Universal processing of molecular-related tasks

Abstract

According to the implementation of the invention, a general processing scheme for molecular related tasks is provided. According to the scheme, input including text and indicating a target molecular system is acquired, and the text indicates a target task related to the target molecular system; based on a target molecular graph corresponding to the target molecular system, generating molecular features of the target molecular system, the molecular graph comprising a plurality of nodes connected by edges, each node representing an atom set in the corresponding molecular system; and determining a result of the target task based on the text features of the text and the molecular features of the target molecular system. In this way, the task related to the molecular system can be obtained through the natural language, the result of the task is determined based on the molecular features and the text features, and the generalization ability of processing the task related to the molecular system is improved.

Description

Background Art

[0001] Chemical science and applications involve a vast array of molecular tasks, such as designing new molecules, predicting their properties and reactions, discovering potential drugs, and understanding the interactions of molecules with biological systems. Recent advances in artificial intelligence (AI), particularly deep learning, have enabled significant breakthroughs in these tasks, producing models that achieve state-of-the-art performance and, in some cases, surpass human experts. However, most of these models are expert chemistry AI models, meaning they are trained on specialized datasets and are tailored for specific domains and tasks. Summary of the Invention

[0002] According to the implementation of the present disclosure, a general processing scheme for molecule-related tasks is proposed. In this scheme, an input including text and indicating a target molecular system is obtained, and the text indicates a target task related to the target molecular system. Based on a target molecular graph corresponding to the target molecular system, molecular features of the target molecular system are generated. The molecular graph includes a plurality of nodes connected by edges, and each node represents a set of atoms in the corresponding molecular system. Based on the text features of the text and the molecular features of the target molecular system, the result of the target task is determined. According to the scheme of the present disclosure, tasks related to molecules expressed in natural language can be identified, and the results of the tasks can be determined accordingly. In this way, the text modality is associated with the molecular modality, thereby being suitable for processing various types of molecule-related capabilities, that is, this improves the generalization ability of processing molecule-related tasks.

[0003] This section is provided to introduce a selection of objects in a simplified form, which will be further described in the detailed description below. This section is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0004] Figure 1 A block diagram illustrating an example environment in which various implementations of the present disclosure can be implemented;

[0005] Figure 2 A schematic block diagram illustrating the architecture of a general processing model for molecule-related tasks according to some implementations of the present disclosure;

[0006] Figure 3 A schematic diagram illustrating a molecular model pre-training process according to some implementations of the present disclosure is shown;

[0007] Figure 4 shows an envelope curve of molecular model pre-training according to some implementations of the present disclosure;

[0008] Figure 5 A schematic diagram illustrating a training process of a general processing model according to some implementations of the present disclosure is shown;

[0009] Figure 6 A flowchart illustrating a process for handling tasks related to molecular systems according to some implementations of the present disclosure;

[0010] Figure 7 A schematic block diagram of an electronic device capable of implementing various implementations of the present disclosure is shown. DETAILED DESCRIPTION

[0011] The present disclosure will now be discussed with reference to several example implementations. It should be understood that these implementations are discussed only to enable those skilled in the art to better understand and implement the present disclosure, rather than to imply any limitation on the scope of the present disclosure.

[0012] As used herein, the term "including" and its variations are to be interpreted as open-ended terms meaning "including but not limited to." The term "based on" is to be interpreted as "based, at least in part, on." The terms "an implementation" and "an implementation" are to be interpreted as "at least one implementation." The term "another implementation" is to be interpreted as "at least one other implementation." The terms "first," "second," and so on may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0013] It should be noted that the titles of any section / subsection provided herein are not limiting. Various implementations are described throughout this document, and any type of implementation may be included under any section / subsection. Furthermore, the implementation described in any section / subsection may be combined in any manner with any other implementation described in the same section / subsection and / or in different sections / subsections.

[0014] Herein, unless explicitly stated otherwise, executing a step “in response to A” does not mean executing the step immediately after “A” but may include one or more intermediate steps.

[0015] As used herein, the term "model" can learn the association between the corresponding input and output from the training data, so that after the training is completed, the corresponding output can be generated for a given input. The generation of the model can be based on machine learning technology. Deep learning (DL) is a machine learning algorithm that processes input and provides corresponding output by using multiple layers of processing units. A neural network model is an example of a model based on deep learning. In this article, "model" may also be referred to as "machine learning model", "learning model", "machine learning network" or "learning network", and these terms are used interchangeably in this article.

[0016] Generally speaking, machine learning can be roughly divided into three stages, namely the training stage, the testing stage, and the use stage (also called the inference stage). In the training stage, a given model can be trained using a large amount of training data, and iterated continuously until the model can obtain consistent inferences that meet the expected goals from the training data. Through training, the model can be considered to be able to learn the association between input and output (also called input-to-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing stage, the test input is applied to the trained model to test whether the model can provide the correct output, thereby determining the performance of the model. In the inference stage, the model can be used to process the actual input based on the parameter values ​​obtained through training to determine the corresponding output.

[0017] Sample Environment

[0018] Figure 1 1 shows a schematic diagram of an example environment 100 in which implementations of the present disclosure can be implemented. Figure 1 As shown, the environment 100 includes an electronic device 110. It is desirable to use such an electronic device 110 to output the results of tasks related to molecular systems (also referred to as "molecule-related tasks"). To this end, in some implementations, a general processing model 140 for molecule-related tasks may be deployed in the electronic device 110 to output the results of the tasks related to molecular systems.

[0019] like Figure 1 As shown, electronic device 110 may determine result 130 based on input 120. In an implementation of the present disclosure, input 120 may include text and indicate a molecular system. The text is used to indicate a task related to the molecular system, which may include a single molecule or multiple molecules.

[0020] exist Figure 1 In the present invention, electronic device 110 can be any system with computing capabilities, such as various computing devices / systems, terminal devices, servers, etc. Terminal devices can be any type of mobile, fixed, or portable terminal, including mobile phones, desktop computers, laptop computers, notebook computers, netbook computers, tablet computers, media computers, multimedia tablets, or any combination thereof, including accessories and peripherals of these devices or any combination thereof. Servers include but are not limited to mainframes, edge computing nodes, computing devices in cloud environments, etc.

[0021] It should be understood that Figure 1 The components and arrangements in the depicted environment are merely examples, and a computing system suitable for implementing the implementations described herein may include one or more different components, other components, and / or different arrangements.

[0022] It should be understood that the structure and function of the various elements in the environment 100 are described for illustrative purposes only and do not imply any limitation on the scope of the present disclosure.

[0023] As briefly mentioned above, models used to handle molecular-related tasks are typically expert models. Taking the field of chemistry as an example, existing chemical AI expert models have two major limitations. First, data scarcity is a common challenge in chemistry, as the experimental cost and complexity of obtaining high-quality molecular data are often unaffordable. As a result, most chemical AI expert models are unable to leverage the rich knowledge embedded in other fields and tasks and may suffer from overfitting and poor generalization. Second, the diversity and complexity of molecular tasks require flexible and scalable AI systems that can adapt to new scenarios and solve new problems. However, most chemical AI expert models are rigid, require extensive retraining or fine-tuning for each new task or domain, and may not be able to transfer their learned skills across tasks and domains.

[0024] On the other hand, general-purpose techniques and scaling paradigms have demonstrated remarkable capabilities in AI, particularly in language and multimodality, compared to AI specialists. For example, large language models (LLMs) have demonstrated the ability to perform a variety of natural language tasks through few-shot or zero-shot learning, using natural language as a unified interface and source of knowledge. Similarly, large multimodal models have demonstrated the ability to integrate and interpret multiple modalities, such as text, vision, and robotics, and perform a variety of embodied reasoning tasks using natural language instructions. These models can be considered general-purpose processing models because they can handle a wide range of tasks using a single model and generalize well to unseen tasks and data. Furthermore, these models exhibit strong scaling laws, meaning that their performance improves with increasing model size and data, often significantly outperforming specialist models. Scaling laws also indicate that these models can potentially capture more complex and subtle patterns and relationships in data and learn more general and robust representations.

[0025] In view of this, according to the implementation of the present disclosure, a general processing scheme for molecular-related tasks is proposed. In this scheme, an input including text and indicating a target molecular system is obtained, and the text indicates a target task related to the target molecular system. Based on a target molecular graph corresponding to the target molecular system, molecular features of the target molecular system are generated, and the molecular graph includes a plurality of nodes connected by edges, and each node represents a set of atoms in the corresponding molecular system. Based on the text features of the text and the molecular features of the target molecular system, the result of the target task is determined. According to an embodiment of the present disclosure, multimodal data is included in the input of the task, and tasks related to the molecular system are processed based on natural language. In this way, the generalization capability of task processing related to molecular systems can be enhanced.

[0026] Example architecture for a common processing model

[0027] Some example implementations of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0028] While molecular models can encode and interpret molecular information, they cannot leverage the rich world and scientific knowledge embedded in natural language, nor can they generalize to new tasks. Therefore, language models are needed to complement molecular models, and language models can provide a natural language interface for general processing models. Language models are generative models pre-trained on large corpora that perform various natural language tasks through zero-shot or few-shot learning. Natural language instructions can be used as unified task specifications and guidance for generating results. In order to effectively combine molecular models and language models, a comprehensive framework is needed to align the representations and goals of the two models, thereby enabling seamless communication and collaboration on molecular tasks.

[0029] In the implementation of the present disclosure, a new general processing model is proposed. Figure 2 Schematic block diagram of the architecture of the general processing model 200 according to some implementations of the present disclosure. The various components in the general processing model 200 can be implemented by hardware, software, firmware, or any combination thereof. The general processing model 200 can be implemented in Figure 1 In the electronic device 110. Note that Figure 2 FIG. 2 shows only a partial structure of the entire general process model 200 , and the model may also have more other components. The general process model 200 can be regarded as an example implementation of the general process model 140 .

[0030] The general processing model 200 generally includes a molecular model 210, a feature adaptation model 220, a language model 230, and a numerical decoding model 240. These models can be implemented based on machine learning techniques, such as a neural network architecture.

[0031] The general processing model 200 can generate a result 260 based on an input 250. The input 250 includes text 252 and indicates a target molecule system 254. The target molecule system 254 can be a single molecule or include multiple molecules. For example, in the case of predicting whether an intermolecular reaction can occur, the target molecule system 254 can include two or more molecules. The text 252 indicates a target task related to the target molecule system 254. The target task can include various suitable types of molecule-related tasks.

[0032] In some implementations, the target task can be a predictive task, such as a property (e.g., solubility, toxicity, etc.) of the predicted target molecule system 254, a reaction of the predicted target molecule system 254, etc. In this article, description will be mainly made with the predictive task as an example, but it should be understood that this is only exemplary. The scheme of the present disclosure is also applicable to generative tasks, such as generating molecules with target properties. In this case, input 250 may not indicate a molecular system, but may include text describing the desired target property.

[0033] After having obtained input 250, universal processing model 200 generates the molecular features of target molecule system based on the target molecule graph corresponding to target molecule system.Because target molecule system 254 can be a single molecule or include multiple molecules, accordingly, target molecule graph can include single graph or multiple isolated subgraphs, each subgraph corresponding to a molecule.Molecular graph includes multiple nodes connected by edges, and each node represents the atomic set in the corresponding molecular system.Atomic set can be a single atom or multiple atoms (for example, atomic clusters).In particular, in some implementations, each node represents an atom.Hereinafter, will be mainly described by taking each node representing an atom and the molecular system as a single molecule as an example.

[0034] The molecular model 210 generates features (also called intermediate features) of the target molecular system from the target molecular graph and can be represented as follows:

[0035]

[0036] in represents the molecular model 210, g represents the target molecular graph, The intermediate features capture the fundamental chemical and physical properties of the molecular system, enabling the general processing model 200 to perform various tasks in the fields of chemistry and materials science with high fidelity and adaptability.

[0037] After generating the intermediate features, the feature adaptation model 220 in the general processing model 200 can be used to convert the intermediate features into a text feature space to obtain molecular features. The feature adaptation model 220 is configured (e.g., trained) to convert the molecular features into a text feature space. Without intending any limitation, the feature adaptation model 220 can be in the form of a multi-layer perceptron (MLP). In this way, the molecular features can be in the same dimension as the text features and aligned with the text features in the shared feature space.

[0038] The feature adaptation model 220 converts the intermediate feature into the text feature space, and the obtained molecular feature can be expressed as follows:

[0039]

[0040] in represents a feature adaptation model 220, represents the intermediate features, Represents molecular characteristics.

[0041] After generating the molecular features, the general processing model 200 determines the result of the target task based on the text features of the text and the molecular features of the target molecular system. In some implementations, the text features and the molecular features can be provided as input to the language model 230 to obtain the output of the language model.

[0042] After obtaining the molecular features aligned with the text features, the language model 230 can process the multimodal features (i.e., including text features and molecular features). Any suitable language model can be used, such as any suitable large language model. By way of example and without any limitation, the Llama 2 model can be used as the language model 230. The Llama 2 model covers different topics and fields including scientific literature and is optimized for conversational use cases. Specifically, the Llama 2-70B model can be selected, which has 70 billion parameters and supports few-shot learning or zero-shot learning. When combined with other modalities (e.g., molecular systems), larger language models (i.e., models with more parameters) are more robust to language capability degradation, so the Llama 2-70B model can better handle multimodal input.

[0043] The language model 230 processing multimodal features can be expressed as follows:

[0044]

[0045] in Represents molecular characteristics, Represents text features, represents the output of the language model 230, and i represents the input word.

[0046] These multimodal features enable language model 230 to generate outputs that encode linguistic context, meaning, and relevant molecular information. By utilizing a shared feature space and MLP-style feature adaptation model 220, universal processing model 200 integrates the capabilities of molecular model 210 and language model 230, effectively leveraging their strengths and improving the performance of universal processing model 200 across a variety of molecular tasks.

[0047] After obtaining the output of the language model 230, the numerical representation can be determined from the output of the language model 230 based on the position of the predetermined representation corresponding to the numerical value in the output of the language model 230. In some implementations, the numerical representation can be determined by a certain word unit (e.g., word unit <num>) to represent numerical values, based on word units <num>At a position in the output of the language model 230, a numerical representation may be determined from the output of the language model 230. The numerical representation may be the output hidden state of the last decoding layer in the language model 230.

[0048] Since the language model 230 is designed to process language patterns and structures, it is not good at numerical reasoning and calculation. In some implementations, after determining the numerical representation, the numerical decoding model 240 can be used to obtain a numerical result from the numerical representation and use the numerical result as at least part of the result. For example, the numerical decoding model 240 can be in the form of an MLP. Figure 2 As shown, input 250 is "What is the solubility of the molecule <mol> < / mol> ? (What is a molecule <mol> < / mol> The solubility is <num> -1.52< / num> (Solubility is <num> -1.52< / num> )", so the numerical result -1.52 is the numerical result obtained by the numerical decoding model 240 from the numerical representation, and the numerical result can be used as at least a part of the final output result. In this way, by extracting valuable numerical predictions, the numerical decoding model 240 further enhances the performance of the general processing model 200 in various molecular tasks.

[0049] Using the numerical decoding model 240, obtaining a numerical result from a numerical representation can be expressed as follows:

[0050]

[0051] in Representation language model 230, Represents the numerical decoding model 240, y num Indicates a numerical result.

[0052] Example pre-training process for a molecular model

[0053] The molecular model 210 is used to encode and interpret molecular information. In some implementations, a Transformer-based graph neural network (e.g., Graphormer) can be selected as the backbone architecture of the molecular model 210. The Transformer-based graph neural network is a pure transformer architecture that can operate on graphs and introduces minimal inductive bias to ensure that the model can capture the topological and geometric information of molecules. The Transformer-based graph neural network has been shown to achieve advanced performance on multiple molecular tasks such as quantum property prediction and average distribution prediction. The Transformer-based graph neural network is designed to be scalable and general because it can process various types of graphs, such as three-dimensional molecular graphs, protein graphs, reaction graphs, etc. Therefore, the Transformer-based graph neural network is an ideal candidate model for the molecular model 210 because it can provide a powerful and general representation for molecules.

[0054] In some implementations, the molecular model 210 may be pre-trained. During the pre-training process of the molecular model 210, a loss function may be set, which may be used as an objective function for optimizing the molecular model 210. By minimizing the loss function, the parameters of the molecular model 210 may be updated, thereby optimizing the molecular model 210.

[0055] In some implementations, a denoising masked auto-encoder (DMAE) loss for molecules can be used, which generalizes the traditional MAE loss for self-supervised learning of molecular representations while being able to directly serve multiple tasks. Additionally, in some implementations, supervised losses for predicting energy and force can also be used to cover periodic material systems and regularize the learned representations to support the molecular properties of interest. In some implementations, the molecular model 210 can be pre-trained by alternating between batches of data from different fields (e.g., the chemical field and the materials field). In this way, the molecular model 210 is helped to learn shared representations and enables the molecular model 210 to be effectively generalized in these fields.

[0056] Figure 3 A schematic diagram of a molecular model 210 pre-training process 300 according to some implementations of the present disclosure is shown. In process 300, a mask may first be applied to a pre-trained molecular graph 310 corresponding to a pre-trained molecular system to obtain a masked molecular graph 320. Information related to at least one set of atoms in the pre-trained molecular system is masked in the masked molecular graph. The mask may be used to indicate that certain data points or features are to be ignored and replaced with a predetermined representation.

[0057] After obtaining the masked molecular graph, the molecular model 210 can be used to generate mask features 330 of the pre-trained molecular system from the masked molecular graph 320. In some implementations, the molecular model 210 can encode the masked molecular graph 320 to obtain the mask features 330 of the pre-trained molecular system. Next, atomic information of at least one atom set can be predicted based on the mask features 330.

[0058] In some implementations, the atom information may include the type of the atom, that is, which chemical element the atom is. Accordingly, the mask feature 330 may be input into the first decoder 340 to predict the atom type 350 of each atom in the at least one atom set.

[0059] Alternatively or additionally, the atomic information may include the position of the atom, such as the coordinates of the atom in the molecule, etc. Accordingly, the mask feature 330 may be input into the second decoder 342 to predict the atomic position 352 of each atom in the at least one atomic set.

[0060] After the predicted atomic information is obtained, model parameters of the molecular model may be updated based on a difference between the predicted atomic information and reference atomic information of at least one atom set.

[0061] In some implementations, the difference between the predicted atomic information and the reference atomic information can be calculated using a loss function. The DMAE loss is inspired by the similarity between MAE (Masked Auto-Encoder) training of molecular structures and the denoising form of the diffusion model, where the latter can be viewed as performing the former at multiple noise levels. Following the diffusion model formula, the noise level is marked by the evolution time t∈[0,T] of the noise injection diffusion process and is added to the model input. An example of the first loss function 360 for atomic positions is as follows:

[0062]

[0063]

[0064] where R msk Represents the predicted atomic information, R vis represents the reference atom information, φ(R msk , R vis , t) represents the output of the molecular model 210, D pos (R msk , R vis , φ) represents the predicted atomic positions. t~Unif[0,T] and is a random variable drawn for each evaluation, and σ t ∈(0,1) is the noise schedule that defines the diffusion process, which increases with the increase of t. This makes it possible to msk |R vis ) conditionally generates high-quality samples, following the diffusion formula of the noise insertion process. An example of this diffusion formula is as follows:

[0065]

[0066] where the simulation is performed at uniform discrete time steps i∈{0, 1, ..., N diff }, with a step size h = T / N diff , σ i :=σ t＝ih ,as well as

[0067] In order to enhance the understanding of the molecular model 210 for the atom type input, the corresponding MAE loss can be retained. type (φ) outputs a logical value for the mask atom type. An example of the second loss function 362 for the atom type is as follows:

[0068]

[0069] Where msk represents the "mask" word that replaces the mask atom, D type [A msk ] is a logical value evaluating to the true type of the masked atom, and the atom position input is omitted in the expression.

[0070] To implement the two loss functions mentioned above efficiently, they can be evaluated on the same set of mask inputs, i.e., the atom type and molecular graph inputs used for Equation (5) are used as mask inputs, and the atom position inputs used for Equation (7) are used as perturbation inputs.

[0071] Additionally, in some implementations, loss functions for energy and force may also be calculated. In this case, no mask or noise may be added to the pre-trained molecular graph 310. The pre-trained molecular graph 310 may be input into the molecular model 210. The features output by the molecular model 210 may be input into a third decoder 343 to predict the energy and force 354 of the molecular system.

[0072] The energy and force prediction loss is a supervised loss that trains the molecular model 210 to predict the total energy and atomic forces of the molecular graph given its atomic coordinates and lattice vectors. An example of the third loss function 364 for energy and force is as follows:

[0073]

[0074] in and Indicates the ability and power of prediction, and D energy and D force Denote the decoders for energy and force prediction, respectively. This loss captures the physical properties and interactions of the molecular graph and aligns with downstream molecular tasks such as structure optimization and phase transition prediction. These tasks can be viewed as minimizing the energy and forces of the molecular graph.

[0075] In this way, a combination of self-supervised and supervised learning objectives is used to provide a comprehensive pre-training method for prediction and generation tasks in both molecular and materials domains. By adopting an interactive training process that alternates between batches of molecular and materials data, the molecular model 210 is able to learn shared representations and generalize effectively in both domains.

[0076] As model parameters, training data, and computational cost increase, the molecular model 210 may have better performance. Figure 4 FIG. 2 shows an envelope curve of a molecular model 210 pre-trained according to some implementations of the present disclosure. Figure 4 As shown in curves 410 and 420 in FIG, as the number of model parameters and the number of floating-point operations per second (FLOPS) increase, the training loss of molecular model 210 decreases. Therefore, molecular model 210 can predictably benefit from the scaling up of the model and data.

[0077] Example training process for a general processing model

[0078] Figure 5 FIG. 5 is a schematic diagram illustrating a general processing model 200 training process 500 according to some implementations of the present disclosure. Figure 5 As shown, the process 500 may include at least one of feature alignment training 510 , knowledge injection training 520 , and instruction fine-tuning training 530 .

[0079] Since the molecular model 210 and the language model 230 have different input modalities (i.e., molecular systems and text) and feature spaces, this means that the two models cannot effectively exchange information and may suffer from information loss or noise. In some implementations, during the training phase, feature alignment training 510 for the feature adaptation model 220 can be performed, and the goal of the feature alignment training 510 includes causing the feature adaptation model 220 to map the output features of the molecular model 210 to the input feature space of the language model 230. Exemplarily, the feature adaptation model 220 is composed of some linear layers with nonlinear activation functions and has learnable parameters that can be adjusted during training.

[0080] Specifically, first, first training data can be obtained. The first training data includes a first training molecular graph corresponding to the first training molecular system and a reference description for the first training molecular system. The reference description for the first training molecular system may include a description of the properties of the molecule, such as a description of the molecular solubility or toxicity, etc. Alternatively or additionally, the reference description for the first training molecular system may include a description of the composition of the molecule, such as a description of the functional groups included in the molecule, etc. Further, the first training data may include one or more appropriate data sets, such data sets may include descriptions of the properties, composition, etc. of the molecule.

[0081] After obtaining the first training data, feature adaptation model 220, molecular model 210, and language model 230 can be used to generate a predictive description of the first training molecular system based on the first training molecular graph. For example, these models can be used to generate the number of functional groups in the first training molecular system and descriptions of each functional group. For another example, these models can be used to generate descriptions of properties of the first training molecular system, such as solubility and toxicity.

[0082] Next, the model parameters of the feature adaptation model 220 can be updated based on the difference between the predicted description and the reference description. For example, the predicted description for the first training molecular system is "having 3 functional groups, namely functional groups A, B, and C", while the reference description for the first training molecular system is "having 4 functional groups, namely functional groups C, D, E, and F". Based on the difference between the predicted description and the reference description, the model parameters of the feature adaptation model 220 are updated. Exemplarily, a loss function can be constructed between the predicted description and the reference description, and the model parameters of the feature adaptation model 220 can be updated by minimizing the loss function and maximizing the mutual information between the molecular system modality and the text modality. During the training of the feature adaptation model 220, the parameters of the molecular model 210 and the language model 230 can be frozen, and only the parameters of the feature adaptation model 220 can be adjusted.

[0083] The molecular model 210 is pre-trained in common molecular tasks, such as predicting molecular properties and generating molecular graphs. However, the molecular model 210 does not have sufficient scientific knowledge to perform tasks that require cross-domain integration and reasoning, such as predicting the therapeutic effects of molecules or planning the synthesis routes of molecules. Therefore, in some implementations, during the training phase, knowledge injection training 520 can be performed for the molecular model 210. The goal of knowledge injection training 520 includes enabling the molecular model to extract features associated with the properties of the corresponding molecular system from the molecular graph. After such training, the molecular model 210 can associate the extracted features with the properties of the molecular system.

[0084] Specifically, first, second training data can be obtained. The second training data includes a second training molecular graph corresponding to the second training molecular system and reference values ​​for properties of the second training molecular system. Furthermore, the second training data can include various suitable datasets. By way of example and not limitation, such datasets can include scientific corpora, datasets providing numerical values ​​for physicochemical properties, biological activities, and the like of molecules, and the like.

[0085] After obtaining the second training data, feature adaptation model 220, molecular model 210, and language model 230 can be used to generate predictions for properties of the second training molecular system based on the second training molecular graph. For example, these models can be used to generate the biological activity of the second training molecular system.

[0086] Next, the model parameters of the molecular model 210 can be updated based on the difference between the predicted value and the reference value. For example, a loss function (e.g., a mean square error loss function or a cross entropy loss function) can be constructed between the predicted value and the reference value, and the model parameters of the molecular model 210 can be updated by minimizing the loss function. During the process of training the model parameters of the molecular model 210, the parameters of the language model 230 can be ignored, and only the parameters of the molecular model 210 can be adjusted.

[0087] Alternatively or additionally, in some implementations, during the training phase, instruction fine-tuning training 530 is performed on the language model 230. The goal of instruction fine-tuning training 530 is to enable the language model to understand tasks related to molecular systems expressed in natural language. After such training, the language model 230 can be better aligned with human instructions, particularly scientifically relevant instructions, enabling the language model 230 to use natural language instructions as a unified task specification and guidance for generating results, and to perform new molecular tasks in zero-shot or few-shot learning situations.

[0088] Specifically, first the third training data can be obtained. The third training data include the third training molecular graph corresponding to the third training molecular system, the reference instruction associated with the third training molecular system, and the reference response for the reference instruction. Further, the third training data can include various suitable data sets. As an example and without any restriction, such data sets can include the instruction-response pairs generated from scientific papers, the question-answer pairs (that is, instruction-response pairs) collected from question-answer forums, etc. These data sets provide abundant natural language instructions and responses for various scientific problems, and can help language model 230 learn the corresponding relationship and consistency between instructions and tasks and between tasks and results.

[0089] After the third training data is obtained, the feature adaptation model 220 , the molecular model 210 , and the language model 230 may be used to generate a predicted response to the reference instruction based on the third training molecular graph.

[0090] Next, the model parameters of the language model 230 can be updated based on the difference between the predicted response and the reference response. For example, a loss function (e.g., a cross-entropy loss function) can be constructed between the predicted response and the reference response, and the model parameters of the language model 230 can be updated by minimizing the loss function. During the model parameter training process, the parameters of the molecular model 210 can be frozen, and only the parameters of the language model 230 can be adjusted.

[0091] In some implementations, feature alignment training is performed before knowledge injection training, and knowledge injection training is performed before instruction fine-tuning training. Feature alignment training can ensure compatibility between the molecular model 210 and the language model 230, so feature alignment training can be performed before knowledge injection training and instruction fine-tuning training. Because knowledge injection training imparts the necessary skills for various molecular tasks to the molecular model 210, knowledge injection training can be performed before instruction fine-tuning training.

[0092] Example Process

[0093] Figure 6 FIG. 6 is a flow chart illustrating a process 600 for processing tasks related to molecular systems according to some implementations of the present disclosure. The process 600 may be implemented in Figure 1 At the electronic device 110, for example, it can be implemented using the general processing model 140 in the electronic device 110.

[0094] like Figure 6 As shown, in block 610 , the electronic device 110 obtains input including text and indicating a target molecular system, where the text indicates a target task related to the target molecular system.

[0095] At block 620 , the electronic device 110 generates a molecular feature of the target molecular system based on a target molecular graph corresponding to the target molecular system. The molecular graph includes a plurality of nodes connected by edges, each node representing a set of atoms in the corresponding molecular system.

[0096] In block 630 , the electronic device 110 determines a result of the target task based on the text features of the text and the molecular features of the target molecular system.

[0097] In some implementations, generating molecular features of the target molecular system includes: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; and converting the intermediate features into a text feature space using a feature adaptation model to obtain the molecular features.

[0098] In some implementations, determining a result of a target task includes: providing text features and molecular features as inputs to a language model to obtain an output of the language model; determining a numerical representation from the output of the language model based on a position of a predetermined representation corresponding to the numerical value in the output of the language model; and obtaining a numerical result from the numerical representation using a numerical decoding model as at least a part of the result.

[0099] In some implementations, molecular features are generated using a molecular model and a feature adaptation model, and the result of the target task is determined using a language model, and in the training phase, the method further includes at least one of the following: performing feature alignment training for the feature adaptation model, the goal of the feature alignment training including enabling the feature adaptation model to map the output features of the molecular model to the input feature space of the language model, performing knowledge injection training for the molecular model, the goal of the knowledge injection training including enabling the molecular model to extract features associated with the properties of the corresponding molecular system from the molecular graph, or performing instruction fine-tuning training for the language model, the goal of the instruction fine-tuning training including enabling the language model to understand tasks related to the molecular system expressed in natural language.

[0100] In some implementations, performing feature alignment training includes: obtaining first training data, the first training data including a first training molecular graph corresponding to a first training molecular system and a reference description for the first training molecular system; generating a predicted description for the first training molecular system based on the first training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the feature adaptation model based on a difference between the predicted description and the reference description.

[0101] In some implementations, performing knowledge injection training includes: obtaining second training data, the second training data including a second training molecular graph corresponding to the second training molecular system and reference values ​​of properties of the second training molecular system; generating predicted values ​​of properties of the second training molecular system based on the second training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the molecular model based on a difference between the predicted values ​​and the reference values.

[0102] In some implementations, performing instruction fine-tuning training includes: obtaining third training data, the third training data including a third training molecular graph corresponding to a third training molecular system, a reference instruction associated with the third training molecular system, and a reference response to the reference instruction; generating a predicted response to the reference instruction based on the third training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the language model based on a difference between the predicted response and the reference response.

[0103] In some implementations, feature alignment training is performed before knowledge injection training, and knowledge injection training is performed before instruction fine-tuning training.

[0104] In some implementations, the molecular features are generated using a molecular model, and in the pre-training stage of the molecular model, process 600 also includes: obtaining a masked molecular graph by applying a mask to a pre-trained molecular graph corresponding to the pre-trained molecular system, in which information related to at least one atomic set in the pre-trained molecular system is masked; generating masked features of the pre-trained molecular system from the masked molecular graph using the molecular model; predicting atomic information of at least one atomic set based on the masked features; and updating model parameters of the molecular model based on the difference between the predicted atomic information and the reference atomic information of the at least one atomic set.

[0105] Example device

[0106] Figure 7 Schematic block diagram of an electronic device capable of implementing multiple implementations of the present disclosure is shown. It should be understood that Figure 7 The illustrated electronic device 700 is merely exemplary and should not be construed as limiting the functionality and scope of the implementations described in the present disclosure.

[0107] like Figure 7 As shown, electronic device 700 comprises an electronic device in the form of a general purpose computing device 700. Components of electronic device 700 may include, but are not limited to, one or more processors or processing devices 710, memory 720, storage device 730, one or more communication units 740, one or more input devices 750, and one or more output devices 760.

[0108] In some implementations, the electronic device 700 may be implemented as a computing device, a computing system, a server, a mainframe, or other device with computing capabilities.

[0109] The processing device 710 may be a real or virtual processor and may be capable of performing various processes according to a program stored in the memory 720. In a multi-processor system, multiple processing units execute computer-executable instructions in parallel to increase the parallel processing capabilities of the electronic device 700. The processing device 710 may include a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor, a controller, and / or a microcontroller, among others.

[0110] The electronic device 700 typically includes a plurality of computer storage media. Such media can be any available media accessible to the electronic device 700, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 720 may include volatile memory (e.g., registers, cache, random access memory (RAM)), non-volatile memory (e.g., read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory) or some combination thereof. The storage device 730 may include removable or non-removable media and may include computer-readable media, such as memory, flash drive, disk or any other media that can be used to store information and / or data and can be accessed within the electronic device 700.

[0111] The electronic device 700 may further include additional removable / non-removable, volatile / non-volatile storage media. Figure 7 As shown in FIG, a magnetic disk drive for reading or writing from a removable, nonvolatile magnetic disk and an optical disk drive for reading or writing from a removable, nonvolatile optical disk can be provided. In these cases, each drive can be connected to a bus (not shown) by one or more data media interfaces.

[0112] The communication unit 740 enables communication with another computing device via a communication medium. Additionally, the functions of the components of the electronic device 700 can be implemented as a single computing cluster or multiple computing machines that can communicate via a communication connection. Thus, the electronic device 700 can operate in a networked environment using logical connections to one or more other servers, personal computers (PCs), or another general network node.

[0113] The input device 750 may be one or more various input devices, such as a mouse, a keyboard, a data import device, etc. The output device 760 may be one or more output devices, such as a display, a data export device, etc. The electronic device 700 may also communicate with one or more external devices (not shown) through the communication unit 740 as needed, such as a storage device, a display device, etc., with one or more devices that allow a user to interact with the electronic device 700, or with any device that allows the electronic device 700 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0114] In some implementations, in addition to being integrated on a single device, some or all of the various components of the electronic device 700 may also be arranged in the form of a cloud computing architecture. In a cloud computing architecture, these components can be remotely located and can work together to implement the functionality described in this disclosure. In some implementations, cloud computing provides computing, software, data access, and storage services that do not require the end user to know the physical location or configuration of the systems or hardware providing these services. In various implementations, cloud computing uses appropriate protocols to provide services over a wide area network (such as the Internet). For example, a cloud computing provider provides applications over a wide area network, and they can be accessed through a web browser or any other computing component. The software or components of the cloud computing architecture and the corresponding data can be stored on servers at a remote location. The computing resources in a cloud computing environment can be consolidated at a remote data center location or they can be dispersed. The cloud computing infrastructure can provide services through a shared data center, even if they appear to be a single access point for users. Therefore, the components and functionality described herein can be provided from a service provider at a remote location using a cloud computing architecture. Alternatively, they can be provided from a conventional server, or they can be installed directly or otherwise on the client device.

[0115] The electronic device 700 can be used to implement molecular modeling in various implementations of the present disclosure. The memory 720 may include one or more modules having one or more program instructions, which can be accessed and executed by the processing unit 710 to implement the functions of various implementations described herein. For example, the memory 720 may include a molecular modeling module 722 for performing molecular modeling in one or more of the above implementations. Figure 7 As shown, electronic device 700 can obtain input required for molecular modeling through input device 750 and provide output of molecular modeling, such as predicted properties, through output device 760. In some implementations, electronic device 700 can also receive input from other devices (not shown) via communication unit 740.

[0116] Example Implementation

[0117] Some example implementations of the present disclosure are listed below.

[0118] In one aspect, the present disclosure provides a computer-implemented method comprising: obtaining input comprising text indicating a target molecular system, wherein the text indicates a target task related to the target molecular system; generating molecular features of the target molecular system based on a target molecular graph corresponding to the target molecular system, wherein the molecular graph comprises a plurality of nodes connected by edges, each node representing a set of atoms in the corresponding molecular system; and determining a result of the target task based on the textual features of the text and the molecular features of the target molecular system.

[0119] In some implementations, generating molecular features of the target molecular system includes: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; and converting the intermediate features into a text feature space using a feature adaptation model to obtain the molecular features.

[0120] In some implementations, determining a result of a target task includes: providing text features and molecular features as inputs to a language model to obtain an output of the language model; determining a numerical representation from the output of the language model based on a position of a predetermined representation corresponding to the numerical value in the output of the language model; and obtaining a numerical result from the numerical representation using a numerical decoding model as at least a part of the result.

[0121] In some implementations, molecular features are generated using a molecular model and a feature adaptation model, and the result of the target task is determined using a language model, and in the training phase, the method further includes at least one of the following: performing feature alignment training for the feature adaptation model, the goal of the feature alignment training including enabling the feature adaptation model to map the output features of the molecular model to the input feature space of the language model, performing knowledge injection training for the molecular model, the goal of the knowledge injection training including enabling the molecular model to extract features associated with the properties of the corresponding molecular system from the molecular graph, or performing instruction fine-tuning training for the language model, the goal of the instruction fine-tuning training including enabling the language model to understand tasks related to the molecular system expressed in natural language.

[0122] In some implementations, performing feature alignment training includes: obtaining first training data, the first training data including a first training molecular graph corresponding to a first training molecular system and a reference description for the first training molecular system; generating a predicted description for the first training molecular system based on the first training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the feature adaptation model based on a difference between the predicted description and the reference description.

[0123] In some implementations, performing knowledge injection training includes: obtaining second training data, the second training data including a second training molecular graph corresponding to the second training molecular system and reference values ​​of properties of the second training molecular system; generating predicted values ​​of properties of the second training molecular system based on the second training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the molecular model based on a difference between the predicted values ​​and the reference values.

[0124] In some implementations, performing instruction fine-tuning training includes: obtaining third training data, the third training data including a third training molecular graph corresponding to a third training molecular system, a reference instruction associated with the third training molecular system, and a reference response to the reference instruction; generating a predicted response to the reference instruction based on the third training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the language model based on a difference between the predicted response and the reference response.

[0125] In some implementations, feature alignment training is performed before knowledge injection training, and knowledge injection training is performed before instruction fine-tuning training.

[0126] In some implementations, the molecular features are generated using a molecular model, and in the pre-training stage of the molecular model, the method further includes: obtaining a masked molecular graph by applying a mask to a pre-trained molecular graph corresponding to the pre-trained molecular system, in which information related to at least one atomic set in the pre-trained molecular system is masked; generating masked features of the pre-trained molecular system from the masked molecular graph using the molecular model; predicting atomic information of at least one atomic set based on the masked features; and updating model parameters of the molecular model based on the difference between the predicted atomic information and the reference atomic information of the at least one atomic set.

[0127] In another aspect, the present disclosure provides an electronic device. The electronic device includes: a processor; and a memory coupled to the processor and containing instructions stored thereon, wherein when executed by the processor, the instructions cause the device to perform the following actions, the actions comprising: obtaining input including text and indicating a target molecular system, wherein the text indicates a target task related to the target molecular system; generating molecular features of the target molecular system based on a target molecular graph corresponding to the target molecular system, wherein the molecular graph includes a plurality of nodes connected by edges, each node representing a set of atoms in the corresponding molecular system; and determining a result of the target task based on the text features of the text and the molecular features of the target molecular system.

[0128] In some implementations, generating molecular features of the target molecular system includes: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; and converting the intermediate features into a text feature space using a feature adaptation model to obtain the molecular features.

[0129] In some implementations, determining a result of a target task includes: providing text features and molecular features as inputs to a language model to obtain an output of the language model; determining a numerical representation from the output of the language model based on a position of a predetermined representation corresponding to the numerical value in the output of the language model; and obtaining a numerical result from the numerical representation using a numerical decoding model as at least a part of the result.

[0130] In some implementations, molecular features are generated using a molecular model and a feature adaptation model, and the result of the target task is determined using a language model, and in the training phase, the method further includes at least one of the following: performing feature alignment training for the feature adaptation model, the goal of the feature alignment training including enabling the feature adaptation model to map the output features of the molecular model to the input feature space of the language model, performing knowledge injection training for the molecular model, the goal of the knowledge injection training including enabling the molecular model to extract features associated with the properties of the corresponding molecular system from the molecular graph, or performing instruction fine-tuning training for the language model, the goal of the instruction fine-tuning training including enabling the language model to understand tasks related to the molecular system expressed in natural language.

[0131] In some implementations, performing feature alignment training includes: obtaining first training data, the first training data including a first training molecular graph corresponding to a first training molecular system and a reference description for the first training molecular system; generating a predicted description for the first training molecular system based on the first training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the feature adaptation model based on a difference between the predicted description and the reference description.

[0132] In some implementations, performing knowledge injection training includes: obtaining second training data, the second training data including a second training molecular graph corresponding to the second training molecular system and reference values ​​of properties of the second training molecular system; generating predicted values ​​of properties of the second training molecular system based on the second training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the molecular model based on a difference between the predicted values ​​and the reference values.

[0133] In some implementations, performing instruction fine-tuning training includes: obtaining third training data, the third training data including a third training molecular graph corresponding to a third training molecular system, a reference instruction associated with the third training molecular system, and a reference response to the reference instruction; generating a predicted response to the reference instruction based on the third training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the language model based on a difference between the predicted response and the reference response.

[0134] In some implementations, feature alignment training is performed before knowledge injection training, and knowledge injection training is performed before instruction fine-tuning training.

[0135] In some implementations, the molecular features are generated using a molecular model, and in the pre-training stage of the molecular model, the action also includes: applying a mask to a pre-trained molecular graph corresponding to the pre-trained molecular system to obtain a masked molecular graph, in which information related to at least one atomic set in the pre-trained molecular system is masked; using the molecular model, generating masked features of the pre-trained molecular system from the masked molecular graph; based on the masked features, predicting atomic information of at least one atomic set; and updating model parameters of the molecular model based on the difference between the predicted atomic information and the reference atomic information of the at least one atomic set.

[0136] On the other hand, the present disclosure provides a computer program product, which is tangibly stored in a computer storage medium and includes computer-executable instructions, which, when executed by a device, cause the device to perform the following actions, the actions including: obtaining input including text and indicating a target molecular system, the text indicating a target task related to the target molecular system; generating molecular features of the target molecular system based on a target molecular graph corresponding to the target molecular system, the molecular graph including multiple nodes connected by edges, each node representing a set of atoms in the corresponding molecular system; and determining a result of the target task based on the text features of the text and the molecular features of the target molecular system.

[0137] In some implementations, generating molecular features of the target molecular system includes: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; and converting the intermediate features into a text feature space using a feature adaptation model to obtain the molecular features.

[0138] In some implementations, determining a result of a target task includes: providing text features and molecular features as inputs to a language model to obtain an output of the language model; determining a numerical representation from the output of the language model based on a position of a predetermined representation corresponding to the numerical value in the output of the language model; and obtaining a numerical result from the numerical representation using a numerical decoding model as at least a part of the result.

[0139] In some implementations, molecular features are generated using a molecular model and a feature adaptation model, and the result of the target task is determined using a language model, and in the training phase, the method further includes at least one of the following: performing feature alignment training for the feature adaptation model, the goal of the feature alignment training including enabling the feature adaptation model to map the output features of the molecular model to the input feature space of the language model, performing knowledge injection training for the molecular model, the goal of the knowledge injection training including enabling the molecular model to extract features associated with the properties of the corresponding molecular system from the molecular graph, or performing instruction fine-tuning training for the language model, the goal of the instruction fine-tuning training including enabling the language model to understand tasks related to the molecular system expressed in natural language.

[0140] In some implementations, performing feature alignment training includes: obtaining first training data, the first training data including a first training molecular graph corresponding to a first training molecular system and a reference description for the first training molecular system; generating a predicted description for the first training molecular system based on the first training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the feature adaptation model based on a difference between the predicted description and the reference description.

[0141] In some implementations, performing knowledge injection training includes: obtaining second training data, the second training data including a second training molecular graph corresponding to the second training molecular system and reference values ​​of properties of the second training molecular system; generating predicted values ​​of properties of the second training molecular system based on the second training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the molecular model based on a difference between the predicted values ​​and the reference values.

[0142] In some implementations, performing instruction fine-tuning training includes: obtaining third training data, the third training data including a third training molecular graph corresponding to a third training molecular system, a reference instruction associated with the third training molecular system, and a reference response to the reference instruction; generating a predicted response to the reference instruction based on the third training molecular graph using a feature adaptation model, a molecular model, and a language model; and updating model parameters of the language model based on a difference between the predicted response and the reference response.

[0143] In some implementations, feature alignment training is performed before knowledge injection training, and knowledge injection training is performed before instruction fine-tuning training.

[0144] In some implementations, the molecular features are generated using a molecular model, and in the pre-training stage of the molecular model, the action also includes: applying a mask to a pre-trained molecular graph corresponding to the pre-trained molecular system to obtain a masked molecular graph, in which information related to at least one atomic set in the pre-trained molecular system is masked; using the molecular model, generating masked features of the pre-trained molecular system from the masked molecular graph; based on the masked features, predicting atomic information of at least one atomic set; and updating model parameters of the molecular model based on the difference between the predicted atomic information and the reference atomic information of the at least one atomic set.

[0145] In yet another aspect, the present disclosure provides a computer-readable medium having computer-executable instructions stored thereon. When the computer-executable instructions are executed by a device, the device performs one or more example implementations of the method of the above aspects.

[0146] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, example types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), load programmable logic devices (CPLDs), and the like.

[0147] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0148] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0149] In addition, although each operation is described in a specific order, this should be understood as requiring such operation to be performed in the specific order shown or in a sequential order, or requiring that all illustrated operations should be performed to obtain the desired result. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details have been included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate implementation can also be implemented in a single implementation in combination. On the contrary, the various features described in the context of a single implementation can also be implemented in multiple implementations individually or in any suitable sub-combination mode.

[0150] Although the subject matter has been described in language specific to structural features and / or methodological logical acts, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms of implementing the claims.< / num> < / num>

Claims

1. A computer-implemented method comprising: obtaining an input including text and indicating a target molecular system, wherein the text indicates a target task related to the target molecular system; generating molecular features of the target molecular system based on a target molecular graph corresponding to the target molecular system, the molecular graph comprising a plurality of nodes connected by edges, each node representing a set of atoms in the corresponding molecular system; and Based on the text features of the text and the molecular features of the target molecular system, a result of the target task is determined.

2. The method of claim 1 , wherein generating a molecular signature of the target molecular system comprises: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; as well as The intermediate features are converted into a text feature space using a feature adaptation model to obtain the molecular features.

3. The method according to claim 1, wherein determining the result of the target task comprises: providing the text features and the molecular features as inputs to a language model to obtain an output of the language model; determining a numerical representation from an output of the language model based on a position of a predetermined representation corresponding to a numerical value in the output of the language model; as well as A numerical result is obtained from the numerical representation as at least a part of the result using a numerical decoding model.

4. The method of claim 1, wherein the molecular features are generated using a molecular model and a feature adaptation model, and the result of the target task is determined using a language model, and in the training phase, the method further comprises at least one of the following: Performing feature alignment training for the feature adaptation model, wherein the goal of the feature alignment training includes enabling the feature adaptation model to map the output features of the molecular model to the input feature space of the language model, performing knowledge injection training for the molecular model, wherein the goal of the knowledge injection training includes enabling the molecular model to extract features associated with properties of the corresponding molecular system from a molecular graph, or Instruction fine-tuning training is performed on the language model, where a goal of the instruction fine-tuning training includes enabling the language model to understand tasks related to molecular systems expressed in natural language.

5. The method according to claim 4, wherein performing the feature alignment training comprises: Acquire first training data, where the first training data includes a first training molecular graph corresponding to a first training molecular system and a reference description for the first training molecular system; generating a prediction description for the first training molecular system based on the first training molecular graph using the feature adaptation model, the molecular model, and the language model; as well as Based on the difference between the predicted description and the reference description, model parameters of the feature adaptation model are updated.

6. The method according to claim 4, wherein performing the knowledge injection training comprises: Acquiring second training data, the second training data comprising a second training molecular graph corresponding to a second training molecular system and reference values ​​of properties of the second training molecular system; generating a predicted value of a property of the second training molecular system based on the second training molecular graph using the feature adaptation model, the molecular model, and the language model; and Based on the difference between the predicted value and the reference value, a model parameter of the molecular model is updated.

7. The method according to claim 4, wherein performing the instruction fine-tuning training comprises: Acquiring third training data, the third training data including a third training molecular graph corresponding to a third training molecular system, a reference instruction associated with the third training molecular system, and a reference response to the reference instruction; generating a predicted response to the reference instruction based on the third training molecular graph using the feature adaptation model, the molecular model, and the language model; as well as Model parameters of the language model are updated based on the difference between the predicted response and the reference response. 8 . The method according to claim 4 , wherein the feature alignment training is performed before the knowledge injection training, and the knowledge injection training is performed before the instruction fine-tuning training.

9. The method according to claim 1, wherein the molecular features are generated using a molecular model, and in a pre-training stage of the molecular model, the method further comprises: Obtaining a masked molecular graph by applying a mask to a pretrained molecular graph corresponding to a pretrained molecular system, wherein information related to at least one set of atoms in the pretrained molecular system is masked in the masked molecular graph; generating mask features of the pre-trained molecular system from the masked molecular graph using the molecular model; predicting atomic information of the at least one atom set based on the mask features; as well as Model parameters of the molecular model are updated based on the difference between the predicted atomic information and the reference atomic information of the at least one set of atoms.

10. An electronic device comprising: processing unit; as well as a memory coupled to the processing unit and containing instructions stored thereon, the instructions, when executed by the processing unit, causing the apparatus to perform the following actions, the actions comprising: obtaining an input including text and indicating a target molecular system, wherein the text indicates a target task related to the target molecular system; generating molecular features of the target molecular system based on a target molecular graph corresponding to the target molecular system, the molecular graph comprising a plurality of nodes connected by edges, each node representing a set of atoms in the corresponding molecular system; and Based on the text features of the text and the molecular features of the target molecular system, a result of the target task is determined.

11. The apparatus of claim 10, wherein generating a molecular signature of the target molecular system comprises: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; as well as The intermediate features are converted into a text feature space using a feature adaptation model to obtain the molecular features.

12. The apparatus according to claim 10, wherein determining the result of the target task comprises: providing the text features and the molecular features as inputs to a language model to obtain an output of the language model; determining a numerical representation from an output of the language model based on a position of a predetermined representation corresponding to a numerical value in the output of the language model; as well as A numerical result is obtained from the numerical representation as at least a part of the result using a numerical decoding model.

13. The apparatus according to claim 10, wherein the molecular features are generated using a molecular model and a feature adaptation model, and the result of the target task is determined using a language model, and in the training phase, the method further comprises at least one of the following: Performing feature alignment training for the feature adaptation model, wherein the goal of the feature alignment training includes enabling the feature adaptation model to map the output features of the molecular model to the input feature space of the language model, performing knowledge injection training for the molecular model, wherein the goal of the knowledge injection training includes enabling the molecular model to extract features associated with properties of the corresponding molecular system from a molecular graph, or Instruction fine-tuning training is performed on the language model, where a goal of the instruction fine-tuning training includes enabling the language model to understand tasks related to molecular systems expressed in natural language.

14. The apparatus of claim 13, wherein performing the feature alignment training comprises: Acquire first training data, where the first training data includes a first training molecular graph corresponding to a first training molecular system and a reference description for the first training molecular system; generating a prediction description for the first training molecular system based on the first training molecular graph using the feature adaptation model, the molecular model, and the language model; as well as Based on the difference between the predicted description and the reference description, model parameters of the feature adaptation model are updated.

15. The apparatus of claim 13, wherein performing the knowledge injection training comprises: Acquiring second training data, the second training data comprising a second training molecular graph corresponding to a second training molecular system and reference values ​​of properties of the second training molecular system; generating a predicted value of a property of the second training molecular system based on the second training molecular graph using the feature adaptation model, the molecular model, and the language model; and Based on the difference between the predicted value and the reference value, a model parameter of the molecular model is updated.

16. The apparatus according to claim 13, wherein performing the instruction fine-tuning training comprises: Acquiring third training data, the third training data including a third training molecular graph corresponding to a third training molecular system, a reference instruction associated with the third training molecular system, and a reference response to the reference instruction; generating a predicted response to the reference instruction based on the third training molecular graph using the feature adaptation model, the molecular model, and the language model; as well as Model parameters of the language model are updated based on the difference between the predicted response and the reference response. 17 . The apparatus of claim 13 , wherein the feature alignment training is performed before the knowledge injection training, and the knowledge injection training is performed before the instruction fine-tuning training.

18. The apparatus according to claim 10, wherein the molecular features are generated using a molecular model, and in a pre-training stage of the molecular model, the method further comprises: Obtaining a masked molecular graph by applying a mask to a pretrained molecular graph corresponding to a pretrained molecular system, wherein information related to at least one set of atoms in the pretrained molecular system is masked in the masked molecular graph; generating mask features of the pre-trained molecular system from the masked molecular graph using the molecular model; predicting atomic information of the at least one atom set based on the mask features; as well as Model parameters of the molecular model are updated based on the difference between the predicted atomic information and the reference atomic information of the at least one set of atoms.

19. A computer program product tangibly stored in a computer storage medium and comprising computer-executable instructions that, when executed by a device, cause the device to perform the following actions, the actions comprising: obtaining an input including text and indicating a target molecular system, wherein the text indicates a target task related to the target molecular system; generating molecular features of the target molecular system based on a target molecular graph corresponding to the target molecular system, the molecular graph comprising a plurality of nodes connected by edges, each node representing a set of atoms in the corresponding molecular system; and Based on the text features of the text and the molecular features of the target molecular system, a result of the target task is determined.

20. The computer program product of claim 19, wherein generating a molecular signature of the target molecular system comprises: generating intermediate features of the target molecular system from the target molecular graph using a molecular model; as well as The intermediate features are converted into a text feature space using a feature adaptation model to obtain the molecular features.

Images