Multi-target evaluation of Korean herbal compounds against the SRD5A2 / AR / beta-catenin axis for androgenetic alopecia: Boltz-2 cofold, ChEMBL-anchored ranking, ADMET-AI 107-endpoint safety panel, and pilosebaceous single-cell atlas-constrained prioritization of Saponin Re, Emodin, and Biochanin A

Correction (new version, 2026-07-21) — citation correction. Correction (2026-07-21) - calibration-anchor attribution and metric. Background: two calibration anchors had been conflated across this corpus. The 'Spearman |rho| ~ 0.72' anchor (n=15) was computed against the ligand panel found to be fabricated (2026-07-16 audit) and is void. The separate 'Pearson R = -0.453, n=93' anchor is REAL - it comes from the ChEMBL-derived 93-compound cohort of companion preprint #18 (pilot/cpu_meaningful/chembl_boltz2_calibration.csv, carrying per-record ChEMBL document identifiers) and reproduces on recomputation (Pearson -0.4535, Spearman -0.4582). It was, however, being cited incorrectly. In this record the abstract attributed R = -0.453 to 'the parallel MMP-1 calibration set', which reads as the fabricated panel and is wrong on two counts: the figure comes from preprint #18's real cohort, not from this project's withdrawn panel and not from preprint #8 (which never reported it); and it was measured on Boltz-2 affinity_pred_value, whereas this screen ranks on affinity_prob_binary. On the same 93 compounds affinity_prob_binary correlates with pIC50 at Pearson +0.178 / Spearman +0.048 - essentially no potency signal. The abstract now says so and the rankings are presented as relative and uncalibrated. This is separate from, and additional to, the |rho| ~ 0.72 withdrawal note already carried in the ensemble-validation section from the 2026-07-18 correction. No result, ranking or figure value changes; this is a provenance and scope-of-claim correction. Full audit: preprints/_metadata/DANGLING_CITATION_AUDIT_2026_07_21.md; underlying finding: preprints/_metadata/FABRICATED_PANEL_SCOPE_2026_07_16.md.

Correction (new version, 2026-07-18) — integrity correction. Corrected version (2026-07-18). This preprint contains no fabricated panel of its own; one sentence had cited a within-class MMP-1 calibration of |rho| about 0.72 from companion preprint #8 §3.6 as support for the Boltz-2 screen. That calibration has been withdrawn (its panel carried fabricated potencies), so the citation is removed; the downgrade of the Emodin x AR hit rests on the Chai-1 ensemble disagreement, which is unaffected. The n=93 ChEMBL calibration anchor cited elsewhere in the abstract is a different, uncontaminated cohort and is unchanged. In silico only.

Androgenetic alopecia (AGA) is driven by androgen-pathway enzymes (SRD5A1/2 catalyzing testosterone to dihydrotestosterone), the androgen receptor (AR), and the hair-follicle-cycle Wnt / beta-catenin axis, with emerging evidence for mechanotransduction (PIEZO1 / MLCK) as a non-androgen contributor to follicular miniaturization. Korean traditional medicine documents hair-vitalization preparations centered on Polygonum multiflorum (Hasuo; emodin, physcion), Astragalus membranaceus (Hwanggi; astragaloside IV), and Panax ginseng (Insam; ginsenosides Rg1, Rb1, Saponin Re), among others. We evaluate a 15-compound curated Korean herbal library against the SRD5A2 / AR / CTNNB1 (beta-catenin) three-target panel using a four-layer in silico framework: (i) Boltz-2 protein-ligand co-folding with cached MMseqs2 multiple-sequence alignments (n = 39 successful cofolds, 6 SMILES sanitization or fold failures explicitly listed); (ii) a ChEMBL pIC50 calibration anchor carried over from the parallel MMP-1 calibration set (Pearson R = -0.453, n = 93), establishing that Boltz-2 affinity_probability_binary is a relative-ranking predictor and not an absolute potency estimate; (iii) ADMET-AI v2.0.1 safety prediction across 107 endpoints including hERG, skin reaction, AMES, and ClinTox; and (iv) cell-type filtering against an 821k-cell pilosebaceous single-cell atlas (cross-link to the companion PIEZO1 / MLCK repositioning study), restricting downstream prioritization to compounds whose targets are expressed in the dermal papilla and connective-tissue-sheath cell populations relevant to AGA pathobiology. SRD5A1 was not folded (cached MSA absent) and is an explicit limitation. Boltz-2 results identify Saponin Re from Panax ginseng and Emodin from Polygonum multiflorum as joint top-1 by mean affinity (mean 0.675 each), with Saponin Re leading SRD5A2 (0.746) and Emodin leading AR (0.768). Both leads carry explicit trade-offs: Emodin (skin reaction 0.957, AMES 0.711) and Physcion (skin reaction 0.921, AMES 0.734) carry significant safety flags; Saponin Re (MW 947, logP -0.03) is far outside the topical sweet spot. Biochanin A from soy and Astragalus membranaceus emerges as the cleanest combined-profile candidate (mean 0.544, hERG 0.52, AMES 0.24, topical-friendly). Finasteride, the reference SRD5A2 inhibitor, ranks low in the Boltz-2 binary classifier (mean 0.202), reflecting the classifier's known limitation for mechanism-based covalent inhibition; this is a methodological observation, not a refutation of finasteride's clinical efficacy. Open Targets evidence-tier mapping places SRD5A2, AR, and CTNNB1 in the green-evidence tier for AGA. We position Biochanin A, Saponin Re, and Emodin as topical-scalp candidates with explicit safety-versus-affinity trade-offs whose forward path is dermal papilla cell culture and three-dimensional scalp organoid validation. All findings are in silico; no experimental potency, no clinical claim, and no commercial product is reported. Code, raw cofold JSON, ADMET predictions, and atlas-filter outputs are released under Apache-2.0.

Keywords: androgenetic alopecia, 5-alpha-reductase, SRD5A2, androgen receptor, Wnt beta-catenin, Korean traditional medicine, multi-target in silico evaluation, Boltz-2 cofold, ChEMBL calibration anchor, ADMET-AI, pilosebaceous single-cell atlas, Biochanin A, Saponin Re, Emodin.

---