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 modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / maspr
ProteinSmall molecule
Carnegie Mellon UniversityUniversity of PittsburghMohamed bin Zayed University of Artificial IntelligenceReleased January 2025

MASPR

Adenylation domain substrate specificity prediction from frozen ESM-2 embeddings, with zero-shot calls on substrates absent from training.

The short version

  • —Predicts NRPS adenylation domain substrates from sequence alone, no structure required
  • —Scores substrates it never saw in training by predicting their chemical fingerprint
  • —Trained on bacterial domains and still generalizes to fungal ones
  • —Drives Seq2Hybrid to propose whole NRP-polyketide structures from microbial genomes
36Openness

Where to run it

No providers recorded yet. Browse all providers

Non-ribosomal peptides and their polyketide hybrids supply a large share of clinically used antibiotics, antifungals, immunosuppressants, and anticancer agents. They are assembled by modular megaenzymes in which each adenylation domain selects one building block, so reading a biosynthetic gene cluster off a genome and predicting the compound it makes reduces, in large part, to predicting what substrate each A-domain will pick. Existing predictors handle the substrates they were trained on and degrade sharply on anything unfamiliar — which is precisely the regime that matters, since the point of genome mining is to find compounds nobody has characterised.

MASPR — modelling A-domain specificity using unsupervised pretrained representations — is a specificity predictor built on frozen ESM-2 embeddings by the Mohimani, Behsaz, Koes, and Xing groups at Carnegie Mellon University and the University of Pittsburgh. It is not a general protein language model in its own right; it is a small supervised head over one, and its contribution lies in what that head predicts. Instead of classifying directly into a fixed substrate vocabulary, MASPR predicts a molecular fingerprint of the substrate, which lets it rank candidates that never appeared in training and makes its reasoning inspectable at the level of chemical substructures.

The method was posted to bioRxiv in January 2025 and is awaiting peer review.

#Key Features

  • Chemistry-space output rather than a fixed label set: The network predicts a 296-dimensional substrate fingerprint, so any molecule that can be fingerprinted becomes a scoreable candidate — the mechanism behind its zero-shot behaviour.
  • Frozen backbone, tiny head: ESM-2 weights are never updated. Only the fingerprint predictor and classifier train, which is why the released checkpoint is a few megabytes and inference is cheap.
  • Focused on the specificity-conferring residues: Embeddings are extracted at the ten Stachelhaus code positions rather than pooled over the whole domain, concentrating the signal on the binding pocket.
  • Interpretable predictions: Because the intermediate output is a fingerprint, the substructural features driving a call can be read off directly instead of inferred from attention maps.
  • Feeds a full genome-mining pipeline: Seq2Hybrid uses MASPR's A-domain calls together with AT-domain annotations to enumerate candidate structures for hybrid NRP–polyketide clusters.

#Technical Details

MASPR takes per-residue embeddings from the 650-million-parameter ESM-2 model at 1,280 dimensions, selects the ten Stachelhaus positions to give a 10 × 1,280 representation per A-domain, and passes it through two heads trained concurrently. The fingerprint predictor is a stack of linear layers with ELU activations and layer normalisation, supervised by a cosine-distance loss against the true substrate fingerprint; the classifier head is two linear layers over the predicted fingerprint that emit a distribution over 41 known substrates, supervised by cross-entropy. Substrate fingerprints concatenate 167 MACCS keys, a 128-bit Morgan fingerprint at radius 2, and the average partial charge. Training used 2,294 annotated A-domain–substrate pairs from MIBiG 3.0, with each domain–monomer pair from a promiscuous domain treated as a separate instance.

On an out-of-distribution split bucketed by Hamming distance from the training sequences, MASPR reaches 63.1% top-5 accuracy against 47.5% for AdenPredictor, with the margin widening on the most distant bucket, and reaches 72.2% versus 67.8% on promiscuous domains. In leave-one-substrate-out evaluation — the substrate is removed from training entirely and must be recovered from its fingerprint — more than half of held-out substrates exceed 50% top-5 accuracy, and 34% exceed 75%. Trained on bacterial domains and tested on fungal ones, it scores 15% higher top-5 accuracy than AdenPredictor.

#Applications

The direct users are natural-product discovery groups doing genome mining: given a sequenced microbial genome, MASPR annotates the A-domains of its NRPS clusters and Seq2Hybrid assembles those calls into candidate structures, enumerating gene assembly orders, applying pre- and post-assembly tailoring reactions combinatorially, and filtering against paired mass spectrometry when it is available. The authors used the pipeline to propose biosynthetic gene clusters for four orphan natural products — Octaminomycin A, Dityromycin, SW-163B, and JBIR-39 — whose producing clusters had not been assigned. The same specificity calls are useful upstream of NRPS engineering, where swapping a domain for one with a desired selectivity is the standard route to analogues.

#Impact

MASPR's transferable idea is that predicting the substrate's chemistry, rather than its label, converts a closed-vocabulary classification problem into an open one — a pattern that applies wherever a protein selects a small molecule from an incompletely enumerated set. Its evidence is computational: the proposed gene cluster assignments are hypotheses supported by mass spectrometry evidence, not by heterologous expression, and top-5 accuracy in the sixties leaves room for the ranking to be wrong. The repository ships the pretrained checkpoint, an inference server, the training table, and scripts reproducing the standard, zero-shot, and bacteria-to-fungi evaluations; it carries no licence file, which limits reuse in commercial settings.

At a glance

Released
January 2025
Category
Protein
Organizations
Carnegie Mellon University / University of Pittsburgh / Mohamed bin Zayed University of Artificial Intelligence

Links

GitHub RepositorybioRxiv Preprint

Tags

genome_miningsubstrate_specificity_predictiontransfer_learningzero_shot

Something wrong?

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