Topology-Aware Generation and Activity-Based Filtering: A Computational-Experimental Framework for Data-Scarce Quaternary Ammonium Compound Discovery

Quaternary ammonium compounds (QACs) are widely used
antimicrobial
disinfectants whose efficacy is threatened by increased bacterial
resistance. Artificial intelligence-guided development of novel QACs
is constrained by historically sparse structure–activity data
and methods to generate novel chemical entities with bioactivity.
This paper presents a comparative experimental study of two computational
workflows designed to accelerate QAC discovery under data-limited
conditions. Both workflows employ a topology-aware variational autoencoder
to generate novel candidates. In Workflow 1, generated QAC structures
were directly subjected to expert evaluation within a fixed time constraint
through the systematic application of chemistry-domain decision criteria.
In Workflow 2, generated candidates were first computationally filtered
using predictive models trained to anticipate antimicrobial activity,
advancing only molecules projected to be highly active against at
least one bacterial strain for expert evaluation. This predictive
filtering enabled the assessment of a larger, higher-quality candidate
pool within the same time constraint. Comparative assessment of the
compound sets resulting from the two workflows revealed substantial
improvements in candidate quality: compounds deemed synthesis-worthy
increased from 9% to 38%, while invalid outputs decreased from 21%
to 0%. Experimental characterization of 29 selected compounds across
both workflows yielded 11 novel QACs with experimentally validated
minimum inhibitory concentrations of 1–32 μM against
four bacterial pathogens. These results demonstrate that topology-aware
generation coupled with computational prefiltering enables systematic
navigation of data-scarce chemical spaces while respecting practical
constraints on expert evaluation time.