Research map / Probabilistic and scientific modeling / Molecules and proteins

Predicting directional flexibility in proteins

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein interface and docking prediction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BackFlip-2, an SE(3)-equivariant IPA-based graph neural network that directly predicts per-residue anisotropic covariances and pairwise residue couplings from a backbone structure, matching large ensemble generators at a fraction of the cost.

Problem

Predicting directional (anisotropic) protein flexibility and residue-residue correlations from an equilibrium structure, as a fast alternative to MD and ensemble generative models.

Approach

An SE(3)-equivariant covariance head (local SPSD matrix A A^T rotated by residue frame R_i) combined with an edge-embedding-based pairwise coupling head, trained on ATLAS MD data, so that predicted covariances transform correctly under global rotation unlike the invariant DynaProt head.

Key results

• On the ATLAS test set, BackFlip-2 reaches per-target RMSF r = 0.87 with MAE 0.58 Å, beating BackFlip (0.84, 0.62 Å) and FlexPert/Pegasus, with 965K parameters and ≤0.02 s inference.
• For covariance and coupling prediction on ATLAS, BackFlip-2 improves over DynaProt (RMWD 0.88 vs 1.18; DCCM r 0.80 vs 0.66) and is competitive with AFMD-T (RMWD 0.84) and BBFlow, while being orders of magnitude faster.
• Ablation: replacing SE(3)-equivariance with invariance raises RMWD from 0.88 to 1.05 and lowers ellipsoid IoU from 0.49 to 0.42.

Surprising findings

• The invariant covariance variant predicts nearly isotropic (spherical) covariances, showing equivariance is key for anisotropy.
• The ATLAS-trained model transfers to mdCATH (500 ns, higher temperature) and NMR ensembles without training on them, while FlexPert and BackFlip degrade.

Limitations

• Slightly less accurate than large ensemble generation models (AFMD-T, BBFlow) and underperforms them on residue couplings.
• Restricted to nanosecond-to-sub-microsecond timescales and Gaussian-approximable fluctuations; cannot describe large slow conformational changes.
• Ground truth from ATLAS/mdCATH only; joint ATLAS/mdCATH training slightly degrades ATLAS performance.

Experimental setup

Datasets: ATLAS mdCATH de novo proteins (RFdiffusion/FrameFlow) NMR ensembles from PDB

Benchmarks: ATLAS test set mdCATH de novo proteins NMR ensembles

Metrics: Pearson r (RMSF) MAE (RMSF) RMWD variance symmetric KL variance DCCM Pearson r DCCM MAE ellipsoid IoU Dice

Baselines: FlexPert Pegasus BackFlip DynaProt AlphaFlow-MD-Templates BBFlow BioEmu NMA (ANM)

Builds on: BackFlip DynaProt AlphaFold2 invariant point attention AlphaFlow

Closest papers by meaning

• Boosting Protein Graph Representations through Static-Dynamic Fusion (ICML 2025 · similarity 0.89)
• Flexibility-conditioned protein structure design with flow matching (ICML 2025 · similarity 0.88)
• Learning to engineer protein flexibility (ICLR 2025 · similarity 0.88)
• Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.88)
• Ensembits: an alphabet of protein conformational ensembles (NeurIPS 2026 · similarity 0.87)
• TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.87)