Choosing among Competing Pockets in Oligomeric Proteins: An OIPS-Assisted, Traceable Multi-Evidence Analysis

Technical Summary: Choosing among Competing Pockets in Oligomeric Proteins

Problem Statement
In structure-based drug discovery, identifying a binding pocket is often insufficient; the critical challenge lies in prioritizing among multiple geometrically plausible candidates, particularly in oligomeric proteins. These systems present unique complexities: pockets may be formed by multiple chains, reshaped by assembly, exist only in specific complex states, or be located near interfaces without being centered on them. Furthermore, a crystallographic ligand may occupy a compact orthosteric or internal allosteric site distinct from dominant interchain surfaces. Existing methods often struggle to reconcile outputs from different pocket-detection tools that use varying boundaries, descriptors, and ranking conventions. Selecting the wrong region can lead to technically coherent but biologically irrelevant hypotheses, directing resources toward incorrect mechanisms (e.g., competitive inhibition vs. allosteric modulation vs. interface disruption).

Methodology: The OIPS Framework
The authors developed OIPS (Oligomerization-associated Pocket Interpretation Score), a traceable, multi-evidence framework designed to reconstruct, prioritize, and interpret competing pocket hypotheses without reference leakage. The workflow consists of four main stages:

1. Reference-Blind Candidate Reconstruction (Cluster-v2):

   • Five distinct pocket-detection workflows (CavityPlus, DoGSiteScorer, DoGSite3, CASTpFold, and SiteMap) were run on ligand-depleted oligomeric structures.
   • A two-stage clustering algorithm consolidated predictions. First, same-tool duplicates were grouped based on residue Intersection over Union (IoU) and center distance. Second, cross-tool grouping was performed using medoid-constrained clustering (limiting transitive chaining) to ensure all members remained compatible with a cluster medoid and a maximum center diameter of 12 Å.
   • This process generated 733 formal candidates from 1,742 initial records across 21 curated oligomeric protein-ligand systems.

2. Static Prioritization ($OIPS-P_{static}$):

   • Candidates were ranked using a weighted score derived from five modules, normalized to a 0–100 scale. Crucially, reference ligand positions were not used in this calculation.
   • Modules:
     • Cross-tool consensus ($C_{cons}$): Weighted support from contributing tools.
     • Geometric evidence ($G_{geo}$): Volume, depth, enclosure, etc.
     • Ligandability evidence ($P_{lig}$): Druggability estimates and scores (e.g., SiteScore, Dscore).
     • Oligomeric relevance ($O_{rel}$): A novel module assessing structural association with the retained assembly. It calculates interface overlap, chain entropy, center-to-interface distance, and assembly context.
     • Evidence completeness ($Q_{evidence}$): Based on the number of tools, availability of centers/residues, and tool diversity.
   • The final score is a weighted sum where missing modules are excluded and weights are renormalized, rather than treating missing data as zero.

3. Automated Quality Control (QC) and Post-Hoc Annotation:

   • After static ranking, the Top-3 candidates underwent automated QC to flag issues like over-merging, splitting, or boundary sensitivity.
   • Evidence Layers: Only after the static rank was frozen were additional evidence layers appended:
     • Reference Association ($R_{auto}$): Based on distance to crystallographic ligand center (DCC) and residue overlap.
     • Assembly/MD Support ($A_{auto}$): Candidates supported by interface relationships or persistent molecular dynamics (MD) contacts.
     • Uncertainty ($U_{auto}$): Candidates lacking sufficient support.

4. Validation and Sensitivity Analysis:

   • The framework was tested on 21 systems.
   • Ablation: The $O_{rel}$ module was removed to assess its impact on reference recovery versus assembly preservation.
   • Perturbation: Module weights were varied by ±20% to test ranking stability.
   • Dynamic Evidence: Redocking (Glide) and MD simulations (Desmond, 100 ns) were performed to evaluate local pose recoverability and contact persistence, serving as post-static qualifiers rather than re-scorers.

Key Results

• Candidate Landscape: The reconstruction process separated spatially heterogeneous regions, creating distinct candidates where previous methods might have merged or fragmented them.
• Prioritization Performance:
  • The first reference-associated candidate ranked 1st in 12/21 systems, Top-3 in 18/21, and Top-5 in 19/21 (Mean Reciprocal Rank [MRR] = 0.723).
  • When expanding the definition to the "first-supported" candidate (including those supported by assembly/MD evidence even if not reference-associated), 20/21 systems had a Top-1 candidate, and all 21 were within the Top-3 (MRR = 0.976).
  • OIPS did not uniformly outperform individual tools on retrospective Top-1 recovery (e.g., DoGSite3 achieved 12/14 Top-1 in the complete subset vs. OIPS's 7/14), but it maintained high Top-3 retention (12/14) and provided a unified decision layer.
• Impact of Oligomeric Relevance ($O_{rel}$):
  • Removing $O_{rel}$ slightly improved reference Top-1 recovery (0.571 $\to$ 0.667) but significantly reduced the preservation of assembly-supported alternatives (First-supported Top-1 dropped from 0.952 to 0.905).
  • This confirms a trade-off: $O_{rel}$ sacrifices some retrospective reference recovery to preserve structurally plausible assembly-associated candidates that may represent valid alternative mechanisms.
• Static vs. Dynamic Evidence:
  • MD and redocking often revealed boundary shifts or conflicts rather than simple confirmations. For example, in system 4W9H, the static Top-1 (assembly-associated) conflicted with the MD-supported region (local reference), highlighting that static and dynamic evidence address different levels of uncertainty.
  • Redocking RMSD was generally low ($\le$ 2 Å) but failed in cases of cross-layer conflict (e.g., 4W9H RMSD = 3.544 Å).

Significance and Claims
The paper positions OIPS not as a universal "correct" predictor or an affinity estimator, but as a traceable decision-support framework that transforms pocket prediction into a ranked set of testable structural hypotheses.

• Methodological Contribution: It demonstrates that candidate reconstruction is an integral part of pocket interpretation. By using medoid-constrained clustering, OIPS separates competing hypotheses (e.g., a compact orthosteric site vs. a broader interface-associated region) rather than forcing them into a single score.
• Conceptual Shift: The framework explicitly retains disagreement. Instead of collapsing boundary ambiguity or cross-layer conflicts into a single accuracy metric, OIPS outputs distinct layers:
  • OIPS-P: Prioritization score.
  • OIPS-T: Working pocket type (orthosteric, allosteric, interface, complex-dependent).
  • OIPS-C: Evidence state (supported, boundary-sensitive, conflicting, unresolved).
• Practical Utility: The authors claim this approach guides downstream design choices. For instance, an interface-associated candidate suggests preserving the full biological assembly for docking, while a compact core might allow for focused local grids. The framework supports a "modality-aware" progression, potentially guiding the selection of small molecules, peptidomimetics, or antibodies based on the structural nature of the pocket.
• Limitations: The authors acknowledge that the study is a curated case series, not a population-level benchmark. The $O_{rel}$ module is a structural approximation and does not include energetic or evolutionary data. The framework is presented as a tool for generating explicit audit trails and decision pathways, requiring prospective validation to confirm its generalizability and experimental utility.