bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / rp3net
ProteinDNA & Gene
EMBL-EBIAstraZenecaReleased May 2025

RP3Net

Recombinant protein expression predictor scoring soluble E. coli production from construct sequence, with AUROC 0.83 on prospective constructs.

82Openness

Where to run it

No providers recorded yet. Browse all providers

RP3NetProteinEMBL-EBI

RP3Net (Recombinant Protein Production Prediction Network) predicts whether a given protein construct will express solubly at small scale in Escherichia coli. Recombinant expression is one of the most common hard stops in protein science: a target that will not express cannot be purified, assayed, crystallised, or fed into a structure-based discovery campaign, and each failed attempt costs weeks of bench time. RP3Net turns that gamble into a score — a single probability per construct, computed from sequence alone before any wet-lab work begins.

The model was developed at EMBL-EBI in collaboration with AstraZeneca by Evgeny Tankhilevich, Sergio Martinez Cuesta, Ian Barrett, Carolina Berg, Lovisa Holmberg Schiavone, and Andrew Leach. It was posted as a preprint in May 2025 and published in Bioinformatics in January 2026. Rather than training a bespoke sequence encoder from scratch on a small expression dataset, RP3Net treats the problem as a probing task over existing pretrained protein and nucleotide models, adding a learned aggregation head and a label-denoising procedure on top of frozen or lightly fine-tuned foundation-model representations.

What distinguishes RP3Net from earlier solubility and expressability predictors is the evidence behind it. Most such tools are evaluated on retrospective splits of the same corpus they were fitted to. RP3Net was tested prospectively on 97 constructs that were manually selected, expressed, and scored after the model was frozen — and it is already in routine use by protein scientists at AstraZeneca.

#Key Features

  • Foundation-model features: Sequence representations come from pretrained models rather than a task-specific encoder, with the 650M-parameter ESM-2 protein language model and the codon model CaLM selected after a systematic comparison.
  • Set Transformer Pooling: Instead of averaging per-residue embeddings, a multi-head attention layer with a trainable seed vector learns which residues matter for expressability, giving a context-dependent construct-level representation.
  • Meta Label Correction: A teacher network trained on abundant but noisy large-scale purification outcomes rewrites the labels used to train the student on the smaller, cleaner small-scale expression data, recovering signal that would otherwise be discarded.
  • Prospective experimental validation: On 97 independently chosen constructs, RP3Net reached an AUROC of 0.83, gave a correct call in 77% of cases, and correctly flagged 92% of the constructs that did express.
  • Deployable and permissively licensed: The trained model ships as an MIT-licensed PyPI package with a command-line interface, a Python API, a Docker image, and a one-click Colab notebook that runs inference straight from a FASTA file.

#Technical Details

RP3Net was trained on a curated corpus of 67,055 unique sequences spanning 5,077 target proteins, drawn from internal AstraZeneca small-scale expression screens, small-scale expression and one-step purification data from SGC Stockholm, and large-scale purification records from SGC Toronto. Candidate backbones were benchmarked before selection: on the protein side ProtTrans ProtBert and ProteinBERT alongside ESM-2, and on the nucleotide side DNABERT, HyenaDNA, and CaLM. Among the DNA and codon models, only CaLM improved on the baseline.

Performance climbs steadily with each architectural component. An XGBoost baseline reaches an AUROC of 0.62 on the SGC Stockholm test data; a frozen ESM-2 encoder with mean pooling reaches 0.70; and the full model, combining fine-tuning, Set Transformer Pooling, and Meta Label Correction, reaches 0.77 — the 0.15 improvement over baseline reported in the abstract. The prospective AUROC of 0.83 on the 97-construct external set is higher still. Inference is cheap: a batch of 16 constructs under 500 amino acids completes in roughly three minutes on CPU using under 5 GB of RAM.

#Applications

The immediate use is construct triage. Protein production groups routinely design several constructs per target — different boundaries, tags, fusion partners, and codon optimisations — and must choose which to take to the bench. RP3Net ranks those options in seconds, letting teams drop low-probability designs before they consume expression slots. That matters most for reagent supply in drug discovery, where structural biology, biophysics, and assay development all queue behind a purified protein, and for high-throughput structural genomics pipelines where thousands of targets compete for finite capacity.

#Impact

RP3Net is a rare example of a protein-sequence model that has been validated prospectively and then actually adopted: it is deployed for AstraZeneca protein scientists rather than remaining a benchmark result. It also makes a broader methodological point — that pairing a protein language model with a codon model captures signal neither carries alone, since expressability depends on both the folded product and the nucleotide sequence being translated. The main limitation is scope: the model is fitted to small-scale soluble expression in E. coli and does not speak to yield at manufacturing scale, to eukaryotic expression hosts, or to whether the expressed protein is functional.

At a glance

Released
May 2025
Category
Protein
Organizations
EMBL-EBI / AstraZeneca

Links

GitHub RepositoryResearch PaperbioRxiv PreprintGoogle ColabDataset

Tags

foundation_modeltransfer_learningtransformer

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.