Editing-site selection for constrained QED improvement through frozen GenMol remasking
- 1Independent Researcher, Tsahal 32, Ness Ziona, Israel, 7407732
Abstract
Shallow network, boosted decision-tree and graph-attention models were trained on a remaskingsite selection task utilizing the GenMol generator and a subset of bioactive molecules (30,000 training and 5,000 validation) from the ChEMBL 37 dataset. Models learned to optimize fragment selection policies by maximizing the fraction of successful remasked candidates, with success contingent on an increase in the RDKit quantitative estimate of drug-likeness (QED) under several chemical constraints. All three models show a marked increase in generation of at least one successful derivative out of 16 remasking runs on a held-out test set comprising 5,000 bioactive molecules: 66.0%–66.9%, 67.0%–67.4% and 66.3%–66.7% for the shallow network, boosted decision-tree and graph models, compared to 54.8%–55.6% for a fixed policy selecting the largest fragment (one-sided bootstrap 𝑝 < 10−4 for all models). A trivial baseline, selecting fragments randomly, displays 47.3% success. The fraction of valid candidates produced by all policies is about 99%. Analysis of learned policies showed consistent trends among the three models. Selection is biased toward larger pendants. Additionally, pendant selection shows several trends conditioned on the pendant-side, but not core-side, attachment atom. Larger carbon-and oxygen-attached fragments are selected more often (two-sided binomial 𝑝 < 10−10 in every model); a reversed tendency, albeit not statistically significant (𝑝 ≈ 0.02–0.30), is observed for nitrogen attachments. Controlling for fragment size, learned policies show preference for pendant-side nitrogen over oxygen and carbon, and for both chlorine and bromine over fluorine and carbon. Only LightGBM displays a statistically significant (two-sided exact binomial 𝑝 < 0.01) preference for sp2 over sp3 hybridization of pendant-side carbon atoms. At equal size, carbonyl-attached pendants are preferred to hydrocarbons.
Related
- Chemical Editing of DNA-Mediated Nanoparticle Superlattices through a CRISPR-inspired Strategy
- Protein-State Editing through Proximity-Induced Targeted Delactylation
- Ipso-selective pyridine-to-aniline editing through amine-induced fragmentation
- Skeletal Editing through Reprogramming of the Gabriel–Colman Rearrangement Enables Formation of a Ring-Expanded Spirocycle
- Reversible Site-Selective Editing of Atomically Precise Au6 Nanoclusters through Controlled Oxidative Addition/Reductive Elimination
BibTeX
@misc{brenner2026editingsite,
title = {{Editing-site selection for constrained QED improvement through frozen GenMol remasking}},
author = {Tom Brenner},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010340/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010340/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Editing-site selection for constrained QED improvement through frozen GenMol remasking
AU - Tom Brenner
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010340/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010340/v1
AB - Shallow network, boosted decision-tree and graph-attention models were trained on a remaskingsite selection task utilizing the GenMol generator and a subset of bioactive molecules (30,000 training and 5,000 validation) from the ChEMBL 37 dataset. Models learned to optimize fragment selection policies by maximizing the fraction of successful remasked candidates, with success contingent on an increase in the RDKit quantitative estimate of drug-likeness (QED) under several chemical constraints. All three models show a marked increase in generation of at least one successful derivative out of 16 remasking runs on a held-out test set comprising 5,000 bioactive molecules: 66.0%–66.9%, 67.0%–67.4% and 66.3%–66.7% for the shallow network, boosted decision-tree and graph models, compared to 54.8%–55.6% for a fixed policy selecting the largest fragment (one-sided bootstrap 𝑝 < 10−4 for all models). A trivial baseline, selecting fragments randomly, displays 47.3% success. The fraction of valid candidates produced by all policies is about 99%. Analysis of learned policies showed consistent trends among the three models. Selection is biased toward larger pendants. Additionally, pendant selection shows several trends conditioned on the pendant-side, but not core-side, attachment atom. Larger carbon-and oxygen-attached fragments are selected more often (two-sided binomial 𝑝 < 10−10 in every model); a reversed tendency, albeit not statistically significant (𝑝 ≈ 0.02–0.30), is observed for nitrogen attachments. Controlling for fragment size, learned policies show preference for pendant-side nitrogen over oxygen and carbon, and for both chlorine and bromine over fluorine and carbon. Only LightGBM displays a statistically significant (two-sided exact binomial 𝑝 < 0.01) preference for sp2 over sp3 hybridization of pendant-side carbon atoms. At equal size, carbonyl-attached pendants are preferred to hydrocarbons.
ER -