TL;DR
This paper introduces TopoPalmTree, a novel machine learning tool that leverages protein topology to predict S-palmitoylation sites on transmembrane proteins (TMPs). By overcoming the limitations of traditional bottom-up proteomics, TopoPalmTree successfully identifies previously unknown S-palmitoylation sites and even guides the rational design of such sites, offering a powerful new approach for studying this crucial post-translational modification.
What Problem Does This Paper Address?
S-palmitoylation is a reversible lipid modification that plays a critical role in regulating the function and localization of transmembrane proteins (TMPs), impacting processes from viral replication to inflammatory responses. While many small-scale studies have pinpointed S-palmitoylation sites, particularly in juxtamembrane regions, large-scale discovery efforts have been hampered. Traditional methods like bottom-up proteomics, which rely on trypsin digestion, struggle to effectively analyze the hydrophobic regions of TMPs, leading to an underrepresentation of these crucial S-palmitoylation sites. This creates a significant gap in our understanding of where and why TMPs are palmitoylated, limiting our ability to study their roles in health and disease.
Methods
The core of this work is the development of TopoPalmTree, a gradient-boosted tree (GBT) machine learning model. Unlike previous approaches that primarily relied on sequence information, TopoPalmTree incorporates topological features derived from UniProt annotations. These features capture the spatial arrangement of amino acids relative to transmembrane domains, including proximity to the membrane, distance from termini, and hydrophobicity gradients. The model was trained on a curated dataset of experimentally validated S-palmitoylation sites, prioritizing methods like site-directed mutagenesis and radiolabeling over trypsin-based proteomics. For validation, TopoPalmTree was tested on a dataset of viral TMPs and benchmarked against GPS-Palm, a deep learning-based predictor. The researchers also employed Acyl-RAC (Acyl-Resin Assisted Capture), a chemical proteomics technique, to experimentally confirm predicted S-palmitoylation sites in various proteins.
Datasets
- Training Dataset: A curated set of experimentally validated S-palmitoylation sites on transmembrane proteins. This dataset was specifically filtered to include sites identified through methods less prone to missing hydrophobic regions, such as site-directed mutagenesis and 3H-palmitate radiolabeling.
- Viral Holdout Dataset: A dataset of 30 viral S-palmitoylated proteins, chosen for their lack of sequence homology to the training data to rigorously assess the model's generalization capabilities.
- Murine TMP Proteome: Used for large-scale prediction of S-palmitoylation sites.
- UniProtKB: A primary source for topological and sequence information used in feature engineering.
- SwissPalm: A database of known S-palmitoylation sites, used for both training data curation and identifying novel sites.
Key Insights and Ideas
- Hydrophobic regions are S-palmitoylation "hot spots": The study strongly supports the hypothesis that juxtamembrane regions, which are hydrophobic and difficult to analyze with standard proteomics, are indeed preferred sites for S-palmitoylation.
- Topology is key: Incorporating topological features, rather than just sequence, significantly improves the prediction of S-palmitoylation sites on transmembrane proteins. This aligns with experimental observations that proximity to the membrane is crucial.
- Traditional proteomics misses important sites: The findings highlight how trypsin-based proteomics can miss a substantial fraction of bona fide S-palmitoylation sites due to the inherent difficulty in digesting and analyzing hydrophobic transmembrane regions.
- Machine learning can bridge experimental gaps: TopoPalmTree demonstrates the power of machine learning to overcome experimental limitations and uncover biological insights that are otherwise hidden.
- Predictive models can guide experimental design: The successful rational design of an S-palmitoylation site on KdelR2 showcases the potential of TopoPalmTree not just for discovery but also for engineering protein function.
Key Results
- TopoPalmTree identified 1591 novel putative S-palmitoyl sites on the mouse TMP proteome that were not previously reported in SwissPalm or UniProt.
- The model achieved high performance, with an AUC of 0.950 on 10-fold cross-validation and a final model MCC of 0.9859.
- Experimental validation confirmed S-palmitoylation at Cys69 of Vamp5 and Cys6 of Aquaporin-5, sites that were highly scored by TopoPalmTree but not readily identified by other methods or databases.
- TopoPalmTree accurately predicted experimentally confirmed S-palmitoylation sites in three Type I transmembrane proteins: Cadm4 (Cys347), Chodl (Cys238/239), and Havcr2 (Cys220).
- The tool successfully guided the rational design of an S-palmitoylation site by introducing a Cys residue at position 200 in KDEL Receptor 2 (KdelR2), which was experimentally confirmed.
Limitations and Future Work
- TMP Specificity: TopoPalmTree is currently limited to transmembrane proteins due to its reliance on topological features. It cannot predict S-palmitoylation on soluble or membrane-associated proteins.
- Data Availability: The model's performance is dependent on the availability of topological data in UniProt. A significant portion of TMPs have incomplete annotations, making them inaccessible to the current TopoPalmTree.
- Future Work: The authors suggest merging TopoPalmTree with chemoproteomic or other post-translational modification (PTM) datasets to explore the interplay of S-palmitoylation with drug binding or other PTMs. Continued topological annotation of TMPs is also crucial to expand the model's applicability. The authors emphasize that TopoPalmTree is a hypothesis-generating tool that complements experimental validation.
Summary updated