PRISM

PRISM (Protein inverse folding through fine-grained Retrieval on structure-sequence multimodal representations) addresses inverse folding: designing an amino acid sequence that will fold into a given three-dimensional backbone. Inverse folding is a core step in de novo protein and enzyme design, and models such as ProteinMPNN and ESM-IF established structure-conditioned sequence generation as a workhorse tool. PRISM improves on this line of work by treating sequence design as a retrieval-augmented generation problem rather than relying solely on the parameters learned during training.

The central idea is that useful local sequence-structure patterns—motifs—recur across the known protein universe, and that retrieving them at design time supplies evidence that a purely parametric decoder would have to memorize. PRISM retrieves fine-grained candidate motif representations from a corpus of known proteins and integrates them with the query backbone through a hybrid self- and cross-attention decoder. The method is cast as a latent-variable probabilistic model and implemented with an efficient approximation, combining a principled generative formulation with practical scalability.

PRISM was developed by researchers at Carnegie Mellon University, the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Intel, and was accepted to ICLR 2026.

#Key Features

• Fine-grained motif retrieval: Rather than retrieving whole proteins, PRISM retrieves granular sequence-structure motif representations from known proteins and conditions generation on them, injecting relevant priors at design time.
• Multimodal structure-sequence representations: Retrieval operates over a joint representation space, so structural context and sequence evidence are matched together instead of through structure alone.
• Hybrid self-cross attention decoder: A decoder that interleaves self-attention over the query backbone with cross-attention to retrieved motifs fuses target geometry and retrieved evidence into per-residue predictions.
• Latent-variable probabilistic formulation: The retrieval-augmented objective is derived as a latent-variable model with an efficient approximation, giving the approach a theoretical grounding while remaining tractable to train and sample.
• Trained once, evaluated zero-shot: A single model trained on CATH-4.2 generalizes to held-out benchmarks without task-specific retraining.

#Technical Details

PRISM is trained on the CATH-4.2 dataset, the standard split used to benchmark inverse folding, and evaluated across CATH-4.2, TS50, TS500, CAMEO 2022, and PDB date-based splits. It reports state-of-the-art sequence perplexity and native amino acid recovery across these benchmarks, and also improves foldability metrics—self-consistency RMSD, TM-score, and pLDDT—computed by refolding designed sequences. The architecture combines a retrieval index over multimodal motif representations with a hybrid attention decoder, and the latent-variable objective is optimized with an efficient variational-style approximation so that retrieval can be incorporated without prohibitive cost.

#Applications

PRISM is aimed at protein engineers and computational biologists who need to design sequences for a fixed target backbone—for example, scaffolding a functional motif, redesigning a protein for stability, or generating diverse candidate sequences for a de novo structure. Higher native sequence recovery and improved refolding metrics translate into candidate sequences that are more likely to adopt the intended fold, which reduces the number of designs that must be screened experimentally. The retrieval component also makes the model's behavior more interpretable, since the motifs informing a design can be inspected.

#Impact

By reframing inverse folding as retrieval-augmented generation, PRISM demonstrates that non-parametric memory can push recovery and foldability past strong parametric baselines on widely used benchmarks. Its acceptance at ICLR 2026 places it among recent efforts to bring retrieval methods, which have reshaped natural language modeling, into structural biology. As a preprint-stage method, its practical advantage over established tools such as ProteinMPNN will ultimately be judged by experimental validation of its designs; the reported gains are so far in-silico.