💻 bioinformatics

EnsPlex: Integrative Protein Complex Prediction through Multi-source Complementary Structural Sampling and Topology-Aware Candidate Selection

EnsPlex is a multi-source framework that integrates diverse structural sampling methods with a topology-aware quality assessment model (FACET) to significantly improve protein complex prediction accuracy and success rates in antigen-antibody systems compared to existing approaches.

Original authors: Hu, B., Lu, Z., Zaman, K., Sun, Z., He, X.

Published 2026-10-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hu, B., Lu, Z., Zaman, K., Sun, Z., He, X.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Life depends on a constant, intricate conversation between molecules. Proteins, the workhorses of biology, rarely act alone; they must find one another, lock together, and form complex machines to carry out tasks like fighting infection or sending signals between cells. To understand how these biological machines function, scientists need to see their three-dimensional shapes. While predicting the shape of a single protein has become remarkably accurate in recent years, predicting how two different proteins fit together remains a formidable challenge. The space of possibilities is vast, and the correct way they join is often hidden among millions of incorrect guesses. Without a clear picture of these partnerships, designing new medicines or engineering better antibodies becomes a process of blind trial and error.

A team of researchers has developed a new strategy called EnsPlex to solve this specific puzzle. Instead of relying on a single method to guess the shape of a protein complex, they combined several different approaches to cast a wider net. Imagine trying to find a lost key in a dark room; using just one flashlight might miss the spot, but using four flashlights from different angles increases the chance of finding it. The researchers used four distinct computational tools to generate thousands of possible shapes for protein pairs. One tool, known for its speed, predicts structures directly from genetic sequences. Another uses a more traditional physics-based approach to simulate how proteins might bump into and stick to each other. By running all four tools simultaneously, the team ensured they covered a much broader range of potential shapes than any single tool could achieve on its own.

However, generating many guesses is only half the battle; the real difficulty lies in identifying the one correct shape among the thousands of wrong ones. Each of the four tools produces its own internal score to rank its guesses, but these scores are not compatible with one another. A high score from one tool does not mean the same thing as a high score from another. To fix this, the researchers trained a new artificial intelligence model, named FACET, to act as a universal judge. This model learned to look at the geometric details of the protein interfaces—the specific points where the two proteins touch—and predict how well they fit, regardless of which tool originally generated the shape. It effectively translated the different languages of the four tools into a single, reliable ranking system.

The team tested this combined system on 102 challenging cases involving antibodies and the antigens they target, which are known for being particularly difficult to predict due to their flexible nature. When they compared the results of their new system against AlphaFold3 with its built-in ranking under matched output budgets, they found a significant improvement. The new approach successfully identified the correct structure for up to 27.8% more targets than AlphaFold3 with its built-in ranking used alone. The study showed that the different tools often found correct shapes that the others missed, and the new judging model was crucial for spotting these hidden successes. The researchers demonstrated that their method worked not just on the specific data it was trained on, but also on completely new, unseen datasets, proving its ability to generalize.

The findings suggest that the future of protein structure prediction lies not in building a single, perfect tool, but in integrating multiple imperfect ones. By combining diverse ways of sampling the possible shapes with a smart, unified way of selecting the best candidates, scientists can navigate the complexity of protein interactions more effectively. This approach does not claim to have solved every problem in the field, but it provides a robust framework for handling the most difficult cases. The work highlights that in the quest to map the molecular machinery of life, diversity in approach and precision in selection are equally important.

Technical Summary: EnsPlex

Problem Statement

Protein complex structure prediction faces a dual challenge: generating a broad enough set of candidate conformations to include near-native structures, and effectively identifying those accurate models within a limited output budget. Existing approaches often treat these stages in isolation. End-to-end models (e.g., AlphaFold-Multimer, AlphaFold3, Boltz-1) and molecular docking workflows (e.g., HADDOCK) focus on candidate generation but rely on disparate, non-calibrated internal scoring functions that struggle to consistently rank accurate models. Conversely, quality-assessment models typically rerank a predefined pool from a single source, failing to address the need for complementary sampling across different generators. This is particularly critical in antigen-antibody systems, where weak coevolutionary signals and flexible interfaces make it difficult for any single method to simultaneously provide broad coverage and high-confidence selection.

Methodology

The authors present EnsPlex (Ensemble-sourced Sampling and Selection for PPI Complex Structure Prediction), a unified framework that integrates multi-source structural sampling with a topology-aware candidate selection mechanism. The workflow consists of three consecutive stages:

Multi-Source Structural Sampling:
- Generators: The framework independently generates candidates using four distinct sources:
  - AlphaFold-Multimer, AlphaFold3, and Boltz-1: End-to-end deep learning models generating structures from sequences.
  - Conformation-Expanded Docking: A workflow involving short molecular dynamics simulations to expand monomer ensembles, followed by flexible HADDOCK docking with interface restraints.
- Strategy: This approach aims to capture complementary conformational spaces that single generators might miss.
Within-Source Reranking via FACET:
- FACET Model: A unified quality-assessment model named FACET (Fused Atomic-Contact Evaluation via Topology-informed Prediction) is employed to rerank candidates within each of the four sources.
- Architecture: FACET utilizes a structure-first architecture combining:
  - Global $C\alpha$ residue-graph encoding.
  - Interfacial heavy-atom graph encoding.
  - Late fusion of ESM-2 sequence embeddings.
- Training Objective: The model is trained on a large dataset (>10 million candidates from >6,000 PPI systems) to predict three components of the DockQ metric: $F_{nat}$ (fraction of native contacts), $S_i$ (interface RMSD score), and $S_L$ (ligand RMSD score). The final DockQ score is reconstructed as the mean of these three components. This allows the model to distinguish between locally plausible interfaces and correct global assemblies.
Source-Balanced Output:
- Candidates from each source are ranked by FACET.
- A predefined quota is selected from each source to form the final output set, ensuring that the complementarity of the different sampling methods is preserved in the final prediction pool.

Key Contributions

Integrated Framework: EnsPlex couples complementary sampling with a unified quality-assessment model, addressing the disconnect between generation and selection in current pipelines.
FACET Model: A novel, source-agnostic quality assessment tool trained to predict DockQ components using global geometry, interfacial contacts, and sequence information. It outperforms source-specific scoring functions and existing data-driven methods (e.g., DeepRank-GNN-ESM, GDockScore) in reranking tasks.
Conformational Expansion: The inclusion of a conformation-expanded docking branch (monomer expansion + HADDOCK) provides access to conformational regions not sampled by end-to-end deep learning models alone.

Results

The framework was evaluated on a curated test set of 102 antigen-antibody targets, representing a challenging subset of protein-protein interactions.

Target Success Rate: Under matched output budgets, EnsPlex achieved a 27.8% relative improvement in target success rate (defined as having at least one model with DockQ $\ge$ 0.23) compared to AlphaFold3 using its built-in ranking.
Complementary Coverage: Analysis of the Radius of Gyration ( $R_g$ ) and Solvent-Accessible Surface Area (SASA) distributions revealed that the four sources sampled distinct conformational regions. Conformation-expanded docking, in particular, covered broader regions of $R_g$ -SASA space than AlphaFold3, though its native scoring function failed to identify these acceptable models effectively.
Reranking Performance: FACET consistently improved Top-k target success rates across all four candidate sources compared to their original scoring functions.
- In the conformation-expanded docking pool, FACET outperformed the HADDOCK score, van der Waals energy, and PRODIGY $\Delta G$ .
- In AlphaFold-based pools, FACET surpassed built-in confidence metrics (e.g., ipTM).
Generalization: FACET maintained performance on external, homology-filtered datasets (AbSet blind-docking and site-directed subsets) and the FoldBench benchmark, outperforming DeepRank-GNN-ESM, GDockScore, and AlphaFold3's own ranking metrics.
Case Studies: Representative examples (e.g., PDB IDs 7TXT, 8WTD) demonstrated that FACET successfully identified correct global assemblies that were ranked lower by original scores due to locally plausible but globally incorrect interfaces.

Significance and Claims

The paper claims that the success of EnsPlex stems from the necessity of treating sampling and scoring as coupled, rather than independent, components. The authors argue that:

Complementarity is Essential: No single generator provides sufficient coverage for all targets; combining heterogeneous generators expands the search space.
Unified Scoring is Critical: Raw scores from different tools are not comparable. A unified model like FACET is required to calibrate and select the best candidates across these diverse sources.
Context Matters: Accurate prediction requires evaluating local interface contacts within the context of the overall complex assembly, a capability FACET demonstrates by distinguishing between correct and incorrect global orientations.

The authors conclude that integrating complementary sampling with effective candidate selection yields superior results compared to improving either step in isolation. They note that the evaluation is currently limited to antigen-antibody systems and matched output counts, leaving the performance under fixed computational budgets and applicability to other PPI classes as open questions for future work.