Protein inverse folding ensemble that fuses five pretrained sequence designers through a self-attention encoder, reaching 63.1% recovery on CATH4.2.
No providers recorded yet. Browse all providers
Inverse folding — writing an amino acid sequence that will fold into a given backbone — is now served by several strong models, and they disagree with each other in instructive ways. Message-passing designers are trained on multi-chain assemblies and handle interfaces well; autoregressive sequence-structure transformers capture evolutionary signal but were trained on proteins shorter than 500 residues; models optimized on soluble subsets favor hydrophilic surface residues. Each therefore carries a different bias, and none dominates across protein length, solvent exposure, or conservation.
CHIEF (Chimera Ensemble Inverse Folding) treats that heterogeneity as a resource rather than a problem. Developed by Zilong Geng, Yuze Wang, Tingting Liu and colleagues at Shanghai Jiao Tong University with collaborators at several Chinese hospitals and universities, it runs five frozen pretrained designers — ProteinMPNN-vanilla, ProteinMPNN-soluble, ESM-IF, Frame2seq and PiFold — over the same backbone, then learns a transformer encoder that combines their per-residue probability distributions. Unlike averaging or weighted voting, which assign each base model a fixed weight, the attention layer reweights them according to sequence context, so a model can dominate exactly where it is strong. Two of the five base models are variants of ProteinMPNN.
This entry describes a protein design model and is unrelated to CHIEF, the histopathology foundation model that shares the acronym.
Each base model receives backbone N, Cα, C and O atoms and emits a 21-dimensional log-probability per residue (20 amino acids plus a missing token). The five outputs are concatenated into an [N, 105] tensor, augmented with sinusoidal positional encodings, and passed through a three-layer transformer encoder with four attention heads and a hidden dimension of 128, trained with a masked negative log-likelihood objective using Adam at a learning rate of 0.001 for more than 200 epochs on a single A100. Training used PDB structures clustered at 30% sequence identity, split into 23,358 training, 1,464 validation and 1,539 test clusters. Sequence recovery reaches 61.8% on the PDB test set (perplexity 3.4), 63.1% on single-chain CATH4.2, 70.6% on TS50 and 71.2% on TS500 — 16.6–28.0% above the best base model, with perplexity 22.7–34.6% lower. Averaging and weighted-voting ensembles of the same five models reach 56.3% and 59.2%. Designed sequences refolded with ESMFold match or beat the base models on RMSD, TM-score, pLDDT and secondary structure recovery. On ProteinGym deep mutational scanning, CHIEF ranks mid-pack on Spearman correlation, AUC and MCC — below ESM-IF — but leads on NDCG (0.75) and top-10% recall (0.22), the metrics that matter for prioritizing high-fitness variants.
CHIEF is aimed at enzyme engineering and functional protein design, where the goal is a sequence that folds and works rather than one that merely matches a reference. The authors redesigned malate dehydrogenase from an AlphaFold-predicted backbone (UniProt A0A319AA41), sampling 1,000 sequences at temperature 0.15 and synthesizing ten; all ten expressed in E. coli BL21, purified, and showed measurable enzymatic activity, with one comparable to the wild-type enzyme. A second demonstration redesigned the 954-residue nuclease Cas13d, reaching a TM-score of 0.92 against the AlphaFold 3 prediction of the wild type, versus 0.72 for the best single base model.
The result argues that ensembling frozen inverse folding models is worth more than it sounds: the gain over the strongest member is large, and the least-squares attribution analysis recovers the expected division of labor, with the autoregressive model carrying early residues and conserved positions while the multi-chain designers carry interfaces and surfaces. Practical uptake is limited for now. The work is a preprint that has not been peer reviewed, the wet-lab validation rests on ten designed sequences of a single enzyme, and no code or weights have been released — the paper states that the repository is being prepared for public release. Until that lands, CHIEF is a reproducible recipe rather than a runnable tool.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.