MolGPT

MolGPT#

image-20240805210342420

1.  Methods#

The input of the model

  • Molecular SMILES

  • Scaffold

  • Properties:\(\left\{\begin{aligned}&\log{P}\quad \text{ ( The logarithm of the partition coefficient.)}\\ &SAS \quad \text{ (Synthetic Accessibility score) }\\&TPSA \quad \text{ (Topological Polar Surface Area)}\\&QED \quad \text{ (Quantitative Estimate of Drug-likeness )}\end{aligned}\right.\)

Dataset used

MOSES or the GuacaMol

**MOSES: **A data set composed of 1.9 million clean lead-like molecules from the Zinc data set46 with molecular weight ranging from 250 to 350 Da, number of rotatable bonds lower than 7, and XlogP below 3.5.

**GuacaMol: **A subset of the ChEMBL 24 database that contains 1.6 million molecules.

image-20240808150346615

1.1.  The architecture of Transformer#

image-20240718211213841

Detailed tutorial can be seen in: Transformer.md

2.  Training Procedure and Evaluation Metrics#

Model is trained for 10 epochs using the Adam optimizer with a learning rate of \(6 × 10^{−4}\).

Evaluation metrics :

  • Validity: The fraction of a generated molecules that are valid.

  • Uniqueness: The fraction of valid generated molecules that are unique.

  • **Novelty: **The fraction of valid unique generated molecules that are not in the training set.

  • Internal Diversity (\(\mathbf{IntDiV_p}\) ): Measures the diversity of the generated molecules.

  • **Freshet ChemNet Distance(FCD): **Calculated using the features of the generated molecules and the features of molecules in the data set.

  • KL Divergence: KL divergence between two distributions P and Q for any given property is a measure of how well Q approximates P

3.  RESULTS AND DISCUSSION#

  1. Nonconditioned Molecular Generation.

  2. Generation-based on Single and Multiple Properties.

  3. Generation Based on Scaffold.

  4. Generation Based on Scaffold and Property.

Contents