Computational design of synthesizable molecules by imitating reference Chemistry

Promovendus/a: Alan Kerstjens

Promotor(en): Prof. dr. Hans De Winter

De novo molecular design is the practice of generating molecules with desirable properties from scratch. When done computationally the proposed molecules tend to be difficult to synthesize and overall chemically unappealing. In this work we present methods to extract patterns from available data and bias molecular design towards synthetically accessible chemistry. Given a list of known synthesizable compounds, we design molecules with the same chemical features, under the assumption that this resemblance increases the likelihood of them being synthesizable as well.

Molecules were designed using evolutionary algorithms that breed populations of molecules by modifying their molecular graphs. The designed molecules were constrained to be composed of chemical features that are prevalent in reference chemistry. Different ways of defining chemical features were explored, and we determined that mimicking small circular atomic environments allowed us to design reasonably fit and easy to synthesize molecules.

We developed an evolutionary algorithm that constructs molecules with desirable chemical features by assembling molecular fragments in a computationally efficient way, and showed how it outperformed competing algorithms in both the quality of the generated molecules and its ability to navigate chemical space effectively.

We also developed a molecule correction algorithm that can identify flaws in molecules and sanitize them to make the molecules more desirable. Said tool can be used to post-process molecules generated elsehow, or integrated into molecule generators to enforce chemical constraints in a hands-off fashion.