Chemotype-Symmetric Junction-Tree VAE for SARS-CoV-2 Mpro Inhibitor Design
Chidhvilas Tanay Mogalluru · Zenodo (CERN European Organization for Nuclear Research) · 2026
The main protease (Mpro) of SARS-CoV-2 is inhibited by covalent drugs such as nirmatrelvir, whose activity comes from a warhead that bonds to the enzyme. Designing safer analogs is hard because standard drug-likeness objectives optimise the warhead away, and the fused-aromatic scaffolds involved are where most generative models fail. This work presents a framework that improves predicted safety while preserving the warhead and affinity. Built on a junction-tree variational autoencoder, it encodes each molecule, represented as a tree of chemical fragments, into a continuous latent space and generates analogs by decoding points near a known inhibitor. A chemotype-symmetric vocabulary covers the Mpro chemotypes evenly, and a latent-conditioned attachment scorer reassembles fused-ring chemistry, reconstructing 19 of 20 targets faithfully. Over this latent space, NSGA-III, selected by benchmarking, runs a constrained six-objective search including lipophilicity and warhead reversibility, with potency from a deep-ensemble surrogate whose disagreement gates it away from unfamiliar chemistry. From twenty anchors across five chemotype classes it yields 1,013 analogs, retaining the source warhead in 85% of the parents that carry one. Predicted mutagenicity tracks the warhead class, the reversible nitrile as clean as non-covalent chemistry. Its final Pareto front of 138 analogs is predicted as safe as or safer than each source inhibitor on four of eight ADMET endpoints and at parity on mutagenicity, keeps the warhead in 94% of members, and binds within Vina's scoring error of the parents. Its strongest candidate, a dual-warhead analog of nirmatrelvir, is predicted safer than the drug on seven of eight endpoints. This safety advantage, without erasing the warhead, is the central result, one that objectives penalising the warhead cannot reach without discarding the mechanism.