AnoChem: Prediction of chemical structural abnormalities based on machine learning models

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence in Chemistry

Background:

• *De novo* drug design aims to discover novel compounds efficiently.
• Generative models often produce non-synthesizable or unrealistic molecules.
• Accurate assessment of generated chemical structures is crucial for drug design.

Purpose of the Study:

• To present AnoChem, a deep learning framework for assessing the realism of generated molecules.
• To evaluate the performance of generative models using AnoChem.
• To provide a reliable tool for *de novo* drug design.

Main Methods:

• Developed a deep learning computational framework named AnoChem.
• Trained AnoChem to distinguish between real and generated molecules.
• Evaluated AnoChem's performance using Area Under the Receiver Operating Characteristic Curve (AUC).
• Compared AnoChem with other metrics like SAscore and Fréchet ChemNet distance (FCD).

Main Results:

• AnoChem achieved an AUC score of 0.900 in distinguishing real from generated molecules.
• AnoChem demonstrated a strong correlation with SAscore and FCD metrics.
• The framework effectively validated the performance of various generative models.

Conclusions:

• AnoChem is a reliable computational tool for assessing the chemical realism of generated molecules.
• This framework can significantly aid *de novo* drug design by filtering unrealistic compounds.
• AnoChem enhances the development and evaluation of generative models for drug discovery.