MGOL: Molecule Generation With Ordering Loss.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Pharmaceutical drug discovery is a lengthy and expensive process.
• Machine learning and bioinformatics are accelerating drug discovery.
• Current methods use neural networks to optimize molecular representations, but struggle with unknown property formulations and limited data.

Purpose of the Study:

• To propose a novel loss function for optimizing chemical properties in molecule discovery.
• To address limitations of existing methods in handling unknown property functions and data scarcity.

Main Methods:

• Developed a new 'ordering loss' function that enforces molecular ordering based on property values.
• Compared the performance of the ordering loss against conventional mean squared error (MSE) for property optimization.
• Utilized neural networks for optimizing latent space vectors representing molecules.

Main Results:

• The ordering loss significantly outperforms MSE in optimizing black-box property functions.
• The proposed method effectively captures chemical property variations by mirroring patterns of black-box functions.
• Demonstrated superior performance under data scarcity constraints.

Conclusions:

• The ordering loss establishes a new state-of-the-art framework for molecule discovery optimization.
• This approach offers a more effective way to optimize molecular properties, especially when data is limited.
• Shows potential for broader applications in black-box function optimization.