CN109033738A - A kind of pharmaceutical activity prediction technique based on deep learning - Google Patents

Abstract

The invention discloses a drug activity prediction method based on deep learning. The present invention uses the RDkit open source library to calculate the basic characteristics of each atom in a given molecule, including atom type, valence, formal charge, etc. Only calculating the atomic characteristics greatly reduces time consumption. The present invention is a prediction model that combines two models of graph convolution and LSTM (long-term short-term memory network). For the graph convolution model, all molecular features are integrated by treating atoms as nodes and bonds as edges in an undirected graph. Turn it into a graph, extract the molecular structure features, and use the graph convolutional neural network to reduce time consumption and obtain features that cannot be obtained by traditional methods. LSTMs learn complex metrics by exchanging information between evidence and query molecules. So as to achieve higher prediction accuracy under low data volume.

Description

A drug activity prediction method based on deep learning

technical field

The invention relates to a method for predicting drug activity based on deep learning, which belongs to the field of software technology.

background content

The main goal of drug research and the development of the pharmaceutical industry is to discover drug molecules related to the treatment of diseases, and exploring lead discovery methods is the main way to achieve this goal. When biological research finds that a particular molecule has therapeutic activity, the discovered molecule is often discarded for a variety of reasons, including toxicity, low activity, and low solubility. According to the statistics of the American Pharmaceutical Research and Manufacturers Association, the research and development of new drugs in the entire pharmaceutical industry accounted for 12.8% of sales revenue, and 75% of them were due to the failure of new drug research and development, and less than 5 compounds were hit in the primary screening. % can enter the preclinical evaluation. Since there is no sample limitation in computer virtual screening, if computer virtual screening is performed first, and then pharmacological testing is carried out, this research strategy will significantly shorten the development cycle of new drugs and reduce the cost of research and development compared with the traditional strategy of directly conducting pharmacological testing. cost. At present, the mainstream direction of lead discovery lies in the study of quantitative structure and activity relationship (QSAR) of molecules, mainly to quantitatively describe the structure of molecules, that is, the selection of molecular feature description methods and the selection of mathematical functional relationships connecting these molecular features and activity .

The current common practices are mainly divided into the following categories:

Based on the relationship between the topological structure, side chain, skeleton and specific toxic action site of the compound molecule. Wang et al. studied the relationship between the topology, side chains, skeletons and specific toxic sites of about 60,000 toxic compound molecules in the RTECS (Registry of Toxic Effects of Chemical Substances) database (such as skin toxicity , hematological toxicity, and renal toxicity, etc.), and compare the number of occurrences of these topological structures in the entire database with the number of occurrences in the toxicity chemical library. This method requires a large amount of data and many positive samples, and only extracting toxic features will lead to large errors in the judgment of non-toxic molecules.

1. Predict the activity of the drug to be tested based on the support vector machine method. Zhang et al. screened the target genes associated with hereditary diseases from the obtained drug targets according to the obtained genetic disease-related gene information, and obtained the characteristic attributes of each sample drug, which is the drug corresponding to the sample drug. The correlation between the target and the target gene associated with the genetic disease; the characteristic attribute of each sample drug is used as the input vector, and the activity of the sample drug is used as the output, and the model is established by the support vector machine method to predict the activity of the drug to be tested. This method is difficult to obtain molecular features, requires a specific data set, and has poor universality.

2. Combination of supervised and unsupervised algorithms based on deep learning for drug active molecule identification. Gao Shuangyin will support vector machine (Support Vector Machine), artificial neural network (Artificial Neural Network), semi-supervised support vector machine (Semi-supervised support vector machine), cost security semi-supervised support vector machine (Cost security semi-supervised support vector machine) machine), stacked autoencode (StackedAutoEncode), and deep belief network (Deep Belief Network) methods are combined to conduct in-depth exploration of three types of drug active molecules (PLK1PBD, SMAD3, IL-1B). Due to the complex structure of active pharmaceutical molecules, the chemometric software MOE is used for precise calculations to obtain their 2D and 3D molecular descriptors, and the above two types of algorithms are used to identify active pharmaceutical molecules. This method requires large data sets, and calculation of molecular features using chemometric software is time-consuming.

To sum up, various methods for drug activity prediction are limited by their own characteristics. Methods based on big data analysis require a large amount of data and have high requirements for the distribution of samples; Time-consuming; the above machine learning algorithms based on supervised and unsupervised not only require a large amount of data, but also use chemometric software to calculate molecular features also require a lot of time.

Glossary:

LSTM: Long short-term memory network.

Atom degree: The weight value of each atom calculated by RDkit is the number of atoms directly connected to the atom.

Lewis structural formula: a way of writing molecules, such as hydrogen cyanide H-C≡N

Sigmoid: The Sigmoid function is a mathematical function of an S-shaped curve, and its formula is

It has a wide range of applications in logistic regression and artificial neural networks.

Tanh: The hyperbolic tangent function, which is derived from the basic hyperbolic functions hyperbolic sine and hyperbolic cosine:

Contents of the invention

The invention overcomes the shortcomings of the prior art, and discloses a method for predicting drug activity based on deep learning. The present invention uses the RDkit open source library to calculate the basic characteristics of each atom in a given molecule, including atom type, valence, formal charge, etc. Only calculating the atomic characteristics greatly reduces time consumption. For the graph convolutional model, all molecules are characterized as graphs by treating atoms as nodes and bonds as edges in an undirected graph, and molecular structure features are extracted. Using graph convolutional neural networks can reduce time consumption and obtain traditional A characteristic that cannot be obtained by the method. LSTMs learn complex metrics by exchanging information between evidence and query molecules. So as to achieve higher prediction accuracy under low data volume.

In order to solve the problems of the technologies described above, the technical solution adopted in the present invention is:

A method for predicting drug activity based on deep learning, comprising the steps of:

Step 1. Construct a drug activity data set and segment the drug activity data set, wherein a part of the data in the drug activity data set is used as a training set, a part of the data is used as a development set, and a part of the data is used as a test set;

Step 2, extracting atomic features from the molecules in the training set, and converting the molecular structure of the training set into an adjacency matrix;

Step 3. Build a prediction model, which includes five layers of graph convolution and one layer of LSTM;

Step 4, train the data obtained in steps 2 and 3;

Step 5. After graph convolution, pooling, and full connection, the output value is sent to the classifier, the loss function is optimized, and training continues;

Step 6. After iterative calculation, the trained prediction model is obtained;

Step 7: Input the drug to be predicted into the prediction model to obtain the prediction result.

2. The drug activity prediction method based on deep learning as claimed in claim 1, wherein in said step seven, the development set and the test set are first processed through steps two to six, and poured into the prediction model to obtain the test result.

3. the drug activity prediction method based on deep learning as claimed in claim 1, is characterized in that, described step 1 comprises the steps:

1.1 The drug activity data set is divided and disrupted, including 80% of the training set, 10% of the development set and 10% of the test set, and the development set and the test set are fixed for control; Ensure that the data of the training set, development set and test set are evenly distributed in the data set;

1.2 Mark the molecules that have an impact on the receptor in the data set as 1, that is, as a positive sample, and mark the molecules that have no influence as 0, that is, a negative sample, remove the null value without data, and remove the interference data to improve the accuracy.

4. the drug activity prediction method based on deep learning as claimed in claim 1, is characterized in that, in described step 2, extracts atomic feature to the molecule of training set, simultaneously the molecular structure of training set is transformed into adjacency matrix:

2.1 Extract statistical features from molecular data: ['C', 'N', 'O', 'S', 'F', 'Si', 'P', 'Cl', 'Br', 'Mg',' Na', 'Ca', 'Fe', 'As', 'Al', 'I', 'B', 'V', 'K', 'Tl', 'Yb', 'Sb', 'Sn' , 'Ag', 'Pd', 'Co', 'Se', 'Ti', 'Zn', 'H', 'Li', 'Ge', 'Cu', 'Au', 'Ni',' Cd', 'In', 'Mn', 'Zr', 'Cr', 'Pt', 'Hg', 'Pb', '=', '+', '-', '(', ')' , '/', '\', '[', ']', '@', '#', 'Unknown'], the above features ignore numbers, decimal points, and get a dictionary containing all statistical features in the molecule, dictionary values is the number of occurrences of the molecule or the corresponding character of the molecule;

2.2 Extract the degree of the atom in the molecule, ranging from 0 to 10, and the atom degree is defined as the number of atoms directly connected to the atom;

2.3 Extract the number of implicit high spins in the molecule, ranging from 0 to 6, and the angular momentum possessed by the nucleus is called the spin of the nucleus;

2.4 Extract the formal charge of atoms in the molecule;

2.5 Extract the number of free electrons of atoms in the molecule;

2.6 Whether the extracted molecule is an aromatic compound;

2.7 Represent all molecules as a structural graph by treating the atoms in the molecule as nodes and the chemical bonds as edges in the undirected graph, and generate a molecular structure graph represented by an adjacency matrix. The adjacency matrix takes all the atoms in the molecule as a matrix row and The label of the column, when two atoms in the molecule are chemically bonded, the value of the corresponding position in the matrix is 1.

5. the drug activity prediction method based on deep learning as claimed in claim 1, is characterized in that, described step 3 comprises the steps:

3.1 The input x is divided into two parts, one is the atomic feature of the molecule, and the other is the adjacency matrix converted from the molecular structure, and x is a matrix formed by combining the atomic feature of the molecule and the adjacency moment converted from the molecular structure;

3.2 For the real value of the output y, the array [1,0] represents 0, the array [0,1] represents 1, and the result of each training and test is an array [a,b], a,b are two probability values , a+ b=1; a and b represent the probability that the actual value of the output y is the array [1,0], and the other represents the probability that the actual value of the output y is the array [0,1];

3.3 The prediction model uses a five-layer graph convolutional neural network. The graph convolutional neural network has two basic features: one is that each node has its own feature information; the other is that each node in the graph also has structural information; the following formula is the calculation formula of graph convolution, and the central node of graph convolution is v:

u: represents the neighbor nodes of the central node v; h _conv (v): represents the graph convolution feature value of the central node v and node u; M: represents the set of all nodes in the graph convolutional neural network;

Indicates the characteristic parameters, a value will be preset, all of which are 1, and the parameters will be continuously updated during the training process;

σ: represents the pooling function;

Assume

Equation (1) transforms the feature of one edge of the central node v into h _conv (v), and then accumulates the h _conv (v) of all neighbor nodes u, which is the graph convolution of the central node v;

h _conv (G) = [h _conv (v ₁ ), h _conv (v ₂ ), h _conv (v ₃ ), ...] (2)

h _conv (G) represents the collection of h _conv (v) of the currently calculated drug molecule, and G represents the currently calculated molecule G;

Finally, the set of graph convolutions of all nodes v in the molecule is obtained, which is the set of molecular structural features.

6. the drug activity prediction method based on deep learning as claimed in claim 4, is characterized in that, in described step 5, graph convolution process is as follows:

5.1.1 Traverse all nodes in the molecular structure diagram;

5.1.3 Set the central node of the graph convolution to v;

5.1.4 Traverse all the neighbor nodes u of the central node v, and establish a relational dictionary d;

5.1.5 Transform the feature of node u into u′:

in, Indicates the characteristic parameters, a value will be preset, all of which are 1, and the parameters will be continuously updated during the training process;

5.1.6 Add all u';

5.1.7 Return the characteristics of the central node v;

The pooling process is as follows:

5.2.1 Maximum pooling of neighbor nodes u′;

5.2.2 Return the graph convolution feature h _conv (v) of the central node v;

The full connection process is as follows:

5.3.1 Use LSTM to judge whether the graph convolution features of molecules are useful, so as to select useful features;

5.3.2 Connect all the selected useful features and send the output value to the classifier.

7. the drug activity prediction method based on deep learning as claimed in claim 1, is characterized in that, multiple iterative calculations in the step 6, the step of obtaining the model after training is as follows:

6.1 Each time a sample of 128 batchsize is randomly selected from the training set, poured into the model for training, and after the training results are obtained, the gradient descent method is used to optimize the loss function.

As a further improvement, in the third step, the prediction model is a two-category prediction model.

Compared with prior art, adopt the advantage of the present invention as follows:

1. The first and second steps preprocess the data more reasonably, remove the interference data without data, and improve the accuracy of the model; at the same time, adopt a simpler and more effective method for feature extraction, and only calculate atomic features. There is no need to simulate the molecular structure, and the molecular structure is converted into an adjacency matrix, and features are extracted by graph convolution, which greatly reduces time consumption.

2. The third step is to build a more reasonable model. The five-layer graph convolution layer can extract the structural features of molecules more efficiently, while the LSTM layer filters the features to obtain better features.

3. Steps 4 to 7 implement the entire training process and optimize the model. The size of each batch of data for 2000 training sessions is 128, which ensures better optimization of the model while traversing all the training set data. A lower loss function value is obtained.

4. The method of this patent combines graph convolution and LSTM, which greatly reduces the time for feature extraction, and at the same time extracts reasonable and appropriate features for atoms in molecules, without using traditional computational chemistry methods to calculate time-consuming and detailed molecular feature data. It can also obtain more reasonable features that cannot be obtained by traditional methods, so as to achieve better drug activity prediction accuracy under low data volume.

Description of drawings

Fig. 1 is a general flowchart;

Fig. 2 is the adjacency matrix of ethane (C ₂ H ₆ ) molecules;

Figure 3 is a flow chart of LSTM.

Detailed ways

Fig. 1 is the general flowchart of this patent.

The specific technical scheme of this patent is:

The first step is to build a data set:

1.1 Segment and scramble the drug activity data set, including 80% training set, 10% development set and 10% test set, and keep the development set and test set unchanged for control.

1.2 Mark the molecules that have an impact on the receptor in the data set as 1 (positive sample), and mark the molecules that have no influence as 0 (negative sample), remove the null value without data, and eliminate the interference data can significantly improve the accuracy.

1.3 The segmentation of the data ensures that the distribution of the training set, development set and test set is consistent.

The second step is to extract the atomic features of the molecules in the training set, and at the same time convert the molecular structure of the training set into an adjacency matrix:

2.1 Extract statistical features from molecular data: ['C', 'N', 'O', 'S', 'F', 'Si', 'P', 'Cl', 'Br', 'Mg',' Na', 'Ca', 'Fe', 'As', 'Al', 'I', 'B', 'V', 'K', 'Tl', 'Yb', 'Sb', 'Sn' , 'Ag', 'Pd', 'Co', 'Se', 'Ti', 'Zn', 'H', 'Li', 'Ge', 'Cu', 'Au', 'Ni',' Cd', 'In', 'Mn', 'Zr', 'Cr', 'Pt', 'Hg', 'Pb', '=', '+', '-', '(', ')' , '/', '\', '[', ']', '@', '#', 'Unknown']. The above features contain common elements as well as symbols representing special valence bonds, parentheses, special molecules, ions, etc. Numbers, decimal points are ignored. Get a dictionary containing all the statistical features in the molecule, and the dictionary value is the number of occurrences of the molecule or character;

2.2 Extract the degree of the atom in the molecule, ranging from 0 to 10, and the atom degree is defined as the number of atoms directly connected to the atom;

2.3 Extract the number of implicit high spins in the molecule, ranging from 0 to 6. The angular momentum of the nucleus is called the spin of the nucleus, which belongs to the important quantum mechanical properties of the nucleus.

2.4 Extract the formal charge of atoms in the molecule. The formal charge is introduced when writing the Lewis structural formula of the covalent compound in order to judge the stability of each possible species.

2.5 Extract the number of free electrons of atoms in the molecule. Free electrons refer to electrons that are not confined inside a certain atom. The amount of free electrons will affect the electrical conductivity and thermal conductivity of the material.

2.6 Whether the extracted molecule is an aromatic compound, an aromatic compound is a compound with a benzene ring structure, which has a stable structure, is not easy to decompose, and has strong toxicity.

2.7 Characterize all molecules as graphs by treating atoms as nodes and bonds as edges in an undirected graph, generating a molecular topology represented by an adjacency matrix that takes all atoms in a molecule as labels for the rows and columns of the matrix , when two atoms in the molecule are connected by a chemical bond, the value of the corresponding position in the matrix is 1. Figure 2 shows the adjacency matrix form of ethane (C ₂ H ₆ ) molecules

The third step is to build a prediction model (two-category prediction model), including five layers of graph convolution and one layer of LSTM:

3.1 Input x is the adjacency matrix converted from the atomic features of the molecule and the molecular structure;

3.2 For the real value of the output y, use [1,0] to represent 0, [0,1] to represent 1, the result of each training and test is an array [a,b], a,b are two probability values, a +b=1;

3.3 The prediction model uses a five-layer graph convolutional neural network. The graph convolutional neural network has two basic features: one is that each node has its own feature information; the other is that each node in the graph also has structural information; the following formula is the calculation formula of graph convolution, and the central node of graph convolution is v:

u: the neighbor node of the center node v; h _conv (v): the graph convolution feature value of the center node v and node v, M: the set of all nodes in the graph convolutional neural network;

The characteristic parameters will preset a value, all of which are 1, and the parameters will be continuously updated during the training process;

σ: pooling function;

Equation (1) converts the feature of an edge of node v into h _conv (v), and then accumulates h _conv (v) of all neighbor nodes u, which is the graph convolution of node v;

h _conv (G) = [h _conv (v ₁ ), h _conv (v ₂ ), h _conv (v ₃ ), ...] (2)

h _conv (G) represents the set of currently calculated molecules h _conv (v), and G represents the currently calculated molecule G;

Finally, the set of graph convolutions of all nodes v in the molecule is obtained, which is the set of molecular structural features.

3.4LSTM (long short-term memory network):

The difference between LSTM and RNN is that it adds a "processor" to the algorithm to judge whether the information is useful or not (Figure 3. The middle module).

A repeating module in an LSTM consists of four interacting activation functions (three sigmoids, one tanh): each line in the graph represents a full vector, from the output of one node to the input of the other. As shown in Figure 3, circles represent point-wise operations, such as vector addition, while rectangular boxes represent threshold activation functions. Combining lines means concatenation, and dividing lines means copying content and outputting it to different places.

The structure in the storage unit that manages removal or addition to the unit is called a threshold, and there are three types: forget gate, input gate, and output gate. The threshold consists of a sigmoid activation function and a pointwise multiplication operation. The hidden state of the previous time step, one is sent to the forget gate (input node), one is sent to the input gate, and one is sent to the output gate. For the forward pass, the input gate learns to decide when to pass activations into the memory cell, and the output gate learns when to pass activations out of the memory cell. Correspondingly, for the back pass, the output gate learns when to let errors flow into the memory cell, and the input gate learns when to let it flow out of the memory cell.

Use the input x _t , the output h _t-1 of t-1 times to calculate the forgetting rate Decide whether a feature should be forgotten, 0 means completely forgotten, 1 means remember all.

The fourth step is to train the data obtained in steps 2) and 3).

Step 5: After convolution, pooling, and full connection, the output value is sent to the classifier, the loss function is optimized, and training continues. The specific process is as follows:

5.1 Graph convolution process:

for all nodes v in graph

set k = deg(v)

for u in neigh(v)∪{v}

set d = dist(v,u)

transform features u′=W ^k,d u+b ^k,d

sum all u and apply nonlinearity

return new features for v

That is, 5.1.1 traverse all nodes in the molecular structure diagram;

5.1.3 Set the central node of the graph convolution to v;

5.1.4 Traversing all neighbors u of the central node v, and establishing a relational dictionary d;

5.1.5 Transform the feature of node u into u′:

(As explained in Equation (1))

5.1.6 Add all u';

5.1.7 Return the characteristics of node v;

5.2 Pooling process:

max over u in neigh(v)∪{v}

return new features for v

5.2.1 Maximum pooling of neighbor nodes u′;

5.2.2 Return new features for node v.

5.3 Full connection process:

5.3.1 Use LSTM to judge whether the features are useful, so as to select meaningful features.

5.3.2 Connect all the selected features and send the output value to the classifier

The sixth step is to obtain the trained model after multiple iterative calculations:

6.1 Each time a sample of 128 batchsize is randomly selected from the training set, poured into the model for training, and after the training results are obtained, the gradient descent method is used to optimize the loss function.

The seventh step is to process the development set and the test set through the same feature processing, and inject them into the model to obtain the test results.

The eighth step, the experimental results and their discussion.

8.1 The dataset used in this patent is the Tox21 dataset (Tox21 Data Challenge) [https://tripod.nih.gov/tox21/challenge/], the 2014 Tox21 data challenge aims to help scientists understand how chemicals and compounds destroy biological organs The potential of the tox21 data set is that scientists have shown through toxicological analysis that these chemicals and compounds may have toxic effects on organisms;

8.2tox21 data set contains 8013 possible receptors for 12 human body (NR-AR, NR-AR-LBD, NR-AhR, NR-Aromatase, NR-ER, NR-ER-LBD, NR-PPAR-gamma, SR -ARE, SR-ATAD5, SR-HSE, SR-MMP SR-p53) impact data, each receptor has 8000 pieces of data;

8.3 Experiment 1 of this patent will establish models for these 12 kinds of receptor data, and get 12 predicted results:

Training set test set development set activation function epoch number of features Dev set accuracy Test set accuracy NR-AR 5951 744 744 tanh 2 75 0.961 0.962 NR-AR-LBD 5521 691 691 tanh 2 75 0.972 0.972 NR-AhR 5353 669 669 tanh 2 75 0.922 0.921 NR-Aromatase 4752 594 594 tanh 2 75 0.949 0.949 NR-ER 5052 632 632 tanh 2 75 0.867 0.867 NR-ER-LBD 5612 710 710 tanh 2 75 0.949 0.951 NR-PPAR-gamma 5266 658 658 tanh 2 75 0.953 0.956 SR-ARE 4748 593 593 tanh 2 75 0.828 0.829 SR-ATAD5 5786 723 723 tanh 2 75 0.929 0.938 SR-HSE 5275 660 660 tanh 2 75 0.916 0.912 SR-MMP 4736 592 592 tanh 2 75 0.897 0.866 SR-p53 5527 691 691 tanh 2 75 0.935 0.920

Table 1. Tox21 dataset experimental results

8.4 Experiment 2 of this patent will compare the performance of this model with traditional machine learning models, and compare the results of Experiment 1 with five methods of logistic regression, support vector machine and three Bayesian methods (BernoulliNB, GaussianNB and MultinomialNB) :

This method LR SVM Bernoulli NB Gaussian NB Multinomial NB NR-AR 0.962 0.957 0.961 0.858 0.190 0.884 NR-AR-LBD 0.972 0.968 0.965 0.890 0.158 0.907 NR-AhR 0.922 0.861 0.855 0.810 0.228 0.682 NR-Aromatase 0.949 0.933 0.933 0.879 0.143 0.832 NR-ER 0.867 0.859 0.850 0.790 0.185 0.805 NR-ER-LBD 0.957 0.945 0.938 0.904 0.138 0.885 NR-PPAR-gamma 0.956 0.831 0.804 0.775 0.114 0.912 SR-ARE 0.829 0.815 0.815 0.753 0.202 0.724 SR-ATAD5 0.938 0.945 0.798 0.734 0.141 0.840 SR-HSE 0.916 0.930 0.727 0.926 0.139 0.874 SR-MMP 0.897 0.855 0.818 0.796 0.221 0.740 SR-p53 0.935 0.931 0.931 0.883 0.126 0.842

Table 2. Comparison with traditional methods

In the experiment, the same data processing as the model of this patent is used for all compared machine learning methods to ensure the validity of the experimental comparison. The data in the table are the results of the test set. The data in Table 2 shows that, under the Tox21 data set, this patent is better than the traditional machine learning method in 5 out of the 12 receptor data of the traditional method. Compared with the traditional machine learning method, this patent can obtain better and more stable The result proves that the innovation of this patent on the model is effective.

The above example is only a specific embodiment of the present invention, and its simple transformation, replacement, etc. are also within the protection scope of the present invention.

Claims (8)

1. a kind of pharmaceutical activity prediction technique based on deep learning, which comprises the steps of:

Step 1: building pharmaceutical activity data set, carries out cutting to pharmaceutical activity data set, wherein in pharmaceutical activity data set A part of data are as training set, a part of data as development set, some data is as test set；

Step 2: the molecule to training set extracts atomic features, and adjacency matrix is converted by the molecular structure of training set；

Step 3: building prediction model, prediction model includes five layers of picture scroll product, one layer of LSTM；

Step 4: the data that step 2 and three obtain are trained；

Step 5: by picture scroll product, output valve after full connection, is conveyed to classifier, optimizes loss function, continue to instruct by Chi Hua Practice；

Step 6: by iterative calculation, the prediction model after being trained；

Step 7: drug input prediction model to be predicted is obtained prediction result.

2. the pharmaceutical activity prediction technique based on deep learning as described in claim 1, which is characterized in that the step 7 In, development set and test set are first also passed through into step 2 to six processing, prediction model is poured into and obtains test result.

3. the pharmaceutical activity prediction technique based on deep learning as described in claim 1, which is characterized in that the step 1 packet Include following steps:

Pharmaceutical activity data set is carried out cutting by 1.1, is upset, including 80% training set, 10% development set and 10% test set, Development set and test set are immobilized and are used to compare；Wherein, training set, development set and test set are guaranteed to the cutting of data set Data be uniformly distributed in data set；

1.2 regard molecular labeling influential on receptor in data set as positive sample for 1, and the label of no influence is i.e. negative sample This, the not null value removal of data rejects interference data and improves accuracy.

4. the pharmaceutical activity prediction technique based on deep learning as described in claim 1, which is characterized in that the step 2 In, atomic features are extracted to the molecule of training set, while converting adjacency matrix for the molecular structure of training set:

2.1 pairs of molecular datas extract statistical nature: [' C', ' N', ' O', ' S', ' F', ' Si', ' P', ' Cl', ' Br', ' Mg', ' Na', ' Ca', ' Fe', ' As', ' Al', ' I', ' B', ' V', ' K', ' Tl', ' Yb', ' Sb', ' Sn', ' Ag', ' Pd', ' Co', ' Se', ' Ti', ' Zn', ' H', ' Li', ' Ge', ' Cu', ' Au', ' Ni', ' Cd', ' In', ' Mn', ' Zr', ' Cr', ' Pt', ' Hg', ' Pb', '=', '+', '-', ' (', ') ', '/', ' ', ' [', '] ', '@', ' #', ' Unknown'], features above is ignored Number, decimal point obtain the dictionary comprising all statistical natures in molecule, and dictionary value is that molecule or molecule correspond to character and go out Existing number；

2.2 extract the degree of the middle atom of molecule, and range is 0~10, and atom degree is defined as and the direct phase of the atom Atom number even；

2.3 extract the quantity of implicit high-spin in molecules, and range is 0~6, the angular momentum that atomic nucleus has be known as it is nuclear from Rotation；

2.4 extract the formal charge of atom in molecule；

2.5 extract the free electron quantity of atom in molecule；

2.6 extract whether molecule is aromatic compound；

2.7 by being considered as node for the atom in molecule and being expressed as all molecules using chemical bond as the side in non-directed graph Structure chart generates the molecular structure indicated with adjacency matrix, and adjacency matrix is using atoms all in molecule as matrix row and column Label, when two atomic ordering keys are connected in molecule, matrix corresponding position value be 1.

5. the pharmaceutical activity prediction technique based on deep learning as described in claim 1, which is characterized in that the step three guarantees Include following steps:

It is two parts that 3.1 input x, which are divided to, first is that the atomic features of molecule, second is that the adjacency matrix that molecular structure is converted to, x are The adjoining square that the atomic features of molecule and molecular structure are converted to is combined into a matrix being converted to；

3.2 indicate 0 with array [1,0] for exporting the true value of y, and array [0,1] indicates 1, every time the result of training and test For an array [a, b], a, b are two probability values, a+b=1；A and b mono- true value for indicating output y are array [1,0] Probability, another indicates that the true value of output y is the probability of array [0,1]；

3.3 prediction models use five layers of figure convolutional neural networks, and there are two essential characteristics for figure convolutional neural networks tool: first is that each Node has the characteristic information of oneself；Second is that each node in figure also has structural information；The following figure is that the calculating of picture scroll product is public Formula, if the central node of picture scroll product is υ:

U: the neighbor node of central node v is indicated；h_conv(v): indicating the picture scroll product characteristic value of central node v and node u；M: table The set of all nodes in diagram convolutional neural networks；b^v: it indicates characteristic parameter, a value can be preset, be all 1, Parameter is constantly updated during training；

σ: pond function is indicated；

If

The feature on a side of central node υ is converted h by formula (1)_conv(v), then by the h of all neighbor node u_conv(v) tire out Add, the picture scroll product of as central node υ；

h_conv(G)=[h_conv(v₁), h_conv(v₂), h_conv(v₃) ...] (2)

h_conv(G) h of the drug molecule currently calculated is indicated_conv(v) set, G indicate the molecule G currently calculated；

Finally obtain the set of the picture scroll product of all node v in molecule, the as set of molecular characterization.

6. the pharmaceutical activity prediction technique based on deep learning as claimed in claim 4, which is characterized in that the step 5 In, figure convolution process is as follows:

5.1.1 all nodes in molecular structure are traversed；

5.1.3 the central node of setting picture scroll product is v；

5.1.4 all neighbor node u, opening relationships dictionary d of central node v are traversed；

5.1.5 u ' is converted by the feature of node u:

Wherein,b^v: it indicates characteristic parameter, a value can be preset, be all 1, parameter is constantly updated during training；

5.1.6 all u ' are added；

5.1.7 the feature of central node v is returned；

Pond process is as follows:

5.2.1 maximum pond neighbor node u '；

5.2.2 the picture scroll product feature h of central node v is returned_conv(v)；

Full connection procedure is as follows:

5.3.1 judge whether the picture scroll product feature of molecule is useful using LSTM, to pick out useful feature；

5.3.2 all useful features picked out are connected, give output valve to classifier.

7. the pharmaceutical activity prediction technique based on deep learning as described in claim 1, which is characterized in that in the step 6 The step of successive ignition calculates, model after being trained is as follows:

6.1 randomly select the sample of 128batchsize size from training set every time, pour into model and are trained, are trained As a result after, optimize loss function using gradient descent method.

8. the pharmaceutical activity prediction technique based on deep learning as described in claim 1, which is characterized in that the step 3 In, prediction model is the prediction model of two classification.