Can an LLM Agent Select the Right Pose? Analysis of Agentic Pose Selection for PXR Fragments

Can an LLM-based agentic approach close the gap between generating good protein-ligand structure predictions and selecting the correct one? We study this using 184 recent structures from the OpenADMET PXR (Pregnane X Receptor) Blind Challenge Structure Track. While diverse sampling with Boltz and Boltz-Perturb builds pools where an oracle could select 104/184 sub 2 angstrom RMSD poses (average lDDT-PLI of 0.73) and outperform the competition-winning entry (average IDDT-PLI of 0.61), standard model confidence ranking recovers only 45/184 (average lDDT-PLI of 0.54). This highlights a generation-versus-selection bottleneck. We evaluate two agentic designs: (1) target-specific agents that compose literature- and confidence-based scoring functions grounded in PXR data and literature, and (2) a non-target-specific data-driven multi-agent system (RnP Ranker) that discovers symbolic ranking metrics through iterative feedback on independent non-PXR benchmark. Neither approach beats confidence-only selection, degrading the number of sub-2.0 angstrom ligand poses from 45 to a range of 20–31 and 35-45, respectively. We identify two recurring failure modes common to both approaches: first, the agents arbitrarily invented numerical weights for correct domain knowledge without evidence; and second, they come up with new scoring functions without validation against simple baselines. Nevertheless, agents provide highly complementary signals and combining their top selections with confidence-based selection outperforms in top-$k$ selection. For example, a small ensemble combining agent selections with the top confidence score recovers 63/184 correct poses, a 19\% relative improvement over the standard top-5 confidence baseline (53/184). These results show that agents can extract the correct domain knowledge. Turning that knowledge into discriminative scoring remains an open challenge for building agentic scientific systems, especially in flexible, promiscuous binding pockets where correct and incorrect poses satisfy the same key contacts.