MLP-Based Regression Prediction Model For Compound Bioactivity

Area of Science:

• Oncology
• Pharmacology
• Computational Chemistry

Background:

• Breast cancer development is strongly associated with the estrogen receptor alpha (ERα).
• ERα is a significant therapeutic target for breast cancer treatment.
• Antagonizing ERα activity presents a potential drug development strategy.

Purpose of the Study:

• To develop a predictive model for screening potential anti-breast cancer compounds.
• To identify key molecular descriptors influencing biological activity.
• To validate a machine learning approach for drug discovery.

Main Methods:

• Utilized LASSO regression with 10-fold cross-validation to screen 1974 compounds and identify top molecular descriptors.
• Constructed a Multi-Layer Perceptron (MLP) fully connected neural network for bioactivity prediction.
• Employed Mean Squared Error (MSE) as the loss term to evaluate model validity.

Main Results:

• Identified the top 20 molecular descriptors significantly impacting biological activity.
• Developed an MLP-based regression model with a low validation loss of 0.0146, indicating effective training.
• Successfully predicted the bioactivity values of 50 new compounds.

Conclusions:

• The developed MLP model demonstrates a valid and efficient strategy for predicting compound bioactivity.
• This computational approach can aid in the efficient development of novel anti-breast cancer drugs.
• The methodology offers a valuable reference for future drug discovery efforts targeting ERα.