Session: COMP Poster Session:

In this work, we want to address the problem of producing novel molecules as 2D molecular graphs and as low-energy 3D conformations. Deep generative models can be classified by the types of molecular representation generated. One way to represent a molecule is by a string of characters called SMILES string. The main weaknesses of the SMILES string-based model are low interpretability and inability of the model to produce 100% valid SMILES. Another representation is the molecular graph. Unlike SMILES strings, molecular graphs contain explicit information about atoms connectivity. These models are more interpretable and yield a 100% validity rate for produced molecules. The most complicated representation is 3D that encodes not only the 2D topology of a molecule but also the 3D conformation of atoms in space. Accurate prediction of stable 3D conformations for a molecule is a classical problem in computational Chemistry and has wide applications in drug discovery. While there have been many works on SMILES strings and molecular graphs generation, generative models for 3D molecules are just emerging now. However, this representation is the most informative and explicit. 3D generative models have great future potential for generating molecules with desired geometries, such as ones with the desired pharmacophore or that satisfy a specific Schrodinger equation solution.
We introduce a generative model called MolecularRNN and its extension to 3D case called MolecularRNN 3D. This model generates both novel molecules and low energy conformations. MolecularRNN is an auto-regressive, meaning that molecules are generated atom-by-atom, connecting the newly generated atom to the previously generated ones. Starting from a single carbon atom, at every time step, the model unrolls the new molecule by predicting next atoms, bonds and torsion angles in 3D case. Unlike models previously proposed in the literature, MolecularRNN 3D does not have any restrictions on molecule size and tackles both problems of novel molecules and 3D conformations generation. We also used MolecularRNN and MolecularRNN 3D for the goal-directed generation of molecules with Reinforcement Learning. In more detail, we combined proposed generative models with QSAR models and/or shape-based similarity scores such as ROCS to produce virtual hits for CDK1 kinase.

Prof Olexandr Isayev

Carnegie Mellon University

Appearances