PINPOINT

Cysteine and serine proteases form an immune hub in the plant apoplast, and microbial pathogens counteract them with small secreted proteins (SSPs) that inhibit protease activity. Those inhibitors are traditionally found by sequence-similarity searches against annotated inhibitor domains — a filter that discards the SSPs that matter most, since a growing number of fungal SSPs inhibit host proteases despite carrying no annotated inhibitor domain. Sequence-unrelated but structurally similar (SUSS) effectors are the clearest example: they converge on inhibitor-like folds without any detectable sequence homology.

PINPOINT (Protease INhibitor PredictiOn at the plant-pathogen INTerface), from the Indian Institute of Technology Jodhpur, is a multi-level pipeline built to recover exactly those missed candidates. Level 1 screens sequences with two fine-tuned protein language model classifiers: PIP-BERT, derived from the ProtBERT model in ProtTrans, and PIPES-M, derived from the 150M-parameter ESM-2 checkpoint. Level 2 applies StructAE, a structure-aware one-class autoencoder that flags candidates whose structural embeddings fall outside the learned distribution of known inhibitors. Level 3 adds effector prediction with EffectorP 3.0 to retain proteins routed to the apoplast.

The pipeline consumes SSP sequences and monomeric structures, taking either precomputed models from the AlphaFold Protein Structure Database or predictions from the ESMFold public API, and runs as a set of Google Colab notebooks that need no local installation. Because nothing in the pipeline is organism-specific, it can be applied to the proteome of any organism without retraining. The work was posted as a bioRxiv preprint in July 2026 and has not yet been peer reviewed.

#Key Features

• Domain-free detection: Predicts protease-inhibitory function directly from sequence and structure, recovering SSPs — including SUSS effectors — that similarity searches against annotated inhibitor domains discard.
• Two complementary classifiers: PIP-BERT and PIPES-M draw on different pretraining corpora and show different residue-attribution patterns, so agreement between them acts as a confidence signal and disagreement widens candidate recall.
• Structure-aware filtering: StructAE scores AlphaFold or ESMFold models by reconstruction error against ~17,900 curated inhibitor structures, removing candidates without inhibitor-like structural features.
• Screening pre-filter: The pipeline is designed to cut the candidate list ahead of expensive multimer complex modelling, rather than replace it.
• Free-tier deployable: All steps except complex modelling run in Colab notebooks, handling roughly 4,000 sequences on a free GPU session.

#Technical Details

The classifiers were fit on 80% of a curated set of 18,101 sequences (MEROPS and UniProt inhibitors of 60–250 residues as positives, Pfam-screened UniProt non-inhibitors as negatives), evaluated by stratified 5-fold cross-validation and an independent holdout, with cost-sensitive learning for class imbalance. Descriptor-based baselines — tree ensembles, a meta-ensemble, and the TabPFN-2.5 tabular foundation model — topped out below an F1 of 0.90, with TabPFN-2.5 on CTD descriptors reaching accuracy 0.9381 and F1 0.8996. Fine-tuned language models closed that gap: PIP-BERT (420M parameters, BertForSequenceClassification) reached F1 0.9847, MCC 0.9776, and ROC-AUC 0.999, while PIPES-M (150M parameters, EsmForSequenceClassification) reached F1 0.9666, MCC 0.9514, and ROC-AUC 0.9931. Both held F1 above 0.94 across length bins and at 30–40% pairwise sequence identity, and exceeded 0.95 AUROC on an out-of-distribution set of longer inhibitors and non-inhibitor kinases, DNA-binding proteins, and endopeptidases.

StructAE is a PyOD deep autoencoder with Optuna-tuned bottleneck width, activation, learning rate, and dropout, trained as a one-class inlier model on structure-level embeddings computed by the RCSB embedding model, which aggregates residue embeddings from ESM-3. Inputs are truncated to the first 250 N-terminal residues, matching the training length range.

#Applications

Applied to the soil-borne fungal pathogen Macrophomina phaseolina, the pipeline narrowed 13,804 predicted proteins to 1,837 secreted proteins, 640 SSPs of ≤250 residues, and 437 lacking any Pfam annotation. Level 1 retained 218 candidates, Level 2 retained 208, and Level 3 identified 112 with apoplastic or dual localisation. AlphaFold-Multimer screening of those 112 against five soybean cysteine and six serine proteases yielded 43 that formed high-confidence complexes (confidence ≥ 0.70), most contacting the protease catalytic site through a characteristic loop, against 9 hits from a 43-SSP negative control set.

#Impact

PINPOINT addresses a specific and underserved gap in plant-microbe interaction research: the effectors that homology search cannot see. Its practical value is as a pre-filter that shrinks the candidate pool before multimer screening, which is the rate-limiting step in this kind of work. Weights for both classifiers and StructAE are on Hugging Face under CreativeML OpenRAIL-M, the AlphaFold-Multimer screening outputs are on Zenodo under CC BY 4.0, and the notebooks are on GitHub, though the repository carries no license file. Validation to date is computational rather than biochemical, and the authors note that StructAE may reject genuinely novel inhibitor folds absent from its MEROPS training distribution.