연구 기록

We assess cross-model reproducibility over structure panels for the MMP-1 Zn²⁺ active site without asserting compound identities or potency. Three-NNP comparison yields a minimum pairwise Pearson r = 0.9142 (bootstrap 95% CI [0.826, 0.971]; leave-one-out ±0.0104); MACE-OMol25 and Orb-v3 are highly correlated (r = 0.9992) and therefore constitute effectively two independent engines. An upstream GFN2-xTB OPT pre-relaxation reduces the σ outlier from 14.27 to 0.0069 kcal/mol — a 2,039-fold reduction — and structure generation yields a 93.9% PoseBusters v2 pass rate. We report conformal calibration and coarse bound/unbound classification as reproducibility-focused computational assessments; no inhibitor potency or affinity ranking is claimed. Correction review is active; the linked Zenodo v6 record remains the current public version. Sole author. In silico stage; no clinical efficacy claim.

Machine-learning interatomic potentials (MLIPs) are widely assumed to dominate semi-empirical alternatives on molecular tasks. We test this assumption on the conformational energy landscape of fifteen Zn²⁺-binding matrix metalloproteinase-1 (MMP-1) inhibitors drawn from ChEMBL, generated by Boltz-2x cofold sampling with physicality steering. Across thirteen independent diffusion replicates, we evaluate ten energy functions on 100 conformers per ligand per replicate (253,500 conformer single points total). The headline finding is the OMol25 paradox: the same Orb-v3 architecture trained on Meta FAIR's OMol25 molecular dataset drops to r = 0.374 ± 0.025 against GFN1-xTB, versus r = 0.886 ± 0.009 for Orb-v3 OMat — a 0.512 Pearson gap with ≈ 68σ statistical significance across 13 replicates. The v5h revision (2026-05-15) adds a 17-cycle pipeline timing reproducibility table and a chain × concurrent-ADMET fair-share linear regression model. Cluster placement is set by training-data domain, not by architecture. We recommend materials-trained or broadly-trained MLIPs over molecular-only-trained MLIPs for Zn²⁺ metalloenzyme conformer-ensemble work; higher-level DFT validation is the next validation tier.

Boltz-2 cofold sampling for protein-ligand affinity prediction occasionally produces catastrophic outliers — single conformers with unphysical energies that distort downstream affinity ranking. We evaluate six protocol variants on the canary ligand CHEMBL94487 and the full MMP-1 hydroxamate cohort (15 inhibitors × 100 samples each), comparing standard Boltz-2 vs. community-fork bug-patched versions, with and without the --use_potentials physicality steering flag. The full-cohort outlier rate drops from 0.067 % (standard Boltz-2 without flag) to 0.000 % (standard Boltz-2x with --use_potentials, σ_filt = 4.29 kcal/mol). The community fork's metal-ion bug fix alone (without the flag) does not eliminate outliers and shows σ_filt = 6.66 kcal/mol. We recommend standard Boltz-2 + --use_potentials as the publishable protocol for Zn-metalloenzyme cofold work.

De novo enzyme and binder design for Zn²⁺ metallohydrolases (MMP, HDAC, carbonic anhydrase families) requires a sequence-design step that respects the metal-coordination geometry of the active site. We benchmark the canonical four-stage open-source pipeline (RFdiffusion3 backbone generation → LigandMPNN sequence design → FlowPacker side-chain packing → AlphaFold3/Boltz-2x cofold validation) on 1HFC matrix metalloproteinase-1 catalytic domain. The headline finding: LigandMPNN doubles Zn-coordinating residue native recovery from 46.4 % (plain ProteinMPNN) to 95.3 %, with structural-Zn-triad recovery jumping from 0 % to 90.6 %. ESM-C 600M zero-shot pseudoperplexity confirms the LigandMPNN designs as more native-like (2.85 vs 3.03). The pipeline assembly and per-stage failure modes are documented for community reproduction.

This technical record documents absolute binding free energy (ABFE) pipeline components, including thermodynamic-cycle closure across complex- and solvent-decoupling legs, ligand restraints, and analytical standard-state correction. Calibration and quantitative potency-assessment results are in correction review; the linked Zenodo record remains the current public version.

Autonomous virtual-screening pipelines suffer a recurring failure mode: cheap predictors (Boltz-2 affinity, ADMET, xTB conformer scoring) accumulate thousands of candidate scores, yet expensive validation tiers (long molecular dynamics, absolute binding free energy, wet-lab) advance unevenly because the scheduler cannot reason simultaneously about cost, information value, gate-policy, and runtime availability of each…

Topical (transdermal) drug delivery is the dominant route for dermatological therapeutics, yet most AI drug-discovery pipelines omit physiologically-based pharmacokinetic (PBPK) modeling for skin. We present a fully open-source pipeline combining: (1) Dancik 4-layer skin PBPK (stratum corneum → viable epidermis → dermis → systemic); (2) LightGBM logKp head trained on SkinPiX (n=2,326, OECD 428) + NPASS 2026 ADME reco…

Modern AI-driven drug discovery is built on a heterogeneous stack of open-source tools. Specialized application domains require tool selection, integration, and adaptation. We describe Genesis_Medicine, an open-source pipeline for AI-driven Korean traditional medicine (한약·생약) drug discovery. We honestly distinguish: (i) a 7-tool active core that produces all real pipeline outputs in this work (Boltz-2 cofold, REINVEN…

Tight-binding semi-empirical quantum chemistry (xtb-GFNn-xTB) has become a workhorse triage tool for ranking large compound libraries ahead of more expensive structure-based scoring. Whether the resulting ranks are robust to the methodological choices an analyst makes — the GFN parameter set, the conformer-ensemble size, and the implicit solvent model — is rarely tested at scale. We benchmark these three robustness a…

This technical record examines limitations of fixed-charge ZAFF-AMBER absolute binding free energy calculations for zinc metalloenzymes. The correction review covers quantitative accuracy and the real-activity-based ABFE failure panel; the linked OSF record remains the current public version.

Embelin (2,5-dihydroxy-3-undecyl-1,4-benzoquinone) is the principal bioactive of Embelia ribes Burm.f. (Ayurvedic Vidanga; East Asian 자단), an Ayurvedic and East Asian traditional-medicine plant with documented anti-fibrotic activity in liver and pulmonary models but no published investigation in skin fibrosis. We present an in silico case study in which embelin serves as the scaffold-hop seed for an AI-augmented …

Embelia ribes Burm.f. (Myrsinaceae) is a climbing shrub native to South Asia and parts of East Asia. Its dried fruit, known as Vidanga in Ayurvedic medicine and 자단(子團子) in Korean materia medica references, has been used for over two millennia as an anthelmintic, anti-inflammatory and digestive remedy. The principal bioactive constituent is embelin (2,5-dihydroxy-3-undecyl-1,4-benzoquinone), present at 4–5% of dry w…

Topical hyperpigmentation disorders (melasma, post-inflammatory hyperpigmentation, solar lentigines) are mediated by the tyrosinase (TYR) / TYRP1 / DCT (TRP-2) melanin-synthesis network and the master transcription factor MITF. Korean traditional medicine maintains a long-standing repertoire of 비백제 (depigmenting) preparations centered on Glycyrrhiza uralensis (감초), Camellia sinensis (녹차), Morus alba (상백피), *B…

We report the discovery of a pterocarpan-vinyl-polyphenol scaffold family that acts as a universal molecular template across six independent skin-disease verticals (scar regeneration, pigmentation, alopecia, acne, photoaging, fibrosis cross-disease). Six members of this family were identified through six rounds of Bayesian active learning (R9–R14, 4,597 cofold predictions integrated, 14 protein targets, 200-candidate…

R16 optimized the R15 chromanol fragment toward a topical lead hypothesis by increasing skin-window compatibility while preserving a compact chromanol core. We evaluated six chloro/dimethyl analogs across TGFB1, DCT, and TYR using Boltz-2 cofolding, an 18-pair 30 ns OpenMM stability matrix, two 60 ns robustness panels, and 100 ns plus 200 ns anchor-triad follow-up. The top cofold row was r16_03_tgfb1 (R15_chromanol_C…

R15 generated a compact chromanol fragment, OCC1COc2cc(O)ccc2C1, as a safety-first derivative of the broader pterocarpan-vinyl-polyphenol scaffold program. The central question is whether this fragment should be treated as a topical lead or as a systemic-safety fragment hypothesis. The answer is deliberately split. In ADMET/xTB triage, the R15 parent is predicted to be clean for AMES, DILI, and hERG liability, but it…

DOI issued; manuscript metadata is the current public record.

Androgenetic alopecia (AGA) is driven by androgen-pathway enzymes (SRD5A1/2 catalyzing testosterone → DHT), the androgen receptor (AR), and the hair-follicle-cycle Wnt / β-catenin axis. Korean traditional medicine documents hair-vitalization preparations centered on Polygonum multiflorum (하수오; emodin, physcion), Astragalus membranaceus (황기; astragaloside IV), and Panax ginseng (인삼; ginsenosides Rg1, Rb1, Saponi…

Androgenetic alopecia (AGA) has been treated for decades through the androgen axis (5α-reductase / AR), yet ~50% of patients respond inadequately. A 2026 Nature Communications report identified connective tissue sheath (CTS) hypercontractility mediated by PIEZO1 mechanosensation and myosin-light-chain kinase (MLCK) as a non-androgen cause of follicular miniaturization, with ML-7 (MLCK inhibitor) restoring hair grow…

Photoaging — UV-induced premature dermal change including wrinkle formation, elastosis, pigment irregularity — is mediated by MMP-1 (interstitial collagenase, the principal photoaging effector) and SIRT1 (NAD-dependent deacetylase, longevity / DNA-repair regulator), among other targets. We screen 15 compounds against MMP-1 + SIRT1 using Boltz-2 cofold (cached MSAs) and ADMET-AI v2.0.1. FBN1, mTOR, and elastin were NO…

Inflammatory acne pathogenesis includes androgen-driven sebaceous activity (SRD5A1/2 + AR + SREBP1), ductal hyperkeratinization, Cutibacterium acnes colonization with virulence factor expression (RoxP, GehA, sortase), and inflammatory cascades. Korean traditional medicine documents anti-acne preparations centered on 황련 (Coptis chinensis; berberine, palmatine), 황금 (Scutellaria baicalensis; baicalein, baicalin, w…

Pathological fibrosis across organs — skin scarring, idiopathic pulmonary fibrosis (IPF), systemic sclerosis, renal fibrosis, hepatic fibrosis — is widely framed in medicinal-chemistry literature as sharing a converging TGF-β / Smad / MMP / CTGF / collagen-deposition master-switch network. We investigated whether this conceptual axis-sharing translates to evidence-based cross-disease applicability for an AI-derived m…

Korean traditional medicine texts (동의보감 1610, 향약집성방 1433) catalog hundreds of herbal monographs with empirical clinical use across centuries, but systematic cross-reference between these monographs and contemporary molecular target classes is fragmentary. We assemble a scaffold-level atlas mapping 동의보감 / 향약집성방 monograph entries to modern phytochemistry and target classes via 60 ns molecular dynamics evidence. Hypothesis-generating only; no clinical efficacy claim.

Pharmacogenomic (PGx) variation drives substantial inter-individual differences in drug response. The Korean population exhibits distinctive allele frequencies relative to the Caucasian-dominated reference data underlying many published PGx panels — for example, CYP2D6\10 (intermediate metabolizer, ~50% frequency in Koreans vs ~3% Caucasian [1]) and HLA-B\15:02 (severe cutaneous adverse-reaction risk, frequency ~0.…

Korean traditional medicine maintains a centuries-old chronotherapeutic framework — 자오류주(子午流注, Cha-Oh-Ryu-Ju) twelve-meridian timing — in which the twelve principal meridians (간/담/심/소장/비/위/폐/대장/신/방광/심포/삼초) are assigned 2-hour windows of peak qi flow (시진, sì-jin) across the 24-hour cycle. Modern circadian biology has revealed substantial molecular evidence for cell-autonomous and tissue-level circadian rhythms, includ…

Korean medicine (한의학) clinics face a structural gap between traditional pattern-based diagnosis (변증) and modern personalized molecular medicine. We describe an integrated end-to-end workflow assembled across three affiliated entities — HAN PREDICT, Inc. (AI healthcare technology platform), Genesis_Medicine Lab (in silico natural-product drug discovery), and Recover Korean Medicine Clinic (skin-regeneration practice o…