SymFold

An inverse folding pipeline hands off twice. A graph network reads the backbone and writes a draft sequence; a protein language model then edits that draft into something more evolutionarily plausible. But the language model never sees the backbone — it is handed a string of letters and asked whether they read like a protein. So when the draft has already put a hydrophobic residue on the surface, the edit that makes the sequence look more natural can make it fit the structure worse, and the result stays capped by what the encoder handed over.

SymFold keeps that draft and adds a second reader that does see the geometry. Multimodal protein language models — ESM-3 among them — take atomic coordinates as input tokens and emit sequences directly, so they hold structural priors a sequence-only model cannot. Used alone they are poor inverse folders: zero-shot ESM-3 recovers 42.03% of residues on CATH 4.2, below a small graph network trained from scratch, and fine-tuning lifts it only to 48.46%. The paper treats that as a fusion problem, not a capacity problem, running the sequence-side and structure-side models in parallel over the same draft.

How much each should count varies along the chain. A buried residue in a packed core is constrained mostly by local geometry; a surface residue on a long loop, mostly by sequence context. An Adaptive Synergistic Fusion module reads the frozen encoder's structural embedding at each position and emits a scalar gate that scales the structure branch's logits before they are added to the sequence branch's. SymFold comes from Handong Wang and colleagues at the Computer Network Information Center, CAS and the University of Chinese Academy of Sciences.

#Key Features

• Symmetric dual-path refinement: The sequence model and the multimodal model refine the draft in parallel rather than in series, so the structural stream is never downstream of a text-only edit it cannot see.
• Residue-wise adaptive fusion: A small weighting network — layer norm, linear, SiLU, 1D convolution, linear — turns structural embeddings into one gating coefficient per position, which beats uniform averaging, a global scalar, or per-sequence weighting.
• Self-correction training: Training unrolls the iterative refinement that inference performs, mixing both branches' outputs back with the coarse sequence before a final pass. That closes the one-step-train, multi-step-test gap and peaks in fewer iterations.
• Swappable structure encoder: The encoder is a slot, not a fixed part. Wrapping ProteinMPNN instead of PiFold lifts that encoder too, and a stronger encoder gives a stronger fused model.

#Technical Details

The structure encoder is a frozen, pretrained PiFold graph network. The sequence branch is ESM-C at 600M parameters and the structure branch is ESM-3 at 1.4B, co-trained counterparts that pair cleanly. LoRA adapters at rank 8, scaling factor 32 and dropout 0.1 attach to the attention and feed-forward blocks of both, leaving roughly 0.1% of parameters trainable; training runs on a single NVIDIA A800 at batch size 4 and converges within about five epochs, and inference uses two self-correction steps. On CATH 4.2 (18,024 / 608 / 1,120 proteins) SymFold reaches 63.11% median recovery at perplexity 3.23, against 58.59% for Bridge-IF and 51.66% for PiFold, and 62.23% on CATH 4.3. TS50 and TS500 give 66.02% and 70.48%. On CASP15 and CASP16 monomer targets — 45 and 50 structures released after the training data — recovery is 56.57% and 52.18%, against 43.06% and 39.32% for ProteinMPNN. Removing either branch costs several points: 56.62% with the structure path alone, 59.56% with the sequence path alone.

#Applications

Inverse folding is the step between a backbone and a molecule someone can order: sequences for an engineered enzyme scaffold, for a de novo backbone, or for a pocket redesigned after a binding site is grafted in. Because the gate leans on geometry where geometry is decisive, the gains concentrate on structurally constrained targets — 50.00% recovery on short proteins under 100 residues and 55.45% on single-chain proteins, where a sequence-driven editor has least context to work with.

#Impact

The negative result may outlast the architecture: multimodal protein language models read coordinates and write sequences in one step, which makes them look like drop-in inverse folders. SymFold's measurements say otherwise. Evaluation is entirely in silico: two CASP16 designs refolded with AlphaFold 3 give pLDDT 78.24 and 93.15 and TM-scores of 0.85 and 0.97 to the native backbones, and the authors leave functional viability unverified. An earlier version was submitted to ICLR 2026 as DualFold and withdrawn; that name still appears in several appendix captions. Distribution is the sharpest limitation: the authors' repository holds a README reading "Code Coming Soon" and nothing else, with no code, weights or license, and the paper carries no availability statement. Every corpus and frozen component it builds on is public, but there is no SymFold to run.