Efficient exploration of conformational ensembles via protein interaction-informed deep learning

Technical Summary: Efficient exploration of conformational ensembles via protein interaction-informed deep learning

Problem Statement
Accurately characterizing protein conformational ensembles is critical for understanding biological function and advancing drug design. While Molecular Dynamics (MD) simulations serve as the gold standard for exploring these landscapes, they are computationally prohibitive for high-throughput applications or large macromolecular systems. Generative artificial intelligence (AI) models, such as diffusion and flow-based frameworks, offer a faster alternative. However, existing methods often struggle with structural hallucinations, particularly in flexible or information-sparse regions, and rely heavily on expanding training datasets with isolated single-chain structures. The authors posit that artificially decoupling protein chains from their native interaction contexts limits models to learning static structural distributions rather than the underlying physical constraints governing conformational change.

Methodology
The study introduces DynoM, a generative framework designed to leverage protein complex interactions to improve the prediction of monomeric conformational ensembles. The methodology involves a two-stage training strategy:

1. Architecture: DynoM utilizes a conditional diffusion architecture based on the AlphaFold3 (AF3) template-free trunk. It incorporates a positional encoding module to differentiate chain identities in multi-chain inputs. The diffusion module employs Invariant Point Attention (IPA)-based transformers to iteratively reconstruct protein backbones from random noise.
2. Training Data:
   • Pre-training: The model was pre-trained on a curated dataset of 62,044 static protein complex structures (heterodimers and antigen-antibody complexes), representing only ~10% of the volume of available monomer data but offering higher sequence diversity.
   • Fine-tuning: To address the limitations of static data, the authors generated an in-house dataset of 5,502 all-atom MD trajectories (100 ns each) for protein complexes using GROMACS. The pre-trained model was fine-tuned on this dynamic data to act as a "structural anchor," correcting non-physical motions.
3. Loss Functions: The training objective combines denoising score-matching loss with auxiliary terms: direct Mean Square Error (MSE) on backbone coordinates, local neighborhood loss on pairwise atomic distances (≤ 8.5 Å), and a steric clash penalty to prevent unphysical atomic overlaps within and between chains.
4. Evaluation: Performance was benchmarked against state-of-the-art baselines (BioEmu, ConfDiff, AlphaFlow) using an independent test set of 63 proteins with varying lengths and topologies. Metrics included global topological measures (PwD-JSD, Rg-JSD), dynamic consistency (TIC-JSD, TIC-RMWD), and residue-level flexibility (rRMSF).

Key Contributions

• Interaction-Informed Learning: The paper demonstrates that incorporating a relatively small fraction (10%) of complex data significantly outperforms models trained solely on monomers. Specifically, supplementing monomer-trained models with complex data improved performance by 6.64% in PwD-JSD and 22.87% in Rg-JSD.
• Suppression of Structural Hallucinations: DynoM effectively captures long-range physical constraints inherent in complex interactions. This capability allows the model to suppress structural hallucinations in inter-domain positioning for multi-domain monomers, a challenge where previous models often fail.
• Complementary Priors: The study establishes that static complex structures provide essential topological priors, while MD fine-tuning provides the necessary dynamic constraints to approximate equilibrium distributions.
• Data Efficiency: DynoM achieves superior performance using significantly less training data (both static structures and MD trajectories) compared to existing baselines.
• Resource Release: The authors provide the research community with a new dataset of 5,502 MD trajectories of protein complexes and the source code for DynoM.

Results

• Benchmark Performance: On the 63-protein test set, DynoM outperformed BioEmu, ConfDiff, and AlphaFlow across all equilibrium metrics (P < 0.05). It achieved the lowest TIC-JSD and TIC-RMWD, indicating superior coverage of the accessible conformational space while minimizing the generation of non-physical states.
• Multi-Domain Accuracy: For multi-domain proteins, DynoM demonstrated high consistency in inter-domain positioning, significantly reducing deviations in off-diagonal pairwise Cα-Cα distances compared to baselines.
• Conformational Diversity: Analyses of proteins with local structural changes, apo-holo transitions, and rigid-body domain movements showed that DynoM could successfully sample multi-conformational landscapes. Crucially, the model did not over-constrain the conformational diversity of monomeric states; predicted ensembles for unbound antigens covered both bound and unbound conformational spaces.
• Complex Prediction: Preliminary tests on 20 antigen-antibody complexes showed DynoM could distinguish rigid framework regions from flexible CDRs and identify plausible conformations in cases where AlphaFold3 failed, though interface precision (iRMSD) remains a limitation compared to AF3.

Significance and Claims
The authors claim that interfacial priors derived from protein complexes are an indispensable "missing link" in current conformational ensemble prediction. By leveraging these physical constraints, DynoM provides a robust, data-efficient tool that bridges the gap between the need for dynamic insights and the computational burden of physics simulations. The work suggests that data quality (incorporating interaction constraints) is more critical than sheer data quantity. While the model currently focuses on dimers and antigen-antibody complexes and relies on backbone generation rather than end-to-end all-atom prediction, it establishes a proof-of-concept for using complex-derived constraints to enhance the physical plausibility of protein modeling. The authors acknowledge limitations regarding the scope of training data (excluding larger multimeric assemblies) and the need for further optimization of loss functions and atomic-level prediction for complex dynamics.