Original title: PANDA: Protein All-atom Nested-tree Denoising Architecture for End-to-End Generation
Jie Bai, Huifeng Jiang · September 2026
Proteins are tiny molecular machines that do work inside living cells. Designing new ones is hard. Most computer methods split the task into separate steps or avoid working with all the atoms directly in normal physical space. To fix this, researchers built PANDA. It is a new system that generates whole proteins end-to-end. It cleans up noisy all-atom shapes directly in coordinate space and reads out the protein's chemical sequence from where its atoms sit. It links global and local coordinate tracks while keeping side-chain shapes intact. PANDA scored the highest self-consistency design success across protein lengths among tested all-atom tools. For functional proteins, it holds motif distances flexible, achieving much higher scaffolding success on an enzyme benchmark, especially on larger motifs.
Imagine you are building a model car from a kit. Normally you first assemble the frame, then later add the engine and wheels, but sometimes the engine does not fit because the frame was too small. Here, you design the frame, engine, and wheels all at once, so every part fits perfectly from the start. In this paper, PANDA designs all atoms of a protein together, ensuring the backbone and side chains are compatible.
Designing new proteins usually happens in stages—first the backbone shape, then the amino acid sequence—which can miss how side chains actually fit. PANDA designs every atom at once, so the final protein is more likely to fold correctly. This could speed up creating custom enzymes or therapeutic proteins.
The abstract does not discuss failure modes, computational cost, or generalizability beyond the benchmarks tested. It also does not compare against non-all-atom methods or experimental validation in the lab.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
PANDA: Protein All-atom Nested-tree Denoising Architecture for End-to-End Generation — Jie Bai, Huifeng Jiang — bioRxiv — doi:10.64898/2026.09.11.750898
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.