best model for the test points used. This is a general limitation of data modeling techniques, as one should not necessarily expect performance results to be the same for all portions of future unseen data. Note that future performance can also vary significantly depending on the evaluated sample size of the unseen data and any sample biases that may have been introduced in the unseen data. One way to partially address these shortcomings is to predict an accurate theoretical endpoint as a standardized metric for evaluating models. An example of such an endpoint in chemistry is the molecular weight of a molecule.

[0007] The field of digital modeling of chemical species and reactions using neural networks has three main branches:

[0008] The first branch is learning models from graph neural networks (GNNs). They can use any molecular input format to calculate atomic properties. This format is not easily scalable, making it difficult for small datasets like those commonly found in all chemistry.

[0009] The second branch is NLP methods based on line-extended strings (such as the SMILES format), where chemistry is learned exclusively from this syntax. This method has the advantage of data augmentation, since the same molecule can be described by writing new sentences in different rule-based sequences (sentence types).

[0010] The third branch is image convolutional neural networks that learn and predict from molecular images.

[0011] Such techniques require rich data sets, which are rare in the fields of fragrance design and olfactory testing, perfumery, fine fragrance manufacturing, and fragrance design. Without rich data sets, the use of neural network techniques can lead to inefficient models due to the risk of memorization by the network, given the number of parameters that need to be considered.

[0012] Furthermore, the input of such graph neural networks is to convert molecules into a specified input format, usually the SMILES format of molecular structures, which is insufficient for creating efficient chemical species features for chemical reaction prediction models.

[0013] Ensembling is a technique that consists in training several models (usually called foundation models or weak learners) and aggregating their outputs at inference time with some voting mechanism.

[0014] This technique is widely used by experts (especially to obtain winning solutions in many machine learning competitions) and is often a key step in improving final performance.

[0015] While the use of ensembles is very common, finding the best ensemble procedure for building, training, and combining foundation models is generally nontrivial. Traditionally, ensemble techniques have attempted to generate a diverse (or complementary) set of foundation models and combine them using some kind of voting technique, usually meant to reduce bias and / or variance in the resulting system. A number of different techniques can be used to train diverse models. For example, bagging (with bootstrap resampling) introduces diversity through sampling of the training dataset, while boosting sequentially introduces diversity through training models in such a way that each model has an incentive to compensate for errors made by its predecessor. Voting techniques can consist of simple averaging, majority voting (for classification), or stacking, whereby the final prediction is generated by a meta-model trained to combine foundation models on several holdout datasets.

[0016] While these techniques can work very well in practice, they primarily consist in handcrafting heuristics to enhance diversity or complementarity between underlying models. While these heuristics can be used, for example, to make underlying models complementary during training (e.g., in boosting) or inference (e.g., in stacking), the models do not directly learn how to best complement each other. In particular, a model cannot explicitly capture the fact that it is part of an ensemble.

[0017] To address this challenge, proposals have been made to train all the underlying models jointly (or end-to-end). In the context of neural networks, this means considering each of the underlying models as part of a larger neural network and training them all jointly using a common loss.

[0018] Interestingly, this blurs the notion of ensembles and multi-head (or multi-branch) networks, as each of the underlying models can be seen as a separate branch within a single neural network model. While this end-to-end approach is attractive, it is known that blindly optimizing the global loss of the entire system often does not yield the best results, and it is usually better to perform some degree of individual training of the underlying models (often controlled by specific terms in the loss).

[0019] Currently, the most well-known method for training such end-to-end models is often to try different interpolations between the individual loss terms and a global (or distilled) loss term, and in general, the best approach appears to be problem and model dependent.

[0020] Summary of the Invention The present invention aims to address all or some of these drawbacks.

[0021] According to a first aspect, the present invention is directed to a method for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition, said method comprising: - defining, on a computer interface, a digital representation of a chemical structure or composition; - executing, by a computing device, an end-to-end trained ensemble neural network or multi-branch neural network model on the defined digitized representation to predict at least one physicochemical value and / or odor characteristic value of the chemical structure or composition; - providing, on a computer interface, at least one physicochemical value of the chemical structure or composition and / or an odor characteristic value, moreover, - providing an exemplary set of data to an end-to-end ensemble neural network or multi-branch neural network device, the exemplary set of data including at least one set of inputs, the inputs corresponding to digitized representations of chemical structures or compositions, and at least one set of outputs, the outputs corresponding to physicochemical and / or odor properties associated with the set of inputs, the end-to-end ensemble neural network or multi-branch neural network device: several neural network sub-devices, each sub-device configured to provide an independent prediction based on example data; a layer configured to output at least one value based on or representative of the distribution of the independent predictions. providing a layer comprising a sampling device configured to output at least one random value as a function of a probability distribution representing a distribution of independent predictions, the output random value being calculated in a differentiable manner and used for backpropagation within an end-to-end ensemble neural network or multibranch neural network device; - operating an end-to-end ensemble neural network or multibranch neural network device based on a set of exemplary data; - obtaining a trained end-to-end ensemble neural network or multi-branch neural network model configured to predict physicochemical and / or odor properties for an input digitized representation of a chemical structure or composition.

[0022] Such provision allows for accurate prediction of physicochemical and / or odor characteristic values ​​of defined chemical structures or compositions.

[0023] Such provision similarly leads to much higher prediction stability and reliability, improves training speed and overall performance, and also allows for the provision of a variance metric that represents the model uncertainty. Thus, such embodiments allow for resource savings in terms of computation time or power, and in terms of model complexity. Current approaches typically require the use of a large number of models and iterations to obtain a reliable predictive model.

[0024] Furthermore, such provision enables the trained model to reach higher accuracy than competing approaches that either require designing diversity among the underlying models or rely on fine-tuned loss functions to balance the objectives of training individual models in conjunction with the ensemble.

[0025] Such a provision also provides a simple means of regularizing the ensemble by introducing noise. Finally, such a provision allows for more stable training dynamics and better individual underlying models. This approach requires no additional tuning and does not introduce any new learnable parameters.

[0026] In certain embodiments, the at least one input set of exemplary data corresponds to a hash vector of at least one atomic property in the chemical structure or composition, and the method further includes, upstream of the performing step, a step of converting the defined digitized chemical structure or composition into a set of hash vectors of at least one atomic property representing the digitized chemical structure or composition, said set of hash vectors being used as input during the performing step.

[0027] Such provision has proven particularly efficient in increasing the reliability of the prediction results in the context of the prediction of physicochemical and / or odor properties.

[0028] In certain embodiments, at least one hash vector of the atomic characteristic is: - atomic number of the corresponding atom, - atomic symbols of the corresponding atoms, - atomic mass, - explicit map number, - row index in the periodic system, - column index in the periodic system, -total number of hydrogens on the atom, - the implicit number of hydrogens on the atom, - the explicit number of hydrogens on the atom, - atomic degree, -total degree of atoms, - valence state of the atom, - the implied valence of the atom, - the explicit valence of the atom, -formal charge on the atom, - partial charges on atoms, -electronegativity on the atom, - number of bonds by bond type, - atomic number, number of neighboring atoms by wildcard, - Bond type + atomic number, number of neighboring atoms by wildcard - number of neighboring atoms by wildcard, -aromaticity value, - a value indicating an aliphatic atom, - a value indicating the conjugated atoms, - a value indicating a ring atom, - a value indicating a macrocyclic atom, - a value indicating the structural constraint atom, - a value indicating an electron-withdrawing atom, - a value indicating the electron donating atom, - a value indicating the reactive site, - a value indicating the hydrogen bond donor, - a value indicating the hydrogen acceptor, - a value indicating multivalency as a hydrogen bond donor, - a value indicating multivalency as a hydrogen bond acceptor, - number of cycles on the atom, -Atomic ring size, - the hybridization state of the atom, -values ​​indicating atomic structure, - number of electrons in the atomic orbital, - number of lone pair electrons, -radical state, -Isotopes on atoms, -atom-centrosymmetric functions, -Relative stereochemistry values ​​as clockwise chi, counterclockwise chi, - absolute stereochemical value, - absolute stereochemical value, - double bond stereochemistry value, - priority values ​​for determining stereochemistry, a value representing the positive or negative influence of the atom on the determined training target to indicate its enrichment contribution to the knowledge base; - a value representing the positive or negative influence of the atom on the determined training target to indicate its dilution contribution to the knowledge base; and / or - represents one of the ring stereochemistry values.

[0029] In certain embodiments, at least one hash vector of the binding characteristic is: -bond order, -bond type, -bond stereochemistry: -bond orientation in tetrahedral stereochemistry, -bond orientation of double bond stereochemistry or -bond direction of spatial orientation, -Atomic numbers of atoms "from" and / or "to" - atomic symbols for atoms "from" and / or "to", -dipole moment in the bond, -Quantum-chemical properties: -electron density in the bond, -Electron configuration of the bond, -bonding orbital, -Binding energy -Gravity, -repulsion force, - bond distance, - aromatic bond, -aliphatic bond, -Ring properties of bonds: - number of rings on the bond, -bond ring size - the minimum ring size of the bond, -Maximum ring size of the bond, - rotatable bonds, - spatially constrained bonds, -hydrogen bonding properties, -ionic bonding properties, - Bond order of the reaction, including "null" bonds to identify bonds broken / formed during the reaction: - the bond order in the reagent, - the bond order of intermediate products, or -bond order in the transition state.

[0030] In certain embodiments, the at least one output value representing the distribution represents the variance of the distribution.

[0031] In certain embodiments, the end-to-end ensemble neural network or multi-branch neural network device is trained to minimize at least one value representing the variance of the distribution.

[0032] In certain embodiments, the at least one odor characteristic is: - insect repellent performance value, -sensory characteristic values, -biodegradability value, -Antibacterial value, -Odor detection threshold, - odor intensity value, -Top-Heart-Base Values, -hazard value, -Biological activity on taste -Olfactory biological activity, -biological enhancement or modulation of taste activity, - biological enhancement or modulation of olfactory activity, and / or -Describes an olfactory smell.

[0033] In certain embodiments, the at least one physical property is: - boiling point value, - melting point value, - water solubility value, - Henry's constant value, -vapor pressure value, -volatile value, or -Headspace concentration value.

[0034] In certain embodiments, at least one neural network device comprises: -Recursive Neural Network Device, -Graph Neural Network Device, -Variational Autoencoder Neural Network Device, or -Autoencoder neural network device.

[0035] In a particular embodiment, the method object of the present invention includes a step of atom or bond relationship vector expansion upstream of the providing step.

[0036] Such a provision would initially limit the dataset available to those currently required for neural network applications. Indeed, a single molecular structure represented by a single hash or an extended series of hashes can be extended a maximum number of times corresponding to the number of hashes in the series. Thus, a single molecular structure can serve as several inputs in a natural language processing application.

[0037] In certain embodiments, the step of atom or bond relationship vector expansion includes a step of horizontal expansion configured to provide several vectors representing a single digitized representation of a molecular structure or composition, each vector representing a particular representation of the molecular structure or composition of the canonical representation, and each vector being treated as a single input in the providing step.

[0038] Such a provision would initially limit the dataset available to those currently required for neural network applications. Indeed, a single molecular structure represented by a single hash or an extended series of hashes can be extended a maximum number of times corresponding to the number of hashes in the series. Thus, a single molecular structure can serve as several inputs in a natural language processing application.

[0039] In certain embodiments, the step of atomic or bond relationship vector expansion includes a step of vertical expansion to create several groups of several horizontal expansions representing unique molecular structures or compositions, each group being treated as a single input in the providing step.

[0040] Such a provision would initially limit the dataset available to those currently required for neural network applications. Indeed, a single molecular structure represented by a single hash or an extended series of hashes can be extended a maximum number of times corresponding to the number of hashes in the series. Thus, a single molecular structure can serve as several inputs in a natural language processing application.

[0041] According to a second aspect, the present invention is directed to a method for efficiently assembling a chemical structure or composition, the method comprising: - executing a method object of the present invention; - assembling a chemical structure or composition associated with the output obtained during the obtaining step.

[0042] Such provision allows for the realization of chemical structures upon which odor property predictions can be performed.

[0043] According to a third aspect, the present invention is directed to a system for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition, the system comprising: - means for defining, on a computer interface, a digitized representation of a chemical structure or composition; - means for executing, by a computing device, an end-to-end trained ensemble neural network or multi-branch neural network model on the defined digitized representation in order to predict at least one physicochemical and / or odor characteristic value of the chemical structure or composition; means for providing, on a computer interface, at least one physicochemical and / or odor characteristic value of the chemical structure or composition. The system comprises: - means for providing an exemplary set of data to an end-to-end ensemble neural network or multi-branch neural network device, the exemplary set of data including at least one set of inputs, the inputs corresponding to digitized representations of chemical structures or compositions, and at least one set of outputs, the outputs corresponding to physicochemical and / or odor properties associated with the set of inputs, the end-to-end ensemble neural network or multi-branch neural network device comprising: several neural network sub-devices, each sub-device configured to provide an independent prediction based on example data; a layer configured to output at least one value based on or representative of the distribution of said independent predictions; - means for providing, wherein the layer comprises a sampling device configured to output at least one random value as a function of a probability distribution representing a distribution of independent predictions, the output random value being calculated in a differentiable manner and used for backpropagation within an end-to-end ensemble neural network or multibranch neural network device; means for operating an end-to-end ensemble neural network or multi-branch neural network device based on a set of exemplary data; - means for obtaining a trained end-to-end ensemble neural network or multi-branch neural network model configured to predict physicochemical and / or odor properties for an input digitized representation of a chemical structure or composition.

[0044] The advantages of the system objects of the present invention are similar to the advantages of the method objects of the present invention. Furthermore, all embodiments of the method objects of the present invention can be reproduced, mutatis mutandis, in the system objects of the present invention.

[0045] Other advantages, objects, and particular features of the present invention will become apparent from the following non-exhaustive description of at least one particular embodiment or sequence of steps of the invention when taken in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0046] [Figure 1] FIG. 2 is a diagram illustrating a first particular sequence of steps of a method object of the present invention; [Figure 2] FIG. 2 is a diagram illustrating a specific embodiment of a system object of the present invention. [Figure 3] FIG. 1 is a diagram illustrating a general overview of a machine learning system. [Figure 4] FIG. 10 is a diagram illustrating a second particular sequence of steps of a method object of the present invention. [Figure 5] FIG. 10 is a schematic diagram illustrating a detailed view of a particular embodiment of a new neural network layer used during training of an end-to-end ensemble neural network or multibranch neural network device. [Figure 6] FIG. 10 is a diagram illustrating a second particular sequence of steps of a method object of the present invention. [Figure 7] FIG. 10 illustrates, in outline, a particular sequence of steps for obtaining a hash vector used by a system or method object of the present invention. [Figure 8] FIG. 2 is a diagram illustrating an example of an implementation of a training method object of the present invention. [Figure 9] FIG. 1 shows performance results for three odor characteristics. [Figure 10] FIG. 1 shows performance results for three odor characteristics. [Figure 11] FIG. 1 shows performance results for three odor characteristics. [Figure 12] FIG. 1 shows a schematic representation of a training architecture that can be used to select, classify, or predict odor and / or physicochemical properties of chemical structures. [Figure 13] FIG. 1 shows a schematic representation of a training architecture that can be used to select, classify, or predict odor and / or physicochemical properties of chemical structures. [Figure 14] FIG. 1 shows a schematic representation of a training architecture that can be used to select, classify, or predict odor and / or physicochemical properties of chemical structures. [Figure 15] FIG. 1 illustrates a schematic diagram of a particular training architecture that can be used to classify chemical structures.

[0047] MODE FOR CARRYING OUT THE INVENTION This description is not exhaustive, as each feature of one embodiment may be advantageously combined with any other feature of any other embodiment.

[0048] Various inventive concepts may be embodied as one or more methods, examples of which are provided. The actions performed as part of a method may be ordered in any suitable manner. Thus, while exemplary embodiments show actions as sequential, embodiments can be constructed in which actions are performed in a different order than illustrated, which may include performing some actions simultaneously.

[0049] The indefinite articles "a" and "an," as used in this specification and claims, unless expressly indicated otherwise, should be understood to mean "at least one."

[0050] The term "and / or," as used in the specification and claims, should be understood to mean "either or both" of the elements so conjoined, i.e., elements that are present conjunctively in some cases and disjunctively in other cases. Multiple elements listed with "and / or" should be construed in the same manner, i.e., "one or more" of the elements so conjoined. Other elements may optionally be present other than the elements specifically identified by the "and / or" clause, whether related or not to the elements specifically identified. Thus, as a non-limiting example, a reference to "A and / or B," when used in conjunction with open-ended language such as "comprising," can, in one embodiment, refer to A only (optionally including elements other than B); in another embodiment, refer to B only (optionally including elements other than A); in yet another embodiment, refer to both A and B (optionally including other elements); and so on.

[0051] As used in this specification and the claims, "or" should be understood to have the same meaning as "and / or" as defined above. For example, when separating items in a list, "or" or "and / or" should be interpreted as inclusive, i.e., the inclusion of at least one, but including more than one of a plurality of elements or a list of elements, and optionally, additional unlisted items. Only terms clearly indicated to the contrary, such as "only one of" or "exactly one of," or, when used in the claims, "consisting of," refer to the inclusion of exactly one element of several elements or a list of elements. In general, the term "or" as used herein, when used in the claims, when preceded by terms of exclusivity, such as "either," "one of," "only one of," "exactly one of," "consisting essentially of," etc., shall be interpreted only as indicating exclusive alternatives (i.e., "one or the other, but not both") and shall have its ordinary meaning as used in the field of patent law.

[0052] As used in this specification and claims, the phrase "at least one" in reference to a list of one or more elements should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but does not necessarily include at least one of each and every element specifically listed in the list of elements, nor does it exclude any combination of elements in the list of elements. This definition also allows for the optional presence of elements other than those specifically identified in the list of elements to which the phrase "at least one" refers, whether related to the specifically identified elements or not. Thus, as a non-limiting example, "at least one of A and B" (or, equivalently, "at least one of A or B," or, equivalently, "at least one of A and / or B") can refer in one embodiment to at least one, optionally including more than one, A, with no B present (and optionally including elements other than B); in another embodiment to at least one, optionally including more than one, B, with no A present (and optionally including elements other than A); in yet another embodiment to at least one, optionally including more than one, A, and at least one, optionally including more than one, B (and optionally including other elements); and so forth.

[0053] In the claims and the above specification, all transitional phrases such as "comprising," "including," "carrying," "having," "containing," "involving," "holding," "composed of," and the like, are to be understood to be open-ended, i.e., to mean including, but not limited to. Only the transitional phrases "consisting of" and "consisting essentially of," respectively, shall be closed or semi-closed transitional phrases.

[0054] At this point, please note that the drawings are not to scale.

[0055] As used herein, the term "ingredient" refers to any ingredient that preferably exhibits fragrance or perfumery capabilities. The terms "compound" or "ingredient" refer to the same item as "volatile ingredient." An ingredient may be formed from one or more chemical molecules.

[0056] The term composition denotes a liquid, solid or gaseous collection of at least two perfume or fragrance ingredients, or one perfume or fragrance ingredient and an aprotic solvent for dilution.

[0057] As used herein, "aroma" refers to the olfactory perception resulting from the sum of the activation, enhancement, and inhibition (if present) of odorant receptor(s) by at least one volatile component via orthonasal and retronasal olfaction and the activation of taste buds containing taste receptor cells. Thus, by way of example and not intended to limit the scope of the present disclosure in any way, "aroma" results from the olfactory and taste perception resulting from the sum of a first volatile component that activates odorant receptors or taste buds associated with coconut tonality, a second volatile component that activates odorant receptors or taste buds associated with celery tonality, and a third volatile component that inhibits odorant receptors or taste buds associated with hay tonality.

[0058] As used herein, "fragrance" refers to an olfactory perception resulting from the aggregation of the activation, enhancement, and inhibition (if present) of one or more odorant receptors by at least one volatile component. Thus, by way of example and not intended to limit the scope of the present disclosure in any way, a "fragrance" results from the aggregation of a first volatile component that activates odorant receptors associated with coconut notes, a second volatile component that activates odorant receptors associated with celery notes, and a third volatile component that inhibits odorant receptors associated with hay notes.

[0059] As used herein, "odor characteristics" or "olfactory characteristics" refer to any psychophysical property of an ingredient or composition, i.e., such properties refer to how the human body responds to the physical presence of an olfactory ingredient or composition, taking into consideration that such psychophysical properties are directly related to the ability of the ingredient or composition to readily penetrate and, by proximity, contact olfactory receptors present in the human body.

[0060] As used herein, the term "inputting means" refers to, for example, a keyboard, mouse, and / or touchscreen adapted to interact with the computing system to collect user input. In variations, the inputting means is logical in nature, such as a network port of the computing system configured to receive electronically transmitted input commands. Such inputting means may be associated with a GUI (Graphical User Interface) or API (Application Programming Interface) presented to the user. In other variations, the inputting means may be a sensor configured to measure certain physical parameters relevant to the intended use case.

[0061] As used herein, the term "computing system" or "computer system" refers to any electronic computing device, whether integrated or distributed, that can receive numerical input and provide numerical output by any type of interface, digital and / or analog. Typically, a computing system designates either a computer running software with access to data storage, or a client-server architecture in which the client side acts as the interface while the data and / or computations are performed on the server side.

[0062] As used herein, the term "digital identifier" refers to any computerized identifier, such as an identifier used in a computer database, that represents a physical object, such as a fragrance ingredient. A digital identifier can refer to a label that represents the name, chemical structure, or internal reference of the fragrance ingredient.

[0063] As used herein, the term "embodied" is intended to exist outside the digital environment of the present invention. "Embodied" can mean, for example, readily found in nature or synthesized in a laboratory or chemical plant. In either case, an embodied composition exhibits a tangible reality. The terms "formulated" or "formulation" refer to the act of materializing a composition, whether by extraction and assembly of components or by synthesis and assembly of components.

[0064] As used herein, the term "atomic property" refers to a property of an atom and / or a property of a bond attached to any atom, regardless of molecular context. Thus, atomic property refers to an absolute description of an atom's characteristics, as opposed to a relative description of an atom within a molecule in the broader context of the molecule of which such atom is a part.

[0065] As used herein, the term "activation function" defines how, in a neural network, a weighted sum of inputs is transformed into an output from one or more nodes in a layer of the network. These activation functions may be defined by the layers in the network or by the mathematical solution of a loss function.

[0066] As used herein, an "end-to-end ensemble neural network or multi-branch neural network device" refers to a group of independent neural network devices that cooperate to provide an output, as well as a single neural network device that includes independent branches that cooperate to provide an output.

[0067] As used herein, the term "atomic property" refers to a property of an atom and / or a property of a bond attached to any atom, regardless of molecular context. Thus, atomic property refers to an absolute description of an atom's characteristics, as opposed to a relative description of an atom within a molecule in the broader context of the molecule of which such atom is a part.

[0068] The embodiments disclosed below are presented in a general manner.

[0069] Figure 3 shows an overview of the main components of machine learning.

[0070] Figure 5 illustrates a specific embodiment of two layers of the end-to-end ensemble neural network or multibranch neural network training device object of the present invention. Figure 5 also helps to understand the technical contribution of the present invention. The theory underlying the model shown in Figure 5 is presented below.

[0071]

number

[0072] The kth foundation model is k =M k (x)∈R h The output is , and h is the output dimension of the underlying model. The layer inputs o k , k=1...K, and o~D(g(o 1 , ..., o K )), where ~ denotes differentiable sampling, D is a multivariate distribution, and g is a function that maps the vector ok to the parameters of the distribution, o∈R. hhas the same dimension as the individual input vectors. Using the output of this layer, o, the final output of the network is, e.g.

number

[0073] Using reparameterization tricks, sampling can be done in a differentiable way, making it compatible with gradient descent-based neural network training. Thus, in contrast to traditional ensemble methods such as bagging or stacking, which separate the training of each model in the ensemble, our layer ensures that gradients are provided to all underlying models for all training samples, thereby creating an end-to-end form of training.

[0074] There are different options for the calculations performed by this layer, specified by D and g. As an example, a simple variant is shown in Figure 5, where D is a Gaussian distribution parameterized by a diagonal covariance matrix and the function g(o 1 , ..., o K ) is oi~N(μ i , σ 2 i ) so that the parameter μ∈R h and σ∈R h Calculate.

[0075] D(g(o 1 , ..., o K Several different methods for constructing and sampling from ) are disclosed below.

[0076] - Averaging: No sampling is done, the average of all underlying models is taken and used to calculate the output. In essence, we train a multi-head neural network in which the outputs of the heads are simply averaged. More precisely,

number

[0077] - Diagonal: This corresponds to the example shown in Figure 5. During training, the output of the underlying model k from,

number

number

number

[0078] -Diagonal parameterization: This method is similar to diagonal parameterization, but instead

number

number

number

number

[0079] - Full Covariance: This is a diagonal extension, where instead of using a diagonal covariance matrix, sampling is done using the full covariance matrix. Again, during training, the layer

number

number

number

number

[0080] The performance of this architecture allows us to evaluate the performance of our approach on the CIFAR-10 image classification task. Each competing model is trained using 5 random seeds and 120 epochs. The test loss is calculated on the entire test set using a regular split of CIFAR-10, with the columns and test set consisting of 50,000 and 10,000 images, respectively.

[0081] To evaluate sampling as a novel technique for end-to-end ensemble training, the training method object of the present invention is compared with different ensemble methods. All ensemble methods use K=8 base models, which are standard CNNs with ReLU and batch normalization layers. Each base model has 68,906 parameters, resulting in a total of 551,248 parameters for each ensemble. Different variations, as described above, are evaluated, including a parameterized isotropic variation with a multilayer perceptron used for the function l(·). It is observed that when uniform-based weight initialization is used, training can be initially unstable when diagonal sampling is used. This is due to the fact that the initial base model is not sufficiently diverse at the beginning of training, resulting in a standard deviation close to zero that makes Gaussian sampling prone to numerical instability. Therefore, Gaussian- or orthogonal-based initialization can be used, which do not appear to suffer from this issue. In this particular embodiment, a version of bagging based on random initialization of the network weights, coupled with random shuffling of data points, is used. Finally, negative correlation learning ("NCL") is used, i.e.,

number

number

[0082] In addition to the ensemble method, our results can be compared to those obtained by a standalone CNN of similar performance to the ensemble, both without ("Simple") and with dropout ("Simple + Dropout"). This CNN has a similar structure to the CNN used for the base model, but with 506,290 parameters, which can be obtained by increasing the depth and number of channels.

[0083] The validation accuracy of different models may be used as a measure of performance. The coefficient of variation provides a measure of diversity among ensemble members throughout training. 1 , ..., o K Finally, the mean testing accuracy of the underlying models may be used as a performance metric. This measures the degree of distillation during training, i.e., the performance of each independent underlying model on the test set.

[0084] From this comparison, It can be seen that the average performs better than a single head, showing that multiple heads can already provide some advantage even when the outputs of the underlying models are simply averaged during training. However, the average is deterministic and does not inject any noise. By comparing the diagonal with the average, the diagonal clearly provides better results, so the advantage provided by sampling during training compared to simple averaging can be seen.

[0085] -Full covariance sampling ends up having better validation accuracy than the simpler diagonal alternative. The coefficient of variation is larger for full covariance sampling compared to diagonal, which seems to indicate that full covariance benefits from better diversity. Furthermore, comparing the two sampling methods, full covariance sampling provides better aggregate test accuracy starting from about the 30th epoch of training. Note, however, that full covariance only outperforms diagonal for test accuracy from roughly the 60th epoch onward. In other words, even when test accuracy is poor, full covariance provides better averaged individual test accuracy than diagonal. Thus, full covariance offers better distillation properties. Overall, here, sampling appears to perform better with a richer distribution. However, full covariance only outperforms the other methods after a few dozen epochs, and it is more computationally expensive. Overall, the gain comes at the expense of more computation for the same number of parameters.

[0086] By comparing diagonal and diagonal MLPs, we can see that the latter is purely advantageous, and using such MLP parameterization functions, 1 , ..., o K This suggests that computing the parameter D from the concatenated vector of is less preferable than directly computing the element means and standard deviations of the vector. However, it is unclear whether the performance gap is due to the inductive bias of the direct (nonparametric) option or to some other factor. Therefore, it is proposed to compute the parameters of D directly, which also has the advantage of not adding new learnable parameters.

[0087] The training device of the present invention is presumed to function primarily as a regularization mechanism, thereby enabling end-to-end training of ensembles that would otherwise be prone to overfitting. Indeed, averaging is inferior to bagging in terms of test accuracy and distillation, even when averaging already provides a net gain over unity. Nevertheless, as soon as sampling is used, better test accuracy is obtained, and distillation is obtained in favor of the sampling method. Moreover, injecting noise in this way allows the sampling procedure to avoid overfitting during training. 1 , ..., o K In comparison, relying on a noise schedule is somewhat more complex and requires careful tuning.

[0088] The performance gap between -diagonal and full covariance indicates that the more sophisticated distributions provide better representation to the ensemble, suggesting that the sampling distribution is able to capture part of the data and thus does more than a noise mechanism that provides regularization with variable magnitude during training.

[0089] -Single with dropout is more regularized than full covariance (other methods overfit to the training set). However, full covariance still performs best on the test set. Thus, our sampling-based method appears to be able to navigate various regions of bias-variance tradeoff.

[0090] While NCL and bagging perform similarly in terms of test accuracy, NCL has slightly better distillation and more diversity, which demonstrates the benefits of end-to-end ensembling. Nevertheless, in addition to these advantages, our method only requires choosing the correct distribution, rather than tuning new hyperparameters (as in the case of NCL).

[0091] -Training a standalone CNN suffers from overfitting, which can be mitigated by dropout. Nevertheless, single+dropout is more unstable than its diagonal and full-covariance counterparts, which can be problematic when stability is important (as in industrial deployments).

[0092] Finally, in the table below, we can see that total covariance has an advantage over competing methods in terms of test accuracy. [Table 1]

[0093] Our technique is therefore particularly useful for combining multiple branches of neural networks and can be thought of as a way to perform end-to-end training of an ensemble of neural networks. It consists of a new neural network layer that takes as input several individual predictions from separate underlying models (or branches) and produces a single output using differentiable sampling, while providing regularization and distributing gradients across all underlying models. This technique has several advantages:

[0094] First, our method achieves higher accuracy than competing methods that either require designing diversity among the underlying models or rely on fine-tuned loss functions to balance the focus of training individual models in conjunction with the ensemble.

[0095] Second, it provides a simple means of regularizing the ensemble by introducing noise.

[0096] Third, it leads to more stable training dynamics and better individual underlying models. The approach requires no additional tuning and does not introduce new learnable parameters.

[0097] Figure 1 shows a specific sequence of steps aimed at the method 100 of the present invention. This method 100 for predicting at least one physicochemical and / or odor characteristic value for a chemical structure or composition comprises: - a step 105 of defining, on a computer interface, a digital representation of the chemical structure or composition; - a step 110 of executing, by a computing device, an end-to-end trained ensemble neural network or multi-branch neural network model on the defined digitized representation in order to predict at least one physicochemical and / or odor characteristic value of the chemical structure or composition; - providing 115, on a computer interface, at least one physicochemical and / or odor characteristic value for the chemical structure or composition, The objective of the method 100 of the present invention is to providing 120 an exemplary set of data to an end-to-end ensemble neural network or multi-branch neural network device, the set of data including at least one set of inputs, the inputs corresponding to digitized representations of chemical structures or compositions, and at least one set of outputs, the outputs corresponding to physicochemical and / or odor properties associated with the set of inputs, the end-to-end ensemble neural network or multi-branch neural network device: several neural network sub-devices, each sub-device configured to provide an independent prediction based on example data; a layer configured to output at least one value based on or representative of the distribution of said independent predictions; a step 120 of providing, in which the layer comprises a sampling device configured to output at least one random value as a function of a probability distribution representing a distribution of independent predictions, the output random value being calculated in a differentiable manner and used for backpropagation within an end-to-end ensemble neural network or multibranch neural network device; - operating 125 an end-to-end ensemble neural network or multi-branch neural network device based on the set of exemplary data; and obtaining 130 a trained end-to-end ensemble neural network or multi-branch neural network model configured to predict physicochemical and / or odor properties for an input digitized representation of a chemical structure or composition.

[0098] It should be noted that a layer configured to output at least one value based on or representative of the distribution of said independent predictions may be understood as either a layer providing a value representative of the distribution used by the sampling device or a layer providing a value obtained from the sampling device.

[0099] By "differentiable way" we mean a way of drawing samples from a distribution that allows us to calculate the gradient of the layer output with respect to the parameters of the distribution. It also means that these parameters are calculated using a differentiable function of the outputs of the neural network sub-devices. This allows us to obtain a "proper" neural network layer that can calculate the gradient of its output with respect to its input, allowing us to embed it into any larger neural network trained using backpropagation.

[0100] The defining step 105 is performed, for example, by using an input device 240 coupled to an I / O subsystem 220 as disclosed with respect to FIG.

[0101] During this defining step 105, the chemical structure or composition is defined.

[0102] Chemical structure is defined as the molecular structure and, optionally, the electronic structure of a target molecule. Molecular structure refers to the spatial arrangement of atoms in a molecule and the chemical bonds that connect the atoms and can be represented by a molecular model using a structural formula, and a complete electronic structure description includes specifying the occupancy of the molecule's molecular orbitals. Structure determination can be applied to a variety of targets, from very simple molecules (e.g., diatomic oxygen or nitrogen) to very complex ones (e.g., proteins or DNA).

[0103] Composition is defined as the sum of molecules or compounds typically called fragrance or perfume ingredients.

[0104] During this defining step 105, for example, a user can access the GUI and select an existing chemical structure or design a chemical structure by specifying constituent atoms and associated sequences. Alternatively, a user can access the GUI and select existing fragrance or aroma ingredients, each of which is associated with at least one chemical structure. Such selection or definition of a chemical structure or composition is performed using a digital representation of a tangible equivalent of said chemical structure or composition. The representation may be presented as text or may be associated with an entry in a computer database that stores several parameters for each representation.

[0105] The performing step 110 is performed by one or more hardware processors 210, such as those shown in Figure 2, configured to execute a set of instructions representing a trained end-to-end ensemble neural network or multi-branch neural network model. Particular embodiments for implementing the performing step 110 are disclosed above, particularly with reference to Figure 5.

[0106] The inputs of executing step 110 depend on the parameters that the end-to-end ensemble neural network or multibranch neural network device is manipulated to obtain the end-to-end ensemble neural network or multibranch neural network model. For example, such parameters may be: - A list of atoms that make up a chemical structure such as a molecule, - a list of atomic properties in chemical structures, - a list of ingredients in the composition, - a list of molecules in the composition and / or - a list of molecules corresponding to components in the composition.

[0107] The end-to-end ensemble neural network or multi-branch neural network model is configured to provide output for a standardized input format. This standardized input format may correspond to digital representations of the atoms, atomic properties, molecules, components, compositions, and / or chemical structures. Such digital representations may correspond to character strings. Such character strings may be concatenated to form a single input representing a larger tangible item, such as several atoms forming a molecule.

[0108] Examples of such inputs are shown with respect to Figures 7 and 12-14.

[0109] The providing step 115 is performed, for example, by using an output device 235 coupled to an I / O subsystem 220 as disclosed with respect to FIG.

[0110] In certain embodiments, this providing step 115 displays on a GUI the results of the model's predictions based on the defined chemical structures or compositions provided to the model.

[0111] The providing step 120 may be performed via a computer interface, such as an API or any other digital input means. This providing step 120 may be initiated manually or automatically. The set of exemplary data may be assembled from a larger set of exemplary data manually on a computer interface or automatically by a computing system.

[0112] Exemplary data may be, for example: - at least one digitized representation of a chemical structure or composition, and -values ​​representing odor characteristics or physicochemical properties related to chemical structure or composition.

[0113] Such odor characteristics may be, for example, the tonality of a chemical structure, the odor detection threshold of a chemical structure, the odor intensity of a chemical structure (e.g., categorizing olfactory intensity into four ranges: odorless, weak, medium, and strong classes of ingredients or compositions), and / or the top-heart-base value of a chemical structure (e.g., categorizing the persistence of an ingredient or composition during evaporation into three classes: top, heart, and base classes of ingredients or compositions, where "top" represents an ingredient or composition that can be measured by bromide or gas chromatography analysis within 15 minutes of evaporation, "heart" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "base" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "base" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "base" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "top" represents an ingredient or composition that can be measured by bromide or gas chromatography analysis within 15 minutes of evaporation, and "heart" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "base" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "base" represents an ingredient or composition that can be measured by 15 minutes to 2 hours, and "top ...

[0114] The odor characteristics are - insect repellent performance value, -sensory characteristic values, -biodegradability value, -Antibacterial value, -Odor detection threshold, -odor intensity value, -Top-Heart-Base Values, -hazard value, - biological activity on taste, -Olfactory biological activity, -biological enhancement or modulation of taste activity, - biological enhancement or modulation of olfactory activity, and / or - Olfactory odor description, which can correspond to

[0115] Physical chemistry is - boiling point value, - melting point value, - water solubility value, - Henry's constant value, -vapor pressure value, -volatile value, or - Headspace concentration value, which can correspond to.

[0116] The operating step 125 may be performed, for example, by a computer program running on a computing system. During this operating step 125, an end-to-end ensemble neural network or multi-branch neural network device is configured to train based on input data. During this operating step 125, each neural network subdevice of the end-to-end ensemble neural network or multi-branch neural network device configures coefficients of a layer of artificial neurons to provide outputs, which form a distribution of outputs. Values ​​of statistical parameters representing the distribution may be obtained and used in the activation function to be minimized.

[0117] Each neural network subdevice in the ensemble may be of the same type or of a different type.

[0118] In certain embodiments, at least one neural network subdevice comprises: -Recursive Neural Network Device, -Graph Neural Network Device, -Variational Autoencoder Neural Network Device, or -Autoencoder neural network device.

[0119] In certain embodiments, at least two of the activation functions are: -means of statistical distribution of multiple independent forecasts, - represents the variance of the statistical distribution of multiple independent predictions, and optionally extended with additional activation functions, - skew in the statistical distribution of multiple independent predictions, and / or - represents the kurtosis of the statistical distribution of multiple independent forecasts.

[0120] In certain embodiments, the at least one output value representing the distribution represents the variance of the distribution.

[0121] Such a value may correspond, for example, to the standard deviation of the outputs of the neural network sub-devices.

[0122] In certain embodiments, the end-to-end ensemble neural network or multi-branch neural network device is trained to minimize at least one value representing the variance of the distribution.

[0123] The obtaining step 130 may be performed via a computer interface, such as an API or any other digital output system. The obtained trained model may be stored in a data storage device, such as a hard drive or a database.

[0124] In certain embodiments, the neural network device obtained during the obtaining step 130 is further configured to provide at least one value representing the statistical variance of the output.

[0125] In certain embodiments, the at least one input set of exemplary data corresponds to a hash vector of at least one atomic property in the chemical structure or composition, and the method further comprises, upstream of the performing step 110, a step 135 of converting the defined digitized chemical structure or composition into a set of hash vectors of at least one atomic property representing the digitized chemical structure or composition, said set of hash vectors being used as input during the performing step.

[0126] The hash corresponds to the result of a hash function, which corresponds to any function that can be used to map data of any size to a fixed-size value. Many such functions are known to those skilled in the art, such as SHA-3, Skein, or Snefru.

[0127] Such hash values ​​may be organized into vectors that can be used by an end-to-end ensemble neural network or multi-branch neural network device.

[0128] To obtain such a hash vector representing the atomic properties of a chemical structure, - receiving, by a computing system, at least one atomic feature digital identifier and a digitized representation of a chemical structure comprising the at least one atomic feature digital identifier of said at least one atomic feature digital identifier; - determining by a computing system at least one value corresponding to at least one atomic or bond property of at least one atom or bond of the digitized representation of the chemical structure; hashing, by a computing system, the at least one determined value to form a unique string fingerprinting the at least one atomic feature digital identifier and the at least one associated atomic feature; - providing, by a computing system, at least one hash to an ensemble of neural network devices to be trained.

[0129] The receiving step is performed, for example, by any input device 240 suited to the particular use case. For example, during this receiving step, at least one digitized representation of a chemical structure, an atomic structure, or a bonding structure is input into the computer interface. Such input may be entirely logical, for example, by using an API (Application Programming Interface) or by interfacing the computing system with another computing system via a computer network. Such input may also rely on a human-machine interface, for example, a keyboard, a mouse, or a touchscreen. The mechanism used for the receiving step is not important for the scope of the present invention.

[0130] Ultimately, a digitized representation of a chemical structure essentially consists of two types of data: - atom identifiers, typically represented by at least one letter (e.g., "C" for carbon or "H" for hydrogen), corresponding to atoms that are part of the molecular structure; - at least one relationship of at least one atom identifier that defines whether and how at least one atom is connected to other atoms of the molecule; and

[0131] This digitized representation can take many forms, depending on the system. For example, the SMILES (for "Simplified Molecular Input Line Entry System") format is a line representation of a molecular structure that provides the two types of data mentioned above. Another example is a molecular graph representation of the molecule. Another representation is the SDF (for "Structural Data File") format, which defines atoms with their properties and bond tables. Another representation is a full molecular matrix, consisting of atom numbers and an adjacency matrix that defines bonds.

[0132] Typically, the primary digitized representation used for chemical reaction modeling and feature prediction is the SMILES format.

[0133] The determining step is performed by, for example, one or more hardware processors 210, as shown in Figure 2, configured to execute a set of instructions representing computer software. During this determining step, the at least one atomic property may be read from a digitized representation of the chemical structure, or may be determined through execution of a dedicated algorithm, or may be obtained from dedicated third-party software.

[0134] The hashing step is performed by, for example, one or more hardware processors 210 as shown in FIG. 2, configured to execute a set of instructions representing computer software.

[0135] The output of the hashing step is a given number of hashes, each of which is a representation of one atom identifier and at least one associated atomic property of the identified atom. Thus, a chemical structure containing several atoms is represented by a sequence of several hashes. Each hash acts as a unique fingerprint, which is particularly useful for neural network applications. This means that within a dataset, each atom can be represented by a corresponding hash key (a unique fingerprint).

[0136] Such hashing enforces sparsity in the network in an inherent way.

[0137] During this hashing step, the hash may be constructed from repeated values ​​of properties to define property values ​​of reagents, intermediates, transition states, and products.

[0138] In certain embodiments, at least one of the hashed atomic properties is: - atomic number of the corresponding atom, - atomic symbols of the corresponding atoms, - atomic mass, -explicit map number - row index in the periodic system, - column index in the periodic system, -total number of hydrogens on the atom, - the implicit number of hydrogens on the atom, - the explicit number of hydrogens on the atom, - atomic degree, -total degree of atoms, - valence state of the atom, - the implied valence of the atom, - Explicit valence of the atom -formal charge on the atom, - partial charges on atoms, -electronegativity on the atom, - number of bonds by bond type, - atomic number, number of neighboring atoms by wildcard, - Bond type + atomic number, number of neighboring atoms by wildcard, - number of neighboring atoms by wildcard, -aromaticity value, - a value indicating an aliphatic atom, - a value indicating the conjugated atoms, - a value indicating a ring atom, - a value indicating a macrocyclic atom, - a value indicating the structural constraint atom, - a value indicating an electron-withdrawing atom, - a value indicating the electron donating atom, - a value indicating the reactive site, - a value indicating the hydrogen bond donor, - a value indicating the hydrogen acceptor, - a value indicating the multivalency of the hydrogen bond donor; - a value indicating multivalency as a hydrogen bond acceptor, - number of cycles on the atom, -Atomic ring size, - the hybridization state of the atom, -values ​​indicating atomic structure, - number of electrons in the atomic orbital, - number of lone pair electrons, -radical state, -Isotopes on atoms, -atom-centrosymmetric functions, -Relative stereochemistry values ​​as clockwise chi, counterclockwise chi, - absolute stereochemical value, - absolute stereochemical value, - double bond stereochemistry value, - priority values ​​for determining stereochemistry, - a value representing the positive or negative influence of the atom on the determined training target, to indicate its contribution to the richness of the knowledge base; - a value representing the positive or negative influence of the atom on the determined training target, which value indicates the knowledge-based dilution contribution and / or ring stereochemistry.

[0139] Regarding values ​​representing the positive or negative influence of atoms on the determined training target, such values ​​may be initialized by the user or trained as part of the supplemental training method. Such values ​​may be used at the atomic or molecular level.

[0140] An alternative approach to hashing involves assigning a letter to each atom property and concatenating the letters into "words." Such letters may correspond to the letters in a SMILES string with all characters not identified as chemical atom letters removed.

[0141] In certain embodiments, at least one of the hashed binding properties is: -bond order, -bond type, -bond stereochemistry: -bond orientation in tetrahedral stereochemistry, -bond orientation of double bond stereochemistry or -bond direction of spatial orientation, - atomic numbers of atoms "from" and / or "to", - atomic symbols for atoms "from" and / or "to", -dipole moment in the bond, -Quantum-chemical properties: -electron density in the bond, -Electron configuration of the bond, -bonding orbital, -Binding energy -Gravity, -repulsion force, - bond distance, - aromatic bond, -aliphatic bond, -Ring properties of bonds: - number of rings on the bond, -bond ring size - the minimum ring size of the bond, -Maximum ring size of the bond, - rotatable bonds, - spatially constrained bonds, -hydrogen bonding properties, -ionic bonding properties, - Bond order of the reaction, including "null" bonds to identify bonds broken / formed during the reaction: - the bond order in the reagent, - the bond order of intermediate products, or - represents one of the bond orders in the transition state.

[0142] The obtaining step is performed, for example, by using any output device 235 associated with I / O subsystem 220 as shown in FIG.

[0143] In a particular embodiment, a method for obtaining a hash vector comprises: - constructing, by a computing system, a chemical structure string fingerprint by associating in a single string at least two hashes corresponding to at least two atomic characteristics; - at least one step of expanding at least one chemical structure string fingerprint by a computing system, wherein said expanded chemical structure string fingerprint is used during the providing step.

[0144] The constructing step is performed, for example, by one or more hardware processors 210, such as shown in Figure 2, configured to execute a set of instructions representing computer software. During this constructing step, at least two hashes corresponding to at least two atom identifiers and associated features are typically related by concatenation of the respective hashes. The order of concatenation can follow concatenation rules that prevent misinterpretation of the neural network.

[0145] Figure 7 illustrates a schematic representation of a particular embodiment of this method for obtaining a hash vector representing a chemical structure. a step 505 of receiving an input chemical structure; - determining 510 the atom and / or bond properties in the received chemical structure, and then annotating the vector of input properties, where the properties used here are [atomic number, degree, number of hydrogens]; a step 515 of hashing the property vector into a single hash key, where alternatively the hash key identifying the atom type may be formed by simply concatenating the values, e.g., [06, 1, 1] becomes "0611"; a step 520 of constructing a vector of hashed keys for the atoms specified by the atomic order; The step 525 of vector data expansion can be applied by changing the order of the atoms, where the order may include a vector of canonical ordering of atoms.

[0146] In a particular embodiment, the objective of the method 100 of the present invention includes a step 140 of atomic or bond relationship vector expansion upstream of the step 120 of providing input data to an end-to-end ensemble neural network or multibranch neural network device.

[0147] At least one step 140 of expansion is performed by computer software running on a computing device such as a microprocessor, for example. During this step 140 of expansion, the order of the hashes of the constituent hashes of a given molecular structure is shifted by one or more positions in the order of the constituent hashes, i.e., the last hash becomes penultimate, the penultimate becomes penultimate, the first becomes last, or vice versa, depending on the desired expansion order.

[0148] Such an extension allows for an increase in the sample size from the same chemical structure, thereby significantly improving the quality of the output of the neural network device.

[0149] In certain embodiments, the step of atomic or bond relationship vector expansion 140 includes a step of horizontal expansion 145 configured to provide several vectors representing a single digitized representation of a molecular structure or chemical reaction, each vector representing a particular representation of the molecular structure or chemical reaction in a canonical representation, and each vector being treated as a single input during the providing step.

[0150] In certain embodiments, the step 140 of atom or bond relationship vector expansion includes a step 150 of vertical expansion, creating several groups of several horizontal expansions representing unique molecular structures or chemical reactions, each group being treated as a single input during the providing step.

[0151] Such a step 150 of vertical expansion may be performed, for example, by computer software executed by a computing system. This step 150 of vertical expansion may be performed by grouping horizontal expansions with single entries, typically by concatenation of hash keys representing the atoms and / or bonding properties of the chemical structure. Such single entries may be identical or different, for example, by changing the order of concatenation.

[0152] FIG. 6 shows another representation of a particular embodiment of the object of the method 600 of the present invention.

[0153] FIG. 6 illustrates, inter alia, the training of a model, which includes a providing step 120, an operating step 125, and an obtaining step .

[0154] During the providing step 120, a digitized representation of a chemical structure 605 and known odor or physicochemical property values ​​610 are used as input.

[0155] During the operating step 125, the end-to-end ensemble neural network or multi-branch neural network device 615 is trained to output two values ​​620 and 625 that represent the distribution of the individual outputs of the neural network sub-devices that make up the end-to-end ensemble neural network or multi-branch neural network device 615, such as the mean and standard deviation.

[0156] 2 represents a schematic representation of a particular embodiment of the object of the system 200 of the present invention. This system 200 for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition comprises a means 205, which comprises: - means for defining, on a computer interface, a digitized representation of a chemical structure or composition; - means for executing an end-to-end trained ensemble neural network or multi-branch neural network model on the defined digitized representation to predict, by a computing device, at least one physicochemical value and / or odor characteristic value of the chemical structure or composition; means for providing, on a computer interface, at least one physicochemical and / or odor characteristic value of the chemical structure or composition, The means 205 further comprises: - a means for providing an end-to-end ensemble neural network or multi-branch neural network device with an exemplary set of data including at least one set of inputs, said inputs corresponding to digitized representations of chemical structures or compositions, and at least one set of outputs, said outputs corresponding to physicochemical and / or odor properties associated with the set of inputs; several neural network sub-devices, each sub-device configured to provide an independent prediction based on example data; a layer configured to output at least one value based on or representative of the distribution of said independent predictions; - means for providing, wherein the layer comprises a sampling device configured to output at least one random value as a function of a probability distribution representing a distribution of independent predictions, the output random value being calculated in a differentiable manner and used for backpropagation within an end-to-end ensemble neural network or multibranch neural network device; means for operating an end-to-end ensemble neural network or multi-branch neural network device based on a set of exemplary data; - means for obtaining a trained ensemble neural network or multi-branch neural network device configured to predict physicochemical and / or odor properties for an input digitized representation of a chemical structure or composition.

[0157] Figure 2 is a block diagram illustrating an exemplary computer system upon which one embodiment may be implemented. In the example of Figure 2, computer system 205 and instructions for implementing the disclosed techniques in hardware, software, or a combination of hardware and software are represented schematically, for example, as boxes and circles, at the same level of detail commonly used by those skilled in the art to which this disclosure pertains to communicate about computer architecture and computer system implementations.

[0158] Computer system 205 includes an input / output (IO) subsystem 220, which may include buses and / or other communication mechanisms for communicating information and / or instructions between components of computer system 205 via electronic signal paths. I / O subsystem 220 may include an I / O controller, a memory controller, and at least one I / O port. Electronic signal paths are represented schematically in the drawings, for example, as lines, single-headed arrows, or double-headed arrows.

[0159] At least one hardware processor 210 is coupled to I / O subsystem 220 for processing information and instructions. Hardware processor 210 may include, for example, a general-purpose microprocessor or microcontroller and / or an embedded system or a special-purpose microprocessor such as a graphics processing unit (GPU) or digital signal processor or ARM processor. Processor 210 may include an integrated arithmetic logic unit (ALU) or may be coupled to a separate ALU.

[0160] Computer system 205 includes one or more units of memory 225, such as main memory, coupled to I / O subsystem 220 for electronically and digitally storing data and instructions to be executed by processor 210. Memory 225 may include volatile memory, such as various forms of random access memory (RAM) or other dynamic storage devices. Memory 225 may also be used to store temporary variables or other intermediate information during execution of instructions to be executed by processor 210. Such instructions, when stored on a non-transitory computer-readable storage medium accessible to processor 210, may turn computer system 205 into a special-purpose machine customized to perform the operations specified in the instructions.

[0161] The computer system 205 further includes non-volatile memory, such as read-only memory (ROM) 230 or other static storage device, coupled to the I / O subsystem 220 for storing information and instructions for the processor 210. The ROM 230 may include various forms of programmable ROM (PROM), such as erasable programmable read-only memory (EPROM) or electrically erasable programmable read-only memory (EEPROM). A unit of persistent storage 215, which may include various forms of non-volatile random access memory (NVRAM), such as flash memory, or solid-state storage, magnetic disks, or optical disks, such as CD-ROMs or DVD-ROMs, may also be coupled to the I / O subsystem 220 for storing information and instructions. The storage 215 is an example of a non-transitory computer-readable medium that, when executed by the processor 210, may be used to store instructions and data that cause computer-implemented methods to perform the techniques herein.

[0162] The instructions in memory 225, ROM 230, or storage 215 may include one or more sets of instructions organized as modules, methods, objects, functions, routines, or calls. The instructions may be organized as one or more computer programs, operating system services, or application programs, including mobile apps. The instructions may include operating system and / or system software; one or more libraries supporting multimedia, programming, or other functionality; data protocol instructions or stacks for implementing TCP / IP, HTTP, or other communications protocols; file formatting instructions for parsing or rendering files encoded using HTML, XML, JPEG, MPEG, or PNG; user interface instructions for rendering or interpreting commands for a graphical user interface (GUI), command line interface, or text user interface; application software such as an office suite, Internet access application, design and manufacturing application, graphics application, audio application, software engineering application, educational application, game, or other application. The instructions may implement a web server, web application server, or web client. The instructions may be organized as a presentation layer, an application layer, and a data storage layer, such as a relational database system using Structured Query Language (SQL) or without SQL, an object store, a graph database, a flat file system, or other data storage.

[0163] The computer system 205 may be coupled to at least one output device 235 via the I / O subsystem 220. In one embodiment, the output device 235 is a digital computer display. Examples of displays that may be used in various embodiments include a touchscreen display or a light-emitting diode (LED) display or a liquid crystal display (LCD) or an electronic paper display. The computer system 205 may include other types of output devices 235 instead of or in addition to a display device. Examples of other output devices 235 include a printer, a ticket printer, a plotter, a projector, a sound or video card, a speaker, a buzzer or piezoelectric or other audible device, a lamp or LED or LCD indicator, a tactile device, an actuator, or a servo.

[0164] At least one input device 240 is coupled to the I / O subsystem 220 for communicating signals, data, command selections, or gestures to the processor 210. Examples of input device 240 include a touch screen, a microphone, a still and video digital camera, alphanumeric and other keys, a keypad, a keyboard, a graphics tablet, an image scanner, a joystick, a clock, a switch, a button, a dial, a slide.

[0165] Another type of input device is the control device 245, which can perform cursor control or other automatic control functions, such as navigation within a graphical interface on a display screen, instead of or in addition to input functions. The control device 245 may be a touchpad, mouse, trackball, or cursor direction keys for communicating directional information and command selections to the processor 210 and controlling cursor movement on the display 235. The input device may have at least two degrees of freedom in two axes, a first axis (e.g., x) and a second axis (e.g., y), that allow the device to specify a position within a plane. Another type of input device is a wired, wireless, or optical control device, such as a joystick, wand, console, steering wheel, pedals, gear shift mechanism, or other type of control device. The input device 240 may include a combination of multiple different input devices, such as a video camera and a depth sensor.

[0166] In another embodiment, computer system 205 may comprise an Internet of Things (IoT) device that omits one or more of output device(s) 235, input device(s) 240, and control device(s) 245. Alternatively, in such an embodiment, input device(s) 240 may comprise one or more cameras, motion detectors, thermometers, microphones, seismic detectors, other sensors or detectors, measuring devices or encoders, and output device(s) 235 may comprise a dedicated display, such as a single-line LED or LCD display, one or more indicators, display panels, meters, valves, solenoids, actuators, or servos.

[0167] Computer system 205 may implement the techniques described herein using customized hardwired logic, at least one ASIC or FPGA, firmware, and / or program instructions or logic that, when loaded and used or executed in conjunction with the computer system, cause the computer system to operate as a special-purpose machine or program. According to one embodiment, the techniques herein are performed by computer system 205 in response to processor 210 executing at least one sequence of at least one instruction contained in main memory 225. Such instructions may be read into main memory 225 from another storage medium, such as storage 215. Execution of the sequences of instructions contained in main memory 225 causes processor 210 to perform the process steps described herein. In alternative embodiments, hardwired circuitry may be used in place of or in combination with software instructions.

[0168] As used herein, the term "storage medium" refers to any non-transitory medium that stores data and / or instructions that cause a machine to operate in a specific manner. Such storage media may comprise non-volatile media and / or volatile media. Non-volatile media include, for example, optical or magnetic disks, such as storage 215. Volatile media include dynamic memory, such as memory 225. Common forms of storage media include, for example, hard disks, solid-state drives, flash drives, magnetic data storage media, any optical or physical data storage media, memory chips, etc.

[0169] Storage media are distinct from, but may be used in conjunction with, transmission media. Transmission media involves transferring information to and from storage media. For example, transmission media include coaxial cables, copper wire, and fiber optics, including the wires that comprise a bus in I / O subsystem 220. Transmission media can also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications.

[0170] Various forms of media may be involved in carrying at least one sequence of at least one instructions to processor 210 for execution. For example, the instructions may initially be carried on a magnetic disk or solid state drive of a remote computer. The remote computer may load the instructions into its dynamic memory and send the instructions over a communications link, such as a fiber optic or coaxial cable or a telephone line, using a modem. A modem or router local to computer system 205 may receive the data on the communications link and convert the data into a format readable by computer system 205. For example, a receiver such as a radio frequency antenna or infrared detector may receive the data carried in a radio or optical signal, and appropriate circuitry can provide the data to I / O subsystem 220, such as placing the data on a bus. I / O subsystem 220 carries the data to memory 225, from which processor 210 retrieves and executes the instructions. The instructions received by memory 225 may optionally be stored in storage 215 either before or after execution by processor 210.

[0171] Computer system 205 also includes a communications interface 260 coupled to bus 220. Communications interface 260 provides a two-way data communication coupling to network link 265, which is directly or indirectly connected to at least one communications network, such as network 270 or a public or private cloud on the Internet. For example, communications interface 260 may be an Ethernet networking interface, an Integrated Services Digital Network (ISDN) card, a cable modem, a satellite modem, or a modem for providing a data communication connection to a corresponding type of communications line, e.g., an Ethernet cable or any type of metal cable, or an optical fiber line or a telephone line. Network 270 broadly represents a local area network (LAN), a wide area network (WAN), a campus network, an internetwork, or any combination thereof. Communications interface 260 may comprise a LAN card for providing a data communication connection to a compatible LAN, or a cellular radiotelephone interface wired to transmit or receive cellular data in accordance with a cellular radiotelephone radio network standard, or a satellite radio interface wired to transmit or receive digital data in accordance with a satellite radio network standard. In any such implementation, communication interface 260 sends and receives electrical, electromagnetic or optical signals over signal paths that carry digital data streams representing various types of information.

[0172] Network link 265 typically provides electrical, electromagnetic, or optical data communication directly or through at least one network to other data devices using, for example, satellite, cellular, Wi-Fi, or BLUETOOTH technology. For example, network link 265 may provide a connection to host computer 250 through network 270.

[0173] Further, network link 265 may provide connectivity to other computing devices through network 270 or through an Internet network and / or computers operated by an Internet Service Provider (ISP) 275. ISP 275 provides data communication services through a worldwide packet data communication network represented as Internet 280. Server computer 255 may be coupled to Internet 280. Server 255 broadly represents any computer, data center, virtual machine, or virtual computing instance, with or without a hypervisor, or a computer running a containerized program system such as DOCKER or KUBERNETES. Server 255 may represent an electronic digital service implemented using multiple computers or instances and accessed and used by sending web service requests, uniform resource locator (URL) strings with parameters in HTTP payloads, API calls, application service calls, or other service calls. Computer system 205 and server 255 may form elements of a distributed computing system that includes other computers, processing clusters, server farms, or other organizations of computers that collaborate to perform tasks or run applications or services. The server 255 may comprise one or more sets of instructions organized as modules, methods, objects, functions, routines, or calls. The instructions may be organized as one or more computer programs, operating system services, or application programs, including mobile apps.The instructions may include operating system and / or system software; one or more libraries supporting multimedia, programming, or other functionality; data protocol instructions or stacks for implementing TCP / IP, HTTP, or other communication protocols; file formatting instructions for parsing or rendering files encoded using HTML, XML, JPEG, MPEG, or PNG; user interface instructions for rendering or interpreting commands for a graphical user interface (GUI), a command line interface, or a text user interface; and application software such as an office suite, Internet access applications, design and manufacturing applications, graphics applications, audio applications, software engineering applications, educational applications, games, or other applications. Server 255 may comprise a web application server that hosts a presentation layer, application layer, and data storage layer, such as a Structured Query Language (SQL) or non-SQL relational database system, object store, graph database, flat file system, or other data storage.

[0174] Computer system 205 can send messages and receive data and instructions, including program code, through the network(s), network link 265 and communication interface 260. In the Internet example, a server 255 might transmit a requested code for an application program through Internet 280, ISP 275, local network 270 and communication interface 260. The received code may be executed by processor 210 as it is received, and / or stored in storage 215, or other non-volatile storage for later execution.

[0175] Execution of instructions as described in this section may implement a process, in the form of an instance of a computer program consisting of program code and its current activity. Depending on the operating system (OS), a process may consist of multiple threads of execution that execute instructions simultaneously. In this context, a computer program may be a passive collection of instructions, and a process may be the actual execution of those instructions. Multiple processes may be associated with the same program; for example, opening multiple instances of the same program often means that multiple processes are running. Multitasking may be implemented to allow multiple processes to share the processor 210. Although each processor 210 or processor core executes a single task at a time, the computer system 205 may be programmed to implement multitasking to allow each processor to switch between executing tasks without having to wait for each task to finish. In one embodiment, the switch may be performed when a task performs an input / output operation, when the task indicates that it is available to switch, or upon a hardware interrupt. Time sharing may be implemented to enable fast response to interactive user applications by rapidly executing context switches to provide the appearance of concurrent execution of multiple processes simultaneously. In one embodiment, for security and reliability, the operating system may prevent direct communication between independent processes and provide strictly mediated and controlled inter-process communication facilities.

[0176] A particular application of the intended system 200 of the present invention is disclosed with respect to FIG.

[0177] 4 shows a schematic sequence of steps aimed at the method 400 of the present invention. This method 400 for efficiently assembling chemical structures or compositions includes: a step 405 of carrying out a method as shown in FIG. 1; - a step 410 of assembling a chemical structure or composition associated with the output obtained during the obtaining step 115.

[0178] This assembling step 410 is configured to embody the composition. Such assembling step 410 may be performed in a variety of ways, such as in a laboratory or a chemical plant.

[0179] 8 shows a schematic diagram of a specific implementation example for the purpose of the inventive method 800. This method 800 for training an ensemble neural network or multi-branch neural network device is similar to the training performed by the end-to-end ensemble neural network or multi-branch neural network device used for the purpose of the inventive method 100. This method 800 includes: Step 805 of inputting the atomic property and / or bond property hash keys expanded from an exemplary set of chemical structure digital identifiers associated with known outputs representing at least one physicochemical property and / or odor property associated with the atomic property and / or bond property hash keys; a step 810 of embedding or tokenizing the input; a step 815 of operating an ensemble of recursive neural network devices on the input; a step 820 of operating an attention layer on the output of the step 815 of operating an ensemble of recurrent neural network devices; a step 825 of operating a flattening layer at the output of the step 820 of operating an attention layer; a step 830 of running a multi-layer perceptron ("MLP") layer on the output of the flattening layer; and - outputting 835 the values ​​of the target odor characteristics and / or physicochemical characteristics.

[0180] The execution parameters for this particular embodiment may be:

[0181] [Table 2]

[0182] In this table, the number N represents the number of points (e.g., input batch size). In this architecture, chemical structures are represented as augmented 2D chips, which are then transformed using embedding and recurrent neural network layers. The attention layer performs feature selection. The MLP portion of the network is a fully connected neural network with activations.

[0183] 9-11 show the performance of the architecture shown in FIG. 8 for three different targets, namely: Figure 9 shows the performance of the Odor Detection Threshold ("ODT"); - Figure 10 shows the volatility performance, - Figure 11 shows the performance of LogVP (logarithmic form of vapor pressure).

[0184] As will be appreciated, the present invention is also directed to a computer-implemented ensemble neural network or multi-branch neural network device, which may be obtained by any variation of the object of the computer-implemented method 300 of the present invention.

[0185] As will be appreciated, the present invention is also directed to a computer program product comprising instructions for carrying out the steps of the method 300 of the present invention when executed on a computer.

[0186] As will be appreciated, the present invention is also directed to a computer-readable medium storing instructions for performing the steps of the method 300 of the present invention when executed on a computer.

[0187] FIG. 12 illustrates a schematic training architecture 1200 for selecting chemical structures from a set of chemical structures that provide a particular characteristic, such as an insect repellency value above a determined threshold.

[0188] Such an architecture 1200 may include: as input 1205, a hash vector of at least one atomic property in a chemical structure, and at least one set of outputs, said outputs corresponding to odor properties associated with the set of inputs; an ensemble neural network or multibranch neural network 1210 including a set of recurrent neural networks for generating embeddings 1215; Two alternative, compatible routes can then be implemented, In the first route, a multivariate statistical algorithm 1220 can be used, complemented by a numerical domain eccentricity assessment algorithm 1225, - In the second route, a temporal distribution of embedding 1230 is performed to obtain alternative inputs 1235 complemented by the use of an ensemble 1240 neural network using Tanimoto neural networks as disclosed in the present invention.

[0189] FIG. 13 shows a schematic of a training architecture 1300 for classifying chemical structures that provide a particular characteristic, such as a biodegradability value, from a set of chemical structures.

[0190] Such an architecture 1300 may include: as input 1305, a hash vector of at least one atomic property in a chemical structure, and at least one set of outputs, said outputs corresponding to odor properties associated with the set of inputs; an ensemble neural network or multibranch neural network 1310 including a set of recurrent neural networks for generating embeddings 1315; Two alternative, compatible routes can then be implemented, namely: In the first route, a multivariate statistical algorithm 1320 can be used, complemented by a numerical domain eccentricity assessment algorithm 1325, - In the second route, a temporal distribution of the embedding 1330 is carried out, complemented by the use of an ensemble 1335 neural network using a classification neural network as disclosed in the present invention.

[0191] FIG. 14 illustrates a schematic of a training architecture 1400 for predicting the value of a particular feature, such as odor detection threshold, for a set of chemical structures.

[0192] Such an architecture 1400 may include: as input 1405, a hash vector of at least one atomic property in a chemical structure, and at least one set of outputs, said outputs corresponding to odor properties associated with the set of inputs; an ensemble neural network or multibranch neural network 1410 including a set of recurrent neural networks for generating embeddings 1415; Two alternative, compatible routes can then be implemented, namely: In the first route, a multivariate statistical algorithm 1420 can be used, complemented by a numerical domain eccentricity assessment algorithm 1425, In the second route, the ensemble 1430 neural network uses a single or multitask recurrent neural network, such as those disclosed in the present invention.

[0193] As can be appreciated, the present invention may be used to function as a filtration technique, using any predicted physicochemical and / or odor properties to label molecule or component digital identifiers in a database, said molecules or components being selected as valuable points of exploration by fragrance and perfume manufacturers.

[0194] As can be seen, the present invention can take a pair of molecules as input and perform regression or classification by predicting the proximity of the molecules in the pair or by using the observed differences in the pair.

[0195] As can be appreciated, the present invention may be used as a classifier used in relation to physicochemical and / or odor characteristic values ​​of chemical structure or composition.

[0196] FIG. 15 shows a particular architecture that highlights the performance of such a classifier.

[0197] In machine learning, the chi-squared test is often used to evaluate the performance of a classification model. For example, suppose one has a binary classification problem where one wants to predict whether a patient has a disease or not. By comparing the predicted class distribution with the expected class distribution, one can use the chi-squared test to determine whether our model is performing better than chance.

[0198] In ensemble learning, which combines multiple models to improve overall performance, the performance of the ensemble can be evaluated using a chi-squared test. Ensemble learning is a common technique in machine learning, where multiple models are trained and combined to improve overall performance. Using multiple models can reduce the risk of overfitting and improve the robustness of the model.

[0199] In an ensemble of classification models, each model makes an independent prediction for the input data, and the final prediction is made by combining the predictions of all models. The performance of the ensemble can be evaluated by comparing the predicted class distribution of the ensemble with the expected class distribution using a chi-squared test. If the ensemble performs better than any individual model, it can be concluded that the ensemble is effective.

[0200] Overall, the chi-squared test is a powerful tool for assessing the performance of machine learning models and ensembles. By using the chi-squared test, we can make informed decisions about which models to use and how to improve them.

[0201] Forced-choice modeling is an example of a control classification task, where the goal is to identify the correct example from a set of options. This type of task is commonly encountered in many real-world scenarios, such as identifying the correct answer in a multiple-choice test or recognizing a specific object from a set of similar objects. In science, results are often evaluated in a relative setting by comparing two or more candidates with each other. Therefore, it is hypothesized that a control neural network trained to select the more promising entry from a set of options may provide a valuable model.

[0202] The data can be expanded by creating pairs or triplets or other alternatives in the contrast neural network. Indeed, for regression tasks, it is possible to consider only the lower or upper half of the matrix to reduce the number of N to N. 2 -N pairs, or (N 2The data can be expanded to (N) / 2 pairs. Alternatively, for problems with small hit rates, hits can be combined with one or more non-hits in a forced-choice classification. In the latter experiment, the model is trained to detect hit molecules from the proposed alternatives. Another advantage of these control networks includes the creation of a balanced set. In practice, lower values ​​can be expected to be evenly distributed across the number of alternatives.

[0203] To address this issue, ensemble neural networks with independent voting may be used, with each model in the ensemble making an independent prediction on the input data. A final prediction is then made by combining the predictions of all the models. Using an ensemble of models can reduce the risk of overfitting and improve the robustness of the model.

[0204] After making a prediction, we can use a chi-squared test to measure the statistical significance of the decision. In this case, we can compare the predicted class distribution to the expected class distribution, which is a uniform distribution across the three examples. If the chi-squared test shows that the predicted class distribution is significantly different from the expected class distribution, we can conclude that the ensemble is performing well and can correctly identify the correct examples from the input X.

[0205] Overall, using ensemble neural networks with individual voting and chi-squared tests is a powerful technique for contrastive classification tasks and can help improve the accuracy and robustness of models. This technique allows for informed decisions about which examples are correct and which are incorrect, improving the ability to recognize and classify objects in real-world scenarios.

[0206] To perform such calculations, we can use SMILES strings containing explicit-implicit hydrogen atoms. For example, consider the molecule toluene. The explicit SMILES of toluene, written as "[CH3][c]1[cH][cH][cH][cH][cH]1", can be tokenized by grouping the atoms defined by the characters in square brackets, from to [until]. All other characters, such as ring index 1, can be tokenized as individual characters. Thus, the tokenized SMILES of toluene is "[CH3][c]1[cH][cH][cH][cH][cH]1". Similarly, the explicit SMILES of glutamic acid, "[NH2][CH]([CH2][CH2][C](=[O])[OH])[C](=[O])[OH]", can be tokenized by tokenizing bonds individually, e.g., double bonds, i.e., =, and branches, i.e., (and). The tokenized SMILES for glutamic acid is "[NH2][CH]([CH2][CH2][C](=[O])[OH])[C](=[O])[OH]".

[0207] Forced-choice classification is performed using a network layout where the same embedding, GRU, attention, and latent layers are applied to all input entries, followed by a learnable contrast layer to create the differences between all pairs.

[0208] Figure 15 shows the structural layout of a control classifier asked to select molecules with the lowest molecular weight. The input is a tokenized vector with integer-based tokens, followed by a Keras embedding layer, a Keras GRU layer, and an attention layer. The equality sign between these layers indicates that the same layer is applied to both entries. A trainable control layer creates the difference between the output of attention 1 and attention 2. A multilayer perceptron with dropout is used for the classification task. The model is repeated N times to create an ensemble neural network. The values ​​(None, x) and (None, x, y) indicate the output shape of the layers.

[0209] Such a control classifier can be trained using a dataset obtained from NIST. The data can be divided into a training set of 8,518 molecules, a validation set of 819 molecules, and a test set of 772 molecules. The training and validation datasets can be used to track performance during training at each epoch. To train the network, 45,458 pairs of molecules can be constructed with a maximum difference of 14.02 g / mol, corresponding to the mass of one CH2 group in the molecule. The classifier can be trained to detect molecules with the highest molecular weight. Note that any numerical target can be trained, including linear retention index, volatility, or odor detection threshold. The validation set can include several pairs with a maximum difference between molecules of 14.02 g / mol. Each epoch can be trained using 46 iterations with a batch size of 1,000 pairs per iteration. The model is trained using an average binary crossover method calculated across all models in the ensemble.

[0210] Upon completion, the performance can be tested on a test set consisting of several pairs, with the maximum difference between the molecules being 14.02 g / mol. The performance results are displayed in Table 1 below. Given that the model is performing a relative classification task and is required to identify the location of the lowest molecular weight, only the accuracy results are reported. In Table 1, the p-values ​​are calculated using a chi-square test on the votes generated by the ensemble. If the p-value of the vote proportion falls below 0.05, the results are considered conclusive. From Table 1, it can be clearly seen that the results for conclusive entries are significantly better than those for inconclusive entries. [Table 3]

[0211] In summary, using ensemble models for classification significantly improves the ability to convey the confidentiality of the results. Predictions are formed by combining multiple votes for a class, along with an indication of the confidence in the prediction (Table 2). [Table 4]

[0212] In conclusion, the presented methodology can be used for both relative and absolute classification tasks. The example above shows a relative task: learning to select molecules with higher molecular weights. In such classification, the regression task is transformed into a control classification task. In absolute classifiers, the ensemble is asked to predict classes defined in data, such as those performed in the MNIST dataset to detect digits in images.

Claims

1. A method (100) for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition, comprising: defining (105) on a computer interface a digital representation of a chemical structure or composition; Executing an end-to-end trained ensemble neural network or multi-branch neural network model on the defined digitized representation to predict (110) at least one physicochemical value and / or odor characteristic value of the chemical structure or composition by a computing device; and providing (115) on a computer interface said at least one physicochemical and / or odor characteristic value for the chemical structure or composition, The following steps: providing (120) a set of example data to an end-to-end ensemble neural network or multi-branch neural network device, the set of example data including at least one set of inputs corresponding to digitized representations of chemical structures or compositions and at least one set of outputs corresponding to physicochemical and / or odor characteristics associated with said set of inputs, the end-to-end ensemble neural network or multi-branch neural network device comprising: a number of neural network sub-devices, each sub-device configured to provide an independent prediction based on said example data; a layer configured to output at least one value based on or representative of a distribution of the independent predictions; providing (120), wherein the layer comprises a sampling device configured to output at least one random value as a function of a probability distribution representing the distribution of independent predictions, the output random value being calculated in a differentiable manner and used for backpropagation within the end-to-end ensemble neural network or multibranch neural network device; operating (125) the end-to-end ensemble neural network or multi-branch neural network device based on the set of example data; and obtaining (130) the trained end-to-end ensemble neural network or multi-branch neural network model configured to predict physicochemical and / or odor properties for an input digitized representation of a chemical structure or composition.

2. 2. The method (100) of claim 1, wherein at least one set of input exemplary data corresponds to a hash vector of at least one atomic property in a chemical structure or composition, and further comprising, upstream of said performing step (110), a step (135) of converting said defined digitized chemical structure or composition into a set of hash vectors of at least one atomic property representing said digitized chemical structure or composition, said set of hash vectors being used as input during said performing step.

3. At least one hash vector of the atomic property is the atomic number of the corresponding atom, the atomic symbols of the corresponding atoms, atomic mass, Explicit Map Number row index in a periodic system, Column index in a periodic system, the total number of hydrogens on the atom, the implicit number of hydrogens on the atom, the explicit number of hydrogens on the atom, atomic degree, The total degree of atoms, the valence state of the atom, the implied valence of the atom, Explicit valence of the atom Formal charge on an atom, partial charges on atoms, electronegativity on an atom, Number of joins by join type, Atomic number, number of neighboring atoms by wildcard, Bond type + atomic number, number of neighboring atoms by wildcard Number of neighboring atoms by wildcard, Aromaticity value, A value indicating an aliphatic atom, A value indicating the conjugated atoms, a value indicating the ring atom, a value indicating the macrocyclic atom, A value indicating the structural constraint atom, a value indicating the electron-withdrawing atom, a value indicating the electron donating atom, A value indicating the reactive site, a value indicating the hydrogen bond donor; A value indicating the hydrogen acceptor, a value indicating multivalency as a hydrogen bond donor; a value indicating multivalency as a hydrogen bond acceptor; the number of cycles on the atom, ring size on atoms, the hybridization state of the atom, A value indicating the atomic structure, the number of electrons in the atomic orbital, number of lone pair electrons, Radical state, Isotopes on atoms, Atomic centrosymmetric functions, Relative stereochemistry values ​​as clockwise chi and counterclockwise chi, Absolute stereochemical values, Absolute stereochemical values, Double bond stereochemistry values, Priority values ​​for determining stereochemistry, a value representing the positive or negative influence of the atom on the determined training target to indicate its contribution to the richness of the knowledge base; a value representing the positive or negative influence of the atom on the determined training target to indicate the contribution of the dilution in the knowledge base; and / or 3. The method of claim 2, wherein the value of the ring stereochemistry represents one of:

4. At least one hash vector of the binding characteristic is bond order, Join type, Bond Stereochemistry: Bond orientation in tetrahedral stereochemistry, Double bond stereochemistry, or spatial orientation of the bond direction, Atomic numbers of atoms "from" and / or "to", Atomic symbols for atoms "from" and / or "to", Dipole moment of the bond, Quantum-chemical properties: electron density in the bond, The electron configuration of the bond, bond orbital, Binding energy gravity, Repulsion force, bond distance, aromatic bond, aliphatic bond, Ring properties of bonds: The number of rings on the bond, Bond ring size The minimum ring size of the bond, Maximum ring size of the bond, rotatable bonds, spatially constrained bonds, hydrogen bonding properties, ionic bonding properties, Bond sequence of the reaction, including "null" bonds to identify bonds broken / formed during the reaction: Bond order in the reagent, the bond order of the intermediates, or 4. The method (100) of claim 2 or 3, wherein the bond order in the transition state represents one of:

5. The method (100) of any one of claims 1 to 4, wherein at least one output value representative of the distribution represents the variance of the distribution.

6. 6. The method (100) of claim 5, wherein the end-to-end ensemble neural network or multi-branch neural network device is trained to minimize at least one value representing the variance of the distribution.

7. At least one odor characteristic is Insect repellent performance value, Sensory characteristic values, Biodegradability value, Antibacterial value, odor detection threshold, Odor intensity value, Top-Heart-Base value, Hazard value, Biological activity on taste Olfactory biological activity, biological enhancement or modulation of taste activity; biological enhancement or modulation of olfactory activity, and / or The method (100) according to any one of claims 1 to 6, wherein the olfactory odor description is expressed.

8. At least one physical property is boiling point value, Melting point value, Water solubility value, Henry's constant value, Vapor pressure value, a volatile value, or The method (100) of any one of claims 1 to 7, wherein the headspace concentration value is expressed as a

9. At least one neural network device Recurrent Neural Network Device, Graph Neural Network Device, Variational Autoencoder Neural Network Device, or The method (100) of any one of claims 1 to 8, wherein the method (100) is an autoencoder neural network device.

10. 10. The method (100) of any one of claims 1 to 9, comprising a step (140) of atomic or bond relationship vector expansion upstream of the step (120) of providing input data to an end-to-end ensemble neural network or multibranch neural network device.

11. 11. The method of claim 10, wherein the step of atom or bond relationship vector expansion (140) comprises a horizontal expansion step (145) configured to provide several vectors representing a single digitized representation of a molecular structure or composition, each vector representing a particular representation of a canonical molecular structure or composition, and each vector being treated as a single input during the providing step.

12. 12. The method of claim 11, wherein the step of atomic or bond relationship vector expansion (140) includes a step of vertical expansion (150) to create several groups of horizontal expansions representing unique molecular structures or compositions, each group being treated as a single input during the providing step.

13. A method (400) for efficiently assembling a chemical structure or composition, comprising: Implementing (405) a method according to any one of claims 1 to 12; and assembling (410) a chemical structure or composition associated with the output obtained during the obtaining step (115).

14. A system (200) for predicting at least one physicochemical and / or odor characteristic value of a chemical structure or composition, comprising: means for defining, on a computer interface, a digitized representation of a chemical structure or composition; means for executing, by a computing device, an end-to-end trained ensemble neural network or multi-branch neural network model on the defined digitized representation to predict at least one physicochemical and / or odor characteristic value of the chemical structure or composition; and means for providing, on a computer interface, said at least one physicochemical and / or odor characteristic value of the chemical structure or composition, The means (205) 1. A means for providing an exemplary set of data to an end-to-end ensemble neural network or multi-branch neural network device, the exemplary set of data including at least one set of inputs, the inputs corresponding to digitized representations of chemical structures or compositions, and at least one set of outputs, the outputs corresponding to physicochemical and / or odor properties associated with the set of inputs, the end-to-end ensemble neural network or multi-branch neural network device comprising: a number of neural network sub-devices, each sub-device configured to provide an independent prediction based on said example data; a layer configured to output at least one value based on or representative of the distribution of the independent predictions; means for providing, wherein the layer comprises a sampling device configured to output at least one random value as a function of a probability distribution representing the distribution of independent predictions, the output random value being calculated in a differentiable manner and used for backpropagation within the end-to-end ensemble neural network or multibranch neural network device; means for operating the end-to-end ensemble neural network or multi-branch neural network device based on the set of exemplary data; and means for obtaining the trained end-to-end ensemble neural network or multi-branch neural network model configured to predict physicochemical and / or odor properties for an input digitized representation of a chemical structure or composition.

Images