PUMBA

PUMBA is a scoring model that evaluates the quality of protein-protein interaction (PPI) interfaces, distinguishing native-like binding poses from incorrect ones. In protein docking, generating candidate complexes is only half the problem; a reliable scoring function is needed to rank thousands of decoy poses and identify those that resemble the true biological interface. This ranking step is a persistent bottleneck in computational studies of protein complexes.

PUMBA builds directly on PIsToN, which frames interface evaluation as an image-classification problem: the contact region between two proteins is encoded as a stack of image-like feature maps capturing geometric and chemical properties, then scored with a Vision Transformer. PUMBA's key change is to replace that Vision Transformer with Vision Mamba, a state-space-model architecture that processes the sequence of image patches with efficient long-range modeling. This lets the scorer integrate both global and local interface patterns more effectively while keeping computation tractable.

Developed by researchers at Florida International University and released as a preprint in October 2025, PUMBA continues a line of specialized PPI interface scorers and demonstrates that recent state-space vision backbones can improve on transformer-based predecessors for this task.

#Key Features

• Vision Mamba interface scorer: PUMBA swaps PIsToN's Vision Transformer for a Vision Mamba backbone, applying efficient state-space sequence modeling to image-encoded interface patches.
• Global and local pattern capture: The long-range modeling of Vision Mamba helps the model detect both broad and fine-grained features of a binding interface.
• Frozen-backbone efficiency: PUMBA uses frozen weights and precomputed embeddings to lower computational overhead while retaining accuracy.
• Consistent gains over PIsToN: Across several large public docking benchmarks, it outperforms its Vision Transformer-based predecessor.

#Technical Details

PUMBA represents a protein-protein interface as a set of image-like patches encoding geometric and physicochemical features of the contact region, then processes these patches with a Vision Mamba architecture rather than a Vision Transformer. It builds on the PIsToN and PInet frameworks and uses frozen backbone components with precomputed embeddings to reduce cost. The model is trained on protein-protein docking decoys and evaluated on standard interface-quality benchmarks including MaSIF-test and CAPRI-derived and FireDock decoy sets, where it consistently exceeds the PIsToN baseline across the reported metrics.

#Applications

PUMBA is useful to structural biologists and computational biophysicists who need to rank docking poses and identify native-like protein-protein complexes. As a scoring component, it can be paired with docking generators to filter and prioritize candidate interfaces, supporting studies of protein complexes, interaction mechanisms, and interface-focused analyses where accurate discrimination of near-native poses is essential.

#Impact

PUMBA provides evidence that Vision Mamba and related state-space architectures can outperform Vision Transformers on image-encoded structural-biology tasks, extending the reach of these efficient sequence models beyond natural images. By improving on PIsToN for interface scoring, it contributes an incremental but practical advance to the protein docking pipeline. As a recent preprint awaiting peer review, its generalization to unseen protein families and its performance relative to the broader landscape of interface scorers remain to be established.