A chemical reaction-sensing multimodal synthetic molecule generation method

Abstract

This invention discloses a multimodal synthetic molecule generation method based on chemical reaction perception, relating to the field of drug design technology. The method includes the following steps: obtaining the SMILES sequence of an initial inactive molecule, generating atomic role information, and then capturing the two-dimensional topological structure through graph attention network encoding; simultaneously converting the SMILES sequence into a token sequence and generating word embeddings to form a hybrid representation that combines structural information and syntactic logic; converting the ID of the candidate template into a global feature vector, integrating it into the hybrid representation, and constructing a decoder context memory; then predicting and outputting the multiple reactant sequences with the highest probabilities through beam search to solve for multiple synthetic paths; evaluating each synthetic path, eliminating molecules with invalid structures or excessive molecular weights, and outputting the optimal synthetic path. This invention addresses the problem that existing sequence paradigms neglect rich two-dimensional molecular topological structures, making it difficult to simultaneously obtain candidate molecules with high synthetic success rates and structural diversity that meet drug-likeness requirements.

Description

Technical Field

[0001] This invention relates to the field of drug design technology, and in particular to a method for generating multimodal synthetic molecules with chemical reaction sensing. Background Technology

[0002] Drug discovery is a lengthy and costly endeavor, with the optimization phase from hit compounds to lead compounds being a key bottleneck. In this phase, researchers need to modify the chemical structure of the initial active molecules obtained from high-throughput screening to enhance their bioactivity against specific targets and simultaneously optimize pharmacokinetic properties. However, localized chemical exploration around hit compounds often encounters combinatorial explosion problems. Therefore, accurately identifying molecules that both enhance target interactions and possess desirable drug-like characteristics while meeting stringent drug-likeness criteria remains a core challenge in modern medicinal chemistry.

[0003] To accelerate this process, Artificial Intelligence-Assisted Drug Design (AIDD)—especially deep learning-based de novo design—has become a powerful paradigm for hit-to-leader optimization. Unlike traditional rule-driven modification methods, deep generative models can efficiently sample the local chemical space around the hit molecule in a data-driven manner using large-scale datasets. Representative works include: reinforcement learning frameworks based on graph neural networks (GNNs) that can utilize protein-protein interaction networks for prospective molecule design; and variational autoencoders (VAEs) that achieve multi-target optimization through continuous latent variable spaces and active learning.

[0004] Another approach combines Monte Carlo Tree Search (MCTS) with reinforcement learning to achieve reaction-aware molecular optimization through targeted exploration in a modular reaction space. However, most existing implementations still use one-dimensional SMILES sequences as state representations. This sequence-only paradigm ignores the rich two-dimensional molecular topologies and the functional roles of atoms in specific reaction templates, limiting the model's ability to characterize complex reaction logics. Consequently, it is difficult to simultaneously obtain candidate molecules with high synthetic success rates and structural diversity that meet drug-likeness requirements. Summary of the Invention

[0005] To address the aforementioned issues, particularly the lack of structure awareness in sequence generation models, this invention proposes a multimodal synthetic molecule generation method based on chemical reaction awareness. This method proposes a CRAFT (ChemicalReaction-Aware Forward Transformer) model, integrating a Graph-to-Sequence conditional generator into a reinforcement learning scheme driven by MCTS. It employs a Graph Attention Network (GAT) encoder to explicitly capture the two-dimensional topological features of molecules and designs an atomic-level role embedding mechanism to inject reaction template semantics into the representation of each atom. Furthermore, it introduces a global feature vector of the template to achieve a deep fusion of reaction conditions and molecular structure.

[0006] This invention provides a method for generating chemically reactive, multimodal, synthesizable molecules, comprising the following steps:

[0007] S10: Obtain the SMILES sequence of the initial inactive molecule, select multiple candidate templates with the highest fitness from the high-frequency reaction templates, establish the correspondence between the candidate templates and the atoms of the initial inactive molecule through pattern matching, and generate atom role information.

[0008] S20: Based on atomic role information, extract atomic physicochemical descriptors and chemical bond attributes, construct the node and edge feature matrix of the molecular graph, and then capture the two-dimensional topological structure through graph attention network encoding; at the same time, convert the SMILES sequence into a token sequence and generate word embeddings, and fuse graph features and sequence features to form a hybrid representation that combines structural information and syntactic logic;

[0009] S30: Transform the ID of the candidate template into a global feature vector, incorporate it into the hybrid representation, and build the decoder context memory library;

[0010] S40: Based on the decoder context memory, it predicts and outputs the multiple reactant sequences with the highest probabilities through beam search;

[0011] S50: Solving multiple synthetic pathways based on reactant sequences;

[0012] S60: Evaluate each synthetic route, eliminate molecules with invalid structures or excessive molecular weights, and output the optimal synthetic route.

[0013] In summary, the present invention has at least the following beneficial effects:

[0014] 1. This invention employs a multimodal encoder to fuse the two-dimensional topological structure of molecules with atomic-level reaction role information, enabling the CRAFT model to gradually construct molecules in a manner similar to that of a medicinal chemist, under constraints that conform to the logic of real reactions, thereby realizing a naturally synthesizable generation process at the mechanistic level.

[0015] 2. A key feature of this invention is the complete transparency of the generation process: the CRAFT model outputs a complete forward synthetic route, explicitly providing the reaction template and co-reactants, allowing medicinal chemists to directly assess experimental feasibility without complex retrosynthetic analysis. Regarding optimization effectiveness and structural diversity, this invention not only significantly improves the yield of highly active molecules but also effectively alleviates the structural homogenization problem common in sequence generators through graph-based topological reasoning, resulting in a candidate molecule library with higher internal diversity and stronger backbone novelty. In terms of generalization ability, the CRAFT model demonstrates strong out-of-distribution exploration capabilities: even in the zero-sample testing scenario of ZINC, this invention can still optimize inactive fragments into novel backbones with high predicted activity using general reaction rules, and the molecular docking results further support their potential physical binding ability.

[0016] 3. This invention focuses on constructing a multimodal generation framework around forward molecular generation and optimization. In view of the shortcomings of existing sequence paradigms that ignore the two-dimensional topological structure of molecules, it proposes a graph-sequence fusion multimodal coding modeling method to generate candidate reactants and synthetic paths under forward synthesis logic, so as to achieve molecular generation with synthetic feasibility, structural diversity and drug-like properties. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a molecular structure similarity analysis diagram targeting DRD2;

[0019] Figure 2 This is a structural similarity analysis diagram of the generated molecules targeting AKT1;

[0020] Figure 3 This is a structural similarity analysis diagram of the generated molecules targeting CXCR4;

[0021] Figure 4 It is a novel skeletal diagram with high integration potential;

[0022] Figure 5 These are attention heatmaps of the CRAFT and TRACER baselines on four representative response types;

[0023] Figure 6 This is a diagram showing the docking results using DRD2 as the ligand.

[0024] Figure 7 This is a diagram showing the docking results using AKT1 as a ligand.

[0025] Figure 8 This is a diagram showing the docking results using CXCR4 as the ligand. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention provides a method for generating chemically reactive, multimodal, synthesizable molecules, comprising the following steps:

[0028] I. Reaction Template Prediction and Atomic-Level Role Allocation

[0029] For a given SMILES sequence of reactant molecule M The predictor then uses the set of all known reaction templates. Output a probability distribution. Select the K templates with the highest probabilities to form the initial candidate set. :

[0030] ;

[0031] In the formula, Indicates the predictor parameters Given the SMILES sequence of the reactants Select template when The conditional probability; Indicates the set of reaction templates The K reaction templates with the highest probability.

[0032] For candidate templates Its SMARTS string contains a set of atomic mapping numbers, denoted as In the preprocessing stage, a lookup table is constructed. Number each mapping It is associated with a specific role ID. During reasoning, reactant M is represented as a molecular diagram. RDKit was used for substructure matching to attempt to find a mapping from template mapping numbers to atomic indices in the molecule. If a match is successful, then number each mapping. They all have corresponding atoms ,satisfy Templates that fail to match substructures will be removed from subsequent candidates.

[0033] It should be noted that this invention first searches the candidate template pool for templates that match... The reaction template is compatible with the chemical characteristics of the target material. Subsequently, atomic role vectors are constructed based on SMARTS substructure matching, explicitly encoding information such as reaction centers and atomic functional groups, thereby injecting template semantics into the atomic representation. The predictor uses the paper titled "Molecular optimization using a conditional transformer for reaction-aware compound exploration with reinforcement learning," published in *Communications Chemistry* in 2025 (article 40, Volume 8, Issue 1).

[0034] Based on the above matching results, for each atom in the molecule Assign a discrete role label to obtain an atomic role vector. The allocation rules are defined as follows:

[0035] ;

[0036] In the formula, Z represents the set of integers, and V represents the total number of atoms. This represents the i-th atom in the molecule. This represents the role vector of the i-th atom.

[0037] If atoms A mapping number was matched during substructure alignment. Then set its character tag to the corresponding character ID. Otherwise, treat the atom as a non-reactive atom and assign it a background role of 0. The resulting vector The functional role of each atom in the reaction is explicitly encoded.

[0038] II. Multimodal Graph Sequence Encoding for Role Perception Fusion

[0039] For each atom Extract its inherent physicochemical descriptors. and compare it with the atomic character tags obtained in the previous step. Fusion is performed. Discrete labels. Mapped to dense vectors through learnable embedding layers. Then, the two are concatenated and linearly projected to obtain the initial node representation. :

[0040] ;

[0041] in, This indicates that the encoder uniformly adopts a dimension of The hidden feature space, This represents a linear projection layer. This design enables the graph encoder to perceive the reactive functional role of each atom under a specific template from the outset. The feature vectors of all atoms are then stacked into a node feature matrix. .

[0042] For each chemical bond Extract its chemical properties (such as bond order, conjugation state, and whether it belongs to a ring structure) and construct edge feature vectors. .

[0043] To consider both the states of neighboring atoms and the chemical bond properties connecting them during node updates, the initial node feature matrix is... Edge index list With edge feature matrix The input is fed into a multi-layer graph attention network (GAT). Unlike the standard GAT, this variant explicitly incorporates edge features when calculating attention coefficients. The update rule for a layer can be written as:

[0044] ;

[0045] In this process, edge features By incorporating an attention mechanism, the CRAFT model can distinguish between different neighbors connected by single bonds, double bonds, or aromatic bonds, thereby generating a final graph representation containing rich structural information and semantic roles. .

[0046] It should be noted that the edge index list It records the start and end atom indices of each chemical bond, and all such index pairs are combined to form the edge feature matrix. It is formed by stacking the eigenvectors of all chemical bonds in sequence.

[0047] SMILES string of reactants Consider it as a token sequence Each token First, it is mapped to standard word embeddings. Since SMILES is essentially a linearized representation of the underlying molecular graph, there is an inherent correspondence between sequence tokens and graph nodes. To utilize structural information during sequence encoding, a token-to-atom alignment mapping is constructed. Used to record sequence positions With graph node index Correspondence:

[0048] ;

[0049] This mapping is determined during the preprocessing stage using RDKit's SMILES parsing function.

[0050] To achieve deep cross-modal fusion, graph-encoded features are injected into the sequence embedding. For sequence position... If a valid alignment exists Then from the matrix Extract the corresponding atomic feature vector from And merge it with the token embedding at that position; otherwise, only retain the sequence embedding:

[0051] ;

[0052] in, A mixed feature vector representing a single sequence position. express Zhongyu The row vector corresponding to the atom being pointed to. This represents a learnable linear projection used to map graph features to the sequence embedding space. The resulting hybrid sequence representation... While preserving the SMILES syntax structure, it enhances structure awareness. Finally, it will... Input a standard multilayer Transformer encoder to obtain the final encoder memory. .

[0053] It should be noted that its Transformer encoder uses the paper titled "Attention is all you need" published in Advances in Neural Information Processing Systems in 2017, which is located in Volume 30.

[0054] III. Template-aware feature fusion

[0055] Although previous modules have injected atomic-level reaction roles into local node features, the CRAFT model still requires a clear global conditional signal to guide the overall forward synthesis direction. Therefore, this invention upgrades discrete reaction template IDs into continuous global feature vectors and fuses them as prefix conditions into the encoder memory.

[0056] Given the input reaction template ID, it is first transformed into a dense vector through an embedding layer. To match the encoder's dimensional space, the vector undergoes linear projection, Dropout regularization, and layer normalization to generate the final template global feature vector. :

[0057] ;

[0058] In the formula, Indicates the normalization layer. Indicates a random deactivation layer. Let denote the linear projection matrix, and b denote the bias vector.

[0059] A prefix concatenation strategy is used to add this global feature vector as a special vector to the mixed sequence encoding output in the previous section. Previously, this operation built the context memory that would ultimately be used for decoding. :

[0060] ;

[0061] In the formula, Indicates the dimension of the hidden layer.

[0062] This design allows subsequent decoders to simultaneously focus on the global reaction intent (template features) and local structural details (graph-sequence hybrid features) through an attention mechanism when generating each token.

[0063] IV. Conditional Transformer Decoder

[0064] To convert the encoded chemical information into the corresponding reactant SMILES sequences, this invention deploys a standard autoregressive Transformer decoder. This module utilizes a cross-attention mechanism and multimodal contextual information to predict the token distribution of the target molecule.

[0065] The decoder is It consists of stacked Transformer blocks. At each time step... The decoder receives the generated subsequence. Its core computational process relies on multi-head cross-attention, a mechanism that links the decoder's queries with the encoder's memory. Interacting (as Key and Value):

[0066] ;

[0067] In the formula, This represents the multi-head cross-attention mechanism, where Q represents the query matrix, K represents the build matrix, and L represents the value matrix.

[0068] In this way, the CRAFT model can dynamically extract relevant information from the graph structure, atomic roles, and global template to predict the probability distribution of the next token. .

[0069] During the training phase, this invention aims to minimize the true target sequence. Negative log-likelihood loss:

[0070] ;

[0071] In the formula, Represents the real target sequence The Tth token, Represents the real target sequence The true prefix of all tokens before the T-th position. express The sequence length, This represents a probability distribution.

[0072] During the inference phase, to ensure the high validity and diversity of the generated reactants, this invention employs a bundle search strategy. The CRAFT model simultaneously maintains B optimal hypothetical paths and ultimately outputs the Top-K reactant sequences with the highest probabilities as single-step prediction results.

[0073] V. Describing the problem using Markov decision processes

[0074] This invention models the molecular optimization problem for a specific target as a Markov decision process (MDP), consisting of quaternions. definition:

[0075] state space :state This indicates the molecular structure generated in the current step, uniquely identified by its SMILES sequence. The root node of the search. The initial input is an inactive seed compound.

[0076] Action space :action This means applying a specific reaction template and combining it with a set of reactants. Due to the vastness of the chemical space, this invention does not presuppose a fixed action space, but instead uses a CRAFT model as a policy network to dynamically generate a high-probability action candidate set.

[0077] State transition State transitions are defined as deterministic chemical transformation functions. Given a precursor molecule and a reaction action, a product molecule is generated based on a selected template and reactants.

[0078] reward function :award It is a scalar feedback used to assess the activity potential of the generated molecule.

[0079] VI. Tree Search Algorithm

[0080] To efficiently find the optimal synthetic route within a vast chemical space, this invention employs the Monte Carlo Tree Search (MCTS) algorithm. Each node in the search tree... Represent a molecular state and store the number of accesses. and cumulative rewards A single iteration of MCTS consists of the following four consecutive steps:

[0081] Selection: From the root node Initially, the algorithm recursively selects the optimal child node until a scalable leaf node is reached. To balance the utilization of high-scoring paths with the exploration of under-visited regions, this invention employs the UCT (Upper Confidence Bound applied to Trees) strategy as the action value evaluation function for the tree. Specifically, for the current node... any child node Its UCB value is calculated as follows:

[0082] ;

[0083] In the formula, Represents child nodes The cumulative reward value, Represents child nodes Total number of visits, This represents the historical average return of the current node. (Constant) This is used to control the trade-off between exploration and exploitation. Specifically, it applies to nodes that have not yet been visited. Its UCB value is set to infinity to ensure that the node will be selected at least once.

[0084] Extension: When the selected leaf node When the expansion conditions are met, this invention calls the CRAFT model to generate new child nodes. Candidate reactant sequences are generated using Beam Search (B=10-50). After effective pruning based on RDKit's structural validity verification and molecular weight constraints (MW≤600), valid candidate molecules are instantiated as new child nodes and added to the search tree.

[0085] Simulation: To quickly assess the long-term potential of new expansion nodes, this invention employs a fast Rollout strategy. Starting from the current node, the CRAFT model serves as the simulation strategy, using a greedy decoding approach to continuously predict multiple reaction steps until a preset maximum depth is reached or no effective molecules can be generated. This strategy significantly reduces the computational overhead of the simulation phase while maintaining a certain level of accuracy.

[0086] Backtracking: After the simulation ends, the highest reward value obtained along the path is recorded. The statistics are propagated back to the current node and all its ancestor nodes. Node statistics are updated according to the following rules:

[0087] ;

[0088] ;

[0089] This maximum value backpropagation mechanism ensures that the value of a parent node reflects the potential best chemical outcome in its subtree.

[0090] VII. Reward Function Design

[0091] The reward function is designed to guide the MCTS search towards chemical regions that are target-specific and meet drug-like criteria. To ensure the comparability of the small molecule set generated in this study with existing baseline models and to guarantee the objectivity and fairness of the evaluation process, this invention directly adopts a pre-trained QSAR model as the activity evaluation engine.

[0092] It should be noted that the pre-trained QSAR model uses the paper titled "Molecular optimization using a conditional transformer for reaction-aware compound exploration with reinforcement learning" published in Communications Chemistry in 2025, specifically article 40 in Volume 8, Issue 1.

[0093] The activity prediction models for the three targets DRD2, AKT1, and CXCR4 were all implemented using random forest classifiers. Each classifier takes a 2048-bit ECFP6 molecular fingerprint as input and outputs the probability that a given molecule is identified as active. These models showed strong predictive performance on the DRD2 and AKT1 test sets. For each node in the search tree... The corresponding activity probability returned by the model It is used as a primary reward signal to guide MCTS to prioritize the exploration of high-potential synthetic pathways.

[0094] To prevent the formation of excessively large molecular weights that could lead to difficulties in synthesis and drug development, this invention applies the same hard constraints as the baseline model in the reward calculation. The final reward function... Defined as a combination of activity score and molecular weight limit:

[0095] ;

[0096] In the formula, This represents the activity probability score. Indicates chemical validity test, Indicates molecular weight. The molecular weight threshold is set with reference to drug-likeness standards. Any path that fails chemical testing or exceeds the molecular weight limit will be given a significant negative penalty reward value. This design ensures that while the search algorithm pursues high target binding affinity, it is inherently constrained by drug-likeness space and chemical effectiveness, thereby generating molecules that have both theoretical activity and practical value.

[0097] Experimental verification

[0098] To comprehensively evaluate the performance of CRAFT in reaction-sensing molecular generation and multi-step optimization, this invention compares it with five representative synthesis-guided generation models. These baseline methods cover different paradigms, ranging from latent space optimization to flow cytometry model generation:

[0099] TRACER is the most direct baseline in this comparative experiment. Similar to CRAFT, TRACER uses MCTS for path planning and employs a conditional Transformer as the policy network to predict reaction products. However, TRACER relies only on a one-dimensional SMILES sequence representation and does not explicitly encode molecular graph topology or atomic-level reaction role information. In the experiment, the same single-step Transformer inference setting as TRACER was used.

[0100] DoG-Gen is the only model in this comparative experiment that explicitly learns chemical reactions and supports multi-step transformations. This method utilizes reaction trees constructed from the USPTO dataset, encoding these trees through a two-stage gated graph neural network (GGNN) to learn the latent spatial mapping. In the benchmark experiments, 30 rounds of optimization were performed following the original settings: each round sampled 7,000 reaction trees from DoG-Gen, then molecules were sorted based on QSAR values, and the top 1,500 molecules were selected for two rounds of fine-tuning.

[0101] CasVAE, a method based on variational autoencoders (VAEs), constructs independent latent spaces for compound structures and reaction trees, and uses Bayesian optimization to search for high-scoring molecules within these latent spaces. In experiments, the model was set to generate five batches of molecules, 3,000 per batch.

[0102] In Molecule Chef, reactants are vectorized using a graph neural network (GNN), and these vectors are summed to form a bag of reactants. The model learns to map this bag of reactants to a latent space and connects it to a property predictor, which guides the optimization process by updating the latent coordinates via gradients.

[0103] SynFlowNet is a model based on GFlowNet that aims to sample samples with a probability proportional to the reward. It uses neural networks to predict available reaction templates and partners in bimolecular reactions and models the structure extensions of chemical reactions as Markov processes. The experimental settings are consistent with the original implementation, including a batch size of 64, a temperature parameter of 32, and 1,000 training steps.

[0104] To ensure fairness in the comparison, all baseline models were trained on the same preprocessed USPTO dataset and used the same QSAR model as the evaluation reward function.

[0105] To evaluate CRAFT's ability to understand chemical reaction logic, this invention first examines its single-step forward prediction performance on a test set. Key evaluation metrics include: Top-K accuracy (measuring whether the predicted results match the actual products) and backbone retention rate (measuring whether the generated molecules retain the core backbone of the reactants). The comparative results are summarized in Tables 1 and 2.

[0106] Table 1. Top n accuracy achieved by the model

[0107]

[0108] As shown in Table 2, the baseline model TRACER-unconditional, which does not use reaction templates, has a Top-1 accuracy of 26.6%, indicating that relying solely on reactant structures is insufficient to reliably infer specific chemical transformations when reaction condition information is lacking. Introducing reaction templates significantly improves performance: TRACER-conditional achieves a Top-1 accuracy of 55.1%, demonstrating that template information effectively constrains the reaction space. Furthermore, CRAFT of this invention outperforms TRACER-conditional in all reporting metrics, achieving a Top-1 accuracy of 56.3% and a Top-10 accuracy of 72.4%. These improvements are primarily attributed to the proposed multimodal encoder: it explicitly captures the two-dimensional topological structure of molecules through GAT and injects fine-grained reaction semantics through atomic-level role embedding, thereby more accurately locating reaction centers and generating chemically consistent product predictions.

[0109] Table 2. Retention capability of the model's core framework

[0110]

[0111] In hit-to-leader optimization, preserving the core skeleton of the starting compound while introducing new substituents is crucial. Therefore, this invention compares the Murcko skeleton retention rates of various models in Table 3. CRAFT achieved the highest retention rate of 96.4%, superior to TRACER-conditional's 94.5%. This result indicates that sequence-based generators are more prone to losing structural constraints when generating longer strings; while by introducing structural signals from the molecular graph, CRAFT can more stably preserve key topological features and reduce the incidence of structural inconsistencies.

[0112] In the MCTS-based molecular optimization framework, the beam width used in the expansion phase It is a key hyperparameter because it directly controls the breadth of the reaction space explored by the policy network at each decision step. To examine... Regarding the impact on the generated molecular weight and optimization efficiency, this invention conducted sensitivity analysis on three target sites. Set the value to The results were compared with the TRACER baseline under the same settings. The results are summarized in Tables 3, 4 and 5.

[0113] Table 3 Comparison of structural optimization results for DRD2 under different beam widths

[0114]

[0115] Table 3 Comparison of structural optimization results for DRD2 under different beam widths (continued)

[0116]

[0117] Table 4 Comparison of AKT1 structural optimization results under different beam widths

[0118]

[0119] Table 4 Comparison of AKT1 structural optimization results under different beam widths (continued)

[0120]

[0121] Table 5. Comparison of structural optimization results of CXCR4 under different beam widths

[0122]

[0123] Table 5 Comparison of structural optimization results for CXCR4 under different beam widths (continued)

[0124]

[0125] First, the molecules generated by CRAFT and TRACER are almost entirely novel molecules relative to the USPTO training set, with a uniqueness approaching 100%. This high novelty mainly stems from the multi-step nature of the MCTS framework: the search process iteratively introduces building blocks and continuously performs chemical transformations, enabling the construction of skeletons that are unlikely to appear in single-step reaction records, demonstrating the framework's ability to perform de novo design outside the coverage of existing databases.

[0126] Beyond evaluating the novelty of the relative training distribution, this invention further examines the structural heterogeneity of the discovered hit molecules to rule out potential structural homogenization issues. Specifically, the internal diversity is calculated for the generated candidate subsets with predicted QSAR activities greater than 0.5. This metric is defined as the average Tanimoto distance between pairwise molecules within the high-activity set. The calculation is as follows:

[0127] ;

[0128] in, This represents the cardinality of the deduplicated generation set that satisfies QSAR > 0.5; Represents compounds and The Tanimoto similarity between them was calculated using ECFP4 fingerprints; index Setting it to 1 is equivalent to taking the arithmetic mean of the pairwise similarities.

[0129] Although both are similar in terms of novelty, CRAFT consistently outperforms baseline methods in terms of internal diversity. For example, on the AKT1 target, CRAFT achieves an internal diversity of 0.768, while TRACER only reaches 0.564. As a sequence-only model, TRACER often compresses information such as reaction sites, reaction partners, and structural editing into a single reaction template token. Furthermore, its strong dependence on sequence likelihood during multi-step generation easily induces structural homogenization, causing the model to repeatedly generate relatively conserved molecules with highly similar structures. In contrast, CRAFT explicitly encodes molecular topology through graph neural networks and uses atomic role embedding to locate reaction centers. This chemical mechanism-based representation allows the policy network to explore regions with greater structural variation while maintaining reaction plausibility, thus effectively reducing homogenization generation.

[0130] CRAFT also demonstrated higher efficiency in generating highly active candidate molecules, reflecting the higher quality of action proposals distributed by its policy network. For example, in the DRD2 task, 23.9% of the molecules generated by CRAFT had QSAR scores exceeding 0.5, compared to only 17.0% for TRACER. As the beam width increased from 10 to 50, the total number of highly active molecules increased, but the ensemble diversity decreased slightly. This trend reflects the exploration-exploitation tradeoff: a larger beam provides more opportunities to exploit high-probability branches and accumulate hit molecules, but it may also cause the search to converge in local regions, thus reducing the structural coverage. Based on these observations, a beam width that maximizes the number of highly active molecules was selected for each target, and this configuration was adopted in the subsequent comprehensive evaluation.

[0131] To comprehensively evaluate the performance of CRAFT in multi-step molecular optimization, this invention employs a summary evaluation strategy. For each target's five starting molecules, the beam width that generates the highest proportion of active candidates is determined, and the corresponding outputs are merged into a unified target-specific molecular library. To rigorously quantify the diversity and distribution characteristics of the generated molecular library, internal uniqueness and Frechet ChemNet Distance (FCD) are further used as key indicators. Internal uniqueness is defined as the proportion of non-repeating canonical smiles in the generated set, directly reflecting the breadth of coverage of the optimization search. FCD is used to evaluate the distribution distance between the generated set and the training dataset in the ChemNet feature space. To ensure statistical reliability, FCD calculation is strictly limited to generated sets containing at least 5,000 unique compounds; sets below this threshold are not included in this analysis.

[0132] CRAFT was compared with five synthetic guidance baseline methods: TRACER, DoG-Gen, MoleculeChef, CasVAE, and SynFlowNet. Detailed results are shown in Table 6.

[0133] Table 6. Comparison of molecular generation models for each protein

[0134]

[0135] Table 6. Comparison of molecular generative models for each protein (continued)

[0136]

[0137] The results show significant differences in chemical space exploration among different paradigms. Regarding novelty relative to the USPTO training set, CRAFT and TRACER achieved near-perfect uniqueness, close to 100.0%, significantly higher than DoG-Gen (41.3%) and Molecule Chef (91.1%). This indicates that explicit simulation of multi-step reaction combinations allows CRAFT to extend beyond the training distribution and construct previously unseen frameworks; while DoG-Gen, due to sampling reaction trees in the latent space, exhibits a stronger tendency to memorize and more easily reproduces the training set structure. Furthermore, CRAFT maintains stable and superior FCD performance across all targets, suggesting that its generated molecules are not limited to narrow local searches around known chemistry, but rather achieve effective exploration of new regions beyond the training distribution.

[0138] In terms of active molecule generation efficiency, CRAFT demonstrates a balance between hit rate and structural diversity on targets with relatively abundant training data, such as DRD2 and AKT1. Taking DRD2 as an example, 23.9% of the molecules generated by CRAFT had predicted QSAR activities exceeding 0.5, significantly outperforming the main baselines TRACER (17.0%), Molecule Chef (0.90%), and SynFlowNet (0.09%). Although DoG-Gen also achieved an activity rate of 23.9%, and CasVAE even higher (32.7%), the diversity of the high-activity subsets for these two methods was significantly lower (0.425 for DoG-Gen and 0.526 for CasVAE). Furthermore, CasVAE only generated 615 molecules in total, limiting its practical usability. In contrast, CRAFT maintains a high hit rate while still achieving an internal diversity of 0.684 within the active candidate set. This advantage is even more pronounced on AKT1: CRAFT achieves an activity rate of 32.7% and an internal diversity of 0.768, outperforming all compared methods. For CXCR4, CRAFT's activity rate is 8.72%, slightly lower than TRACER (13.5%) and DoG-Gen (24.6%). This phenomenon is more likely to reflect a mismatch between the exploratory generator and the small-sample reward model, rather than a lack of generator capability itself. CXCR4's QSAR model is trained on only 853 active molecules, resulting in a narrow applicability domain. CRAFT tends to explore more novel chemical regions, exhibiting a higher FCD (17.0) and high diversity (0.792); such distributed exoskeletons are more easily given lower predictive activity by overfitted reward models. In contrast, more conservative generators like DoG-Gen tend to generate structures closer to the training distribution (FCD 4.42, diversity 0.603), making them more likely to fall within the comfort zone of QSAR predictors and thus achieve higher scores. It is important to emphasize that the molecular assemblies generated by CRAFT on CXCR4 still significantly outperform the structural diversity of all baselines. This indicates that even with limited prior data on targets, CRAFT can still maximize the chance of discovering new skeletons through a wider range of chemical space exploration, thereby potentially circumventing existing patent constraints and enhancing novelty.

[0139] Besides bioactivity and structural diversity, the balance between molecular weight and synthetic accessibility (SA score) is also a key factor in evaluating the translational potential of de novo design methods. Therefore, this invention compares the physicochemical properties and synthetic characteristics of CRAFT-generated molecules with various baseline models, and the summary statistics are shown in Table 7.

[0140] Table 7 Comparison of SA scores and molecular weight for different models

[0141]

[0142] Overall, across all targets, the average molecular weight of molecules generated by CRAFT was slightly higher than that of major baselines such as TRACER and DoG-Gen. This trend suggests that CRAFT tends to introduce additional building blocks or substituents in the optimized trajectory, potentially increasing the number of sites interacting with the target. For example, with DRD2, CRAFT achieved an SA score of 2.54, comparable to TRACER (2.52) and DoG-Gen (2.55), while outperforming Molecule Chef (2.67), CasVAE (2.75), and SynFlowNet (2.84). This indicates that although CRAFT often introduces richer pharmacophore information, its synthetic difficulty does not increase significantly, likely because the candidate molecules are constructed stepwise under explicit reaction template constraints. A similar pattern emerges on AKT1: CRAFT's SA is 2.76, superior to CasVAE (2.81) and SynFlowNet (2.79), and comparable to MoleculeChef (2.71) and DoG-Gen (2.74). Even under the most challenging CXCR4 setting, CRAFT's SA reaches 2.60, slightly higher than DoG-Gen (2.35) and TRACER (2.45), but consistent with Molecule Chef (2.60), and significantly better than CasVAE (2.70) and SynFlowNet (3.41).

[0143] In summary, these results demonstrate that CRAFT can explore a wider chemical space while maintaining synthetic feasibility to a large extent. Although this usually results in molecules with higher molecular weights, CRAFT achieves overall performance comparable to or better than most baseline methods on different targets.

[0144] To assess the plausibility of high QSAR molecules generated by CRAFT in terms of medicinal chemistry, this invention calculates pairwise Tanimoto similarities between generated molecules with predicted activities higher than 0.5 and known active ligands for each target. It is important to emphasize that these reference ligands are not included in the model's USPTO training set; therefore, this analysis can directly measure the model's generalization ability and scaffold rediscovery capability. Similarity distributions and representative cases are shown in [the table / section]. Figure 1 , Figure 2 and Figure 3 The information is provided in the text.

[0145] At different targets, CRAFT was able to generate compounds with significant structural homology to known active ligands. Taking DRD2 as an example, the molecule with the highest similarity in the generated set showed a similarity of 0.933 to a known active ligand, Tanimoto, with a highly consistent core skeleton and substituent pattern. Furthermore, the docking score of this analog was −9.9 kcal / mol, superior to the reference ligand's −7.9 kcal / mol, indicating that the model not only captured the key pharmacophore of the DRD2 inhibitor but may also further enhance the predicted binding ability through subtle structural adjustments. AKT1 and CXCR4 showed similar trends. For AKT1, a molecule generated by CRAFT showed a similarity of 0.524 to a known ligand, retaining key aromatic structural features and the spatial arrangement of hydrogen bond donors / acceptors, with a docking score (−10.0 kcal / mol) close to that of the reference ligand (−10.3 kcal / mol). For CXCR4, despite limited available data for this target, CRAFT generated an analog with a similarity of 0.453, and its docking score (−8.9 kcal / mol) was significantly improved compared to the known ligand (−7.1 kcal / mol). Taken together, these results demonstrate that CRAFT can go beyond the direct training distribution coverage and, by leveraging learned reaction logic, generate candidate molecules de novo that structurally reproduce key fragments of known ligands while maintaining strong predictive binding potential.

[0146] At different targets, CRAFT was able to generate compounds with significant structural homology to known active ligands. Taking DRD2 as an example, the molecule with the highest similarity in the generated set showed a similarity of 0.933 to a known active ligand, Tanimoto, with a highly consistent core skeleton and substituent pattern. Furthermore, the docking score of this analog was −9.9 kcal / mol, superior to the reference ligand's −7.9 kcal / mol, indicating that the model not only captured the key pharmacophore of the DRD2 inhibitor but may also further enhance the predicted binding ability through subtle structural adjustments. AKT1 and CXCR4 showed similar trends. For AKT1, a molecule generated by CRAFT showed a similarity of 0.524 to a known ligand, retaining key aromatic structural features and the spatial arrangement of hydrogen bond donors / acceptors, with a docking score (−10.0 kcal / mol) close to that of the reference ligand (−10.3 kcal / mol). For CXCR4, despite limited available data for this target, CRAFT generated an analog with a similarity of 0.453, and its docking score (−8.9 kcal / mol) was significantly improved compared to the known ligand (−7.1 kcal / mol). Taken together, these results demonstrate that CRAFT can go beyond the direct training distribution coverage and, by leveraging learned reaction logic, generate candidate molecules de novo that structurally reproduce key fragments of known ligands while maintaining strong predictive binding potential.

[0147] Besides reproducing known drug structures, the core value of generative models lies in discovering entirely new chemical spaces that have not yet been fully explored. To this end, this invention screens generative molecules that differ significantly from known active ligand structures (with minimal Tanimoto similarity) but exhibit excellent performance in both predicted activity and docking scores. These molecules can be considered potential lead candidates with novel skeletons. For example... Figure 4 As shown, CRAFT successfully generated novel structures with high potential at all three target sites.

[0148] These results clearly demonstrate that CRAFT is not merely a simple imitation of existing chemical knowledge, but rather possesses the innovative ability to explore unknown chemical spaces, providing candidate scaffolds with high patent novelty and development value for the early stages of drug discovery.

[0149] To explain the decision-making process behind model predictions, this invention visualizes the atomic importance scores obtained from the cross-attention mechanism of the Transformer decoder. Specifically, this invention extracts attention weights from the multi-head cross-attention layer, which characterize the intensity of the decoder's attention to different regions remembered by the encoder when generating the product sequence. To obtain a stable global attention pattern, this invention averages the weights of all decoding layers and attention heads. Subsequently, it utilizes the strict token-atomic alignment mapping constructed in the multimodal encoder. The importance scores of all source-end tokens aligned to the same atom are aggregated, and the aggregated scores are normalized to the [0,1] interval. The final distribution is displayed as a heatmap to characterize the region of interest of the model on the reactant molecule.

[0150] Figure 5 A comparison of the attention heatmaps of CRAFT and TRACER baselines across four representative reaction classes reveals a fundamental difference in their approaches to resolving chemical structures. CRAFT exhibits a sharper attention distribution with clear chemical interpretability. In alcohol oxidation (template 356) and amidation (template 599), CRAFT focuses its attention on the key atoms actually involved in bond breaking and formation, namely the hydroxyl oxygen and carboxyl carbon. Similarly, in nitrile synthesis (template 175) and urea formation (template 230), CRAFT can stably localize to the reactive nitrogen atom. These phenomena indicate that an explicit role-embedding mechanism can effectively guide the model to identify the true reaction center specified by the input template.

[0151] In contrast, TRACER's attention distribution is more diffuse and more susceptible to structural biases. Lacking explicit topological guidance, this sequence-only baseline tends to focus more on dominant skeletal segments (such as aromatic rings or highly connected atoms) rather than actual reactive sites. For example, in the urea formation case (template 230), CRAFT focuses on the reactive amine within the piperidine ring, while TRACER's attention shifts to the neighboring but chemically inert aromatic ring. This suggests that TRACER relies more on memorizing global skeletal patterns than learning local causal rules governing chemical reactivity. Overall, this interpretability analysis suggests that CRAFT's performance improvement stems from its greater ability to capture chemically meaningful causal relationships, thereby reducing the risk of spurious transformations at non-reactive sites.

[0152] To further evaluate the binding potential of the generated molecules from a structural perspective and demonstrate the interpretability of CRAFT in route planning, this invention performed molecular docking simulations on some candidate molecules and traced their corresponding forward synthetic routes. To screen for more promising compounds while ensuring computational efficiency, a hierarchical screening process was adopted. First, each target was sorted according to the activity probability predicted by QSAR, and the top 200 candidate molecules were selected. Then, AutoDock Vina was used for molecular docking, searching for ligand conformations and calculating binding energies. Finally, the candidate molecule with the lowest docking energy for each target was selected for visualization analysis, such as... Figure 6 , Figure 7 and Figure 8 As shown. Overall, the selected molecules exhibited strong predicted binding potential at all three target sites.

[0153] The above are merely preferred embodiments of the invention and are not intended to limit the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A method for generating chemically reactive, multimodal synthetic molecules, characterized in that, Includes the following steps: S10: Obtain the SMILES sequence of the initial inactive molecule. Select multiple candidate templates with the highest fitness from the high-frequency reaction templates. Establish the correspondence between the candidate templates and the atoms of the initial inactive molecule through pattern matching to generate atom role information, specifically: The SMARTS string of the candidate template contains a set of atomic mapping numbers, denoted as... ; Construct a lookup table This makes each atom mapping of the candidate template numbered. Associate it with a specific role ID and the reactants of the candidate templates. Represented as a molecular diagram ; Substructure matching is performed using RDKit to find mappings from atom mapping numbers to atom indices in the molecule. ; If a match is successful, number each atom mapping. The corresponding atoms exist in all of them. ,satisfy If a match fails, the candidate template is discarded. For each atom in the molecule of a successfully matched candidate template Assign role tags to obtain atomic role vectors. : ； In the formula, Z represents the set of integers, and V represents the total number of atoms. This represents the i-th atom in the molecule. This represents the role vector of the i-th atom; In substructure alignment, if a certain mapping number is matched... Then set its character tag to the corresponding character ID. Otherwise, the atom Treat them as non-reactive atoms and assign them a background role of 0, generating atom role information; S20: Based on atomic role information, extract atomic physicochemical descriptors and chemical bond attributes to construct node and edge feature matrices of a molecular graph. Then, capture the two-dimensional topological structure through graph attention network encoding. Specifically: For each atom Extracting inherent physicochemical descriptors and will be associated with atomic role vectors Perform fusion and map to the first dense vector. ; After concatenating the two, a linear projection is performed to obtain the initial node representation. : ； In the formula, This indicates that the encoder uses a dimension of... The hidden feature space, Indicates a linear projection layer; Stacking the feature vectors of all atoms into a node feature matrix ; For each chemical bond Extract chemical properties and construct edge feature vectors ; Node feature matrix and edge feature vectors The input is fed into a multi-layer graph attention network model to generate a final graph representation containing a two-dimensional topological structure. ; Simultaneously, the SMILES sequence is transformed into a token sequence and word embeddings are generated. Graph features and sequence features are then fused to form a hybrid representation that combines structural information and syntactic logic, specifically: The SMILES sequence S of the initial inactive molecule was used as the token sequence. Then, each token sequence is mapped to a standard word embedding. ; Constructing token sequences and atomic alignment maps Record sequence position With graph node index The correspondence; For sequence position If a valid alignment exists Then from the final diagram representation Extract the row vector corresponding to the atom from the... And integrate it with the token embedding at that location; Otherwise, only sequence embeddings are retained: ； In the formula, A mixed feature vector representing a single sequence position. express Zhongyu The row vector corresponding to the atom being pointed to. Represents a learnable linear projection; Mixed feature vectors Stacking by sequence position yields the mixed sequence representation. ; Will Input a multi-layer Transformer encoder to obtain a hybrid representation encoder memory. ; S30: The ID of the candidate template is transformed into a global feature vector, incorporated into the hybrid representation, and used to construct the decoder context memory. Specifically: The role IDs of the candidate templates are input into the encoder's memory, and then converted into a second dense vector through an embedding layer. ; The second dense vector After linear projection, Dropout regularization, and layer normalization, the final template global feature vector is generated. : ； In the formula, Indicates the normalization layer. Indicates a random deactivation layer. Let denote the linear projection matrix, and b denote the bias vector; Then, by using a prefix concatenation strategy, the global feature vector is... Add to encoder memory Previously, the context memory was generated. : ； In the formula, Indicates the dimension of the hidden layer; At each time step T, the generated subsequence is received by the Transformer decoder. The decoder query is compared with the context memory. Interact to generate decoder context memory : ； In the formula, This represents the multi-head cross-attention mechanism, where Q represents the query matrix, K represents the construct matrix, and L represents the value matrix. S40: Based on the decoder context memory, it predicts and outputs the multiple reactant sequences with the highest probabilities through beam search; S50: Solving multiple synthetic pathways based on reactant sequences; S60: Evaluate each synthetic route, eliminate molecules with invalid structures or excessive molecular weights, and output the optimal synthetic route.

2. The method for generating multimodal synthetic molecules based on chemical reaction sensing according to claim 1, characterized in that, The specific steps for obtaining the SMILES sequence of the initial inactive molecule and selecting multiple candidate templates with the highest fitness from the high-frequency reaction templates are as follows: Given the initial SMILES sequence S of inactive molecules, the reaction template set is predicted by a predictor. Output probability distribution; Select the K templates with the highest probabilities to form a candidate template set. : ； In the formula, Indicates the predictor parameters Under the given reactant SMILES sequence S, a candidate template is selected. The conditional probability, Represents the set of reaction templates Among them, the K reaction templates with the highest probability.

3. The method for generating multimodal synthetic molecules based on chemical reaction sensing according to claim 2, characterized in that, Based on the decoder context memory, the specific steps for predicting and outputting the multiple reactant sequences with the highest probabilities through beam search are as follows: decoder context memory Perform training and minimize the true target sequence. negative log-likelihood loss : ； In the formula, Represents the real target sequence The Tth token, Represents the real target sequence The true prefix of all tokens before the T-th position. express The sequence length, Represents a probability distribution; Post-trained decoder context memory Output the Top-K reactant sequences with the highest probabilities.

4. The method for generating multimodal synthetic molecules with chemical reaction sensing according to claim 3, characterized in that, The specific steps for solving multiple synthetic pathways based on reactant sequences are as follows: For the current node of the reactant sequence any child node The UCB value is calculated as follows: ； In the formula, Represents child nodes The cumulative reward value, Represents child nodes Total number of visits, This represents the historical average return of the current node. Represents a constant; After further structural validity verification and molecular weight constraints by RDKit, effective pruning is performed, and legitimate candidate molecules are instantiated as new child nodes and added to the search tree. Starting from the current node, the CRAFT model, as a simulation strategy, uses a greedy decoding method to continuously predict multiple steps of the reaction until the preset maximum depth is reached or no effective molecules can be generated. Then, the highest reward value obtained along the path. Return to the current node and all its ancestor nodes, and output multiple composite paths.

5. The method for generating multimodal synthetic molecules with chemical reaction sensing according to claim 4, characterized in that, The specific steps for evaluating each synthetic route, eliminating molecules with invalid structures or excessive molecular weights, and outputting the optimal synthetic route are as follows: Synthesis path through reward function Evaluation, reward function Defined as a combination of activity score and molecular weight limit: ； In the formula, This represents the activity probability score. Indicates chemical validity test, Indicates molecular weight; The synthesis path with the highest reward value is selected as the optimal synthesis path.

Images