Protriever

Almost everything a model can infer about which mutations a protein tolerates comes from its evolutionary relatives. The standard way to assemble those relatives has not changed in decades: run a homology search, build a multiple sequence alignment, then train or condition a model on it. That pipeline has three structural weaknesses. Alignment-based search misses distant homologs that fall below significance thresholds; sequences with large insertions, deletions, or rearrangements resist reliable alignment even when they are functionally related; and retrieval is decided entirely before modeling begins, with no feedback from whether the retrieved set was actually useful.

Protriever removes the separation. Developed at the University of Oxford and Harvard Medical School with collaborators at the University of Copenhagen, Novonesis, and Ginkgo Bioworks — the group behind ProteinGym and Tranception — it trains a retriever and a downstream reader jointly, so gradients from the modeling objective reshape what counts as a relevant homolog. Retrieval happens in dense embedding space by vector similarity rather than by alignment, which means it can surface relatives that no aligner would score as significant.

On ProteinGym's 217 deep mutational scanning substitution assays, Protriever reaches a Spearman correlation of 0.479 — the best among sequence-based models — while retrieving homologs about two orders of magnitude faster than GPU-accelerated MMseqs2. It was presented at ICML 2025.

#Key Features

• End-to-end differentiable retrieval: The retriever and reader are trained together with an EMDR objective that marginalizes over the top-K retrieved sequences, so retrieval is optimized for the downstream task rather than fixed in advance.
• Alignment-free homology search: Similarity is cosine distance between learned sequence embeddings, letting the model use divergent or indel-heavy relatives that alignment-based search discards.
• Sub-10-millisecond retrieval: Product-quantized Faiss search over the full UniRef50 index returns homologs in 0.0046 seconds per query, against 0.613 seconds for MMseqs2-GPU.
• Architecture- and task-agnostic: The retriever, the index, and the reader are swappable, so the framework can be pointed at a different protein database or repurposed for tasks beyond fitness prediction at inference time.
• Live index refresh: Because the retriever changes during training, all 62 million index embeddings are recomputed ten times over the course of a run to keep them from going stale.

#Technical Details

The retriever is a transformer encoder initialized from ESM-2 at 35M parameters, mean-pooled to a 480-dimensional vector per sequence, and pretrained with dense passage retrieval against BLAST-identified homologs. The index covers roughly 62 million UniRef50 sequences; storing them at full precision would need over 110 GB, so a product quantizer compresses the vectors and an inverted-file structure is sharded across GPUs. The reader is PoET, the sequence-of-sequences decoder transformer later succeeded by PoET-2, initialized from UniRef50 pretraining and trained with a conditional autoregressive objective over retrieved context; fitness is scored in both the N-to-C and C-to-N directions. Across 217 ProteinGym substitution assays aggregated by protein, Protriever posts Spearman 0.479, AUC 0.762, MCC 0.374, and NDCG 0.788, ahead of PoET (0.470), TranceptEVE L (0.458), GEMME (0.459), MSA Transformer (0.432), ESM-1v (0.407), and ProGen2 (0.391).

#Applications

The immediate use is zero-shot variant effect prediction: ranking substitutions for directed evolution, prioritizing candidates in protein engineering campaigns, and interpreting variants of uncertain significance without task-specific training data. The speed advantage changes what is feasible at scale — proteome-wide variant scoring no longer requires building an alignment per protein. Because retrieval is decoupled from the reader, teams can swap in a proprietary or domain-specific sequence database at inference and get homologs conditioned on it.

#Impact

Protriever is a clear demonstration that retrieval-augmented generation, developed for open-domain question answering, transfers to protein modeling once the retriever is trained against the biological objective rather than an alignment heuristic. It also reframes homology as something learned rather than defined, which has implications well beyond fitness prediction — the same framework could condition structure prediction or design models. Practical adoption is currently constrained: the official repository contains no source code, no pretrained weights, and no license, so the results have not been independently reproduced, and running the method requires the memory and GPU capacity to host a quantized index over tens of millions of sequences.