PLiCat

Lipids are not a single ligand class. A protein that binds cholesterol and a protein that binds a phosphoinositide are doing very different biochemistry, yet most sequence-based predictors treat lipid binding as one binary question: does this protein interact with lipids or not. That collapses the distinction that matters for interpreting membrane biology, trafficking, and lipid signaling.

PLiCat (Protein–Lipid interaction Categorization tool) addresses that gap by framing lipid binding as multi-label classification across the eight top-level categories of the LIPID MAPS taxonomy — fatty acyls, glycerolipids, glycerophospholipids, sphingolipids, sterol lipids, prenol lipids, saccharolipids, and polyketides — plus a non-binding class. It was developed by Feitong Dong and Jingrou Wu at Southern University of Science and Technology and published in Briefings in Bioinformatics in November 2025, following a September 2025 bioRxiv preprint.

Architecturally, PLiCat is a supervised head fine-tuned over third-party pretrained backbones rather than a new pretraining run. It pairs ESM Cambrian at the 300M-parameter scale with a BERT encoder, and the released checkpoint totals roughly 0.4B parameters. Beyond category assignment, the authors use residue-level attribution to nominate lipid-binding sites and to score how pathogenic variants perturb predicted binding.

#Key Features

• Eight-way lipid category prediction: Assigns a protein sequence to one or more LIPID MAPS categories, replacing the usual binary lipid-binding call with a specificity prediction.
• Hybrid PLM–BERT architecture: ESMC supplies evolutionary sequence representations while a BERT encoder learns the task-specific contextualization, with all transformer layers fine-tuned end to end.
• Residue-level attribution: Integrated Gradients produces attribution heatmaps that expose which residues drive each category call, giving an interpretable read on sequence-encoded binding signatures.
• Mutation impact scoring: Jensen–Shannon divergence and Wasserstein distance between wild-type and mutant prediction distributions quantify how clinical and cancer variants shift predicted lipid binding.
• Packaged release: Weights are on Hugging Face, the code and processed datasets are MIT-licensed on GitHub, a plicat_model pip package wraps inference, and a Colab notebook runs single sequences without local setup.

#Technical Details

The encoder is ESMC_300m — 30 transformer blocks with rotary position embeddings producing 960-dimensional embeddings — projected to 768 dimensions and passed into a BERT-base encoder, then a fully connected classification head with dropout 0.1 emitting a 9-dimensional logit vector (eight lipid categories plus a "none" class). Training data came from BioDolphin, a curated protein–lipid interaction database, yielding 12,873 positive and 800 negative sequences restricted to 35–500 residues, split into 12,296 training and 1,377 test sequences. Optimization used AdamW at a 2e-5 learning rate, batch size 16, and weighted binary cross-entropy, with 10-fold cross-validation before final selection on the held-out test set.

Per-category AUC-ROC on the test set spans roughly 0.88–0.97, with prenol lipids, fatty acyls, and polyketides at the top of the range. AUC-PR is more revealing of class imbalance: about 0.90–0.91 for fatty acyls, prenol lipids, and polyketides but 0.67 for glycerolipids and 0.73 for saccharolipids. PLiCat outperformed physicochemical-feature baselines (random forest, logistic regression, SVM) and protein language model baselines including ESM-2 and ProtBert. Binding-site recovery from attribution is weaker: F1 of 0.10 under exact residue matching and 0.23 with a ±2-residue tolerance, demonstrated on dehydrosqualene synthase (CrtM) and GID1.

#Applications

The immediate use is annotating uncharacterized proteins with a lipid-binding hypothesis specific enough to design an experiment around — predicting sterol versus sphingolipid preference tells a membrane biologist which pulldown or lipid-strip assay to run. Attribution maps narrow candidate binding residues for mutagenesis, and the variant module offers a mechanistic angle on ClinVar and cBioPortal variants in lipid-associated genes, where standard pathogenicity predictors say nothing about which molecular interaction is disrupted. Protein engineers can also use category predictions as a cheap filter when designing or redesigning lipid-binding modules.

#Impact

PLiCat is the first published tool to predict lipid category from protein sequence alone, and it establishes that pretrained protein language model representations carry enough signal to separate lipid classes without structures or MSAs. Its limitations are explicit and bounded: the training set is roughly 13,000 sequences, small relative to genome-scale resources, so rarer categories such as glycerolipids and saccharolipids remain the weakest predictions; inputs are capped at 500 residues; and the model was trained only on lipid-binding protein data, so it should not be read as a general function predictor. The binding-site F1 scores make clear that attribution is a hypothesis generator rather than a precise localizer. As a compact, MIT-licensed, fully released model with a working demo, it is nonetheless a usable starting point for a question the field previously had no sequence-based tool for.