i-Fold

i-Fold (importanceFold) is a protein structure prediction model that enhances the AlphaFold 2 architecture by directing the network's attention toward the residues that matter most for a given protein. Although AlphaFold 2 was a landmark advance in predicting tertiary structure from sequence, systematic discrepancies persist between its predictions and experimentally determined structures, particularly for difficult protein classes. i-Fold is built on the premise that not all residues contribute equally to a protein's fold and function, and that making the model aware of this can sharpen its predictions.

The model was developed by Yu Liu, Boming Kang, and Qinghua Cui at Peking University (with a joint appointment at Wuhan Sports University), and released as a bioRxiv preprint in September 2025. Its central idea is to compute residue importance scores (RIS) from the ESM protein language model and use them as dynamic positional weights during structure-prediction training, so that functionally critical residues and regions receive proportionally more of the model's representational capacity.

i-Fold joins a growing family of catalog structure predictors that adapt or extend the AlphaFold 2 recipe — including TDFold and DCFold — and is evaluated as a general-purpose predictor: a single fixed checkpoint applied to arbitrary input sequences with no per-target retraining.

#Key Features

• Importance-weighted training: Residue importance scores derived from a protein language model act as dynamic positional weights, steering the network toward functionally critical residues rather than treating all positions equally.
• AlphaFold 2 backbone: i-Fold builds directly on the AlphaFold 2 architecture, inheriting its evoformer-based structure module while adding the importance-weighting mechanism.
• Gains on hard targets: Improvements are most pronounced for protein classes that are historically difficult for AlphaFold 2, including ribosomal proteins, membrane proteins, and orphan proteins with few homologs.
• No added compute burden: The importance weighting improves accuracy without substantially increasing the computational cost of inference relative to the baseline.

#Technical Details

i-Fold augments the AlphaFold 2 architecture with residue importance scores computed from the ESM protein language model, injecting them as dynamic positional weights during training. On a benchmark test set of 3,559 protein structures, i-Fold significantly reduced r.m.s.d. relative to AlphaFold 2 (p = 0) and raised the prediction success rate by 7.6 percentage points, from 55.1% to 62.7%. On a completely independent, temporally held-out set of 167 recently released structures, it again improved the success rate by 6.0 points, from 43.7% to 49.7%, demonstrating that the gains generalize beyond the training distribution rather than reflecting overfitting to a fixed benchmark. Code supporting the method is deposited on GitHub, released without a stated license.

#Applications

i-Fold serves the same broad role as other single-sequence-to-structure predictors: generating three-dimensional models of proteins for structural biology, functional annotation, and downstream design tasks. Its particular value is on targets where AlphaFold 2 struggles — membrane proteins, ribosomal proteins, and orphan proteins lacking deep multiple-sequence alignments — making it a useful complement for researchers working on these historically challenging classes.

#Impact

By showing that explicit residue-importance signals from a protein language model can be folded into the AlphaFold 2 training objective to yield measurable, generalizable accuracy gains, i-Fold illustrates a lightweight route to improving established structure predictors without redesigning them. The reported improvements on independent, recently released structures are its strongest evidence. As a preprint awaiting peer review, with code released under no explicit license and no separately published weights or parameter count, its broader adoption will depend on independent reproduction and a clearer release of trained models.