ProMaya Architecture Overview

The ProMaya framework is a highly specialized pipeline composed of 5 distinct modules that transform raw protein pairs (PDB coordinates and FASTA sequences) into an interaction probability, along with multi-scale attribution maps.

The 5-module ProMaya pipeline: From feature extraction and multimodal alignment through Heterogeneous Graph Transformer (HGT) encoding, cross-protein attention, and hybrid classification.

Pipeline Modules

The computational flow of ProMaya is divided into five sequential steps. Modules 1-3 operate on each protein individually to learn deep structural and contextual representations. Modules 4-5 model the complex interplay between the two proteins.

1. Feature Extraction

Constructs 4 distinct data modalities per protein: Atomic graphs (with LSMD), Residue graphs, Surface Point Clouds, and 1024-d Language Model (ProtTrans) sequence embeddings.

2. Multimodal Alignment

A 4-stage bidirectional cross-attention mechanism that fuses features across scales (e.g., projecting atomic electron density up to residue level, and grounding sequence embeddings in 3D surface geometry).

3. HGT Encoder

A Heterogeneous Graph Transformer acting on 4 node types and 6 edge relations, mapping the aligned multimodal data into a unified, high-dimensional protein embedding.

4. Cross-Protein Interaction

Computes the biophysical binding mechanics between Protein A and Protein B using multi-scale attention (Residue-Residue, Surface-Surface, and LSMD-filtered Atom-Atom sparse attention).

5. Hybrid Classification

A robust classifier merging a 3-layer neural Feed-Forward Network with a 500-tree XGBoost ensemble to predict the final interaction probability and resist overfitting.

---

Next: Deep dive into the modalities in Feature Extraction →