Protein–ligand complexes were predicted for the pregnane X receptor (PXR) ligand-binding domain against a blind set of 185 ligands using AlphaFold 3 (AF3) in its standard configuration, followed by a confidence-based reranking step to select a single pose per ligand. No target-specific training, fine-tuning, or docking refinement was applied.
A single protein chain corresponding to the PXR ligand-binding domain
(290 residues) was used for all predictions. A multiple sequence alignment was
generated once with MMseqs2 (ColabFold protocol) and supplied to AF3 as a
precomputed unpaired MSA (unpairedMsaPath); the paired MSA was left empty, as
the system is a monomer, and no structural templates were provided. Each ligand
was supplied as a SMILES string taken directly from the challenge input table,
with one AF3 job generated per ligand (protein chain A, ligand chain B).
Predictions were run with AlphaFold 3 using the official model weights and
Docker image on a local GPU. The genetic-search data pipeline was disabled
(--norun_data_pipeline) so that the precomputed MSA was used as-is. Each job
used a single random seed (123) and produced the default five diffusion samples,
giving five candidate complexes per ligand.
For each ligand, the five samples were ranked and a single best model retained.
Rather than the global ranking_score — which, for a single small-molecule
system, is dominated by the protein and is a weak proxy for pocket quality —
samples were ranked by the protein–ligand interface ipTM, taken as the
off-diagonal terms of the chain_pair_iptm matrix in each sample's
summary_confidences.json. Ties were broken by the interface PAE, the
corresponding off-diagonal of chain_pair_pae_min (lower is better). This
selection targets confidence in the relative placement of the ligand against the
protein, which is expected to track lDDT-PLI more closely than complex-global
confidence.
The selected mmCIF for each ligand was converted to PDB format with gemmi. The
ligand residue was renamed to LIG, and each file was named by its challenge
structure identifier (<structure_id>.pdb). A per-ligand audit table recording
the chosen seed/sample and its interface ipTM, interface PAE, and ranking score
was retained alongside the final structures.
The pipeline is "vanilla" AF3 with a single seed and confidence-based pose selection; no consensus clustering, physics-based rescoring, or pose-validity filtering (e.g. PoseBusters) was applied. Confidence-based selection is least reliable for targets dissimilar to the training distribution, so per-ligand accuracy may vary with how well each chemotype is represented.