PANDA

What it is

PANDA (Protein All-atom Nested-tree Denoising Architecture) is a generative model from Tianjin University of Science and Technology that designs protein structure and sequence together in a single denoising process, operating directly on atomic coordinates in Cartesian space rather than treating backbone generation and sequence assignment as two separately trained stages, as most prior methods do.

How it works

The model performs denoising in a unified all-atom representation, recovering sequence identity directly from atomic occupancy patterns as generation proceeds. Global and local coordinate tracks are coupled to refine structure jointly, and a dedicated sampling strategy is used to preserve side-chain geometry throughout the process.

What is new

• End-to-end all-atom denoising replaces the usual two-stage backbone-then-sequence pipeline
• Sequence is recovered from atomic occupancy patterns during generation rather than assigned afterward
• Global and local coordinate tracks are coupled for joint structure refinement
• PANDA reports the highest self-consistency design success across protein lengths among evaluated all-atom methods

Why it matters

Most existing approaches separate backbone generation from sequence assignment or rely on latent or torsional representations rather than working directly on all atoms in Cartesian space. Folding both steps into a single atomic model removes a common source of compounding error between separately trained stages, a practical bottleneck in current de novo protein design pipelines; the reported result specifically concerns self-consistency design success rather than experimental validation of designed proteins.

Availability

PANDA is described in a preprint posted to bioRxiv by the Tianjin University of Science and Technology team. The available abstract text does not specify whether code or model weights have been released.