Comprehensive Chemical Space Exploration in De Novo Drug Design: An Algorithmically Constrained, Parallel Dual-Paradigm Framework Validated on BACE1
- 1IDD Research Laboratory
- 2School of Chemistry and Life Science, Hanoi University of Science and Technology, 1 Dai Co Viet, Hanoi
Abstract
Isolated de novo drug design paradigms carry inherent vulnerabilities: ligand-based (LBDD) models risk unreliable extrapolation beyond their training domain, while structure-based (SBDD) genetic algorithms accumulate molecular size-bias across evolutionary generations. Furthermore, on structurally complex targets, deploying either paradigm in isolation risks systematically excluding viable chemical space. Here, we introduce a dual-paradigm framework addressing both failure modes through two novel algorithmic constraints: a Zoned Applicability Domain (AD) integrated as a live reinforcement learning reward gradient, and the Size-Independent Ligand Efficiency (SILE) metric deployed as a generational fitness function. A convergent validation cascade - pharmacokinetic filtering, exhaustive docking, and molecular dynamics - is architecturally positioned to eliminate artifacts these constrained engines may still produce. We evaluated this workflow on β-secretase (BACE1), deliberately selected because its electrostatic catalytic dyad and hydrophobic sub-pockets present divergent reward landscapes. Ablation studies confirmed the Zoned AD suppressed out-of-domain artifacts to enrich extrapolation, while SILE prevented meaningless molecular bloat compared to baseline models. The LBDD pipeline autonomously rediscovered 10 withheld experimental inhibitors, while the SBDD pipeline successfully decoupled docking affinity from molecular size, with both yielding leads with thermodynamic profiles comparable to the clinical reference verubecestat. Critically, the campaign demonstrated a fundamental algorithmic divergence: LBDD generates data-driven compounds inheriting the binding topology of known inhibitors, while SBDD explores pocket-complementary space with substantially different binding mechanisms. This suggests that parallel deployment of constrained paradigms may be necessary for systematic coverage of a target's viable binding space.
Related
- Prioritizing next-day readiness decline in elite female soccer using readiness, workload, and wellness data: A temporally validated and calibrated monitoring framework
- Predicting Blood–Brain Barrier Permeability from Experimental Data: An Interpretable and Externally Validated Machine Learning Framework
- Orthogonally constrained CASSCF framework: Newton–Raphson orbital optimization and nuclear gradients
- A Symmetry-Constrained Fourier--Morse Framework for Compact Anisotropic Interaction Potentials
- Physics-Constrained Conformal Screening: A Certified Machine Learning Framework for the Rational Design of Dual Glycoepitope-Imprinted Polymers Recognizing CA 19-9 and MUC1-Tn
BibTeX
@misc{dang2026comprehensive,
title = {{Comprehensive Chemical Space Exploration in De Novo Drug Design: An Algorithmically Constrained, Parallel Dual-Paradigm Framework Validated on BACE1}},
author = {Thang T. Dang and Mai N. Luong},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv-2025-bj970/v3},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-bj970/v3},
note = {CC BY},
}
RIS
TY - UNPB
TI - Comprehensive Chemical Space Exploration in De Novo Drug Design: An Algorithmically Constrained, Parallel Dual-Paradigm Framework Validated on BACE1
AU - Thang T. Dang
AU - Mai N. Luong
PY - 2026
DA - 2026/09/16
PB - ChemRxiv
DO - 10.26434/chemrxiv-2025-bj970/v3
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-bj970/v3
AB - Isolated de novo drug design paradigms carry inherent vulnerabilities: ligand-based (LBDD) models risk unreliable extrapolation beyond their training domain, while structure-based (SBDD) genetic algorithms accumulate molecular size-bias across evolutionary generations. Furthermore, on structurally complex targets, deploying either paradigm in isolation risks systematically excluding viable chemical space. Here, we introduce a dual-paradigm framework addressing both failure modes through two novel algorithmic constraints: a Zoned Applicability Domain (AD) integrated as a live reinforcement learning reward gradient, and the Size-Independent Ligand Efficiency (SILE) metric deployed as a generational fitness function. A convergent validation cascade - pharmacokinetic filtering, exhaustive docking, and molecular dynamics - is architecturally positioned to eliminate artifacts these constrained engines may still produce. We evaluated this workflow on β-secretase (BACE1), deliberately selected because its electrostatic catalytic dyad and hydrophobic sub-pockets present divergent reward landscapes. Ablation studies confirmed the Zoned AD suppressed out-of-domain artifacts to enrich extrapolation, while SILE prevented meaningless molecular bloat compared to baseline models. The LBDD pipeline autonomously rediscovered 10 withheld experimental inhibitors, while the SBDD pipeline successfully decoupled docking affinity from molecular size, with both yielding leads with thermodynamic profiles comparable to the clinical reference verubecestat. Critically, the campaign demonstrated a fundamental algorithmic divergence: LBDD generates data-driven compounds inheriting the binding topology of known inhibitors, while SBDD explores pocket-complementary space with substantially different binding mechanisms. This suggests that parallel deployment of constrained paradigms may be necessary for systematic coverage of a target's viable binding space.
ER -