UPDD v0.7.1 — Paper 1 v1 release

This release accompanies Paper 1 v1 — a Capability Level 1 methodology release for the UPDD (Universal Peptide Drug Discovery) pipeline, submitted to bioRxiv as a preprint.

What's in this release

Five-module Capability Level 1 suite

• Sequence parser (schema sequence_parser/0.1) — ncAA registry validation and canonical design specification output
• Branched ΔΔG engine (schema branched_ddg/0.3) — paired variant/WT ensembles, advisory detachment_metric field (Light D Hybrid), automated bit-identical WT control
• Iterative cycle manager (schema cycle_manager/0.2) — Cycle 0 / Cycle 1+ orchestration with multi-assay K_d aggregation (inverse-variance weighting) and plateau detection
• User CLI with Stage 1–2 wrappers around RFdiffusion, ProteinMPNN, and AlphaFold2
• HTML report generator (schema html_report/0.2) — per-cycle ΔΔG ranking, σ-decomposition plots, WT control verification, cycle progression view, detachment banner

Three methodological contributions documented in Paper 1 v1

1. σ_btwn / σ_w ensemble-quality decomposition with the Convergence Index dual interpretation (CI = σ_btwn / |⟨⟨ΔG⟩⟩|): binding-mode classifier + automated quality-control auditor
2. Branched ΔΔG architecture with bit-identical wild-type control (intervention isolation verified at floating-point precision)
3. Four-layer onion-peeling verification protocol — closed-shell charge bias (Σq = 0 audit infrastructure), CONECT-record wraparound at the 99 999-atom PDB limit, intra-residue PBC defects on HETATM ncAA residues that mdtraj cannot auto-bond, source-versus-test integrity

Multi-system Definition 3 closure (six ncAA peptide-target families)

1EBP_MTR13 (Tier-1, X.B promoted), 3IOL_NML20 (Tier-1), 7TL8_MTR6 (Tier-1, X.A), 1YCR_NML22 (Tier-1), 2QKH_MTR25 (Tier-3), 2QKI_Cp4 (Tier-3, magnitude consistent with Magotti 2009 SSOT post-CONECT-fix).

Reproducibility infrastructure

• environment.yml — conda environment specification
• Dockerfile — CPU-only validation container
• .github/workflows/ci.yml — continuous integration (lint, unit tests, Docker smoke test)
• tests/ — regression suite (294 pass / 9 skip / 1 pre-existing fail)

Active learning

The simulated AL cycle (scripts/simulated_al_cycle.py) demonstrates Cycle 0 → Cycle 1+ chaining on the 6-target ncAA benchmark with the Magotti 2009 single-anchor convergence demonstration. Single-anchor framing per Paper 1 §3.4 and §4.5: only Cp4 enters as ground truth; the remaining five families are pipeline-self-prediction comparators.

Limitations and roadmap

This is a process-axis release (Capability Level 1: decision support), not an outcome-axis release (Capability Level 3: quantitative match). See the bioRxiv manuscript §4.2 for the six explicit limitations and §4.3 for the Phase γ + γ' roadmap. The Khoury (2014) R² = 0.388 same-system MM-GBSA / experimental ceiling is acknowledged as a defensible field reference, not a target to break.

Citation

Paper: Kang, I. (2026). Iterative Non-Canonical Amino Acid Peptide Drug Discovery via Sign-Direction Reproducible Evaluation: A Multi-Layer Verification Framework. bioRxiv [DOI pending].

Software: Kang, I. (2026). UPDD: Universal Peptide Drug Discovery — v0.7.1 [Computer software]. Zenodo. [DOI pending — Zenodo will fill after publication of this release].

ORCID: 0009-0007-0753-0636.