PANDA

A catalytic site is an arrangement of side-chain atoms — a histidine nitrogen held a few angstroms from a serine oxygen — but the standard de novo design pipeline never generates those atoms. RFdiffusion and Proteina produce a backbone, and an inverse-folding model such as ProteinMPNN then proposes a sequence likely to fold into it; the chemistry-bearing side chains are inferred after the fact, from a scaffold built without them.

PANDA, from the Tianjin Institute of Industrial Biotechnology of the Chinese Academy of Sciences, removes that split by making residue identity a property of geometry rather than a discrete label. Each residue is a nested tree of 28 atomic slots following standard PDB nomenclature, rooted at Cα and branching outward through the β, γ, δ, ε, ζ and η positions. An atom a residue does not possess occupies the coordinates of its immediate parent rather than collapsing onto Cα, so the pattern of coincident and extended slots is itself a fingerprint of the amino acid. Generation produces coordinates; the sequence is read off the occupancy pattern, with no classification head anywhere in the network.

That framing sets PANDA apart from the all-atom generative models it is measured against. Pallatom, Protpardelle, APM and La-Proteina each retain an explicit sequence prediction, colocate absent atoms with Cα, or work in a latent or torsional space.

#Key Features

• Nested-tree atom representation: Twenty-eight slots per residue, ordered by the PDB atom hierarchy, encode which side chain is present without a discrete variable. Nucleic acids anchor at the phosphate instead, and each ligand atom anchors itself.
• Dual-track global–local denoising: Cα is denoised in global coordinates while the remaining 27 atoms live in a Cα-centred local frame. The tracks exchange information through bidirectional cross-attention in a shared trunk and decode into separate velocity fields.
• Tangential coefficient-preserving sampling: Ordinary stochastic sampling on the local track erases the parent-to-child distances that distinguish extended from absent atoms, collapsing side chains toward alanine. TCPS applies the stochastic component only in the subspace orthogonal to the local radial vector, leaving that direction to the deterministic update.
• Motif conditioning without frozen coordinates: Functional motifs enter as a pairwise distance matrix among their atoms rather than as fixed positions, so they move with the scaffold while retaining internal geometry. Because distances are invariant to reflection, a chirality correction resolves the mirror-image ambiguity during sampling.
• Optional burial control: A per-atom relative accessible surface area value, normalised to the unit interval, can be supplied for ligand atoms to steer how deeply each is buried.

#Technical Details

PANDA is a roughly 630-million-parameter dual-track flow-matching denoiser. Residue single and pair channels are 768 and 128 wide, atom and atom-pair channels 128 and 16; the trunk holds 16 dual-track blocks with 12-head residue-level and 4-head atom-level attention, run in BF16. Training drew on PDB-REDO refined experimental structures and confidence-filtered ESMFold atlas models, keeping protein chains of 32 to 384 residues and clustering them jointly with Foldseek 3Di linclust at 50% identity. The objective pairs mask-normalised velocity losses with a SmoothLDDT term; inference runs 400 TCPS steps.

Sweeping the samplers across both tracks, dual-track TCPS reached 96.8% backbone designability while holding the alanine fraction to 15.3%, against 36.2% when coefficient-preserving sampling was used on the local track. Unconditional design success — sequence recovered from the generated coordinates, refolded with ESMFold, and required to reach Cα RMSD below 2 Å and pLDDT above 75 — was the highest among the evaluated all-atom methods in every length band from 64 to 512 residues. Median heavy-atom motif RMSD was near zero for one- and two-residue motifs and rose to roughly 3.5 Å for seven-residue motifs. On the Atomic Motif Enzyme benchmark PANDA scaffolded most targets more successfully than Proteina-Complexa, RFdiffusion3 and RFdiffusion3 paired with LigandMPNN.

#Applications

The design task PANDA is built for is functional rather than purely structural: scaffolding a catalytic or binding constellation whose atoms must sit in a precise relative arrangement. Conditioning on inter-atomic distances rather than absolute coordinates lets a designer specify that geometry without also fixing where it sits in a protein that does not yet exist, which is what makes larger, multi-residue active sites tractable. Nucleic-acid and ligand components fit the same representation, and burial conditioning offers a handle on how a small molecule is packed.

#Impact

PANDA's argument is that sequence and structure need not be separate generative problems — that residue identity can live inside the same Cartesian occupancy pattern already used for coordinates. Its companion finding is more cautionary: sampling schemes tuned on backbone-only models cannot be transplanted to all-atom ones, because isotropic stochastic updates destroy the very geometry that carries sequence information. The evaluation is entirely computational — self-consistency measures whether a generated sequence and structure agree with each other, not whether the protein is stable, soluble or catalytically active. The work is a preprint awaiting peer review, and the announced code repository is a placeholder: nothing has been released to run the model with.