Random forest model training method for classification imbalance data optimization

Abstract

The invention relates to a random forest model training method for classification imbalance data optimization. The method comprises the following steps: S1, calculating descriptors and fingerprints of small molecule compounds; S2, constructing a sampling-machine learning process, and determining candidate parameters; S3, constructing a Bayesian optimization process; and S4, verifying a Bayesian optimization result, and giving a model. The training set generation process aims at a classification imbalance problem and comprises an oversampling method, an undersampling method and the like; random forest model training needs to determine the maximum feature number, the maximum decision tree number, the leaf node sample number, decision tree branch rules (Gini coefficients, information entropy and the like), whether samples are endowed with weights or not and the like.According to the invention, there is a cooperative relationship between successive training set generation and random forest model training, namely, a specific training set generation process corresponds to specific random forest model training, so that required parameters are determined by Bayesian optimization in an integral unified manner.

Description

technical field [0001] The invention relates to the technical field of machine learning in drug small molecule screening, in particular to a random forest model training method optimized for unbalanced classification data. Background technique [0002] Drug development is a process with huge economic costs. According to statistics in 2014, each new drug costs more than 2 billion US dollars on average. Development has a failure rate of up to 90%, with 50% of failures attributed to effectiveness and 30% to security. [0003] In the field of small molecule drugs, confirming drug activity is also costly. In typical cases, after thousands of small molecules have undergone complex demonstrations, only a few can become drug candidates. In addition, the failure rate of designing new compounds is often very high. Therefore, it is of great significance for drug development to simply and quickly select promising small molecules from existing drug-like compounds and reduce unnecessary ...

AI Technical Summary

Problems solved by technology

[0009] Aiming at the deficiencies of the prior art, the present invention provides a random forest model training method optimized for unbalanced classification data, which is convenient for human understanding, less affected by unbalanced classified data, and the number of features is higher than the number of samples of a specific category. Small and other advantages, it solves the problem that the current prediction of small molecule drug activity uses a deep learning model, which is not suitable for human understanding. At the same time, there are orders of magnitude differences between different classifications of drug activity data sets. No matter which classification model, the samples are defaulted to The tendency to enter large sample size types, the result is that although there is a high correct rate, the errors are concentrated in small sample size types, and the number of features is at least an order of magnitude, or even dozens of times, greater than the number of samples of a specific category. The ratio of the number of features to the number of samples is likely to cause overfitting problems that are difficult to troubleshoot

Images