henhaobio
Sign in

JACS AuBioinformaticsPeer-reviewed journal article

PARAS: high-accuracy machine-learning of substrate specificities in nonribosomal peptide synthetases

Barbara R. TerlouwiD, Chuan HuangiD, David MeijeriD, José D D Cediel-BecerraiD, Ruolin HeiD, Marlene L. RotheiD, Matthew JenneriD, Shanshan ZhouiD, Yu ZhangiD, Christopher D. FageiD, Yuta TsunematsuiD, Gilles P. van WezeliD, Serina L. RobinsoniD, Fabrizio AlbertiiD, Lona M. AlkhalafiD, Marc G. ChevretteiD, Gregory L. ChallisiD, Marnix H. MedemaiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Nonribosomal peptides are diverse natural products with important applications in medicine and agriculture. Bacterial and fungal genomes contain thousands of nonribosomal peptide biosynthetic gene clusters (BGCs) of unknown function, providing a promising resource for peptide discovery. Core structural features of such peptides can be inferred by predicting the substrate(s) of adenylation (A) domains in nonribosomal peptide synthetases (NRPSs). However, existing approaches to A domain prediction rely on limited datasets and often struggle with domains selecting large substrates or from less-studied taxa. Here, we systematically curate and computationally analyse 3,653 A domains and present two high-accuracy specificity predictors, PARAS and PARASECT. A type of A domain with unusually high L-tryptophan specificity was identified through the application of PARAS, and intact protein mass spectrometry to the corresponding NRPS showed it to direct the production of tryptopeptin-related metabolites in Streptomyces species. Together, these technologies will accelerate the characterisation of novel NRPSs and their metabolic products. PARAS and PARASECT are available at https://paras.bioinformatics.nl. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=106 SRC="FIGDIR/small/631717v2_ufig1.gif" ALT="Figure 1"> View larger version (35K): org.highwire.dtl.DTLVardef@8e48aaorg.highwire.dtl.DTLVardef@144c78dorg.highwire.dtl.DTLVardef@890c11org.highwire.dtl.DTLVardef@1776de9_HPS_FORMAT_FIGEXP M_FIG C_FIG

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (28 Nov 2025) · CC BY 4.0

Random forest predictors PARAS and PARASECT, trained on 3,653 curated adenylation domains, outperform earlier NRPS substrate predictors and correctly assigned L-Trp to an unusual tryptopeptin synthetase domain.

Earlier adenylation domain predictors were trained on about 1000 labelled points or fewer, mispredicted large substrates and poorly studied taxa, and inherited promiscuity labels from ATP-pyrophosphate exchange assays. This work curates about three times more data and tests structure-based and substrate-fingerprint features.

Key results

  1. ShownOn held-out benchmark sets, PARAS and PARASECT beat all other predictors on fungal domains by a wide margin and matched the best deep learning tool on bacterial domains.

    Fungal (130 domains): PARAS ~70%, PARASECT ~66% vs AdenPredictor ~23% (Mann-Whitney U, p<1e-16). Bacterial (151 domains): PARAS ~77%, PARASECT ~75%, NRPSTransformer ~75% (not significant). · PARAS and PARASECT outperform existing…; Benchmarking

  2. ShownSequence-based PARAS gave its best accuracy on substrate-stratified splits and dropped on taxonomy-stratified splits, especially for fungi. Structural or ESM2 features did not improve on sequence features.

    PARAS: ~88% (~92% bacterial, ~74% fungal) substrate-stratified; ~83% (~89% bacterial, ~48% fungal) taxonomy-stratified. ESM2 features underperformed sequence features by 3-5%. · Structure-based feature extraction provides…

  3. SuggestedTraining set size, not model class, was the dominant driver of accuracy, and manual curation found annotation errors in earlier training sets.

    Accuracy vs training size: R2 0.95 (fungal), 0.80 (bacterial). Dataset of 3,653 domains vs 1089 (SP/AP) and 534 (NRPSPredictor2); 137 [12.6%] and 31 [5.8%] misannotations corrected. · An expanded and highly curated training set of A…; PARAS and PARASECT outperform existing…

  4. SuggestedDomains selecting large substrates such as Phe, Trp and Lys appear to have divergent active site pockets and fall in many clades, which may explain why earlier predictors erred on them.

    Qui18-A1 vs BreC-A3 (both Trp) share 10 of 34 extended signature residues; alignment bitscore 31.0 vs 184.0 for the most similar Trp pair. 41% of Thr domains fall in one clade, 75% in the largest five. · Large-scale analysis of structural models…

  5. ShownPARAS and PARASECT correctly predicted L-Trp for the third A domain of TtpD, a domain absent from training data. Intact-protein MS supported L-Trp as its preferred substrate.

    AdenPredictor predicted L-Asp and NRPSPredictor2 L-Leu. A mixture of L-Trp, L-Phe and L-Leu loaded only L-Trp. The BAP1-purified protein was Trp-loaded. Five tryptopeptin-related metabolites were made in S. coelicolor M1154. · Characterisation of an unusual…; A domain selectivity determination

Methods, data and reproducibility

Curated 3,653 A domains and featurised 34-residue active-site signatures (15 physicochemical properties each), with optional AlphaFold voxel-PCA or ESM2 features and ECFP-4 substrate fingerprints. Trained random forests as single-label PARAS or domain-substrate-pair PARASECT, then benchmarked and tested one prediction by intact-protein MS.

  • Curated A domain training set (MIBiG 3.0 and 4.0, SANDPUMA/AdenPredictor and NRPSPredictor2 sets, literature) with substrate labels · bacteria and fungi · n = 3,653 A domains; 278 substrates
  • Bacterial benchmark set of A domains not seen by any tool in training · bacteria · n = 151 domains · zenodo.org/records/17404295
  • Fungal benchmark set annotated by the authors · fungi · n = 130 domains · zenodo.org/records/17404295
  • AlphaFold models of A domains for active-site voxel PCA · n = 3,254 models
  • Putative tryptopeptin BGC in Streptomyces sparsogenes DSM 40356 draft genome · Streptomyces sparsogenes · NZ_MAXF01000131
  • Tryptopeptin BGC from newly sequenced Streptomyces sp. maeda85 genome · Streptomyces sp. maeda85 · PQ740899
Methods & tools
PARAS, PARASECT, scikit-learn RandomForestClassifier, AlphaFold2 / ColabFold, ESM2 (esm2_t33_650M_UR50D), PIKAChU (ECFP-4), HMMER pHMMs, MUSCLE, AutoDock Vina, AdenPredictor, SANDPUMA, NRPSPredictor2, NRPSTransformer, DeepAden, UHPLC-ESI-Q-TOF-MS
Software & environment
scikit-learn v1.2.0 (v1.1.3 for the guide-alignment test), imblearn v0.10.1, iterstrat v0.1.6, MAFFT v7.508, MUSCLE v3.8.1551, AutoDock Vina v1.2.3, web portal v2.0.0 (React v18.1.0, Flask v3.0.2).
Assay conditions & compute
Speed benchmarks of all tools except NRPSTransformer ran on a CPU server (3T memory, 256 AMD EPYC 9534 CPUs, 2x NVIDIA L40S 48GB). PARAS and PARASECT take a few milliseconds per prediction and under a minute to retrain.
Reproducibility gaps
Supplementary figures and tables are in a separate PDF not included in the text. Two GitHub owner spellings appear in the printed URLs. Training compute for the final models is not reported.
Limitations (5)
  • PARASECT often gives high scores to substrates structurally similar to the true substrate that the domain does not select, which slightly lowers its accuracy relative to PARAS. · stated by the authors
  • On taxonomy-stratified data, PARAS reached only ~48% on fungal domains, and PARASECT dropped for large substrates such as L-Arg and L-Trp, so sequences distant from the training data remain hard to predict. · stated by the authors
  • Promiscuity annotations are of limited quality because ATP-pyrophosphate exchange data can give false positives. Minor substrates were removed from the training set, so promiscuity predictions are less reliable. · stated by the authors
  • The experimental validation covers a single A domain (TtpD-A3) using qualitative intact-protein MS, with no replicate statistics reported. It supports the L-Trp call for that domain but not general performance on novel clades. · noted in this summary
  • The comparison tools differ in training sets and architecture, so data size and model class are confounded. The conclusion that data matter more than architecture rests on a small number of tools and two benchmark sets (151 and 130 domains). · noted in this summary

Read this if you mine NRPS biosynthetic gene clusters and want a curated adenylation domain dataset and predictors that handle large, fungal or phylogenetically unusual substrates.

Citations
38
FWCI
122.21
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHuang '26Li '26Präve '26Calcott '20Messenger '23He '23
75% semantic matchBioinformatics
Jiaquan Huang, Liangjun Ge, Y J Wu+7 more·Nucleic Acids Research·Jul 3, 2026· 7 citations