Method for predicting retention time of liquid chromatography under different chromatographic conditions

A method for predicting a retention time of liquid chromatography under different chromatographic conditions, comprising the following steps:

• S1:
• collecting and cleaning experimental data which include SMLIES information of a compound, the chromatographic conditions, and corresponding retention time values, wherein the chromatographic conditions include chromatographic column information, a mobile phase condition, and a running time of a mobile phase;
• S2:
• extracting structural information of a molecule according to the SMLIES information of the compound, and constructing a molecular graph structure based on the structural information of the molecule, wherein the molecular graph structure includes node characteristics, side characteristics, and an adjacency matrix;
• S3:
• obtaining a physicochemical property of the molecule in the compound, and quantifying the physicochemical property to obtain a physicochemical property description vector;
• S4:
• obtaining chromatographic column data, establishing a chromatographic column characteristic database based on the chromatographic column data, and encoding the chromatographic conditions of the compound based on the chromatographic column characteristic database to obtain a chromatographic encoding vector;
• S5:
• constructing a graph neural network model, comprising a graph network layer, a readout layer and a linear layer;
• S6:
• integrating the molecular graph structure, the physicochemical property description vector, the chromatographic encoding vector, and the retention time values of the experimental data into a dataset, and utilizing the dataset to iteratively train the graph neural network model until a loss function converges to obtain a trained graph neural network model;
• and
• S7:
• obtaining the SMLIES information and chromatographic conditions of the compound to be predicted, inputting them into the trained graph neural network model, and outputting and obtaining a retention time predicted result;
• wherein in step S2, the node characteristics comprise an atom type, aromaticity, a heteroatom or not, in a ring or not, the number of hydrogen atoms, and the number of radical electrons;
• the side characteristics include a bond type, conjugate or not, on a ring or not, and a rotatable bond or not;
• and the adjacency matrix is constructed according to a molecular chemical bond of the compound, which is used to represent atoms, chemical bonds and a connection relationship thereof;
• step S6 comprises:
• S61:
• dividing the dataset into a training set, a verification set, and a test set, wherein each training datum in the dataset contains a molecular graph structure, a physicochemical property description vector, a chromatographic encoding vector, and a retention time value;
• S62:
• utilizing the training set to iteratively train the graph neural network, calculating a value of the loss function according to a difference between an output predicted retention time and retention time data, feeding back and adjusting parameters of the graph neural network model according to the value of the loss function until the loss function converges to a stable value, and obtaining the initially trained graph neural network model;
• S63:
• utilizing the verification set to verify the initially trained graph neural network model, performing hyper-parameter tuning for the initially trained graph neural network model on the verification set, and obtaining the verified graph neural network model;
• and
• S64:
• utilizing the test set to perform a model evaluation on the verified graph neural network model, finely tuning the verified graph neural network model according to an evaluation result, and obtaining the final trained graph neural network model;
• in utilizing the dataset to iteratively train the graph neural network model, a training process of the graph neural network model with an individual training datum is as follows:
• inputting the molecular graph structure into the graph network layer, identifying the atoms, the chemical bonds and the connection relationship thereof in the molecular graph structure according to the adjacency matrix, taking the atoms as nodes, performing an information transfer for the nodes to update node information, aggregating updated node information through an activation function, and outputting and obtaining node information data;
• inputting the node information data, the physicochemical property description vector, and the chromatographic encoding vector into the readout layer to perform merging and encoding, and outputting and obtaining encoded data;
• and
• inputting the encoded data into linear layers, mapping the encoded data to a low dimensional space through a mapping function, extracting a target characteristic related to the retention time in the low dimensional space, and projecting the target characteristic to a scalar value to obtain the predicted retention time, wherein there are two linear layers, and a linear rectification function is set between the two linear layers;
• A formula for the information transfer is expressed as follows:
• V
• i
• (
• t
• +
• 1
• )
• =
• σ
• ⁢
• (
• V
• i
• (
• t
• )
• +
• ∑
• j
• ⁢
• ϵ
• ⁢
• N
• ⁡
• (
• i
• )
• h
• i
• ⁢
• j
• (
• t
• )
• ;
• in the formula, i represents a target node, j represents a source node, N(i) represents all source nodes from which the target node i receives the information transfer during an information transfer process, t and t+1 represent a tth transfer and a (t+1)th transfer, Vi(t+1) is node information of the target node i after a (t+1)th update, Vi(t) is node information of the target node i after a tth update, σ is a sigmoid function, and hij(t) represents a sum of information of the source node j and side information between the source node j and the target node i;
• A formula for aggregating the updated node information through the activation function is expressed as follows:
• y
• m
• ⁢
• o
• ⁢
• l
• ⁢
• e
• ⁢
• c
• ⁢
• u
• ⁢
• l
• ⁢
• e
• =
• ∑
• i
• ⁢
• V
• i
• (
• t
• )
• ;
• wherein ymolecule represents the node information data.