LACAN: Leveraging adjacent co-occurrence of atomic neighborhoods for molecular scoring and generation Authors

Read full text Summarise with VIRAT
Authors: Wim Dehaen
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

LACAN (Leveraging Adjacent Co-occurrence of Atomic Neighborhoods) is a bond-level statistical scoring function and filter based on adjacent atomic environment co-occurrence. When integrated into evolutionary search it enables chemically constrained molecular generation. It scores molecules by asking: how likely is each bond, given the chemical environments on both sides of it? Some common molecular fragments co-occur next to each other less than expected by random chance. We leverage this phenomenon for a data-driven method that assesses the extent to which the local environment around each bond in a molecule conforms to the norms of a reference data set such as ChEMBL. This approach is based on co-occurrence patterns of circular atomic environments centered on the two atoms involved in the bond. We show LACAN can be effectively used as a transparent and interpretable filter for certain unstable and uncommon molecular functionalities such as haloamines and peroxides, and we compare it to other commonly used molecular filters, such as SAScore, QED and MolSkill. We show this filter can be used for molecular mutations and molecular breeding, while providing an effective railguard against unfeasible structures. We apply these operations in an evolutionary algorithm and demonstrate good performance in various generative tasks. Scientific contribution: We present a novel molecular filter which filters atypical bonds using co-occurrence patterns of adjacent atomic environments. We show it enables the application of genetic operations like mutation and crossover with more chemically reasonable outcomes than known genetic algorithms. We present a lightweight molecular generator with a performance similar to state-of-the-art methods.

Fields of Study & Subjects

Computational Drug Discovery Methods Machine Learning in Materials Science History and advancements in chemistry

Publication details

Journal / source
ChemRxiv
Publication date
24 March 2026
Language
EN
Document type
Preprint