Computational pipeline

Overview

The pipeline combines structural prediction, pocket detection, virtual screening, and multiple layers of robustness control. All software is free and open source.

Step 1 — Structural modeling

We obtained two independent structural models for PACS1:

AlphaFold 2 monomer of full PACS1 (Q6VY07) from AlphaFold DB v6. The FBR domain (residues 100–280) had pLDDT = 83.8 (high confidence).
ESMFold complexes for WT and R203W variants. PACS1-FBR (residues 100–280) was fused to the furin cytoplasmic tail (residues 770–807) via a GGGGS×3 flexible linker, yielding a 234-residue construct modeled in both versions.
AlphaFold 3 server: 8 jobs covering WT/R203W ± phosphorylated furin, PACS1 + BICD2 (Krzysiak hypothesis), homodimer (WT/WT), and heterodimer (WT/R203W). All input JSONs are openly available.

Step 2 — Pocket detection

fpocket 4.0 (Le Guilloux et al., 2009) was applied to both WT and R203W ESMFold complexes with default parameters.

70 pockets detected on the apo monomer
22 pockets in R203W complex / 20 in WT complex
Pocket #1 (canonical): highest druggability, 0.464 in R203W vs 0.403 in WT. Located far from the R203 residue itself.
Pocket #19 (cryptic, R203W-specific): includes W203 directly and furin tail contacts.

Step 3 — Ligand library preparation

ChEMBL Database v34 was queried for FDA-approved drugs (max_phase = 4):

Step	Count
Initial query	4,225
Filtered to MW 150–700, no mixtures, valid SMILES	1,500
3D PDBQT successfully generated (RDKit ETKDGv3 + MMFF + Meeko)	1,397

Step 4 — Virtual screening with AutoDock Vina

AutoDock Vina 1.2.7 was used throughout. Precomputed grid maps for pocket #1:

Receptor	Box center (Å)	Box size (Å)
R203W	(−8.03, 12.70, 1.00)	16 × 16 × 16
WT	(−6.92, 13.47, −0.09)	16 × 16 × 16

Hierarchical exhaustiveness strategy:

Step	Exhaustiveness	Purpose	Time per dock
Initial screen	1	Throughput on 1,397 drugs	~28 s
Robust check	8	Publication standard	~7 s (cached maps)
Publication grade	16	Convergence verification	~15 s
Gold standard	32	Final reliability	~40 s

Selectivity score: Δ = ΔG(R203W) − ΔG(WT). A compound is robust when Δ < −0.5 kcal/mol (≈ noise floor at exhaust=8).

Step 5 — Control campaigns

Five independent validation experiments specifically tested the hits.

#	Campaign	n compounds	Purpose
1	Robust check top 10	10 + 2 controls	Distinguish real from sampling noise
2	19 AEDs at pocket #1	19	Is selectivity unique among AED class?
3	19 AEDs at pocket #19	19	Comparison cryptic vs canonical
4	7 random FDA drugs at pocket #1 + pocket #19	7 × 2	Discrimination vs permissivity
5	25 chemically-related compounds at pocket #1	25	Pharmacophore stress test

Step 6 — Molecular dynamics setup

OpenMM 8.5.1 with AMBER14 force field, prepared via PDBFixer. The R203W structure with hydrogens has 2,943 atoms. Short vacuum MD (2.4 ps) verified energetic sanity. Full production MD (100 ns × 3 replicas in TIP3P) was deferred to GPU hardware.

Reproducibility

All software is open-source. All structures, ligands, AF3 inputs, scripts, and results are on OSF (DOI: 10.17605/OSF.IO/ED7SP). Total compute on a 2-core sandbox: ~2 days.