SeedProteo: Diffusion All-Atom Protein Designer

• SeedProteo is a diffusion-based, de novo protein design model that directly generates unified all-atom coordinates using an adapted AlphaFold3-like architecture.
• It employs a global MRF sequence decoder with iterative secondary structure conditioning to enhance long-range sequence–structure coherence.
• The model demonstrates superior performance in unconditional generation and binder design, maintaining high success rates and structural diversity across variable protein lengths.

SeedProteo is a diffusion-based, de novo all-atom protein design model that repurposes an AlphaFold3-like folding architecture into a conditional, generative framework. Its central claim is that a near-identical folding network structure—embedder, Pairformer encoder, and coordinate denoiser—can be made generative by feeding noisy coordinates into the encoder, unifying atomistic representation across amino acids via an atom14 schema, and integrating strong self-conditioning signals through a global Markov random field (MRF) sequence module, secondary-structure iterative unmasking with gating, and previous-step structural templates. In this formulation, “de novo all-atom” denotes direct generation of 3D coordinates for both backbone atoms and sidechain atoms in a unified atom14 representation, without relying on a provided amino-acid sequence or homology (Wei et al., 30 Dec 2025).

1. Definition and scope

SeedProteo is explicitly positioned against backbone-only diffusion models and coarse-grained approaches. Backbone-only models produce frames or coordinates only for $N$–$C_\alpha$–$C/O$, while coarse-grained methods operate on residue centroids or rigid-body frames and defer sidechains to downstream packing or inverse folding. SeedProteo instead decodes sequence during generation from learned global pair and single latent states via an MRF, and sidechain atom coordinates are directly modeled and denoised jointly with backbone. This design is presented as enabling tighter sequence–structure coupling and improved designability.

The model therefore combines three commitments that are often separated in earlier workflows: atom-level coordinate generation, sequence decoding during sampling rather than only after backbone completion, and explicit conditioning mechanisms that steer the denoising trajectory. A plausible implication is that SeedProteo treats protein design less as sequential composition of backbone generation, sidechain packing, and inverse folding, and more as a coupled latent inference problem over coordinates, sequence, and fold grammar. The paper’s empirical focus is correspondingly split between unconditional generation and binder design, with extensive benchmarks reported for both settings (Wei et al., 30 Dec 2025).

2. Architecture and self-conditioning mechanisms

The base folding architecture mimics AlphaFold3-like design: an embedder initializes single-residue and pair features; an encoder composed of Pairformer blocks builds deep geometric representations; and a diffusion head denoises coordinates. Because the encoder has cubic complexity in sequence length, the depth was reduced from 48 to 12 Pairformer layers when conditioning on noisy inputs. Unlike standard folding setups that rely on sequence or homology input, SeedProteo additionally feeds noisy atom coordinates into the embedder to extract geometry-derived 1D sequence-like features and pair features.

The sequence decoder is a global MRF. Given learned single features $a$ and pair features $z$ from the encoder, the amino-acid sequence $x = (x_1,\ldots,x_L)$ is sampled from

$$P(X = x \mid a, z) \propto \exp\!\left[\sum_{i=1}^{L} h_i(x_i \mid a_i) + \sum_{i=1}^{L}\sum_{j=i+1}^{L} e_{ij}(x_i, x_j \mid z_{ij})\right].$$

Here, $h_i(\cdot)$ captures site-specific preferences from single features and $e_{ij}(\cdot)$ captures pairwise couplings from pair features. The paper contrasts this with local “atom14-to-AA-type” heuristics and attributes to the global decoder mitigation of isosteric confusions and improved long-range sequence–structure consistency.

Secondary-structure conditioning is implemented through iterative unmasking with gating. Let $S$ denote the secondary-structure sequence over $C_\alpha$0 with masks $C_\alpha$1. At iteration $C_\alpha$2,

$C_\alpha$3

$C_\alpha$4

$C_\alpha$5

This gradually refines secondary-structure constraints and steers denoising towards physically plausible fold grammars. Two inference modes are defined: SeedProteo-R, which prioritizes fewer, longer secondary-structure segments and stronger stability constraints, and SeedProteo-D, which allows shorter, more frequent segments and richer topology.

A third self-conditioning signal is derived from the previous denoised step. The previous-step structure is converted to a $C_\alpha$6–$C_\alpha$7 distance map, binned into one-hot distogram bins, and injected as a pairwise template feature. This is reported to stabilize sampling trajectories and help maintain long-range contacts (Wei et al., 30 Dec 2025).

3. Coordinate diffusion, atom14 representation, and geometric supervision

SeedProteo uses coordinate-space diffusion and standard folding auxiliary losses. The reported formulation uses a coordinate diffusion loss and does not enumerate explicit noise schedules, but the paper gives a typical DDPM-like coordinate diffusion form. For atom coordinates $C_\alpha$8, the forward process is

$C_\alpha$9

equivalently

$C/O$0

with $C/O$1 and $C/O$2. The reverse parameterization is

$C/O$3

with

$C/O$4

The training objective is reported as

$C/O$5

The coordinate diffusion loss is applied without pre-aligning the target and predicted structures. The paper states that this pushes the architecture to learn $C/O$6 invariances from rotation/translation-invariant auxiliary losses, namely smooth lDDT and distogram supervision, together with the diffusion target itself.

The all-atom representation is standardized through an atom14 schema for every amino acid: 4 backbone atoms $C/O$7 and up to 10 sidechain atoms, with virtual atoms overlaid on $C/O$8 when amino-acid identity is unknown during generation. The paper does not report explicit energy terms such as steric clash penalties, rotamer libraries, or Ramachandran priors. It also does not introduce explicit torsion-angle modeling or periodic losses on $C/O$9, $a$0, or $a$1. Physical plausibility is instead enforced through learned geometric reasoning in pair-representation space, rotation/translation-invariant auxiliary losses, and self-conditioning via previous-step distogram templates and secondary-structure constraints. This distinction is important because “all-atom” in SeedProteo refers to direct atom-level coordinate denoising, not to explicit molecular mechanics or torsion-space parameterization (Wei et al., 30 Dec 2025).

4. Data curation and training curriculum

The monomer training corpus was assembled from AFDB and ESMAtlas. For ESMAtlas sequences, structures were re-predicted with AF2 to mitigate ESMFold artifacts. Inclusion criteria were sequence length 50–768, average pLDDT $a$2, and coil fraction $a$3. Redundancy was removed via Foldseek and MMseqs2, yielding a final curated monomer set of approximately $a$4 million structures. Sequences that SeedFold can fold from single sequence were reserved for Stage 1 training.

For multimers, the paper used experimental PPIs from Pinder, filtered to two-chain complexes with coil $a$5 per chain and interfacial $a$6 distances $a$7 Å, then clustered via Foldseek-Multimer to obtain approximately $a$8k cluster representatives. Augmented DDIs were taken from HumanPPI as AFDB-derived domain-domain interactions, similarly filtered, with approximately $a$9 million interaction pairs. The stated rationale is that intrachain DDIs resemble interchain PPIs in physicochemical and co-evolutionary properties.

The training pipeline is reported as follows:

Stage	Configuration	Data
Stage 1 (Initial)	crop size 384; batch size 128; motif percentage 0%; steps 50k	strictly filtered monomer 100%
Stage 2 (Fine-tuning 1)	crop size 768; batch size 64; motif percentage 20%; steps 20k	expanded monomer 80%, strictly filtered monomer 20%
Stage 3 (Fine-tuning 2)	crop size 768; batch size 64; motif percentage 20%; steps 30k	expanded monomer 40%, strictly filtered monomer 10%, multimer 50%

Motif conditioning was introduced in monomer training during Stages 2 and 3 by injecting partial structures and sequences at 20%, with the stated goal of improving responsiveness to conditional inputs such as secondary structure or templates. Optimizer, learning rate, and hardware are not specified. A plausible implication is that the curriculum is designed to move from strict monomeric folding regularities toward longer-context and interface-aware design without changing the core denoising objective (Wei et al., 30 Dec 2025).

5. Unconditional generation performance

For unconditional generation, the evaluation protocol redesigns one sequence per generated backbone via ProteinMPNN and refolds the redesigned sequence with SeedFold in single-sequence mode. A design is considered successful if the $z$0-RMSD between generated backbone and refold is $z$1 Å and the refolded average pLDDT is $z$2. Diversity is measured as the number of unique structural clusters by Foldseek, and novelty is defined as the maximum TM-score relative to the PDB on successful designs, with lower values indicating greater novelty.

The reported length generalization result is that SeedProteo maintains robust success rates across 100–1000 residues, achieving $z$3 even at length 1000, whereas La-Proteina, Proteina, and RFDiffusion degrade sharply, with near-zero success beyond 600 residues. The paper also reports that SeedProteo sustains high unique cluster counts as length increases, while baselines exhibit mode collapse.

Topology-stratified results are more specific. For HHH proteins, SeedProteo sustains high success and unique clusters up to 1000 residues; at length 800 the reported values are 65% success, 39 unique clusters, novelty 0.79, and at length 1000 they are 63%, 41, and 0.77. For HEL proteins, SeedProteo reaches 57% success at 1000 residues, with 20 unique clusters and novelty 0.74. For EEE proteins, described as notoriously challenging, SeedProteo is reported as the only method to generate diverse, valid, novel EEE structures up to 500 residues; at length 500 the values are 50% success, 1 unique cluster, and novelty 0.68.

These benchmarks support two of the paper’s broader claims. First, the architecture shows superior length generalization despite Pairformer cubic complexity and reduced encoder depth. Second, the combination of self-conditioning and global sequence decoding appears especially consequential for long-range topological coherence, particularly in $z$4-rich regimes where baselines either fail or collapse (Wei et al., 30 Dec 2025).

6. Binder design formulation and benchmarked results

Binder design conditions on full atomic target structures in mmCIF format together with hotspot residues defined according to AlphaProteo benchmarks. The model conditions on the target through encoder pair and single features and secondary-structure constraints, while the previous-step binder structure supplies self-conditioning templates through $z$5 distograms. For each of 10 targets, approximately 1,000 candidates are sampled, covering binder-length ranges via dense 5-residue grids; the paper gives the example that SC2RBD length 80–120 corresponds to 900 candidates.

For structure-generating models, including SeedProteo, RFDiffusion variants, PXDesign, and BoltzGen, binder sequences are redesigned via ProteinMPNN or SolubleMPNN with $z$6 and two sequences per backbone. Designs are validated in silico by SeedFold in single-sequence, target-template mode. The AlphaProteo success criteria are:

$z$7

The reported SeedProteo success counts are:

Target	SeedProteo-R	SeedProteo-D
PD-L1	380	265
Insulin	303	181
TrkA	232	143
VEGF-A	127	45
SC2RBD	92	80
IL-7RA	100	52
BHRF1	296	139
H1	16	25
IL-17A	47	17
TNF$z$8	1	3

Against open-source baselines, the paper states that SeedProteo achieves state-of-the-art performance in success rates, structural diversity, and novelty. BoltzGen is noted as having target-specific strength—for example, BHRF1 627—but also zero on SC2RBD and low or zero on many other targets. RFDiffusion3 is reported as having single-digit successes on BHRF1 and zero elsewhere. A notable case is TNF$z$9, where ProteinMPNN-based pipelines fail at 0–1 success, whereas SeedProteo co-design yields valid binders; the paper interprets this as suggesting that co-design can escape “safe” local minima of inverse folding on difficult interfaces.

On diversity and novelty, SeedProteo-D attains the highest unique success clusters in 8 of 10 targets and the best novelty in 8 of 10, with lower TM-score to PDB typically around 0.80–0.84 versus $x = (x_1,\ldots,x_L)$0 for PXDesign. The paper gives examples including TrkA novelty 0.829, PD-L1 0.832, and IL-17A 0.806. SeedProteo-R shows higher novelty numbers, interpreted in the paper as less novel, consistent with its stability bias, but remains competitive; SC2RBD 0.858 is given as an example. Qualitative case studies include multichain targets, specifically the dimers H1 and VEGF-A and the trimer TNF$x = (x_1,\ldots,x_L)$1, where displayed binders satisfy the success criteria (Wei et al., 30 Dec 2025).

7. Ablations, limitations, and place among related methods

The ablation study on secondary-structure constraints compares SeedProteo-D and SeedProteo-R against SeedProteo-M, which uses fully masked secondary structure during design. The reported result is that secondary-structure conditioning significantly increases success counts. The paper interprets this as evidence that secondary-structure constraints prune the search space and guide diffusion toward de novo-like folds with better designability.

A second ablation compares MRF sequence decoding against atom14 local decoding. At length 100, atom14 local decoding has scRMSD 1.11, scTM 0.94, pLDDT 88.56, success rate 0.90, while MRF has scRMSD 1.63, scTM 0.93, pLDDT 88.65, success 0.84. At length 200, atom14 gives scRMSD 3.87, scTM 0.86, pLDDT 78.06, success 0.68, whereas MRF gives scRMSD 2.08, scTM 0.92, pLDDT 84.36, success 0.80. At length 300, atom14 gives scRMSD 4.71, scTM 0.83, pLDDT 73.40, success 0.52, whereas MRF gives scRMSD 2.30, scTM 0.91, pLDDT 80.33, success 0.76. The paper’s interpretation is that local “atom14-to-AA” decoding collapses for larger proteins because of limited receptive field and difficulty maintaining global consistency, while the global MRF leverages pair features to enforce long-range sequence–structure compatibility.

In positioning, the paper describes RFdiffusion as backbone-only and PXDesign as accelerated backbone binder generation, with both deferring sidechains and sequence design to inverse folding. RFDiffusion3 and other all-atom models such as BoltzGen and Latent-X are described as typically inferring amino-acid identities from local atom14 geometry, sometimes by removing virtual atoms. Flow-based generative methods, specifically Proteina and La-Proteina, are described as scaling generation but exhibiting mode collapse and length-dependent degradation beyond approximately 600 residues. Chroma and multimodal LLMs such as DPLM-2 are described as exploring co-generation but often showing lower sequence consistency than redesigning with ProteinMPNN. Within this landscape, SeedProteo’s distinctives are the AF3-like encoder adapted to accept noisy coordinates, strong self-conditioning, all-atom coordinate diffusion with rotation/translation-invariant auxiliary losses, and a training curriculum mixing curated monomers with PPIs and DDIs.

The paper also states several limitations. The reported results rely on in-silico metrics, including SeedFold pTM, pAE, and RMSD, and no wet-lab validation is provided. True binding affinity and specificity remain to be tested experimentally. Off-target risks and immunogenicity are not assessed. Future work is framed around possible integration of explicit physicochemical terms such as clash penalties and rotamer priors, enhanced $x = (x_1,\ldots,x_L)$2-equivariant modules, torsion-space periodic modeling such as von Mises losses on $x = (x_1,\ldots,x_L)$3 angles, learned binding-energy proxies such as $x = (x_1,\ldots,x_L)$4, and downstream screening pipelines including MD and Rosetta. Reproducibility is supported through a project page and appendix-level documentation of data curation and training stages, while optimizer, learning rate, precise inference speed, hardware footprint, and hardware configuration are not reported (Wei et al., 30 Dec 2025).