RIGA-Fold: A General Framework for Protein Inverse Folding via Recurrent Interaction and Geometric Awareness

Protein inverse folding, the task of predicting amino acid sequences for desired structures, is pivotal for de novo protein design. However, existing GNN-based methods typically suffer from restricted receptive fields that miss long-range dependencies and a “single-pass” inference paradigm that leads to error accumulation. To address these bottlenecks, we propose RIGA-Fold, a framework that synergizes Recurrent Interaction with Geometric Awareness. At the micro-level, we introduce a Geometric Attention Update (GAU) module where edge features explicitly serve as attention keys, ensuring strictly SE(3)-invariant local encoding. At the macro-level, we design an attention-based Global Context Bridge that acts as a soft gating mechanism to dynamically inject global topological information. Furthermore, to bridge the gap between structural and sequence modalities, we introduce an enhanced variant, RIGA-Fold*, which integrates trainable geometric features with frozen evolutionary priors from ESM-2 and ESM-IF via a dual-stream architecture. Finally, a biologically inspired ``predict-recycle-refine’’ strategy is implemented to iteratively denoise sequence distributions. Extensive experiments on CATH 4.2, TS50, and TS500 benchmarks demonstrate that our geometric framework is highly competitive, while RIGA-Fold* significantly outperforms state-of-the-art baselines in both sequence recovery and structural consistency.

💡 Research Summary

Protein inverse folding—predicting an amino‑acid sequence that adopts a given backbone—has become a cornerstone of de novo protein design, yet existing graph‑neural‑network (GNN) approaches suffer from limited receptive fields, over‑smoothing, and a single‑pass inference that propagates early errors. In this work the authors introduce RIGA‑Fold, a novel framework that couples recurrent interaction with explicit geometric awareness to overcome these bottlenecks, and an enhanced variant, RIGA‑Fold*, that integrates frozen large‑scale protein language models (PLMs) in a dual‑stream architecture with an iterative “predict‑recycle‑refine” loop.

Core technical contributions

1. Geometric Attention Update (GAU). Traditional graph attention networks compute attention solely from node embeddings. GAU treats edge features (distance, angles, radial‑basis‑function encodings) as the attention keys. For a central node i and neighbor j, the query is derived from the node, the key from the edge, and the value from the concatenation of node‑i, edge‑ij, and node‑j. This “edge‑as‑key” design directly injects spatial proximity into the attention scores while preserving SE(3) invariance, enabling the model to capture subtle geometric cues such as side‑chain stacking.
2. Dynamic Edge Update. After each GAU step, edge embeddings are refined via an MLP that takes the current node representations and the original edge, ensuring that geometric information evolves together with node features and does not vanish in deeper layers.
3. Global Context Bridge. To compensate for the locality of k‑NN graphs, a global pooling vector is computed with feature‑wise attention, then passed through a two‑stage gating mechanism: an input‑dependent gate that fuses the pooled vector with local node features, followed by an output gate that modulates the final node update. This hierarchical gating creates a “virtual bridge” that selectively injects long‑range information, mitigating over‑smoothing while preserving local specificity.
4. Dual‑Stream Fusion with PLMs. RIGA‑Fold* incorporates frozen embeddings from ESM‑IF (structure‑aware) and ESM‑2 (sequence‑aware). Geometric node embeddings from the GAU‑based encoder are concatenated with these PLM vectors, then processed by a tuning module that outputs per‑residue amino‑acid probabilities. The PLMs remain frozen, providing stable evolutionary priors while the geometric encoder learns task‑specific refinements.
5. Iterative Self‑Correction (Predict‑Recycle‑Refine). Inspired by AlphaFold’s recycling, the model stacks T identical RIGA‑Fold modules (typically T = 3). In the first pass, only the structural PLM (ESM‑IF) is active; the sequence PLM receives a generic token. After each pass, the predicted probability distribution is decoded into a candidate sequence, which is fed back into the frozen ESM‑2 encoder, producing an updated semantic prior for the next pass. Losses are summed across all passes, encouraging consistent improvement at every iteration.

Experimental evaluation
The authors train on the CATH 4.2 dataset (≈18 k chains) and evaluate zero‑shot generalization on TS50 and TS500. Metrics are perplexity (lower is better) and sequence recovery (percentage of exact matches).

• On CATH 4.2 (overall, short, and all‑chain splits), RIGA‑Fold* achieves perplexities of 4.25 (overall) and recovery rates of 49.59 %, surpassing the previous best ScFold (5.80/5.99 PPL, 41.66/40.10 % recovery) and other strong baselines such as PiFold, VFN‑IF, and Knowledge‑Design.
• On TS50, RIGA‑Fold* reaches a perplexity of 2.95 and recovery of 65.74 %, beating the prior state‑of‑the‑art PiFold (3.86 % PPL, 58.72 % recovery) and Knowledge‑Design (3.10 % PPL, 62.79 % recovery).
• On TS500, it records a perplexity of 2.77 and recovery of 70.15 %, again outclassing all listed competitors.

Model configuration details: five message‑passing layers (hidden dimension = 128, k‑NN = 48), Global Context Bridge enabled, and three recycling steps. RIGA‑Fold* adds frozen ESM‑IF and ESM‑2 embeddings fused via the dual‑stream module. Training employs noise‑augmented inputs, AdamW optimizer, and runs on multi‑GPU hardware; full hyper‑parameters are provided in the appendix.

Strengths

• The edge‑as‑key GAU is a principled way to embed geometry directly into attention, preserving SE(3) invariance while enriching the message‑passing signal.
• The Global Context Bridge’s double‑gating offers a flexible, learnable pathway for long‑range information without overwhelming local features.
• Integration of frozen PLMs supplies rich evolutionary priors without the need for costly fine‑tuning, and the recycling loop enables progressive denoising of the sequence distribution.
• Empirical results are comprehensive, covering both in‑distribution (CATH) and out‑of‑distribution (TS50/TS500) benchmarks, with clear gains over a wide range of strong baselines.

Potential weaknesses and open questions

• The architecture introduces multiple gating layers and a recycling loop, increasing memory consumption and inference latency, which may limit scalability to very large proteins or high‑throughput design pipelines.
• Frozen PLMs, while stable, cannot adapt to task‑specific nuances; future work could explore lightweight fine‑tuning or adapter modules to further boost performance.
• The evaluation focuses on standard benchmarks; real‑world design tasks (e.g., enzyme active‑site engineering, antibody affinity maturation) remain to be tested.
• Ablation studies on the number of recycling steps, the impact of each gating component, and the contribution of each PLM are mentioned but not detailed in the main text; deeper analysis would clarify the relative importance of each module.

Impact and future directions
RIGA‑Fold establishes a new paradigm for protein inverse folding by tightly coupling geometric attention with dynamic global context and evolutionary priors. Its ability to capture long‑range dependencies while iteratively refining predictions positions it as a strong candidate for next‑generation de novo design tools. Future extensions could include: (i) model compression or sparse attention to improve scalability, (ii) incorporation of additional modalities such as co‑evolutionary contacts or functional annotations, and (iii) experimental validation on designed proteins to demonstrate practical utility. Overall, the paper makes a substantial methodological contribution and sets a solid foundation for further advances in computational protein design.