Automatic drug design method, system, computing device and computer readable storage medium
By decomposing the compounds into fragments of synthesizable modules and reassembling them, the L2F2L framework solves the shortcomings of the L2L framework in terms of effectiveness, uniqueness and novelty, achieving the ability to efficiently generate high molecular weight and high novel compounds, and has cross-target Point reusability improves the overall performance of drug design.
Patent Information
- Application Number
- CN202011020214.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-09-24
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2040-09-24
AI Technical Summary
The existing molecular generation model under the framework of Lead-to-Lead (L2L) has shortcomings in terms of effectiveness, uniqueness, novelty, molecular weight and cross-target reusability, making it difficult to generate high-fidelity and strong synthetic compounds. Especially in high molecular weight and different target scenarios, applications are limited.
Using the Lead-to-Fragment-to-Lead (L2F2L) framework, the target compound is broken down into fragments of synthesizable modules, sampled through trained drug design machine learning models, and reassembled to generate new compounds, using the VAE model and The RDKIT toolkit performs fragment decomposition and assembly to improve generation efficiency and quality.
It significantly improves the effectiveness, uniqueness and novelty of molecular generation, can generate novel compounds with high molecular weight, and achieve cross-target reuse, simplifies the optimization of the local structure of the compound, and improves the efficiency and quality of drug design.
Smart Images

Figure CN112116963B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to an automatic drug design method, system, computing device and computer-readable storage medium. Background Art
[0002] Designing lead compounds with ideal properties is a core task in the drug discovery stage. In fast-follow and me-too drug design scenarios, the traditional process requires collecting a large number of papers and patents, and designing novel, synthesizable, and druggable compounds based on the reading and understanding of medicinal chemists, which are then verified through chemical synthesis and biological characterization.
[0003] Molecular Generation is an automatic drug design method based on deep generative learning that has developed rapidly in recent years. By letting the model learn the SMILES (a compound structure represented in the form of a string) or Molecular Graph (atom and chemical bond connection represented by a molecular graph) data of the input compound, and mastering the statistical laws, it can automatically generate compounds with new structures, which can greatly improve the efficiency of lead compound design. Common molecular generation algorithms include Recurrent Neural Network (RNN), Generative Adversarial Networks (GAN), Variation Autoencoders (VAE), etc. Regardless of the algorithm, a large number of molecular structure data examples need to be input into the model to fully train the neural network in order to master the essentials of compound structure design.
[0004] refer to Figure 1A and 1B The existing molecular generation algorithms that use SMILES strings as input generally adopt the Lead-to-Lead (L2L) framework: the overall lead compound structure data is used as input in the training phase; the same is true in the sampling phase, in order to obtain new structural lead compounds with similar properties to the target molecule. Although the application model under this framework, such as GENTRL developed by Insilico Medicine, has achieved remarkable success, it has several obvious problems, including:
[0005] Problem 1: The validity and uniqueness of molecules generated by the L2L framework model are low.
[0006] Question 2: The high novelty of molecules generated by the L2L framework model is low, and it is difficult to produce structurally normal compounds in a chemical space far away from the target molecule (molecular similarity Tc < 0.4).
[0007] Question 3: The L2L framework model has difficulty in generating structurally normal compounds with a molecular weight (MW) greater than 500, and therefore cannot be applied to some lead compound design scenarios that require high molecular weight.
[0008] Question 4: The L2L framework model cannot be reused across targets. When applied to different targets, known active compounds of the target targets must be added for training to generate new structural compounds, resulting in a waste of time, manpower and material resources.
[0009] Question 5: The L2L framework does not allow the model to fix a certain part of the substructure of the compound and automatically sample other parts.
[0010] The above problems limit the practical application value of molecular generation models using the L2L framework. Summary of the invention
[0011] The present invention provides an automatic drug design method, comprising: decomposing a target lead compound into fragments with synthesizable modules, and sequentially inputting the fragments into a trained drug design machine learning model for sampling; and reassembling new fragments output by the drug design machine learning model to obtain a new lead compound.
[0012] In one embodiment of the present invention, the training data set of the drug design machine learning model is obtained in the following manner: filtering the active compounds in the CHEMBL 25 data set according to predetermined rules to obtain an initial data set; decomposing each initial compound in the initial data set into fragments with synthesizable modules and removing duplicates to obtain multiple non-redundant fragments; using a randomly generated SMILES method to amplify the non-redundant fragments by a predetermined multiple, and using the multiple SMILES strings obtained after amplification as the training data set.
[0013] In one embodiment of the present invention, the predetermined rules include: the target of the active compound belongs to one of the human protein families GPCR A, Hydrolase, Kinase, Ligand-gated Ion Channel, Oxidoreductase, Protease, Transferase, Transporter, Voltage-gated Ion Channel; the activity test type is SINGLE PROTEIN; compounds containing fragments are removed; compounds with a molecular weight greater than 500 are removed; and low-activity compounds with PCHEMBL<6 are removed.
[0014] In one embodiment of the present invention, the input training SMILES string is converted into a string with a fixed length of 120: if the SMILES string is less than 120, it is padded with spaces; if the length exceeds 120, it is filtered; and, based on the characters used to encode the SMILES string as flags, each character is converted into a one-hot vector. Finally, a SMILES string is converted into a 120×43 matrix as the input of the drug design machine learning model.
[0015] In one embodiment of the present invention, the VAE model is based on the open source Molecular VAE model, and the molecular structure inspection, fragment decomposition and fragment assembly use the RDKIT toolkit; the deep learning framework uses Pytorch 1.5.1, the CUDA version is 10.1.105, the operating system is Ubuntu LTS 18.04, and all computing work is completed on a 4×Geforce RTX-2080-TI GPU server.
[0016] The present invention also provides an automatic drug design system, comprising: a drug design machine learning model; an input module, used to decompose the target lead compound into fragments with synthesizable modules, and sequentially input the fragments into the trained drug design machine learning model for sampling; and an output module, used to reassemble the new fragments output by the drug design machine learning model to obtain new lead compounds.
[0017] The present invention also provides a computing device, including a memory and a processor, wherein the memory stores a program, and the processor implements the above-mentioned automatic drug design method when executing the program.
[0018] The present invention also provides a computer-readable storage medium on which a program is stored, and when the program is executed by a processor, the above-mentioned automatic drug design method is implemented.
[0019] The automatic drug design of the present invention has greatly improved the performance in terms of effectiveness and uniqueness of molecule generation, and can generate molecules with high novelty, strong synthesizability and strong drugability; it can easily generate molecules in the high molecular weight region; it only needs to be trained once with a specific data set and can be reused in different target lead compound generation scenarios; it can easily fix the local structure of the compound and optimize the rest. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1A and 1B This is a schematic diagram of model input, training and sampling under the L2L framework of the prior art.
[0021] Figure 2A and 2B It is a schematic diagram of model input, training and sampling under the L2F2L framework of an embodiment of the present invention.
[0022] Figure 3A and 3B is a comparison of the statistical distribution of the properties of the training data set under the L2L and L2F2L frameworks; among them, Figure 3A is a comparison of the probability density distribution of the SMILES string lengths of the two. Figure 3B It is a comparison of molecular weight distribution.
[0023] Figure 4 This is a structural diagram of the VAE model used in an embodiment of the present invention. The model consists of three parts: an encoder, a latent space, and a decoder.
[0024] Figure 5A and 5B The changes of the VAE model loss function with the number of training epochs are shown in the L2L framework and the L2F2L framework during training.
[0025] Fig. 6A -F shows the distribution of KRAS compounds, novel and high novel compounds generated by the L2L model, and novel and high novel compounds generated by the L2F2L model in the synthesizability (SA) and drugability (QED) space; the area marked by the dotted box is the area with SA<5 and QED>0.2. The larger the SA value, the more difficult it is to synthesize; the larger the QED value, the easier it is to drug.
[0026] Fig. 7A and 7B The novel molecular size distribution is shown by sampling the L2L framework and L2F2L framework VAE model; Fig. 7A shows the SMILES string length distribution, Figure 7B The molecular weight distribution is shown.
[0027] Figure 8 The comparison of the sampling and fixing of local structures by the L2L and L2F2L models is shown. It can be seen that the L2F2L model can fix the α,β-unsaturated amide and not sample it, so that the structure of this part of the newly designed molecule can be retained. On the contrary, the L2L model cannot do this.
[0028] Fig. 9 It is a flow chart of the automatic drug design method according to an embodiment of the present invention.
[0029] Fig.10 It is a structural diagram of the automatic drug design system according to an embodiment of the present invention.
[0030] Fig.11 It is a diagram of the internal structure of a computing device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0031] The inventor believes that under the premise of limited training data and computing power, it is a difficult task to try to make the machine model correctly output a grammatically correct SMILES string (i.e. a complete lead compound) with dozens or even hundreds of characters. Even for a giant generative model like GPT-3, after almost inputting a 45TB corpus, using 175 billion parameters, and trained on a high-performance computing platform, it is still extremely difficult to generate text paragraphs with correct words, grammar, content, and logic. Therefore, in response to the problems existing in the prior art and combining the practical experience of medicinal chemists in drug design, this application takes a different approach and proposes a new molecular generation algorithm framework that uses molecular fragments as input, which is called Lead-to-Fragment-to-Lead (L2F2L). The difference from the L2L framework is that, with reference to Figure 2AAnd B. In the training stage, the present application breaks the whole into parts and decomposes the lead compound into fragments with synthesizable modules (Fragment, which refers to a fragment that can be synthesized through existing chemical raw material intermediates and simple reactions) through the BRICS algorithm (this algorithm decomposes according to whether the bond can be synthesized, and it can return a deduplicated list, and the serial number on the atom corresponds to a specific reaction type). The VAE model is allowed to learn these shorter SMILES strings and master their correct syntax; in the sampling stage, the target compound is also decomposed into fragments, and then the fragments are input into the trained VAE model for sampling in turn, and finally the new fragments are reassembled into a lead compound with a new structure. During the implementation process, the L2F2L model showed obvious advantages over the L2L model: (1) The L2F2L model has greatly improved the performance in terms of the effectiveness and uniqueness of molecule generation; (2) The L2F2L model can generate molecules with high novelty, strong synthesizability, and strong drugability; (3) The L2F2L model can easily generate molecules in the high molecular weight region; (4) The L2F2L model only needs to be trained once with a specific data set and can be reused in different target lead compound generation scenarios; (5) The L2F2L model can easily fix the local structure of the compound and optimize the rest.
[0032] Methods, Materials, and Datasets
[0033] This application uses the CHEMBL 25 dataset and filters the active compounds according to the following rules: (1) The target of the active compound belongs to one of the human protein families GPCR A, Hydrolase, Kinase, Ligand-gated IonChannel, Oxidoreductase, Protease, Transferase, Transporter, Voltage-gated IonChannel; (2) The activity test type is SINGLE PROTEIN; (3) Compounds containing fragments (such as Na+, Cl-, OH- ions) are removed; (4) Compounds with a molecular weight greater than 500 are removed; (5) Low-activity compounds (PCHEMBL<6, PCHEMBL is the value characterizing the activity of the compound in the ChEMBL chemical drug library) are removed. Finally, 153,498 highly active compounds were obtained as the initial dataset and named CHEMBL25L.
[0034] Under the L2L framework, this application uses a randomly generated SMILES method to expand the CHEMBL25L data volume by 10 times, reaching 1,534,980 SMILES strings as a training data set. Under the L2F2L framework, this application decomposes each initial compound in CHEMBL25L into several fragments, and after deduplication, a total of 22,581 non-redundant fragments are obtained. Similarly, this application also uses a randomly generated SMILES method to expand the above-mentioned non-redundant fragments by 10 times, reaching 225,810 SMILES strings as a training data set. The SMILES string length distribution and molecular weight distribution of the training data under the two model frameworks are significantly different. The SMILES string length of the L2L framework training data set is significantly higher than that of the L2F2L training data set (such as Figure 3A ); Similarly, the molecular weight of compounds in the L2L framework training dataset is significantly higher than that in the L2F2L training dataset (e.g. Figure 3B ). In addition, the present application obtained 267 compounds targeting KRAS target from the KRAS target active compounds, of which 84 compounds had a molecular weight below 500, and 183 compounds had a molecular weight above 500. Using the same method, the present application processed the 267 active compounds targeting KRAS collected in the above steps so as to add training data as needed.
[0035] The SMILES string input for training is converted into a string with a fixed length of 120: if the SMILES string is less than 120, it is padded with spaces; if the length exceeds 120, it is filtered. Furthermore, based on the characters used to encode the SMILES string (the characters that constitute the SMILES string, such as C / N / H / O, etc.) as tokens, each character is converted into a one-hot vector (one-bit valid vector). Finally, a SMILES string is converted into a 120×43 matrix as the input of the model.
[0036] VAE model structure, loss function and sampling strategy
[0037] For a fair comparison, this application uses exactly the same VAE model in both L2L and L2F2L frameworks (e.g. Figure 4 ). The model consists of three parts: encoder, latent space, and decoder. The encoder consists of three convolutional layers plus a linear layer. The mean (μ) and variance (σ) of the encoder output are represented by a 1×292 vector. The decoder consists of a GRU layer plus two linear layers.
[0038] This application uses the classic VAE model loss function, which consists of two parts: generation loss and latent loss (Formula 1). The generation loss compares the model output and input, reflecting the accuracy of the autoencoder. Here, this application uses Binary cross entropy as generation loss. Latent loss reflects the difference between the latent space vector (latent vector) and the standard Gaussian distribution. Here, this application uses the KLdivergence function. The model uses Adam as a parameter optimizer.
[0039] Loss=Binary cross entropy(output smiles vector, inputsmiles vector)+KL Divergence(latent vector,unit gaussian)
[0040] (Formula 1)
[0041] The pre-trained VAE model has three sampling strategies: (1) random vector decoding; (2) known molecular perturbation; (3) intermolecular interpolation. Here, in order to facilitate the comparison of the two model frameworks, this application uses the second strategy, namely known molecular perturbation. This is also the way most human medicinal chemists design drugs, that is, modifying existing drug structures.
[0042] Model Evaluation
[0043] In order to evaluate the performance differences of models under different frameworks, this application uses the following indicators to evaluate the performance of model-generated molecules: validity, uniqueness, novelty, and high novelty.
[0044] Validity=(#of valid compounds) / (#of sampling)
[0045] (Formula 2)
[0046] Uniqueness=(#of unique compounds) / (#of valid compounds)
[0047] (Formula 3)
[0048] Novelty=(#of novel compounds(Tc<1.0)) / (#of unique compounds)
[0049] (Formula 4)
[0050] High novelty=(#of|ig|novel compounds(Tc<0.4)) / (#of novel compounds(Tc<1.0))
[0051] (Formula 5)
[0052] Among them, validity is the proportion of generated valid molecules (i.e., syntactically correct SMILES strings) to the number of sampling times (Formula 2); uniqueness is the proportion of unique molecules to valid molecules (Formula 3); novelty is the proportion of molecules with novel structures (similarity to known compounds used as sampling input is less than 1.0) to unique molecules (Formula 4); high novelty is the proportion of highly novel molecules (similarity to known compounds used as sampling input is less than 0.4) to novel molecules (Formula 5).
[0053] In order to evaluate the similarity between the model-generated molecules and the aforementioned 267 active compounds targeting KRAS, the present application uses RDKIT to calculate the Morgan fingerprint of the compound, and then measures the similarity between the two molecules by the Tanimoto Coefficient (Tc) (Formula 6).
[0054]
[0055] (Formula 6)
[0056] Where fp1 is the fingerprint of compound 1, and fp2 is the fingerprint of compound 2.
[0057] In order to evaluate the synthesizability of model molecules, this application uses the synthetic accessibility (SA) index to quantitatively estimate the difficulty of synthesizing a given compound. The larger the SA value, the more difficult the compound is to synthesize; conversely, the lower the difficulty. In addition, this application also uses the quantitative drug-likeness parameter (Quantitative Estimate of Drug-likeness, QED) to estimate the drugability of the generated molecule.
[0058] Software and Hardware
[0059] The VAE model starting code used in this application comes from the open source Molecular VAE model (https: / / github.com / topazape / molecular-VAE), and has been modified on this basis (SMILES string->one-hot vector encoding has been modified, and special characters required for encoding fragment connections have been added). The RDKIT toolkit was used for molecular structure inspection, fragment decomposition, and fragment assembly. The deep learning framework used Pytorch 1.5.1, CUDA (Compute Unified Device Architecture, a computing platform launched by graphics card manufacturer Nvidia) version 10.1.105, and the operating system was Ubuntu LTS 18.04. All calculations were completed on a 4×Geforce RTX-2080-TI (Nvidia) GPU server.
[0060] result
[0061] 1) Both L2F2L and L2L model training can converge well
[0062] First, this application uses CHEMBL25L+KRAS as the training data set, and trains the VAE model under the L2L and L2F2L frameworks respectively. After a certain number of iterations, the model can converge well. Under the L2L framework, only a small number of epochs are needed to achieve convergence stability, and the loss function is stable at around 30 (such as Figure 5A ). Under the L2F2L framework, the model requires a large number of epochs to achieve convergence stability, and the loss function is stable at around 10 (such as Figure 5B ). Since the amount of training data of L2F2L is small, about one tenth of the training data of the L2L model, the time required for the two to reach convergence is similar. In order to further compare the difference in the molecular generation capabilities of the two, this application uses the KRAS dataset as the initial molecular input and obtains new structural compounds through molecular perturbation sampling.
[0063] 2) L2F2L model generates molecules with better effectiveness, uniqueness and novelty
[0064] Using the trained L2L and L2F2L models, this application used 267 compounds from the KRAS dataset as input and performed molecular perturbation sampling. In order to eliminate the interference of random factors, the two models each performed 3 batches of independent sampling; in each batch of sampling, each input molecule was sampled 10 times, for a total of 2,670 times. In terms of various sampling performance indicators, the L2F2L model significantly surpassed the L2L model (Table 1). Among them, the sampling molecule validity of the L2F2L model was 94.59±0.37%, an increase of 2080% compared with the L2L model; the sampling molecule uniqueness of the L2F2L model was 99.95±0.02%, an increase of 62% compared with the L2L model; the sampling molecule novelty of the L2F2L model was 99.76±0.04%, an increase of 41% compared with the L2L model; the high novelty of the sampling molecule of the L2F2L model was 40.93±1.26%, an increase of 265% compared with the L2L model (Table 1). After 2,670 samplings each, the number of novel structural molecules obtained by the L2F2L model was 2518.33±10.60, while the number of novel molecules obtained by the L2L model was only 50.67±3.79. From the last indicator, the efficiency of the L2F2L model in generating new molecules is nearly 50 times higher than that of the L2L model.
[0065] Table 1. Comparison of sampling performance of L2L and L2F2L models trained using the CHEMBL+KRAS dataset.
[0066]
[0067] 3) The L2F2L model generates highly novel molecules that are synthesizable and druggable
[0068] Taking the molecules generated by the first sampling as an example, this application further analyzes the synthesizability and drugability of the highly novel molecules generated by the L2L and L2F2L models. Highly novel molecules are molecules whose similarity (Tc) to any known compounds is less than 0.4, and these compounds have strong innovation. Being able to find synthesizable and druggable compounds among the highly novel molecules generated by the AI generation model is the ultimate goal of AI automatic drug design. Therefore, this application analyzes the distribution of KRAS compounds, as well as novel or highly novel compounds generated by the L2L model in the synthesizability (SA) and druggability (QED) space. Fig. 6A -F, most of the KRAS compounds are distributed in the range of SA < 5 and QED > 0.2 ( Figure 6B ), that is, the spatial range of "easier to synthesize" and "more drug-like". Among the highly novel molecules (Tc < 0.4) generated by the L2L model, only 2 compounds fall into this region, accounting for 4% of all novel molecules ( Fig. 6E). In contrast, 330 of the highly novel molecules generated by the L2F2L model fall into this region ( Fig. 6F ), accounting for 13% of all novel molecules generated. This analysis shows that the L2F2L model not only has a higher efficiency in molecule generation, but is also more likely to generate high-value designed molecules that are highly novel, easy to synthesize (smaller SA), and more drug-like (larger QED).
[0069] 4) The L2F2L model can generate novel molecules with high molecular weight
[0070] The inability to generate novel molecules with high molecular weight (MW>500) is an inherent problem of the L2L model, which is mainly caused by the inability of the VAE model to correctly output SMILES strings of larger lengths. Since in the L2F2L model, the length of the string output by the VAE model is reduced (from the whole molecule to a fragment), the L2F2L model can solve the problem of generating high molecular weight compounds. Using the KRAS dataset as input for perturbation, refer to Fig. 7A -B, this application analyzes the SMILES string length and molecular weight distribution of novel molecules generated by the two models. It can be seen that the L2L model cannot output new structure molecules with SMILES length and molecular weight comparable to the KRAS dataset; on the contrary, the SMILES length and molecular weight distribution of new structure molecules output by the L2F2L model are similar to the KRAS dataset. These results show that the L2F2L model can generate novel molecules with high molecular weight.
[0071] 5) L2F2L model is reusable across targets
[0072] In the L2L framework, in order to ensure sampling performance, for each new target, the corresponding active compound data set must be added to the background data set (such as CHEMBL, ZINC) and the model must be retrained. The lack of cross-target reuse capability makes the use of the L2L model cumbersome. From the perspective of medicinal chemistry, a considerable number of fragments are shared between different drugs, which means that the L2F2L model can be reused across targets. To prove this, this application removed the KRAS data set from the training data set and retrained the L2L and L2F2L models. Then, molecular perturbation sampling was performed using 267 compounds in the KRAS data set as input. Similarly, in order to eliminate the interference of random factors, the two models each performed 3 batches of independent sampling; in each batch of sampling, each input molecule was sampled 10 times, for a total of 2,670 times.
[0073] Table 2. Comparison of sampling performance of L2L and L2F2L models trained using the CHEMBL dataset.
[0074]
[0075] In this experiment, the L2F2L model still significantly surpassed the L2L model in various sampling performance indicators (Table 2). Among them, the sampling molecule validity of the L2F2L model was 68.28±0.21%, an increase of 4700% compared with the L2L model; the sampling molecule uniqueness of the L2F2L model was 99.84±0.17%, an increase of 16% compared with the L2L model; the sampling molecule novelty of the L2F2L model was 99.43±0.03%, an increase of 36% compared with the L2L model; the high novelty of the sampling molecule of the L2F2L model was 33.56±1.91%, which was comparable to the L2L model in this indicator. After 2,670 samplings each, the number of novel structural molecules obtained by the L2F2L model was 1809.67±5.51, while the number of novel molecules obtained by the L2L model was only 24.00±2.65.
[0076] Compared with the L2F2L model trained with the CHEMBL+KRAS dataset, the L2F2L model trained with the CHEMBL dataset alone only decreased in effectiveness (from 94.59±0.37% to 68.28±0.21%), with comparable uniqueness and novelty, and slightly reduced high novelty (from 40.93±1.26% to 33.56±1.9%). In terms of the number of novel molecules finally generated, the L2F2L model trained with the CHEMBL dataset alone produced an average of 1809.67±5.51 new molecules, 28% less than the L2F2L model trained with the CHEMBL+KRAS dataset. The above results show that the L2F2L model is reusable across targets.
[0077] 6) The L2F2L model can perform local sampling optimization on compounds
[0078] In the process of drug design, the following scenario is often encountered: that is, part of the target compound is retained and the rest is modified. This demand usually occurs in the Hit-to-Lead (hit to lead) or Lead Optimization (lead optimization) scenario. Under the L2L framework, since it is impossible to control which part of the SMILES is retained and which part is sampled during the sampling process, the above demand cannot be met. However, under the L2F2L framework, since the sampling object of the model is a fragment, the present application can easily achieve the above demand, that is, in the sampling stage, the fixed fragment part is not sampled, but the rest is sampled and disturbed.
[0079] refer to Figure 8Taking KRAS compound design as an example, for the KRAS G12C mutation, the designed compound needs to have a conservative group, α, β-unsaturated amide (C=CC(=O)N), in order to form a covalent bond with the mutated cysteine (Cysteine). During the L2L model sampling process, the group is easily destroyed because it cannot be fixed locally; while in the L2F2L model sampling process, this application fixes the fragment where the group is located and does not disturb it, thereby obtaining a new structural molecule that meets the needs. The above experimental results show that the L2F2L model has the ability to locally sample compounds.
[0080] The application of artificial intelligence (AI) technology to new drug discovery has been a hot topic in the pharmaceutical industry in recent years, and molecular generation models are one of the focuses. By converting the SMILES string represented by a high-dimensional space into a low-dimensional vector, the generation model constructs a latent space. The vector sampled from the latent space can be restored to a SMILES string in a high-dimensional space by decoding. Merk et al. first demonstrated the application potential of this technology in a paper published in 2018: by training more than 500,000 biologically active compounds from the ChEMBL database, they generated potential agonists of RXR or PPAR receptors. Of the five synthesized compounds, two are PPAR agonists, the other two are PPAR and RXR dual agonists, and the fifth compound is inactive. In 2019, Zhavoronkov et al. published an article in Nature Biotechnology. In this article, the authors trained a DDR1 inhibitor generation model using compounds from literature and patents. Among the six synthesized molecules, four active compounds were found; the best one, compound 1, had an enzyme activity IC50 of 10nM and a cell activity of 10.3nM, and had relatively suitable drug metabolism properties.
[0081] The existing generative models have exposed some obvious deficiencies in the application process. Walters and Murcko pointed out that the biggest problem with the current generative models is the lack of novelty of AI-designed molecules. Judging from the compounds reported so far, they are too similar to known compounds or compounds in the training data set, and the transformation is too simple, which cannot be compared with human medicinal chemists. This application believes that the generative model of the existing Lead-to-Lead (L2L) framework is essentially a language model, which is essentially no different from any natural language processing (NLP) model. Given that the molecular weight of the lead compound is usually between 300-500, the corresponding SMILES string has an average length of about 50 characters. Under the premise of limited training data and computing power, it is obviously a very challenging thing to let the machine model continuously output 50 or even longer characters to form a SMILES string with a synthesizable module. Because of this, the molecules generated by the L2L model can only make slight modifications in the space not far from the training compound, and cannot guarantee that after making greater modifications to the SMILES string, it is still an effective (structurally correct) compound, let alone its synthesizability and drugability.
[0082] This application creatively invented the Lead-to-Fragment-to-Lead (L2F2L) strategy, which decomposes molecules into fragments, samples and innovates the fragments, and then recombines them. Fig. 9 , the automatic drug design method provided by the embodiment of the present invention includes: decomposing the target lead compound into fragments with synthesizable modules, inputting the fragments into the trained drug design machine learning model for sampling in sequence; reassembling the new fragments output by the drug design machine learning model to obtain a new lead compound. In this way, because the average length of the SMILES string corresponding to the fragment is about 20, the difficulty of generating the model is reduced, whether it is training or sampling; accordingly, the effectiveness, uniqueness, and novelty of the molecules generated by the L2F2L model are greatly improved compared with the L2L model.
[0083] Another benefit of using the L2F2L model is that the generative model's ability to generate highly novel molecules (far away from the training compounds) in spatial sampling is greatly enhanced. In the KRAS case presented in this application, after sampling each prototype molecule 10 times, the L2F2L model obtained 330 highly novel compounds with high drugability, while the L2L model only had 2, with a relative efficiency improvement of nearly 165 times. Using the L2F2L generative model, such an improvement is of great significance for drug design in Fast-follow and Me-too scenarios.
[0084] In summary, this application proposes a new molecular generation model framework: L2F2L, to improve the currently commonly used L2L framework. The L2F2L model has the following advantages: (1) excellent molecular sampling effectiveness, uniqueness and novelty; (2) can generate a considerable number of highly novel molecules with strong synthesizability and drugability; (3) the L2F2L model can generate high molecular weight molecules; (4) the L2F2L model can be reused across targets; (5) the L2F2L model can fix the local structure of the compound and perform sampling optimization on the rest, which is suitable for hit-to-lead and lead optimization design scenarios.
[0085] The use of AI molecular generation models to automatically design lead compounds with novel structures, synthesizability and drugability is of great significance to the intelligentization of the pharmaceutical industry. The existing molecular generation model is mainly based on the Lead-to-Lead (L2L) framework. On the basis of learning the SMILES string of the lead compound, a new lead compound SMILES string is generated by implicit space sampling. However, due to the long SMILES string of the lead compound, the model training and sampling performance are poor, the novelty of the generated molecules is not high, high molecular weight molecules cannot be generated, cross-target reuse cannot be made, and local structures cannot be sampled. The problems limit the application value of the L2L model. This application proposes a new generation model framework: Lead-to-Fragment-to-Lead (L2F2L). That is, the generation model is designed to design new structure lead compounds by learning and sampling compound fragments of shorter SMIELS strings. In the case of KRAS inhibitors, the sampling molecule effectiveness of the L2F2L model was 94.59±0.37%, the uniqueness was 99.95±0.02%, the novelty was 99.76±0.04%, and the high novelty was 40.93±1.26%, which is a huge improvement compared to the L2L model. The L2F2L model can generate a considerable number of highly novel molecules that are synthesizable and druggable. The L2F2L model can achieve cross-target reuse. The L2F2L model can modify and innovate local compounds. Therefore, the L2F2L framework has obvious advantages over the L2L framework, and has potential application value in application scenarios such as lead compound design, Hit-to-Lead, and Lead Optimization. The molecular generation model designed in this application has the potential to produce lead compounds with novel structures, synthesizable properties, and good drugability.
[0086] refer to Fig.10The automatic drug design system of an embodiment of the present invention includes: a drug design machine learning model; an input module, used to decompose the target lead compound into fragments with synthesizable modules, and sequentially input the fragments into the trained drug design machine learning model for sampling; an output module, used to reassemble the new fragments output by the drug design machine learning model to obtain a new lead compound.
[0087] The automatic drug design method of the embodiment of the present invention can be implemented in a computing device. An exemplary internal structure diagram of the computing device can be as follows: Fig.11 As shown, the computing device may include a processor, a memory, an external interface, a display, and an input device connected via a system bus. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, an application program, a database, etc. The internal memory provides an environment for the operation of the operating system and programs in the non-volatile storage medium. The external interface includes, for example, a network interface for communicating with an external terminal via a network connection. The external interface may also include a USB interface, etc. The display of the computing device may be a liquid crystal display or an electronic ink display, and the input device may be a touch layer covering the display, or may be, for example, a button, a trackball, or a touchpad provided on the housing of the computing device, or may be an external keyboard, touchpad, or mouse, etc.
[0088] The program stored in the non-volatile storage medium in the computing device can implement the above-mentioned automatic drug design method when executed by the processor. In addition, the non-volatile storage medium can also exist in a separate physical form, such as a USB flash drive. When it is connected to a processor, the program stored on the USB flash drive is executed to implement the above-mentioned automatic drug design method. The method of the present invention can also be implemented as an APP (application) in the Apple or Android application market for users to download to their respective mobile terminals for operation.
[0089] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present application, and does not constitute a limitation on the computing device to which the scheme of the present application is applied. The specific computing device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.
[0090] As described above, a person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0091] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0092] The computer described in the present invention is a computing device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions in a broad sense, and its hardware may include at least one memory, at least one processor, and at least one communication bus. Wherein, the communication bus is used to realize the connection and communication between these elements. The processor may include but is not limited to a microprocessor. The computer hardware may also include an application-specific integrated circuit (ASIC), a programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc. The computer may also include a network device and / or a user device. Wherein, the network device includes but is not limited to a single network server, a server group consisting of multiple network servers, or a cloud composed of a large number of hosts or network servers based on cloud computing (Cloud Computing), wherein cloud computing is a type of distributed computing, a super virtual computer composed of a group of loosely coupled computer sets.
[0093] The computing device may be, but is not limited to, any terminal such as a personal computer, server, etc. that can interact with a user through a keyboard, touchpad, or voice control device. The computing device herein may also include a mobile terminal, which may be, but is not limited to, any electronic device that can interact with a user through a keyboard, touchpad, or voice control device, such as a tablet computer, a smart phone, a personal digital assistant (PDA), a smart wearable device, and other terminals. The network in which the computing device is located includes, but is not limited to, the Internet, a wide area network, a metropolitan area network, a local area network, a virtual private network (VPN), and the like.
[0094] The memory is used to store program code. The memory can be a circuit with storage function in an integrated circuit without a physical form, such as RAM (Random-Access Memory), FIFO (First In First Out), etc. Alternatively, the memory can also be a memory with a physical form, such as a memory stick, TF card (Trans-flash Card), smart media card (smart media card), secure digital card (secure digital card), flash memory card (flash card) and other storage devices, etc.
[0095] The processor may include one or more microprocessors or digital processors. The processor may call program codes stored in the memory to execute related functions. For example, Fig.10 Each module described in the above is a program code stored in the memory and executed by the processor to implement the above method. The processor is also called a central processing unit (CPU), which can be a large-scale integrated circuit, a computing core (Core) and a control core (Control Unit).
[0096] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.
[0097] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0098] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be electrical or other forms.
[0099] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0100] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0101] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.
[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.
[0103] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.
Claims
1. An automatic drug design method, characterized in that: include: Decomposing the target lead compound into fragments with synthesizable modules, and sequentially inputting the fragments into a trained drug design machine learning model for sampling; Reassembling the new fragments output by the drug design machine learning model to obtain new lead compounds; When training the drug design machine learning model, a training data set is obtained through a BRICS algorithm, and an L2F2L framework VAE model is trained based on the training data set to obtain the drug design machine learning model; The training data set is obtained in the following way: Active compounds in the CHEMBL 25 data set were filtered according to predetermined rules to obtain an initial data set; Decomposing each initial compound in the initial data set into fragments with synthesizable modules and removing duplicates to obtain multiple non-redundant fragments; Amplifying the non-redundant fragments by a predetermined multiple using a random SMILES generation method, and using the multiple SMILES strings obtained after amplification as the training data set; The predetermined rules include: The active compound targets one of the human protein families GPCR A, Hydrolase, Kinase, Ligand-gated Ion Channel, Oxidoreductase, Protease, Transferase, Transporter, Voltage-gated Ion Channel; The activity test type is SINGLE PROTEIN; Remove compounds containing debris; Remove compounds with a molecular weight greater than 500; and Remove low activity compounds with PCHEMBL < 6.
2. The automatic drug design method according to claim 1, characterized in that: The SMILES string input for training is converted into a string with a fixed length of 120: if the SMILES string is less than 120, it is padded with spaces; if the length exceeds 120, it is filtered; and, based on the characters used to encode the SMILES string as flags, each character is converted into a one-hot vector. Finally, a SMILES string is converted into a 120×43 matrix as the input of the drug design machine learning model.
3. The automatic drug design method according to claim 1, characterized in that: The VAE model includes: an encoder, a latent space, and a decoder; wherein the encoder includes three convolutional layers and one linear layer, and the mean (μ) and variance (σ) of the encoder output are represented by a 1×292 vector; the decoder includes a GRU layer and two linear layers.
4. The automatic drug design method according to claim 3, characterized in that: The VAE model is based on the open source Molecular VAE model. The RDKIT toolkit is used for molecular structure inspection, fragment decomposition and fragment assembly. The deep learning framework uses Pytorch 1.5.1, the CUDA version is 10.1.105, the operating system is Ubuntu LTS18.04, and all calculations are completed on a 4×Geforce RTX-2080-TI GPU server.
5. An automatic drug design system, characterized in that: include: Machine learning models for drug design; An input module, used to decompose the target lead compound into fragments with synthesizable modules, and sequentially input the fragments into the trained drug design machine learning model for sampling; An output module, used to reassemble the new fragments output by the drug design machine learning model to obtain new lead compounds; When training the drug design machine learning model, a training data set is obtained through a BRICS algorithm, and an L2F2L framework VAE model is trained based on the training data set to obtain the drug design machine learning model; The training data set is obtained in the following way: Active compounds in the CHEMBL 25 data set were filtered according to predetermined rules to obtain an initial data set; Decomposing each initial compound in the initial data set into fragments with synthesizable modules and removing duplicates to obtain multiple non-redundant fragments; Amplifying the non-redundant fragments by a predetermined multiple using a random SMILES generation method, and using the multiple SMILES strings obtained after amplification as the training data set; The predetermined rules include: The active compound targets one of the human protein families GPCR A, Hydrolase, Kinase, Ligand-gated Ion Channel, Oxidoreductase, Protease, Transferase, Transporter, Voltage-gated Ion Channel; The activity test type is SINGLE PROTEIN; Remove compounds containing debris; Remove compounds with a molecular weight greater than 500; and Remove low activity compounds with PCHEMBL < 6.
6. A computing device comprising a memory and a processor, wherein the memory stores a program, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 4 is implemented.
7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the method described in any one of claims 1 to 4 is implemented.
