Molecular optimization method based on phenotypic gene guidance

Abstract

The invention relates to a molecular structure reconstruction method and system based on a gene expression profile, and belongs to the crossing field of computer-aided drug design and artificial intelligence. According to the method, a phenotype-perceived molecular generation model is constructed, and a partially deleted or masked molecular structure is reconstructed by using transcriptome data obtained after drug disturbance. The system comprises a convolution encoder used for extracting phenotypic features from an input drug-induced gene expression profile; the invention also relates to a decoder with a Transform structure, which is pre-trained on large-scale chemical structure data (more than two millions of molecules) so as to realize effective reduction of masked molecular fragments. The methods are trained and validated based on an L1000 public data set comprising more than 74,000 drug perturbed samples of 117 cell lines. In the training process, part of structures input by the molecules are randomly masked, and a model is trained to perform structure reconstruction by taking corresponding gene expression response as a condition, so that a corresponding relation between biological phenotypic characteristics and chemical structure sub-fragments is established. The invention provides a generative method for connecting a gene expression space with a chemical structure space, has good adaptability and interpretability, and provides a new technical path for phenotype-driven molecular design.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and drug design, in particular to a molecular optimization method based on deep learning and gene expression data, especially a personalized molecular structure generation algorithm based on phenotype gene perturbation response guidance, which can be used for phenotype-driven drug design and virtual screening. BACKGROUND

[0002] Traditional drug design methods rely on target structures or known ligand physicochemical properties, and face bottlenecks in the context of unknown targets or complex phenotypes. In recent years, with the development of high-throughput transcriptome sequencing technology, researchers can obtain the overall phenotypic response of cells under drug intervention through gene expression profiling, thereby providing more direct biological clues for new drug design.

[0003] However, existing drug generation methods are mostly based on molecular fragment combination or grammar modeling in chemical structure space, and it is difficult to achieve the reverse generation of molecules that meet the phenotype from the biological phenotype. In addition, due to the extremely large molecular space, the reverse generation based on phenotype has high nonlinearity, sparsity and uncertainty. Therefore, there is an urgent need for a novel and interpretable algorithm framework to realize the optimization and generation of molecules based on gene expression perturbation data. SUMMARY

[0004] The purpose of the present application is to propose a method for guiding the optimization and generation of chemical molecules using the perturbed phenotype gene expression profile as a condition, in order to realize personalized drug design from the gene level.

[0005] The technical solution of the present application is as follows:

[0006] A molecular optimization method based on phenotype gene guidance, comprising the following steps:

[0007] 1) Molecule language model pre-training:

[0008] 1-1 represents the molecular structure using SELFIES (Self-Referencing Embedded Strings), which has syntax constraints and robustness;

[0009] 1-2 pre-train the self-attention Transformer language model on a large-scale molecular structure dataset to learn the combination rules and semantic representation of molecular structures;

[0010] 1-3 use a mask method with future masking to achieve the goal of autoregressive learning, and model the molecular sequence by maximizing the conditional probability.

[0011] 2) Phenotype gene feature encoding:

[0012] 2-1 Obtain the gene expression data of cells after drug treatment, input as a 978-dimensional gene expression vector;

[0013] 2-2 Use a two-layer one-dimensional convolutional neural network (1D-CNN) to extract local co-expression patterns and modular structure features;

[0014] 2-3 The convolution result is mapped to a structured embedding sequence, consistent with the SELFIES label space, as the subsequent conditional generation input.

[0015] 3) Phenotype condition-guided molecule generation:

[0016] 3-1 Divide the molecular structure into three modules: Murcko skeleton, linker and side chain;

[0017] 3-2 Mask processing is performed on any part of the above modules, and the corresponding SELFIES label is replaced with a special [MASK] symbol;

[0018] 3-3 Under the joint condition of molecular context visible information and phenotype encoding, the masked area is reconstructed by the Transformer decoder;

[0019] 3-4 Introduce cross-attention mechanism in each layer of the decoder to fuse biological phenotype features and molecular context, realize modal alignment and condition guidance.

[0020] 4) Training and optimization:

[0021] 4-1 Take the reconstruction accuracy of the masked area as the training target, and use cross-entropy loss function to measure the difference between the predicted label and the real label;

[0022] 4-2 Under the premise of maintaining chemical feasibility, realize diversified, structure-consistent and phenotype-consistent molecule optimization.

[0023] Advantages

[0024] Compared with the prior art, the present application has the following advantages:

[0025] 1. Biological phenotype driven: directly use the transcriptome data after drug perturbation as the generation condition, break through the traditional dependence on target structure or known active molecules, and realize the reverse deduction of potential drugs from cell state.

[0026] 2. Structure module controllable generation: propose a structure-aware mask reconstruction strategy, divide the molecule into functional modules such as skeleton, linker and side chain, and realize independent optimization and combination of different structure regions.

[0027] 3. Cross-modal information fusion: introduce cross-attention mechanism, dynamically integrate biological phenotype and chemical structure in each layer of the model, improve the generation accuracy and biological consistency of the model.

[0028] 4. High-efficiency training and migration ability: significantly improve the model's learning ability of molecular grammar rules through a phased pre-training strategy, reducing the sample size of phenotype data dependence.

[0029] 5. Wide application prospects: can be applied to personalized drug discovery, targeted drug optimization, drug re-design, etc., with important scientific research and industrial transformation value BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is the model framework diagram of the present application;

[0031] Figure 2 is the optimization effect of the present application on known drug molecules based on gene phenotype driving on specific protein targets; DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments.

[0033] The disclosed gene expression data set required by the present application comes from the LINCS L1000 database (https: / / clue.io / ), which provides gene expression profile data of cell lines before and after drug treatment; the molecular structure data comes from the ZINC database (https: / / zinc.docking.org / ), and the SMILES descriptor corresponding to the drug is downloaded; the molecular pre-training corpus comes from the ChEMBL database (https: / / www.ebi.ac.uk / chembl / ), which contains millions of legal drug molecules.

[0034] As shown in Figure 1 , a phenotype gene guided molecular optimization method comprises the following steps:

[0035] 1) Molecular language model pre-training

[0036] 1-1 Use SELFIES (Self-Referencing Embedded Strings) to represent molecular structure. This method ensures that each SELFIES string represents a legal chemical structure, avoiding invalid molecules.

[0037] 1-2 Download about 2 million drug molecule SMILES structures from ChEMBL and convert them to SELFIES format. Build a dictionary to count the size of the vocabulary.

[0038] 1-3 We use an autoregressive Transformer model as the base architecture to maximize the conditional probability P(yt|y1...yt-1) and pre-train the language model with a mask mechanism that hides future tokens.

[0039] 1-4 The model uses an 8-layer Transformer structure, each layer containing multi-head self-attention and a feed-forward neural network, with a hidden dimension of 512 and a word embedding dimension of 256. We use the Adam optimizer with an initial learning rate of 1e-4.

[0040] 2) Phenotype gene feature encoding

[0041] 2-1 We obtain gene expression data before and after drug intervention from real patient samples in the TCGA project (e.g., TCGA-QK-A6IH) and normal samples of the corresponding tissues (e.g., GTEx). We select 978 LINCS general genes as expression features.

[0042] 2-2 For each patient, we calculate the difference between the Z-score standardized before and after intervention to obtain the perturbation vector ΔE as the model phenotype input.

[0043] 2-3 We use a two-layer one-dimensional convolutional network (1D-CNN) encoder to extract local co-expression patterns in ΔE. The convolution kernel size is set to 9, the channel number is 128, and the activation function is ReLU. The output is matched to the SELFIES mask length through linear mapping to obtain the structured phenotype vector C.

[0044] 3) Structure-aware mask reconstruction mechanism

[0045] 3-1 We use RDKit to extract the Murcko skeleton from each molecule. Label this skeleton as a non-replaceable region, and the remaining linker and side chain positions as reconstructable regions.

[0046] 3-2 Label the non-skeleton region in the SELFIES sequence as [MASK] to obtain the masked molecule hat{Y}.

[0047] 3-3 Input the masked molecule hat{Y} and the phenotype vector C into the Transformer decoder, and use the cross-attention mechanism to fuse the molecular context and phenotype features at each layer to achieve conditional reconstruction.

[0048] 3-4 The decoder structure is consistent with the pre-training stage, and the cross-attention mechanism uses 8 heads with a dimension of 64 for each head.

[0049] 4) Loss function and training

[0050] 4-1 Reconstruction loss using mask, i.e. cross-entropy loss function, to compare the error between predicted molecular fragments and real SELFIES fragments:

[0051]

[0052] 4-2 To improve diversity and stability, random sampling is added, and RDkit is used to supervise the legality of generated molecules.

[0053] 5) Generation and testing

[0054] 5-1 After the model training is completed, the parameters are fixed. Input a set of phenotype perturbation vectors (such as the gene expression changes of a certain patient) and a basic molecule (or general skeleton) as the initial mask structure.

[0055] 5-2 Use the model to generate multiple candidate molecules that meet the phenotype, randomly sample different MASK filling results each time, and repeat sampling 100 times.

[0056] 5-3 Verify the legality of the generated molecules using RDKit, calculate the molecular similarity (such as Tanimoto), drug property score (QED), and phenotype consistency score (using the reverse transcriptome simulator of the generated molecule such as CMap / L1000FWD).

[0057] 5-4 Select the molecule with the best structure diversity, synthesis feasibility, and phenotype consistency as the optimization result.

[0058] Example: Phenotype-conditioned molecule generation based on skeleton mask

[0059] In order to evaluate the molecular reconstruction and generation ability of the cross-attention-based phenotype-aware molecule optimization method proposed in the present application under the condition of gene perturbation, the following experimental process is designed.

[0060] I. Experimental object and drug target selection

[0061] The widely used human breast cancer cell line MCF7 in the LINCS L1000 dataset is selected as the experimental object. The experiment is carried out around the following ten common oncogenic targets:

[0062] AKT1, AKT2, AURKB, CTSK, EGFR, HDAC1, MTOR, PIK3CA, SMAD3, TP53

[0063] Each target corresponds to multiple small molecule targeted drugs with clear mechanism of action.

[0064] II. Data preparation and phenotype construction

[0065] 1. Control group data construction:

[0066] Extract 4,000 MCF7 expression profiles without drug treatment from L1000 dataset as Control Group.

[0067] 2. Disturbance Group Data Construction:

[0068] For each target, collect the expression profiles after treatment of drugs known to act on this target, and calculate the mean of the disturbed expression of samples, which is used to represent the average transcriptome response induced by this class of drugs (Phenotype).

[0069] 3. Differential Expression Analysis:

[0070] Based on the law of large numbers, the difference between the average disturbed expression profile and the control group is analyzed (Differential Expression Analysis), and the significantly changed genes are identified as the embedded representation of the drug-induced phenotype (Phenotypic Embedding).

[0071] III. Skeleton Masking and Conditional Reconstruction Task

[0072] 1. Skeleton Extraction and Masking Processing:

[0073] For each targeted drug, use RDKit to extract its Murcko Scaffold (i.e., the skeletal structure of the molecule), and replace the token corresponding to this skeleton part in the SELFIES expression with a special mask token [MASK], to obtain the masked sequence hat{Y}.

[0074] 2. Model Task Definition:

[0075] The task of the model is to predict the missing skeleton part under the condition of input:

[0076] 1. the masked molecular structure sequence hat{Y};

[0077] 2. the phenotype vector C constructed by differential expression; and output the complete molecular structure.

[0078] 1. Sample Generation Settings:

[0079] For each phenotype input (each target), perform the conditional reconstruction task for all related drugs respectively. 50 candidate molecular samples are generated for each masked drug structure, which are used for subsequent evaluation.

[0080] IV. Evaluation Index and Result Analysis

[0081] 1. Structure Restoration Rate:

[0082] The Tanimoto similarity index of the generated results is calculated, and the results show that for more than 90% of the drug samples, at least one sample in the model generated molecule has a Tanimoto similarity of 1.0 with the original structure, indicating that the model can completely restore the skeleton structure.

[0083] 2. Novelty and effectiveness of structure:

[0084] More than 80% of the generated molecules do not appear in the training set, and have significant structural novelty; more than 90% of the generated molecules are effective in syntax and chemical structure, and can be successfully parsed and constructed into a molecular graph by RDKit, indicating that the model has good chemical structure learning and synthesis feasibility.

[0085] 3. Diversity and generalization ability:

[0086] Under the same phenotype input, the generated molecules show high diversity in side chain structure and substituent group; the mask mechanism of controllable skeleton position effectively guides the model to generate drug molecules with high consistency and structural diversity.

[0087] Conclusion

[0088] This embodiment shows that the model proposed in the present application can reconstruct and generate molecules with specific structure and biological function under the gene disturbance induced by a specific disease or target. The method has superior performance in structure controllability, phenotype consistency and chemical feasibility, and can provide an important tool and idea for targeted drug design and individualized medication.

[0089] Description:

[0090] The above only describes the preferred embodiments of the present application, and those skilled in the art can make various modifications or changes without departing from the spirit and essence of the present application. These modifications or changes should also be considered as falling within the protection scope of the present application.

Claims

1. A method for personalized molecular generation based on phenotype-guided, characterized in that, The method comprises the following steps: 1) molecular language model pre-training: 1-1 using SELFIES as the structural representation of molecules, each molecule is represented as an equal-length or padded SELFIES token sequence; 1-2 each token is embedded into a high-dimensional vector, and after adding position encoding, it is input into a language model containing a multi-layer Transformer decoder structure for autoregressive modeling, and the training target is to predict the current token based on the generated left context, and learn the syntactic rules and context dependence of the internal structure of the molecule. 2) perturbed gene expression profile modeling: 2-1 obtain the transcriptome data of cells after drug treatment, input the 978-dimensional gene expression vector into a two-layer one-dimensional convolutional neural network, and extract the local co-expression pattern and gene module structure feature; 2-2 the high-level biological representation after convolution is mapped to a structured embedding sequence consistent with the SELFIES embedding space, which is used for cross-modal fusion. 3) conditional generation and modal fusion: 3-1 introduce gene expression embedding as a conditional signal while keeping the structure of the molecule part visible; 3-2 replace the structure module (such as skeleton, linker, side chain) specified in SELFIES with [MASK] token and input it into the pre-trained Transformer decoder; 3-3 introduce cross-attention mechanism in each layer of the decoder, taking the transcriptome embedding as the key-value vector and the molecule sequence as the query to realize cross-modal fusion; 3-4 the decoder dynamically predicts the content of the masked area according to the visible molecular context and perturbed gene features to realize personalized molecular structure generation under phenotypic conditions. 4) training and optimization: 4-1 the loss function is the cross-entropy loss of the masked position, and the error between the predicted token and the real token is only calculated at the masked position; 4-2 through end-to-end training, the model can generate a molecule fragment that meets the biological perturbation response under the premise of meeting the chemical structure rules.

2. The method of claim 1, wherein: The molecular structure is represented in SELFIES format to ensure the grammatical validity and structural robustness of the generated molecule.

3. The method of claim 1, wherein: The transcriptome input is a one-dimensional vector of length 978, which is encoded using a two-layer one-dimensional convolutional network with convolution kernel sizes of 5 and 3, channel numbers of 128 and 256, and strides of 5 and 2.

4. The method of claim 1, wherein: The masking strategy is based on the Murcko Scaffold structure module, which divides the molecular structure into three parts: skeleton, linker, and side chain, and independently generates conditions for each part.

5. The method of claim 1, wherein: The Transformer decoder contains 6 layers, each containing self-attention, cross-attention, feedforward network, layer normalization, and residual connection, which are used to stabilize training and multi-level modal fusion.

6. The method of claim 1, wherein: The model can be widely used in individualized drug design, targeted molecule optimization, and phenotype-driven candidate compound screening tasks.

Images