PI-Mamba: Linear-Time Protein Backbone Generation via Spectrally Initialized Flow Matching
Summary: arXiv:2603.26705v1
Announce Type: cross
Abstract
Motivation: Generative models for protein backbone design have to simultaneously ensure geometric validity, sampling efficiency, and scalability to long sequences. However, most existing approaches rely on iterative refinement, quadratic attention mechanisms, or post-hoc geometry correction, leading to a persistent trade-off between computational efficiency and structural fidelity.
Results
We present Physics-Informed Mamba (PI-Mamba), a generative model that enforces exact local covalent geometry by construction while enabling linear-time inference. PI-Mamba integrates a differentiable constraint-enforcement operator into a flow-matching framework and couples it with a Mamba-based state-space architecture. To improve optimisation stability and backbone realism, we introduce a spectral initialization derived from the Rouse polymer model and an auxiliary cis-proline awareness head. Across benchmark tasks, PI-Mamba achieves 0.0% local geometry violations and high designability (scTM = 0.91±0.03, n = 100), while scaling to proteins exceeding 2,000 residues on a single A5000 GPU (24 GB).
Key Features of PI-Mamba
- Geometric Validity: Ensures that the generated protein backbones maintain exact local covalent geometries, eliminating the need for post-processing corrections.
- Linear-Time Inference: Achieves rapid generation of protein structures, allowing for efficient exploration of larger protein sequences.
- Optimisation Stability: Incorporates a spectral initialization that enhances the stability of the optimization process, leading to more realistic backbone conformations.
- Cis-Proline Awareness: Features an auxiliary head that accounts for specific structural characteristics related to cis-proline residues, improving design accuracy.
- Scalability: Successfully handles proteins with over 2,000 residues on modern GPU architectures, showcasing its potential for large-scale applications.
Conclusion
PI-Mamba represents a significant advancement in the field of protein design, addressing the critical challenges of geometric validity and computational efficiency. By leveraging a novel approach that combines differentiable constraint enforcement with an innovative architecture, PI-Mamba not only enhances the designability of protein structures but also sets a new benchmark for the scalability of generative models in biochemistry. Future research may explore further enhancements and applications of this groundbreaking model in the realm of synthetic biology and therapeutic development.