Molecular Assembly Enumeration vs. 1D Dictionary Compression Across the Nano-DrugBank Small-Molecule Repository
Assembly Theory (AT) defines the Molecular Assembly Index (MA) as the minimal number of recursive bonddisjoint subgraph-joining steps required to construct a 2D molecular graph [1, 2, 4], whereas critics from Algorithmic Information Theory (AIT) contend that empirical MA rankings collapse to standard 1D dictionary compression applied to molecular strings [5, 6, 15]. Here, we evaluate MA against wrapped zlib (RFC 1950), headerless raw DEFLATE (RFC 1951), and header-free 256-ASCII Welch LZW across the Chodera Lab Nano-DrugBank small-molecule repository (N = 5,653, curated from DrugBank 4.5.0) [20–24] using the C++ branchand-bound solver (assemblycpp-v5) [4] under operating-system process isolation on an Apple M3 Ultra workstation. Applying RDKit valence sanitization, parent-fragment desalting (60 salts), stereochemistry-stripped Kekulé MDL V2000 serialization, and a 600 s per-subprocess wall-clock timeout yielded 5,520 resolved records (97.65%; MA in [1, 36], median MA = 14, median subprocess wall time 15.42 ms, mean 3,051 ms; conditional on the inspected two-argument repository runner pending a hash-logged rerun), 35 RDKit valence rejections (0.62%), and 98 computationally unresolved valid graphs (1.73%): 87 with subprocess runtime right-censored at 600 s (median Nheavy = 77) and 11 macrocycles/polypeptides (median Nheavy = 128) explicitly logging the solver's internal Subgraph enumeration limit reached (ENUM_MAX = 50,000,000) diagnostic between 375.8 and 578.2 s. Among resolved records (median Nheavy = 23), 2,429 (44.00%) satisfy MA ≥ 15 [1, 2]. Per-character compression ratios span more than twofold ranges within fixed-MA slices, and in in-sample linear ordinary least-squares (OLS) models adjusting for heavy-atom count (Nheavy), compressed lengths explain 38.6% (DEFLATE: partial r = +0.6212) and 32.0% (Welch LZW: partial r = +0.5656) of Nheavy-adjusted MA variance on stereochemistrystripped canonical parent SMILES (38.4% and 31.7% across n = 5,274 unique parent strings; 42.9% and 38.8% on Kekulé charge/isotope-stripped V2000 graph SMILES, or 42.3% and 38.1% across n = 5,254 unique keys). Across all tested specifications, 57.1%–68.3% of in-sample Nheavy-adjusted linear MA variance remains unexplained, showing that the evaluated 1D SMILES compressors and 2D MA are positively correlated yet not linearly interchangeable.
Authors
- Chun Tang (ORCID: https://orcid.org/0009-0008-4128-2071)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-20
- DOI
- https://doi.org/10.26434/chemrxiv.15012530/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint