Reinforcement learning driven parent core directed pyroptosis photosensitizer molecule generation method and device

Abstract

The present invention provides a reinforcement learning-based method for generating photosensitizer molecules with nucleus-directed pyroptosis induction, comprising: constructing a molecular library with photosensitizer molecule characteristics based on existing photosensitizer donors or acceptors, and disassembling the molecules into skeleton-decoration pairs according to chemical fragmentation rules, and labeling the connection sites to form a skeleton-decoration dataset; training a priori generation model to master the rules of molecular structure representation and to have the conditional generation ability from the nucleus skeleton to the decorative fragment; simultaneously training a molecular property prediction model based on quantum chemical calculation results; in the reinforcement learning stage, initializing the molecular generation agent with the priori generation model, constructing a normalized multi-objective reward using the property estimates given by the prediction model, and iteratively updating the generation strategy of the agent; by introducing a memory bank and a diversity filtering mechanism, high-scoring and non-repeating molecules are retained and expanded. The present invention can significantly improve the screening efficiency while ensuring structural controllability and property interpretability.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug molecule design technology, and specifically relates to a reinforcement learning-driven method and apparatus for generating nucleus-directed pyrolysis photosensitizer molecules. Background Technology

[0002] Photodynamic therapy (PDT), a minimally invasive treatment, utilizes photosensitizers to generate reactive oxygen species (ROS) under specific wavelengths of irradiation to kill tumor cells. It has gained attention due to its spatiotemporal controllability and minimal damage to normal tissues. Recent studies have further demonstrated that PDT can trigger pyroptosis under certain conditions, thereby amplifying the anti-tumor immune response. However, the discovery of pyroptosis-inducible photosensitizers still primarily relies on researchers' experience-driven, iterative experiments and photophysical property optimization. This process is cumbersome, time-consuming, labor-intensive, and costly, making it difficult to efficiently produce candidate molecules. Therefore, achieving automated and large-scale discovery of pyroptosis-inducible photosensitizers has become a significant challenge in this field.

[0003] Currently, with the development of deep learning, the automation and intelligence of molecular generation and drug discovery are constantly advancing. However, in the pyroptosis-induced photosensitizer scenario, existing automation strategies tend to generate from scratch, and their support for preserving and directionally expanding the validated parent nucleus is still insufficient. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, this invention aims to provide a method and apparatus for generating pyroptosis-induced photosensitizer molecules based on reinforcement learning, addressing the shortcomings of existing technologies in preserving and directionally expanding validated cores and linkage site constraints, resulting in low candidate efficacy. By optimizing multi-objective reinforcement learning around indicators such as singlet-triplet bandgap, charge, and symmetry while keeping the core and linkage sites unchanged, this invention enables automated and batch generation of photosensitizer molecules for pyroptosis induction, improving the reliability and efficiency of drug discovery.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of this invention provides a reinforcement learning-driven method for generating nucleus-directed pyrolysis photosensitizer molecules, comprising:

[0008] Existing photosensitizer donors or acceptors are collected and expanded through fragment combination to construct a molecular library with the characteristics of photosensitizer molecules;

[0009] According to the chemical fragmentation rules, each molecule in the molecular library is disassembled into skeleton-decoration pairs, and the connection sites are labeled to construct a skeleton-decoration dataset.

[0010] The prior generation model is pre-trained based on the skeleton-decoration dataset, enabling the prior generation model to master the molecular structure representation rules and have the conditional generation ability from the parent skeleton to decorative fragments.

[0011] A representative subset is extracted from the molecular library, and the target quantum chemical properties of each molecule in the representative subset are obtained as supervision signals to pre-train the molecular property prediction model.

[0012] Using the prior generation model as an initialization strategy, a molecular generation agent is constructed. During the reinforcement learning loop, under the constraints of a given parent core skeleton and connection sites, the molecular generation agent generates candidate decorative fragments in batches according to the current generation strategy, and combines them with the parent core skeleton to form candidate molecules. The quantum chemical properties of each candidate molecule are evaluated by the molecular property prediction model, and combined with the multiple molecular structure information of the candidate molecules, a weighted multi-objective reward is formed as a signal for updating the generation strategy. Candidate molecules whose weighted multi-objective reward is not lower than a preset threshold and are not repeated after structural standardization are added to the memory bank. The memory bank obtained at the end of reinforcement learning is used as a candidate photosensitizer molecule library for pyroptosis induction.

[0013] In some embodiments, the construction of the skeleton-decoration dataset includes:

[0014] Predefined chemical rules containing multiple functional group fracture modes are used to identify breakable bonds. Combinations of breakable bonds that meet preset quantity and constraints are enumerated, and fracture is performed to obtain multiple fragments. Among the fragments, the fragments that meet the target number of connection sites are selected as the skeleton, and the remaining fragments are used as the decoration set. Segmentation results that do not meet the constraints of the number of connection sites, ring structure integrity, or preset molecular properties are discarded.

[0015] In some embodiments, the prior generation model employs an encoder-decoder sequence generation model conditioned on a preset molecular structure representation format of the parent nucleus backbone, and generates a decorative sequence by fusing parent nucleus backbone information through a global attention mechanism. The expression for the global attention mechanism is:

[0016]

[0017]

[0018]

[0019] in, Indicates the first The distribution of attention weights for each position in the source sequence in each decoding step; The feature dimension representing the encoder output state and the decoder hidden state; The output matrix of the encoder at each time step serves as the key or value in the attention mechanism; For the first The decoder hidden state of each decoding step serves as a query in the attention mechanism; This represents the context vector obtained by weighting and converging the encoder output states according to the attention weights; This represents the fusion function, used to fuse the hidden states of the decoder. With context vector Obtain the decoding state after attention enhancement. It is used to predict the next molecular structure representation label in the output layer.

[0020] In some embodiments, the prior generative model is trained by maximizing the conditional likelihood of the decoration sequence given a parent skeleton; for having The parental skeleton of each connection site is generated by the prior generation model according to the order of the connection sites. A decorative sequence, and the overall generation target under the condition of the mother core skeleton. Represented as:

[0021]

[0022] in, These are the parameters for the prior generation model. This represents the random variable corresponding to the parent core skeleton. This represents the specific structural representation of the currently given target core skeleton; For the first The site decoration at the first A random variable of the step. for The value of , For the first The decoration sequence corresponding to the site is in the . The molecular structure of a step represents a labeled random variable. for The value of , For decoration sequence Length, This represents the conditional probability given by the prior generative model;

[0023] When training the prior generative model, the real decoration sequences in the skeleton-decoration dataset are used as supervision signals to minimize the negative log-likelihood:

[0024]

[0025] in, It is a negative log-likelihood.

[0026] In some embodiments, extracting a representative subset from the molecular library includes:

[0027] Several representative molecules are extracted from the molecular library. Each extracted molecule is parsed using a preset molecular structure representation format and then embedded in three-dimensional coordinates to transform it into a molecular graph. Atoms are used as nodes and covalent bonds as edges. Non-covalent adjacent edges are added when the distance between atoms does not exceed a preset inter-atomic distance threshold to represent short-range interactions. Each node is given a multi-dimensional atomic feature containing elemental and environmental information, and each edge is given a multi-dimensional edge feature containing bond type and geometric information to obtain the representative subset.

[0028] In some embodiments, the molecular property prediction model employs a graph convolution structure incorporating the Kolmogorov–Arnold mechanism. During training, the target is subjected to mean absolute error loss, and a physical consistency constraint term is introduced. The physical consistency constraint term is used to constrain the minimum singlet-triple gap directly predicted by the molecular property prediction model to be consistent with the minimum singlet-triple gap calculated from the energy level differences between multiple singlet and triplet energy levels predicted by the molecular property prediction model.

[0029] In some embodiments, the weighted multi-objective reward employs a normalized scalar reward, which consists of three parts: quantum chemical properties, symmetry, and formal charge, with each part normalized to an interval. And then summed using weighted averages.

[0030] In some embodiments, adding candidate molecules whose weighted multi-objective rewards are not lower than a preset threshold and whose structures are not repeated after standardization to the memory bank includes:

[0031] The candidate molecules generated by the molecular generating agent are queried in the memory. If the structure already exists, its weighted multi-objective reward is set to zero or reduced to a penalty value. If the molecule is new and its weighted multi-objective reward is not lower than a preset reward threshold, the candidate molecule is added to the memory.

[0032] In some embodiments, the optimization objective of reinforcement learning employs an augmented posterior difference strategy based on augmented log-likelihood and prior experience retention:

[0033] Given a target parent skeleton, the molecular generating agent operates according to the current strategy. Sampling candidate molecules Record the candidate molecules Under the current strategy Log-likelihood Log-likelihood under the prior generative model and with reward scaling factor Constructing augmented log-likelihood :

[0034]

[0035] in, as candidate molecules Weighted multi-objective rewards;

[0036] Define the difference term for augmented posterior difference strategies as:

[0037]

[0038] And minimize The square of the result is used to update the strategy parameters of the molecular generating agent.

[0039] A second aspect of the present invention provides a reinforcement learning-driven device for generating photosensitizer molecules with a directional pyrolysis pattern, comprising:

[0040] The molecular library construction module is configured to collect existing photosensitizer donors or acceptors and expand them through fragment combination to construct a molecular library with photosensitizer molecule characteristics.

[0041] The skeleton-decoration dataset construction module is configured to disassemble each molecule in the molecular library into skeleton-decoration pairs according to chemical fragmentation rules and label the connection sites to construct the skeleton-decoration dataset.

[0042] The prior generation model acquisition module is configured to pre-train the prior generation model based on the skeleton-decoration dataset, so that the prior generation model can master the molecular structure representation rules and have the conditional generation ability from the parent skeleton to the decorative fragment;

[0043] The molecular property prediction model acquisition module is configured to extract a representative subset from the molecular library and obtain the target quantum chemical properties of each molecule in the representative subset as a supervision signal to pre-train the molecular property prediction model.

[0044] The generation module is configured to construct a molecular generation agent using the prior generation model as an initialization strategy. During the reinforcement learning loop, the molecular generation agent, under the constraints of a given parent core skeleton and connection sites, generates candidate decorative fragments in batches according to the current generation strategy and combines them with the parent core skeleton to form candidate molecules. The quantum chemical properties of each candidate molecule are evaluated by the molecular property prediction model, and combined with multiple molecular structural information of the candidate molecules, a weighted multi-objective reward is formed as a generation strategy update signal. Candidate molecules with a weighted multi-objective reward not lower than a preset threshold and which are not repeated after structural standardization are added to the memory bank. The memory bank obtained at the end of reinforcement learning is used as a candidate photosensitizer molecule library for pyroptosis induction.

[0045] Compared with the prior art, the present invention has at least the following beneficial effects:

[0046] This invention improves molecule discovery efficiency and increases the effective proportion of candidate molecules by preserving the user-input core structure during the generation process and conducting targeted generation and screening centered on the core. This narrows the originally vast and exhaustive chemical space into a homonuclear local subspace highly correlated with the target optical properties. Furthermore, this method exhibits good versatility and transferability: it can be adapted to experimental scenarios in different laboratories and with different scaffold systems simply by changing the input core according to research needs. For other molecular design tasks whose properties strongly depend on specific structural fragments, exploration can also be rapidly initiated based on existing prior knowledge, reducing the human and time costs of manual trial and error and improving the efficiency of automated screening and iteration. Attached Figure Description

[0047] Figure 1 This is an overall flowchart of a reinforcement learning-driven method for generating photosensitizer molecules with directed pyrolysis of the parent nucleus, provided in the first aspect of this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for explaining this application and are not intended to limit this application.

[0049] Conversely, this application covers any alternatives, modifications, equivalent methods, and schemes made within the spirit and scope of this application as defined by the claims. Furthermore, to provide the public with a better understanding of this application, certain specific details are described in detail below. However, this application can be fully understood by those skilled in the art even without these detailed descriptions.

[0050] According to the first aspect of the present invention, a method for generating nucleus-directed pyroptosis-induced photosensitizer molecules based on reinforcement learning can use a user-input nucleus as a fixed framework and combine multi-objective reinforcement learning to generate and screen candidate structures in a targeted manner, converging the search range to a homonuclear local chemical subspace that is highly correlated with the target photophysical properties; while ensuring framework consistency and property alignment, it realizes the automated and batch production of candidate molecules, thereby improving the efficiency of molecule discovery and increasing the proportion of effective candidates.

[0051] The first aspect of this invention proposes a method for generating nucleus-directed pyroptosis-induced photosensitizer molecules based on reinforcement learning. While keeping the input nucleus unchanged, it introduces linkage site constraints and multi-objective scoring in reinforcement learning to automatically explore synthesizable nucleus-directed pyroptosis-induced photosensitizer molecules and support batch generation, thereby improving the efficiency and quality of candidate nucleus-directed pyroptosis-induced photosensitizer molecule discovery. This embodiment includes the following steps:

[0052] Step S1. Collect existing photosensitizer donors or acceptors, expand them by fragment combination, and construct a molecular library with photosensitizer molecule characteristics to limit the chemical space to be explored to the photosensitizer-related space;

[0053] Step S2. According to the chemical fragmentation rules, each molecule in the constructed molecular library is disassembled into skeleton-decoration pairs, and the connection sites are labeled to construct a skeleton-decoration dataset.

[0054] Step S3. Based on the constructed skeleton-decoration dataset, the prior generation model is pre-trained so that the prior generation model can master the molecular structure representation rules and have the conditional generation ability from the parent skeleton to the decorative fragment;

[0055] Step S4. Extract a representative subset from the constructed molecular library and obtain the target quantum chemical properties of each molecule in the representative subset as a supervision signal to pre-train the molecular property prediction model;

[0056] Step S5. Using the pre-trained prior generation model as the initialization generation strategy, construct a molecular generation agent. In the reinforcement learning loop, under the constraints of the given parent core skeleton and connection sites, the molecular generation agent generates candidate decorative fragments in batches according to the current generation strategy and combines them with the parent core skeleton to form candidate molecules. For each candidate molecule, its quantum chemical properties are evaluated by the molecular property prediction model, and combined with the multiple molecular structure information of the candidate molecule, a weighted multi-objective reward is formed as the generation strategy update signal. Candidate molecules whose weighted multi-objective reward is not lower than the preset threshold and are not repeated after structural standardization are added to the memory bank. The memory bank obtained at the end of reinforcement learning is used as a candidate photosensitizer molecule library for the pyroptosis induction requirement.

[0057] The following is combined Figure 1 The specific implementation process of a method for generating nucleus-directed pyroptosis-induced photosensitizer molecules based on reinforcement learning, provided in the first aspect of the present invention, is described in detail below. The generation method of this example includes the following steps:

[0058] Step S1. Construct a molecular library with photosensitizer molecule characteristics.

[0059] Step S11. Collecting and organizing photosensitizer molecules

[0060] Given the vast chemical space and the inefficiency of direct searches, it is necessary to construct a molecular library with photosensitizer molecule characteristics to limit the explored chemical space to photosensitizer-related areas. Therefore, this embodiment first collects reported pyroptosis-inducible photosensitizer molecules from literature and public databases and stores them in a standardized text format. Considering that existing public molecular datasets and cheminformatics tools commonly use SMILES (Simplified Molecular-Input Line-Entry System) as the text representation of molecular structures, and that SMILES facilitates sequence modeling, word segmentation, and data augmentation for equivalent representations, this embodiment preferably uses SMILES as the molecular structure representation, and stores the collected pyroptosis-inducible photosensitizers uniformly in SMILES format. SMILES is a notation that uses short ASCII strings to linearly describe molecular structures. It uses element symbols and key characters to represent atomic-bond connections, parentheses to indicate branches, numbers to represent ring closures, and square brackets to explicitly represent atomic information, including atomic charge, number of hydrogen atoms, isotopes, chirality, and connection site markers. It can also annotate molecular structural elements such as aromaticity, ring closures, stereochemistry, and isotopes, facilitating indexing and deduplication in large-scale data processing. In other embodiments of this application, SELFIES (SELF-referencing embedded strings), molecular diagrams, or other standardized representations capable of characterizing molecular structures may also be used.

[0061] Given the extreme scarcity of publicly available samples of pyroptosis-inducible photosensitizers, in order to expand the chemical space of similar photosensitizers, structures corresponding to typical chromophore families in traditional photosensitizer research were added and all entered in SMILES format. At the same time, some small organic molecules were selected from standard quantum chemistry datasets such as QM9 to enrich the electronic structure and configuration distribution, and only SMILES expression and source tags were retained to ensure that SMILES is used as the unique molecular identifier and storage field in the molecular library.

[0062] Given that there are often multiple equivalent SMILES notations for the same molecule, after collection, a standardized SMILES expression is generated using RDKit (RationalDiscovery Kit, an open-source cheminformatics toolkit) and deduplicated accordingly, thereby ensuring the uniqueness of molecular records within the molecular library.

[0063] Step S12. Data Filtering

[0064] To reduce the interference of outliers on the quality of the molecular library, after completing molecule collection and SMILES normalization, a hierarchical cleaning process was performed based on elemental composition, formal charge, and relative molecular mass.

[0065] First, the elemental distribution within the molecular library is statistically analyzed, and molecules containing elements with extremely rare occurrences are removed using a preset element set. Next, the formal charge distribution is statistically analyzed, and molecules with abnormally high charge values ​​are removed using a set charge threshold. Finally, based on the overall molecular mass distribution and a preset typical photosensitizer mass range, molecules that significantly deviate from the range are deleted, resulting in a molecular library with photosensitizer molecule characteristics. .

[0066] The above filtering can limit the search space to a chemical space that matches the characteristics of the photosensitizer, and effectively reduce noise and inconsistent molecules, thereby improving the robustness and comparability of subsequent training and evaluation.

[0067] Step S2. Construct the skeleton-decoration dataset

[0068] Step S21. Segmentation Rules and Template Definition

[0069] Based on the number of connection sites in the target nucleus, the desired number of segmentations is defined as follows: (For example This indicates the desired skeleton with four connection sites. Eleven types of cleavable bonds (amide, amine, ester, urea, ether, alkene, quaternary ammonium, aromatic / aliphatic-nitrogenous, lactam-related, aromatic-aromatic, sulfonamide) defined by RECAP (Retrosynthetic Combinatorial Analysis Procedure) were selected as the cleavage rules. Each cleavage type was then programmed into SMARTS (SMILES Arbitrary Target Specification) for subsequent procedural matching and cleavage.

[0070] Step S22. Substructure matching and Key combination enumeration

[0071] molecular library Each molecule in (Read and normalize using SMILES, footer) Representative molecular library The molecule's index, i.e., the first molecule... Template matching is performed on each molecule: one by one, using the 11 classes of SMARTS templates defined by RECAP in the molecular library. We perform a substructure search and record the set of breakable bonds that satisfy the environmental constraints; based on this, we enumerate all possible... The key can be broken and the loop constraint is set to avoid destroying the original loop structure.

[0072] Step S23. Determining Key Breakage and Skeleton / Decoration

[0073] For each breakable bond combination, a bond-breaking operation is performed to obtain a corresponding set of segmented fragments. For the same molecule, different breakable bond combinations may produce different segmentation results. If a certain segmentation result contains exactly... A segment with a connection point is then marked as the skeleton of the segmentation result. (footprint) Represents the sequence number of the segment, i.e., the first segment. (1 segment), the remaining segments constitute the corresponding decoration set. If the segmentation result does not contain exactly... If a segment has no connection sites, the segmentation result is discarded. The above decision-making strategy ensures that in each skeleton-decoration sample, the number of connection sites in the skeleton corresponds to the number of decoration segments, which facilitates subsequent homonuclear-decoration generation.

[0074] Step S24. Fragment Filtering

[0075] To control the quality of segmented data and downstream usability, filtering rules are set for the skeleton and decoration, including: upper and lower limits for the number of rings. and the upper limit of the number of rings Lower limit of molecular weight and upper limit of molecular weight If the skeleton or any decoration does not meet the corresponding threshold, the entire skeleton-decoration sample corresponding to the broken bond combination is discarded, that is, the entire set of segmented fragments obtained by this broken bond operation is discarded. In addition, the method implements further screening restrictions such as element set, formal charge, and aromaticity ratio.

[0076] Step S25. Data Recording and Storage

[0077] For the skeleton that passed the verification and the corresponding decoration set Constituting data pairs Using SMILES as the unique identifier and storage field: a normalized SMILES representation is generated for each segment, and duplicates are removed accordingly. For each data pair... At the connection point, use "[*]" as a placeholder and decorate with "|" to separate, according to The data is stored in the form of a skeleton-decoration dataset, which contains a corresponding sample and all samples are combined to form the skeleton-decoration dataset. The order in which the connection sites in the SMILES expression are written from left to right corresponds to the decorator sequence. One-to-one correspondence.

[0078] Step S3. Obtaining the prior generative model

[0079] This step obtains a prior generative model through autoregressive maximum likelihood pre-training on the large-scale skeleton-decoration dataset constructed in step S2. The training samples are standardized and data augmentation is performed using random SMILES to improve grammatical coverage and robustness. This prior generative model is then used as initialization for subsequent core orientation and reinforcement learning optimization stages. This step specifically includes:

[0080] Step S31. Prior Generative Model Construction

[0081] To achieve decoration generation using a validated master kernel as a skeleton, the prior generation model employs an encoder-decoder conditional sequence generation framework with global attention. The encoder is a bidirectional RNN (containing an embedding layer and 3 LSTM layers, with a hidden state dimension of 512), which sums the forward and backward hidden states in the same dimension and passes the encoded representation to the decoder. The decoder is a unidirectional RNN (also containing an embedding layer and 3 LSTM layers, with a hidden state dimension of 512), which generates decoration sequences step-by-step based on tokens. The attention mechanism matches and weights the decoder's current state with the encoder's output at each decoding step; this global attention mechanism allows the decoder to focus on the region in the input skeleton most relevant to the current decoration at each decoding step. This can be represented as:

[0082]

[0083]

[0084]

[0085] in, Indicates the first The distribution of attention weights for each position in the source sequence in each decoding step; The feature dimension representing the encoder output state and the decoder hidden state; The output matrix of the encoder at each time step serves as the key or value in the attention calculation; For the first The decoder's hidden state in each decoding step acts as a query in the attention mechanism. AND key The correlation score between them is the dot product of their values. The expression is used to measure the degree of matching between the current decoding state and the encoded representation of each position in the input kernel skeleton; the softmax function is currently widely used to obtain normalized attention weights based on the relevance score between the query and the key. This represents the context vector obtained by weighting and converging the encoder output states according to the attention weights; This represents the fusion function, used to fuse the hidden states of the decoder. With context vector Obtain the decoding state after attention enhancement. And used in the output layer to predict the next molecular structure representation label, in a specific embodiment of this application, the fusion function The expression is: ,in This means concatenating the decoder's hidden state with the context vector. This represents the trainable linear transformation parameters.

[0086] Step S32. Lexicon Construction and Word Segmentation

[0087] This embodiment summarizes the molecular structure representation tags of all skeleton fragments and all decoration fragments in the skeleton-decoration dataset obtained in step S25, and constructs a skeleton-side global vocabulary and a decoration-side global vocabulary for model training based on these tags. The vocabulary construction adopts a recursive multi-character-first strategy: first, multi-character tokens are extracted as a whole, and then the remaining content is split into tokens based on single characters to obtain stable word segmentation results consistent with SMILES syntax. Specifically, this includes:

[0088] Step S321. Extracting multi-character tokens

[0089] When segmenting the SMILES strings of each sample in the skeleton-decoration dataset obtained in step S25, first scan the atomic description fragments enclosed in square brackets in the string, such as [nH], [O-], [*], etc., and treat each complete square bracket atomic description as a single token. Square bracket atoms in the SMILES syntax are used to explicitly give information such as elements, isotopes, chirality, hydrogen count, and charge, and can also be used to represent connection site markers. Therefore, they need to be preserved as a whole to avoid fragmenting the molecular semantics after splitting them.

[0090] When the ring closure number exceeds 9, SMILES gives a two-digit ring number in the form of "%nn" (such as %10, %24); such a notation is treated as a single token.

[0091] In the organic subset of SMILES, only Cl and Br are two-character elements that do not require square brackets, so Cl and Br are treated as independent tokens.

[0092] Step S322. Extracting the single-character token

[0093] After completing the three types of multi-character priority matching, the remaining tokens (such as C, N, O, =, #, (,), 1…9 and aromatic lowercase letters c, n, etc.) are segmented into tokens by single character, and the skeleton-side global vocabulary and the decoration-side global vocabulary are constructed accordingly for model training.

[0094] Step S33. Dividing the first training set and the first validation set

[0095] To objectively evaluate the generalization ability of the prior generative model in generating decorations on unseen skeletons, this embodiment uses a skeleton matching method to partition the dataset. Specifically, the dataset is first partitioned by the number of connection points. The skeleton-decoration dataset obtained in step S25 is divided into buckets; one (or several) seed skeletons are randomly selected in each bucket, and then the Bemis-Murcko (BM) skeleton of each molecule in the skeleton-decoration dataset is calculated using RDKit. All samples that share the same Bemis-Murcko skeleton with any seed skeleton are assigned to the first validation set, and the remaining samples are used as the first training set.

[0096] Step S34. Data Augmentation

[0097] To avoid overfitting and to encourage the prior generative model to learn the SMILES grammar and the skeleton-decoration correspondence, thereby improving the model's coverage and robustness against molecular grammar, equivalent structure expressions, and connection site constraints, this embodiment randomizes the skeleton and decoration SMILES descriptions of the samples in each training epoch. Specifically, randomized SMILES expressions are generated for each sample in the first training set and the first validation set, while maintaining the one-to-one mapping of connection sites to decorations, thus ensuring the skeleton... The strings are numbered from left to right Each connection site corresponds to a decoration sequence. .

[0098] Step S35. Prior generative model training

[0099] For the goal of decorative generation using a validated parent kernel as a framework, the prior generation model uses a given set of kernels containing... The target nucleus of each connection site As a condition, generate corresponding connections one by one according to the order of appearance of the connection sites in the skeleton SMILES. Segment decoration sequence. Generative modeling as a conditional language model: when the parent skeleton... At that time, the first Decoration sequences at each site The conditional probability is written as the product of each token, and the mother core skeleton is... Overall goals Set to:

[0100]

[0101] in, These are the parameters for the prior generation model. This represents the random variable corresponding to the parent core skeleton. This represents the specific structural representation of the currently given target core skeleton; For the first The site decoration at the first A random variable of the step. for The value of , For the first The decorative sequence corresponding to the site is in the . The molecular structure of a step represents a labeled random variable. for The value of , For decoration sequence Length, This represents the conditional probability given by the prior generative model.

[0102] Since SMILES are in text sequence form, this embodiment preferably employs an autoregressive maximum likelihood training method. Specifically, during the prior generative model training phase, the real decoration sequences in the skeleton-decoration dataset are used as supervision signals to minimize the negative log-likelihood. This is to obtain the prior generation distribution of legitimate decorative sequences under the constraints of a fixed parent core backbone and connection sites, and to provide a stable, chemically grammatically correct initialization strategy for subsequent reinforcement learning updates based on reward signals. Among these, the negative log-likelihood... The expression is as follows:

[0103]

[0104] The prior generative model is obtained through the above training. .

[0105] Step S4. Obtaining the molecular property prediction model

[0106] Key excited-state properties of photosensitizers (such as the singlet-triplet band gap) directly affect intersystem crossing efficiency and reactive oxygen species generation, and are important criteria for evaluating molecules as pyrogastrophoretic photosensitizer candidates. Although density functional theory (DFT), a relatively mature quantum chemical calculation method, provides accurate calculations, its high computational cost makes it difficult to support real-time scoring in reinforcement learning loops. To meet the real-time feedback requirements of reinforcement learning for candidate molecules, this embodiment extracts a representative subset from the constructed molecular library and obtains key quantum chemical properties such as the singlet-triplet band gap as supervisory signals to train a molecular property prediction model. Step S4 specifically includes:

[0107] Step S41. Construction of the SMILES expression molecular map

[0108] The molecular SMILES obtained in step S1 were parsed into molecular objects using RDKit, generating conformations and extracting geometric information. Subsequently, a molecular graph was constructed using the open-source software framework DGL (Deep Graph Library). Atoms were used as nodes, covalent bonds as edges, and... The distance threshold between atoms is used to truncate and supplement non-covalent interaction edges to characterize short-range interactions. Node features are described using 92-dimensional atomic features (atomic number, radius, electronegativity, etc.), while edge features are multi-dimensional edge features combining 21-dimensional chemical / geometric information (including bond type, bond orientation, whether it is in a loop, bond length, and distance function, etc.).

[0109] Step S42. DFT Annotation and Data Coverage

[0110] To obtain high-quality supervision signals required for training molecular property prediction models, a complete molecular library is constructed in step S1. Approximately 10,000 representative molecules were extracted for DFT calculations, and labels related to quantum chemical properties (such as optical parameters like singlet-triplet band gap) were output. Simultaneously, to improve the extrapolation capability of the molecular property prediction model for subsequent reinforcement learning results, full coverage calculations were performed on all target core-derived series; that is, all compounds corresponding to the target core were DFT-labeled. All the molecules with DFT labels constituted a second dataset, serving as the training and supervision signal data source for the subsequent molecular property prediction model.

[0111] Step S43. Construction of Molecular Property Prediction Model

[0112] Step S431. Construction of KAN layer

[0113] This embodiment adopts the KA-GCN architecture: replacing the MLP (Multilayer Perceptron) used for feature transformation in the traditional GCN (Graph Convolutional Network) with the Kolmogorov-Arnold Network (KAN) layer based on Fourier basis, so that the nonlinearity is changed from the fixed activation on the node to the learnable univariate function on the edge, thereby improving the expressiveness of the molecular property prediction model.

[0114] Let the molecular property prediction model be the first... The input of the layer is Then the first The first in the layer The output of each unit is :

[0115]

[0116]

[0117] in, For the molecular property prediction model The dimension of the layer; For the molecular property prediction model The first layer One input component; index Corresponding to the molecular property prediction model, the first The output unit of the layer; For the action in the A learnable univariate function over each input component, in this embodiment, is a finite-term sine / cosine harmonic (order ). Expand; , These are the trainable coefficients for cosine and sine harmonics. The above form reflects the structure of KAN, where the edges are functions and the points are summed: each input component is first transformed by a univariate function, and then summed at the output unit to obtain the result mapped to the next layer.

[0118] Step S432. Replace the MLP layer of GCN with a KAN layer.

[0119] In the complete KA-GCN architecture, the initialization, message passing and readout of the input molecular graph are all replaced by KAN layers instead of the traditional GCN MLP, achieving unified KAN-based modeling.

[0120] Step S433. Output Head Design and Prediction Target

[0121] In this embodiment, the predicted quantum chemical target includes the singlet-triple bandgap. Absorption wavelength With oscillator strength .in, and Directly corresponds to the first excited singlet state Absorption transition (output single value) ).definition For the first excited singlet state and An excited triplet state The minimum energy difference is used to characterize the proximity of the most likely singlet-triple level:

[0122]

[0123] in, For the first excited singlet state energy, For the first Excitement of triplet state Energy; In this embodiment, the number of excited triplet states involved in the calculation is... ;

[0124] Accordingly, the molecular property prediction model adopts a multi-task output structure, simultaneously outputting multiple quantum chemical property predictions, including singlet state energy levels. , triplet energy levels Minimum singlet-triplet bandgap and the absorption wavelength corresponding to the first excited singlet state. and oscillator strength .

[0125] Step S44. Splitting the second training set and the second validation set.

[0126] To objectively evaluate the predictive ability of the molecular property prediction model on unseen molecules, this implementation first uses RDKit to calculate the Bemis-Murcko skeleton of each molecule in the second dataset obtained in step S42, and then constructs a second training set and a second validation set at a ratio of 9:1. Specifically, several skeletons are selected sequentially in the skeleton space, and all molecules with the same skeletons are included in the second validation set at once until its size reaches 10% of the total data; the remaining samples are used as the second training set.

[0127] Step S45. Loss Function Design

[0128] In this embodiment, the L1 loss function (mean absolute error, MAE) is used as the baseline error metric for each objective, and a physical consistency constraint is added for explicit encoding. The calculation relationship is as follows. MAE can be implemented using PyTorch L1Loss. The total loss function of the design is then as follows:

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] in, This is the total loss function; The loss for predicting the energy level and the minimum singlet-triplet bandgap is used to constrain the prediction of the first excited singlet state energy level. Each triplet energy and the minimum singlet-triplet bandgap Close to the corresponding labeled value; For consistency loss function; The absorption wavelength loss function for the first excited singlet state; Let be the oscillator strength loss function for the first excited singlet state; This represents the predicted value obtained from the molecular property prediction model; They are respectively The corresponding weighting coefficients are used to adjust the impact of each item on the total loss.

[0135] Step S5. Generation of nucleus-directed pyrode photosensitizer molecules based on molecularly generated intelligent agents.

[0136] This step initializes a priori generative model to construct a molecular generative agent. During the reinforcement learning loop, molecules are sampled in batches according to core and site constraints. For each molecule, a molecular property prediction model evaluates key molecular molecular chemistry properties (such as singlet-triplet band gap) to replace costly DFT calculations. RDKit is used to calculate molecular structural information such as charge and symmetry, forming a weighted multi-objective reward as an update signal. Furthermore, to prevent pattern collapse and repetitive generation, a diversity filtering and memory mechanism is introduced: molecules with scores exceeding a threshold are deduplicated before being added to the library, continuously guiding the agent to explore high-quality and diverse chemical spaces, ultimately obtaining a candidate photosensitizer molecule library for pyroptosis induction. This step specifically includes:

[0137] Step S51. Initialization of the molecular-generating intelligent agent

[0138] To enable the molecular generative agent to possess SMILES syntax and skeleton-decoration mapping capabilities from the initial stage, the policy network of the molecular generative agent is initialized using the parameters of the pre-trained generative model, while maintaining the same word segmentation / vocabulary and positional encoding settings as in the prior stage. This ensures that the molecular generative agent adheres to the SMILES syntax and the core-decoration mapping relationship during the optimization process, thereby updating and searching the decoration space under the condition that the core structure and connection site constraints remain unchanged.

[0139] Step S52. Design of the scoring function

[0140] Reinforcement learning evaluates candidate molecules using a normalized total score. This normalized total score consists of three parts: quantum chemical property score, symmetry property score, and formal charge property score. For the total score of a candidate molecule, the weights of the quantum chemical property score, symmetry property score, and formal charge property score are respectively... , and ,in ,and Then the total score of the candidate molecules Represented as:

[0141]

[0142] in, Indicates the quantum chemical score, The score represents the symmetry sub-score. Representation of formal charge quantum score; in this embodiment, , , .

[0143] Since each sub-score is located in The interval is defined, and all the weights mentioned above are non-negative and summed to 1 after normalization. Therefore, the total score of the candidate molecules is also within the interval. Interval.

[0144] The design of each sub-score is described below.

[0145] Step S521. Quantum chemical quantum score design

[0146] The quantum chemical sub-score in this embodiment covers , and Among them, the small single-triple bandgap Facilitates intersystem crossing and increases triplet / singlet oxygen production; absorption wavelength It should be located in the 600–800 nm window; oscillator strength Used to characterize the strength of the allowed transitions of an electric dipole, reflecting its ability to be effectively excited.

[0147] To facilitate stable and adjustable use in reinforcement learning, this embodiment uniformly uses the distance to the target as normalization. Within the specified interval or when a threshold is reached, the score is 1; outside the interval, the score monotonically decreases with distance and is cropped to... . Specifically:

[0148] For quantum chemical quantum score Let the weights of the bandgap sub-score, the absorption wavelength sub-score, and the oscillator intensity sub-score be respectively... , and ,in ,and The quantum chemical quantum score is then expressed as:

[0149]

[0150] in:

[0151] Set bandgap threshold Oscillator strength threshold And define the bandgap deviation. Insufficient strength of the oscillator :

[0152]

[0153]

[0154] The corresponding normalizer score is the gap score. Oscillator strength quantum score ,when or Take at different times , Otherwise, calculate according to the following formula. , :

[0155]

[0156]

[0157] in, , , are the decay scale parameters of the gap quantum score and the oscillator intensity quantum score, respectively, used to control the rate at which the score of the candidate molecule decreases with the degree of deviation when the corresponding threshold condition is not met; and The value of can be preset according to specific task requirements and data distribution. In this embodiment, both are set to 1.

[0158] For the absorption wavelength, first define the absorption wavelength to the setting window. The distance is :

[0159]

[0160] And normalized to the absorption wavelength sub-score When hour ,when When located outside the set window, As the deviation decreases smoothly, it can be calculated using the following formula. :

[0161]

[0162] in, The slope scale of the absorption wavelength sub-score is used to control the attenuation rate outside the threshold.

[0163] This embodiment focuses on quantum chemical quantum scores. The internal weights are set as follows: , , .

[0164] Step S522. Symmetry score design

[0165] This embodiment measures molecular symmetry using decoration equivalence and assigns scores accordingly: first, Morgan fingerprints are generated for each decoration of each molecule generated by the molecular generation agent using RDKit; then, pairwise Tanimoto similarity coefficients are calculated between decorations; a symmetry sub-score is recorded when all decorations are identical, or when two pairs of decorations are identical in the case of four-connection sites (e.g., identical pairs of sites 1–3 and 2–4). The score is 1; if the four-connection sites only have pairwise identical combinations and do not satisfy the symmetry requirement (e.g., sites 1–2 and 3–4 are identical in pairs), then the symmetry sub-score is recorded. The score is 0.5; for other cases, the symmetry sub-score is... It is 0.

[0166] Step S523. Formal charge quantum score design

[0167] Formal charge affects the physicochemical properties and targeting behavior of molecules. This example uses RDKit to calculate the total formal charge of a molecule. Then assign a fraction based on the parity of its absolute value: if Then the formal charge quantum score Set to 0; if And if it is an even number, then the formal charge quantum score is... Set it to 1; if And if it is an odd number, then the formal charge quantum score is... Set it to 0.5.

[0168] Step S524. Diversity Filtering Mechanism

[0169] With the target nucleus and connection sites fixed, the chemical space is significantly shrunk, making reinforcement learning prone to pattern collapse and resampling. To address this, this embodiment introduces a memory-based diversity filtering approach: maintaining a molecular memory bank linked by normalized SMILES expressions and setting a total score threshold. After each generation step, the SMILES expression of the candidate molecules is normalized using RDKit and compared with the molecular memory database; if the same molecule already exists, the total score of the candidate molecule is set to zero; if it does not yet exist and the total score of the candidate molecule is zero, the process is repeated. If so, the candidate molecule is written into the memory bank.

[0170] Step S53. Training of the molecular generative agent

[0171] In this embodiment, reinforcement learning training follows a cyclical step. Iteration is carried out. In the [number]th... In each reinforcement learning step, using the target nucleus as a framework, the molecular generation agent generates 128 candidate molecules in batches according to the current generation strategy, constrained by the target nucleus and connection sites; for each candidate molecule... Using the total score function of the aforementioned design Obtain candidate molecules Total Score And record the current strategy of the molecular generative agent. log-likelihood To prevent the current strategy from deviating from the prior and forgetting the SMILES syntax and skeleton-decoration relationship, an augmented log-likelihood is constructed:

[0172]

[0173] in, Indicates the augmented log-likelihood; Indicates candidate molecules Log-likelihood under the prior generative model These are the parameters for the prior generation model; The scaling factor for the augmentation term is used to adjust the weighting of the total score of candidate molecules on the log-likelihood. In this embodiment, it is taken as... .

[0174] Define the DAP (Difference between Augmented and Posterior) strategy difference term as follows:

[0175]

[0176] The optimization objective is to use its square as the target:

[0177]

[0178] During reinforcement learning training, the parameters of the molecular generative agent are updated. The parameters of the prior generative model and the molecular property prediction model are kept frozen. The prior generative model provides prior log-likelihood constraints, while the molecular property prediction model provides molecular property estimates and reward calculations for candidate molecules. This is achieved by minimizing... The strategy of the molecular generation agent gradually shifts towards high-scoring molecules, while prior constraints are used to avoid forgetting the SMILES grammar and skeleton-decoration mapping. Throughout the reinforcement learning process, a diversity filtering and memory mechanism are combined to remove duplicates from the normalized SMILES: molecules with scores exceeding a threshold and not previously appearing are written into the memory as candidate pyroptosis-inducible photosensitizers. (Completed the first...) After the update, Then proceed to the next round of reinforcement learning iterations until the preset maximum number of iterations is reached. .

[0179] It is understood that the reinforcement learning-based nucleus-directed pyroptosis-induced photosensitizer molecule generation method of this embodiment achieves automation and intelligence: the molecule generation agent automatically samples, generates, scores, and screens candidate molecules, adapts to different input nuclei, reduces the cost of manual trial and error, and improves the efficiency of candidate molecule discovery. Furthermore, under the same target nucleus and the same high-score threshold, the reinforcement learning generation strategy of this embodiment is compared with the random sampling strategy, and the cumulative number of candidate molecules exceeding the high-score threshold is counted as the number of sampling steps increases. The results show that within 500 sampling steps, the cumulative number of high-scoring candidate molecules obtained by the reinforcement learning strategy in this embodiment is 127, significantly higher than the 40 obtained by the random sampling strategy; simultaneously, when reaching the top 5, 10, and 25 high-scoring candidate molecules, the number of sampling steps required by the reinforcement learning strategy is 78, 96, and 127, respectively, while the random sampling strategy requires 88, 159, and 327 steps, respectively. The above results demonstrate that this embodiment can discover candidate molecules that meet the target score criteria more quickly, outperforming random sampling methods in terms of sample utilization efficiency and the speed of discovering high-scoring candidate molecules. This invention significantly improves screening efficiency and reduces experimental and computational costs while ensuring structural controllability and property interpretability. It possesses good versatility and transferability, and can be widely applied to the intelligent design and optimization of core-directed photosensitizers and other functional molecules.

[0180] A second aspect of the present invention provides a reinforcement learning-driven nucleus-directed pyrolysis photosensitizer molecule generation device, comprising:

[0181] The molecular library construction module is configured to collect existing photosensitizer donors or acceptors and expand them through fragment combination to construct a molecular library with photosensitizer molecule characteristics.

[0182] The skeleton-decoration dataset construction module is configured to disassemble each molecule in the molecular library into skeleton-decoration pairs according to chemical fragmentation rules and label the connection sites to construct the skeleton-decoration dataset.

[0183] The prior generation model acquisition module is configured to pre-train the prior generation model based on the skeleton-decoration dataset, so that the prior generation model can master the molecular structure representation rules and have the conditional generation ability from the parent skeleton to the decorative fragment;

[0184] The molecular property prediction model acquisition module is configured to extract a representative subset from the molecular library and obtain the target quantum chemical properties of each molecule in the representative subset as a supervision signal to pre-train the molecular property prediction model.

[0185] The generation module is configured to construct a molecular generation agent using the prior generation model as an initialization strategy. During the reinforcement learning loop, the molecular generation agent, under the constraints of a given parent core skeleton and connection sites, generates candidate decorative fragments in batches according to the current generation strategy and combines them with the parent core skeleton to form candidate molecules. The quantum chemical properties of each candidate molecule are evaluated by the molecular property prediction model, and combined with multiple molecular structural information of the candidate molecules, a weighted multi-objective reward is formed as a generation strategy update signal. Candidate molecules with a weighted multi-objective reward not lower than a preset threshold and which are not repeated after structural standardization are added to the memory bank. The memory bank obtained at the end of reinforcement learning is used as a candidate photosensitizer molecule library for pyroptosis induction.

[0186] It should be noted that the foregoing explanation of the embodiment of a reinforcement learning-driven method for generating photosensitizer molecules with directed pyrolysis of the parent nucleus also applies to the reinforcement learning-driven device for generating photosensitizer molecules with directed pyrolysis of the parent nucleus in this embodiment, and will not be repeated here.

[0187] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0188] Although embodiments of this disclosure have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this disclosure, the scope of which is defined by the claims and their equivalents.

Claims

1. A reinforcement learning-driven method for generating nucleus-directed pyrode photosensitizer molecules, characterized in that, include: Existing photosensitizer donors or acceptors are collected and expanded through fragment combination to construct a molecular library with the characteristics of photosensitizer molecules; According to the chemical fragmentation rules, each molecule in the molecular library is disassembled into skeleton-decoration pairs, and the connection sites are labeled to construct a skeleton-decoration dataset. The prior generation model is pre-trained based on the skeleton-decoration dataset, enabling the prior generation model to master the molecular structure representation rules and have the conditional generation ability from the parent skeleton to decorative fragments. A representative subset is extracted from the molecular library, and the target quantum chemical properties of each molecule in the representative subset are obtained as supervision signals to pre-train the molecular property prediction model. Using the aforementioned prior generative model as an initialization strategy, a molecular generative agent is constructed. In the reinforcement learning loop, the molecular generating agent, under the constraints of a given parent core skeleton and connection sites, generates candidate decorative fragments in batches according to the current generation strategy, and combines them with the parent core skeleton to form candidate molecules. The quantum chemical properties of each candidate molecule are evaluated by the molecular property prediction model, and combined with the multiple molecular structural information of the candidate molecules, a weighted multi-objective reward is formed as a generation strategy update signal. Candidate molecules whose weighted multi-objective reward is not lower than a preset threshold and are not repeated after structural standardization are added to the memory bank. The memory bank obtained at the end of reinforcement learning is used as a candidate photosensitizer molecule library for pyroptosis induction.

2. The generation method according to claim 1, characterized in that, The construction of the skeleton-decoration dataset includes: Predefined chemical rules containing multiple functional group fracture modes are used to identify breakable bonds. Combinations of breakable bonds that meet preset quantity and constraints are enumerated, and fracture is performed to obtain multiple fragments. Among the fragments, the fragments that meet the target number of connection sites are selected as the skeleton, and the remaining fragments are used as the decoration set. Segmentation results that do not meet the constraints of the number of connection sites, ring structure integrity, or preset molecular properties are discarded.

3. The generation method according to claim 1, characterized in that, The prior generation model employs an encoder-decoder sequence generation model conditioned on a pre-defined molecular structure representation format of the parent nucleus backbone, and generates decorative sequences by fusing parent nucleus backbone information through a global attention mechanism. The expression for the global attention mechanism is as follows: in, Indicates the first The distribution of attention weights for each position in the source sequence in each decoding step; The feature dimension representing the encoder output state and the decoder hidden state; The output matrix of the encoder at each time step serves as the key or value in the attention mechanism; For the first The decoder hidden state of each decoding step serves as a query in the attention mechanism; This represents the context vector obtained by weighting and converging the encoder output states according to the attention weights; This represents the fusion function, used to fuse the hidden states of the decoder. With context vector Obtain the decoding state after attention enhancement. It is used to predict the next molecular structure representation label in the output layer.

4. The generation method according to claim 1, characterized in that, The prior generative model is trained by maximizing the conditional likelihood of the decoration sequence given the parent skeleton; for those with The parental skeleton of each connection site is generated by the prior generation model according to the order of the connection sites. A decorative sequence, and the overall generation target under the condition of the mother core skeleton. Represented as: in, These are the parameters for the prior generation model. This represents the random variable corresponding to the parent core skeleton. This represents the specific structural representation of the currently given target core skeleton; For the first The site decoration at the first A random variable of the step. for The value of , For the first The decoration sequence corresponding to the site is in the . The molecular structure of a step represents a labeled random variable. for The value of , For decoration sequence Length, This represents the conditional probability given by the prior generative model; When training the prior generative model, the real decoration sequences in the skeleton-decoration dataset are used as supervision signals to minimize the negative log-likelihood: in, It is a negative log-likelihood.

5. The generation method according to claim 1, characterized in that, A representative subset is extracted from the molecular library, including: Several representative molecules are extracted from the molecular library. Each extracted molecule is parsed using a preset molecular structure representation format and then embedded in three-dimensional coordinates to transform it into a molecular graph. Atoms are used as nodes and covalent bonds as edges. Non-covalent adjacent edges are added when the distance between atoms does not exceed a preset inter-atomic distance threshold to represent short-range interactions. Each node is given a multi-dimensional atomic feature containing elemental and environmental information, and each edge is given a multi-dimensional edge feature containing bond type and geometric information to obtain the representative subset.

6. The generation method according to claim 1, characterized in that, The molecular property prediction model employs a graph convolution structure incorporating the Kolmogorov–Arnold mechanism. During training, the average absolute error loss is applied to the target, and a physical consistency constraint term is introduced. This physical consistency constraint term is used to ensure that the minimum singlet-triple gap directly predicted by the molecular property prediction model is consistent with the minimum singlet-triple gap calculated from the energy level differences between multiple singlet and triplet energy levels predicted by the molecular property prediction model.

7. The generation method according to claim 1, characterized in that, The weighted multi-objective reward adopts a normalized scalar reward, which consists of three parts: quantum chemical properties, symmetry, and formal charge, and each part is normalized to an interval. And then summed using weighted averages.

8. The generation method according to claim 1, characterized in that, The step of adding candidate molecules with weighted multi-objective rewards not lower than a preset threshold and without repetition after structural standardization to the memory bank includes: The candidate molecules generated by the molecular generating agent are queried in the memory. If the structure already exists, its weighted multi-objective reward is set to zero or reduced to a penalty value. If the molecule is new and its weighted multi-objective reward is not lower than a preset reward threshold, the candidate molecule is added to the memory.

9. The reinforcement learning-driven molecular generation method according to claim 1, characterized in that, The optimization objective of reinforcement learning employs an augmented posterior difference strategy based on augmented log-likelihood and prior experience preservation: Given a target parent skeleton, the molecular generating agent operates according to the current strategy. Sampling candidate molecules Record the candidate molecules Under the current strategy Log-likelihood Log-likelihood under the prior generative model and with reward scaling factor Constructing augmented log-likelihood : in, as candidate molecules Weighted multi-objective rewards; Define the difference term for augmented posterior difference strategies as: And minimize The square of the result is used to update the strategy parameters of the molecular generating agent.

10. A reinforcement learning-driven device for generating photosensitizer molecules with a directional pyrode death axis, characterized in that, include: The molecular library construction module is configured to collect existing photosensitizer donors or acceptors and expand them through fragment combination to construct a molecular library with photosensitizer molecule characteristics. The skeleton-decoration dataset construction module is configured to disassemble each molecule in the molecular library into skeleton-decoration pairs according to chemical fragmentation rules and label the connection sites to construct the skeleton-decoration dataset. The prior generation model acquisition module is configured to pre-train the prior generation model based on the skeleton-decoration dataset, so that the prior generation model can master the molecular structure representation rules and have the conditional generation ability from the parent skeleton to the decorative fragment; The molecular property prediction model acquisition module is configured to extract a representative subset from the molecular library and obtain the target quantum chemical properties of each molecule in the representative subset as a supervision signal to pre-train the molecular property prediction model. The generation module is configured to construct a molecular generation agent using the prior generation model as an initialization strategy. In the reinforcement learning loop, the molecular generating agent, under the constraints of a given parent core skeleton and connection sites, generates candidate decorative fragments in batches according to the current generation strategy, and combines them with the parent core skeleton to form candidate molecules. The quantum chemical properties of each candidate molecule are evaluated by the molecular property prediction model, and combined with the multiple molecular structural information of the candidate molecules, a weighted multi-objective reward is formed as a generation strategy update signal. Candidate molecules whose weighted multi-objective reward is not lower than a preset threshold and are not repeated after structural standardization are added to the memory bank. The memory bank obtained at the end of reinforcement learning is used as a candidate photosensitizer molecule library for pyroptosis induction.

Images