Lipid Trafficking in Diverse Bacteria

Conspectus

Lipids are essential for life and serve as cell envelope components, signaling molecules, and nutrients. For lipids to achieve their required functions, they need to be correctly localized. This requires the action of transporter proteins and an energy source. The current understanding of bacterial lipid transporters is limited to a few classes. Given the diversity of lipid species and the predicted existence of specific lipid transporters, many more transporters await discovery and characterization. These proteins could be prime targets for modulators that control bacterial cell proliferation and pathogenesis.

One overarching goal of our research is to understand the molecular mechanisms of bacterial metabolite trafficking, including lipids, and to leverage that understanding to identify or engineer inhibitory ligands. In recent years, our work has revealed two novel lipid transport systems in bacteria: bacterial sterol transporters (Bst) A, B, and C in Methylococcus capsulatus and the TatT proteins in Enhygromyxa salina and Treponema pallidum. Both systems are composed of transporters bioinformatically identified as being involved in the transport of other metabolites, but substrates were never revealed. However, the genetic colocalization of the genes encoding BstABC with sterol biosynthetic enzymes in M. capsulatus suggested that they might recognize sterols as substrates. Also, homologues of TatTs are present in diverse bacteria but are overrepresented in bacteria deficient in de novo lipid synthesis or residing in nutrient-poor environments; we reasoned that these proteins might facilitate the transport of lipids. Our efforts to reveal the substrate scope of two TatT proteins revealed their engagement with long-chain fatty acids.

Enabling the discovery of the BstABC system and the TatT proteins were bioinformatic analyses, quantitative measurements of protein–ligand equilibrium affinities, and high-resolution structural studies that provided remarkable insights into ligand binding cavities and the structural basis for ligand interaction. These approaches, in particular our bioinformatics and structural work, highlighted the diversity of protein sequence and structures amenable to lipid engagement. These observations allowed the hypothesis that lipid handling proteins, in general and especially so in the bacterial domain, can have diverse amino acid compositions and three-dimensional structures. As such, bioinformatics geared at identifying them in poorly characterized genomes is likely to miss many candidates that diverge from well-characterized family members.

This realization spurred efforts to understand the unifying features in all of the lipid handling proteins we have characterized to date. To do this, we inspected the ligand binding sites of the proteins: they were remarkably hydrophobic and sometimes displayed a dichotomy of hydrophobic and hydrophilic amino acids, akin to the ligands that they accommodate in those cavities. Because of this, we reasoned that the physicochemical features of ligand binding cavities could be accurate predictors of a protein’s propensity to bind lipids. This finding was leveraged to create structure-based lipid-interacting pocket predictor (SLiPP), a machine-learning algorithm capable of identifying ligand cavities with physico-chemical features consistent with those of known lipid binding sites. SLiPP is especially useful in poorly annotated genomes (such as with bacterial pathogens), where it could reveal candidate proteins to be targeted for the development of antimicrobials.

Key References

• Zhai L.; Bonds A. C.; Smith C. A.; Oo H.; Chou J. C.; Welander P. V.; Dassama L. M. K.. Novel sterol binding domains in bacteria. eLife 2024, 12, RP90696. 10.7554/eLife.90696 .¹The study discovered and characterized the first putative transporters for bacterial sterols in the sterol-producing bacterium Methylococcus capsulatus.

• Zhai L.; Chou J. C.; Oo H.; Dassama L. M. K.. Structures and Mechanisms of a Novel Bacterial Transport System for Fatty Acids. ChemBioChem 2023, 24( (15), ), e202300156. 10.1002/cbic.202300156 .²This work revealed that some TatT proteins in TRAP transporter systems are fatty acid binding proteins and potentially form a new class of TRAP fatty acid transporters.

• Chou J. C.; Decosto C. M.; Chatterjee P.; Dassama L. M. K.. Rapid proteome-wide prediction of lipid-interacting proteins through ligand-guided structural genomics. bioRxiv 2024, DOI: 10.1101/2024.01.26.577452.³The paper describes the analysis of lipid binding pockets and exploits the physico-chemical properties of pockets to predict lipid-binding proteins on proteome-wide scales.

Introduction

Lipids are biological molecules essential for building barriers and compartments of cells. They serve as signaling molecules to cue biological pathways and can be used as energy sources, etc; they are required in all domains of life. The bacterial domain has a diverse set of lipids to accomplish the above listed tasks.⁴ In addition to the ubiquitous phospholipids, bacteria often incorporate unique lipids such as branched-chain fatty acids, amino-acyl lipids, hopanoids (pentacyclic sterol-like lipids), lipopolysaccharides, and lipoteichoic acids (phosphoglycolipids) into their membranes. Moreover, nonmembrane lipids can serve as signaling molecules for virulence, biofilm formation, and quorum sensing.⁵ Despite the importance of lipids and the diverse roles they play in bacteria, much is still unknown about their synthesis, regulation, signaling, and functional roles. This is particularly acute for bacteria that make and use mammalian lipids.

Our group has a long-standing interest in uncovering the mechanisms of metabolite trafficking in bacteria. We are particularly drawn to bacteria that use unusual lipids like sterols. Overrepresented among these are pathogens that incorporate host sterols in their membranes, use sterols during colonization, or further modify them (e.g., into cholesterol glycolipids⁶).⁷⁻⁹ Also intriguing are commensal gut microbes that acquire and convert host cholesterol to metabolites that profoundly impact host health.¹⁰⁻¹³ While sterol synthesis, regulation, and trafficking mechanisms are conserved and well understood in eukaryotes, this is not the case for bacteria. Many bacteria that engage with or synthesize sterols lack homologues of the eukaryotic sterol handling machinery. With a goal of identifying how pathogens acquire and modify sterols, we reasoned that it might be best to first reveal sterol handling proteins in bacteria that produce sterols de novo.^14,15 This reasoning was partially motivated by the discovery of distinct biosynthetic proteins for sterols in the bacterium Methylococcus capsulatus.¹⁶

Generally, characterized lipid transporters¹⁷ in bacteria have substrate recognition domains that are diverse in their amino acid sequences and their structures, thereby making it challenging to rely on sequence or structural homology to identify lipid handling proteins in pathogens. One example of this divergence is the lipopolysaccharide transport (Lpt) system (Figure 1).¹⁸ The Lpt system has seven unique subunits, where LptA, LptC, and LptD all share one lipopolysaccharide (LPS) recognition domain. The domain has a unique β-taco fold that has not been observed in eukaryotes. Another well-characterized phospholipid transport system in bacteria is the maintenance of lipid asymmetry (Mla) system (Figure 1),¹⁹ which is composed of six subunits and spans the inner membrane, periplasm, and outer membrane. While MlaA is mostly α-helical, MlaC adopts an α/β-mix fold. Furthermore, the inner membrane transporter complex contains an architecture that is unseen in eukaryotic phospholipid transporters.

Figure 1.

Schematic description of the diverse structural folds on lipid transporters in Gram-negative bacteria. The lipopolysaccharide transport (Lpt) system, maintenance of the outer membrane lipid asymmetry (Mla) system, and mammalian cell entry (MCE) system are highlighted as representative systems with structurally distinct individual components and their corresponding PDB IDs.

We reasoned that lipid transporters could represent targets for modulators to control bacteria populations and their chemistries. Examples already exist for targeting lipid transporters with antimicrobial drugs. Scientists at Roche recently identified zosurabalpin as a narrow-spectrum antibiotic that inhibits the LptB₂FGC complex in carbapenem-resistant Acinetobacter baumannii.²⁰ Thanatin is an antimicrobial peptide discovered in 1996 whose mode of action was recently determined to involve destabilizing the Lpt assembly, leading to the accumulation of LPS in the inner membrane.²¹ The Seeliger group also identified an inhibitor for mycobacterial lipid transport protein LprG and demonstrated its antimicrobial activity against Mycobacterium smegmatis(22) and the pathogenic Mycobacterium tuberculosis.²³ These examples demonstrate the feasibility of antibiotics targeting proteins responsible for proper lipid localization. Identifying lipid transporters in bacteria would therefore expand the repertoire of drug targets for novel antibiotics, which is critical with the rise of drug-resistant pathogens.

Because of the aforementioned challenges, current discovery of lipid engagement proteins in bacteria relies heavily on sequence or structural homology. With the premise that sequence encodes function and that function is conserved among proteins with similar sequences, bioinformatic tools were developed to accelerate the pace of identifying functionally homologous proteins. The NCBI’s Basic Local Alignment Search Tool (BLAST)²⁴ and hidden Markov models²⁵ are the two most commonly used bioinformatics approaches to identify homologues using amino acid sequences. Likewise, tools that identify homologous structural folds have been employed, with DALI²⁶ and Foldseek²⁷ accelerating the discovery of structural homologues. With advances in machine learning and deep learning, newer models such as ProteInfer,²⁸ ProLanGO,²⁹ DeepGO,³⁰ DeepGOPlus,³¹ ContactPFP,³² and DeepFRI³³ have been developed. These models, which all rely on homology, have allowed more protein sequences to be functionally annotated. However, lipid binding domains are diverse in their sequences and structures and therefore are not always amenable to discovery via shared homology. Nonetheless, bacteria often cluster the genes with related functions in operons; such clustering allows the assignment of putative functions to proteins that share genomic proximity with characterized proteins. Two examples of how such clustering propelled new discoveries can be found with the BstABC¹ proteins that likely traffic sterols and with PetA,³⁴ a phosphatidylethanolamine (PE) transporter.

In this Account, we recount our discoveries that leveraged traditional sequence and structural homology methods to reveal novel sterol and fatty acid binding proteins. We also describe how our findings enabled the development of a new structural biology tool that can accelerate the pace at which lipid binding proteins are discovered in bacteria. This is particularly suitable for organisms with poor genome annotation or those that diverge substantially in the sequence of functional homologues.

Bacteria Engagement with Sterols

Sterols have long been known for their roles in eukaryotic cells. Cholesterol is a major component of the mammalian plasma membrane and contributes to membrane fluidity.³⁵ Sterols, steroids, and other modified forms are important ligands for multiple receptors and participate in signaling.³⁶ Steranes, which derive from sterols or steroids, are often used as biomarkers for eukaryotes in the geological record because it was presumed that sterol production was exclusive to eukaryotes.³⁷ For a long time, it was thought that bacteria do not make or use sterols.³⁵ However, bioinformatics analyses suggest that more organisms contain key sterol biosynthetic enzymes,^16,38 and it is now appreciated that bacteria can synthesize even complex sterols like cholesterol.¹⁴ In addition, other bacteria acquire sterols for use in their membranes.³⁹ A smaller subset is known to modify sterols, thereby creating metabolites that might have varied roles in microbial communities and in host organisms (Figure 2).¹⁰⁻¹³

Figure 2.

Examples of bacteria–sterol interactions. (A) Commensal bacteria capable of modifying sterol species within the gut. (B) Various pathogenic bacteria incorporate cholesterol into their membranes and use them as metabolites. (C) Bacteria encode the enzymatic machineries for de novo sterol synthesis.

An example of this comes from commensal bacteria in the human gut microbiome that metabolize sterols (Figure 2A). In 2020, Kenny et al. identified IsmA as the cholesterol reductase that converts cholesterol to coprostanol, a molecule that is less readily absorbed by gut epithelial cells and therefore is associated with reduced levels of circulating cholesterol.¹² Le et al.¹¹ and Yao et al.¹⁰ also found that Bacteroides thetaiotaomicron adds sulfates to cholesterol; the role of cholesterol sulfate in host health is still unknown. More recently, McCurry et al. discovered that Eggerthella lenta converts glucocorticoids into progestins and suggested a gut/neuron crosstalk.¹³ These modifications to cholesterol are all performed by putative soluble, cytoplasmic enzymes. How these enzymes access sterols is unknown; also unknown is the impact of sterol-derived metabolites in microbial communities.

Various pathogenic bacteria acquire cholesterol from host cells (Figure 2B). Chlamydia trachomatis hijacks vesicle-mediated lipid transport in host cells to obtain and incorporate cholesterol into its membrane.^7,40,41Chlamydia growth is not affected by the disruption of cholesterol synthesis but is affected upon inhibition of vesicle-mediated cholesterol transport.⁴¹Borrelia burgdorferi, the causative agent of Lyme disease, also directly takes up cholesterol from host cells and incorporates the lipid into its membrane, both as free cholesterol and as a glycolipid; cholesterol is essential in Borrelia.^8,42Mycobacterium tuberculosis acquires and catabolizes cholesterol as a carbon and energy source,⁴³ with the multiprotein complex Mce4⁴⁴ recently identified to mediate cholesterol acquisition. However, Mce transporters are unique to mycobacteria,⁴⁵ which suggests that cholesterol acquisition is disparate in bacteria that engage with them.

Although rare, some bacteria have the ability to synthesize sterols (Figure 2C).³⁸ Sterol production was first observed in the diderm, Gram-negative methanotrophic bacterium Methylococcus capsulatus (M. caps),⁴⁶ which encodes C-4 demethylases distinct from the eukaryotic enzymes.¹⁶ Moreover, unlike cholesterol, M. caps sterols remain singly or doubly methylated at C-4 (Figure 3A). The demethylated sterols are found to be localized to the outer membrane, suggesting the existence of a transport system to shuttle them through the aqueous periplasm. The genomic neighborhood of the two C-4 demethylases (Sterol demethylase A and B) contained three genes of unknown function that were conserved in all methanotrophs shown to produce C-4 demethylated sterols. A cursory bioinformatics analysis suggested that these genes were similar to metabolite transporters; we named these genes bacterial sterol transporters A, B, and C (bstA, bstB, and bstC)¹ and hypothesized that they were involved in shuttling demethylated sterols from the cytoplasm to the outer membrane (Figure 3B).

Figure 3.

BstABC is a novel three-component sterol transport system in M. caps. (A) Sterols synthesized by M. caps. (B) Sterol biosynthetic operon encodes BstABC and the SdmAB demethylases. Protein family IDs are provided below the corresponding genes. (C) Schematic showing the putative localization of BstABC. (D, E) X-ray crystallographic model of (D) BstB (7T1M) and (E) BstC (7T1S).

Bioinformatics suggested that BstABC was a tripartite transport system with components that span the two membranes (Figure 3C). BstA was predicted to be an integral inner membrane protein with a transmembrane domain composed of 11 helices and a soluble periplasmic domain. BstB resembled periplasmic substrate-binding proteins that are seen in many diderm bacteria. Finally, BstC was hypothesized to be a periplasm-facing outer membrane lipoprotein because it contains a lipoprotein signal peptide and shares homology with another periplasm-facing outer membrane lipoprotein.⁴⁷ Protein sequence similarity networks revealed that all three proteins occupied unique clusters in their respective families and had no characterized proteins within their clusters.¹

Attempts to obtain gene deletion mutants of bstABC failed and perhaps suggested that sterol localization is highly regulated in M. caps. To assess whether these proteins display any affinity for sterols, we produced them recombinantly. Incubation of pure proteins with the total lipid extract or a fraction enriched for polar lipids like sterols confirmed that all three proteins can recognize the sterols produced by M. caps.¹ Purification of the sterols allowed us to 1) quantify the strength of the interaction and 2) ascertain specificity for the C-4 monomethylated and dimethylated sterols. We then used microscale thermophoresis (MST) to measure the equilibrium affinity between the proteins and the sterols; MST is a biophysical method that measures the motion of biomolecules (which is inferred from diffusion properties) upon induction of a temperature gradient.⁴⁸ These experiments revealed that all proteins showed a preference for binding to the monomethylated sterols, with BstA and B displaying equilibrium dissociation constants (K_ds) of 4.4 and 2.4 μM for monomethyl sterols and greater than 200 μM for the dimethyl sterols. The difference in equilibrium affinity is slightly smaller for BstC (2.5 vs 80 μM).¹ Additionally, neither BstA nor BstB showed an interaction with any other sterols tested; BstC did show considerable engagement with lanosterol and cholesterol. Finally, both BstA and BstB displayed Hill coefficients consistent with positive cooperativity during monomethyl sterol binding, indicating the presence of multiple ligand binding sites or the oligomerization of the proteins.¹

Efforts to unravel the molecular basis of substrate recognition and selectivity were aided by modest- to high-resolution crystal structures of BstB and BstC (despite numerous efforts over many years, no diffraction-quality crystals of BstA appeared) (Figure 3D, E). These structures allowed substrate docking studies and molecular dynamics (MD) simulations, which together informed a working model for the molecular recognition mechanisms of the BstB and BstC. BstB (PDB 7T1M) adopts a bilobed structure typical of periplasmic substrate binding proteins. We identified two large hydrophobic cavities (one in each lobe) that could potentially accommodate sterols. Molecular docking placed sterols into the lobe A cavity, with the polar head placed within hydrogen bonding distance of Y120;¹ this residue is conserved in all close homologues of BstB. However, variants of the protein where Y120 was replaced with non-hydrogen-bonding residues did not show an appreciable impact on ligand affinity. We posited that the binding of sterols in this protein is largely modulated by the extensive hydrophobic interaction between the majority of the surface in the pocket and the tetracyclic core of sterol. Moreover, MD simulations revealed that the polar head of the substrate engages in H-bonding with various amino acids, and that the persistence of the H-bonds is different for the mono- and dimethyl substrates.¹ These data also showed conformational dynamics of the protein, with the presence of the ligand inducing a closed state where the N- and C-terminal lobes are closer together. This dynamicity might also allow a sampling of H-bonding partners when Y120 is missing.

For BstC (PDB 7T1S), initial docking studies produced a ligand-bound model where the sterol molecule occupied a central tunnel with a part of the alkyl tail remaining solvent-accessible. However, MD simulations demonstrated that the sterol moves ∼8 to 9 Å further into the tunnel to place its hydroxyl headgroup 2.88 Å away from the E86 side chain.¹ Replacing E86 with tryptophan and glycine disrupts ligand binding and therefore supports the hypothesis that hydrogen bonding between E86 and the hydroxyl headgroup is essential for sterol ligand binding.¹ While BstC can tolerate other sterol substrates, the MD simulations with those ligands showed that most do not move far enough into the central tunnel to form hydrogen bonds with protein side chains. This discrepancy could account for the lower affinity for the dimethyl sterol substrate.

Bioinformatics revealed that BstA belongs to the MmpL family, which is known to transport molecules for membrane assembly in mycobacteria (e.g., polyacyltetrahalose).⁴⁹ The closest structural homologue of BstA is HpnN,⁵⁰ which in Burkholderia multivorans is responsible for hopanoid transport; hopanoids are pentacyclic molecules that are structurally similar to sterols. However, the discovery of BstA revealed the first MmpL family member implicated in sterol transport and adds to the substrate scope of this large family of transporters.^49,51 Similarly, the closest characterized structural homologue of BstB is the phosphonate transporter PhnD⁵² from E. coli. Whereas BstC was assigned to the TRAP-T transport system, it was only the second characterized member of the family. The first TRAP-T to be studied was Tp0956⁵³ from Treponema pallidum; despite the structure of this protein being available, its substrate was never identified. The key questions this posed were 1) what is the full range of molecular substrates for these families and 2) how might the structural insights into the first bacterial sterol transporters be leveraged to identify functional homologues, especially in pathogens that hijack host sterols and commensal gut microbes that metabolize host cholesterol? The second question is bolstered by the comparison of BstB and BstC to other characterized sterol sensing domains in eukaryotes (Figure 4). Most eukaryotic sterol binding domains involved in nonvesicular transport have a unique helix-grip fold, where β-sheets wrap around a long α-helix to form a half-barrel structure with a hydrophobic core (Figure 4A). BstB instead has a bilobed structure while BstC is all α-helical (Figure 4B). In structure and sequence, these two proteins diverge substantially from each other and from all previously characterized sterol binding domains. This lack of conservation in folds of sterol binding proteins demonstrates the plasticity of sterol binding pockets and further suggests there might not be a motif that is generalizable to all sterol binding domains.

Figure 4.

Structural comparison of characterized sterol binding domains of (A) eukaryotes and (B) bacteria to highlight the distinct folds in bacterial sterol binding domains. The alpha helices are colored in gold, while beta sheets are colored in green. The PDB codes are labeled below each model.

The work with BstABC demonstrated that members of the same protein family could have vastly different substrate preferences, and the distinct structures of BstABC, despite recognizing the same substrate, highlighted the diversity that exists in sequence and structures of bacterial sterol binding proteins. These two points launched our most recent investigations that are described below.

Discovery of Novel Fatty Acid Binding Proteins

BstC belongs to the TatT superfamily, a family of proteins where the substrates and physiological roles of all members remained elusive since their discovery. TatT was discovered as the fourth component of some TRAP systems (Figure 5A).⁵³ TRAP systems are tripartite transporters composed of a periplasmic (P) component and inner membrane proteins (Q and M) that exist either as a heterodimer or a fusion.⁵⁴ TRAP proteins are involved in the uptake of small hydrophilic molecules such as succinate, malate, glucuronate, ethanolamine, 2-acetolactate, and more.⁵⁵ One well-known TRAP system is SiaPQM⁵⁶ from Haemophilus influenzae that is involved in sialic acid import and supports H. influenzae virulence and colonization. Approximately 2% of TRAP systems encode an additional periplasmic-facing lipoprotein called the T component, or TatT. Prior to our work, the only characterized TatT was Tp0956 from Treponema pallidum (Figure 5A). In 2012, Deka et al.⁵³ reported the crystal structure of Tp0956, confirmed that Tp0956 is within the same operon as other TRAP transporter genes (P, Q, M), and demonstrated through formaldehyde cross-linking that Tp0956 interacts with Tp0957, the P component of the TRAP transporter. While the substrate of Tp0956 was not identified, it was implied that Tp0956 and Tp0957 might accommodate a hydrophobic ligand because of the hydrophobic nature of their putative ligand binding sites.⁴⁷ Given our discovery that BstC is a sterol binding protein and the fact that BstC is a distant homologue of TatT, we hypothesized that Tp0956 and a homologue from Enhygromyxa salina (a cholesterol-producing marine myxobacterium),¹⁴Es_TatT, were lipid binding proteins (Figure 5A).

Figure 5.

TatT is a component of the bacterial fatty acid transport system. (A) Architecture of the TatT-associated TRAP system in T. pallidum. (B) X-ray crystallographic model of Es_TatT bound to palmitate (8G52).

Bioinformatics analysis of the TatT superfamily revealed that TatTs are mostly present in spirochaetes and marine microbes.² The spirochaetes encoding TatTs are in the Treponema genus and the Leptospira genus;⁵⁷ both contain organisms auxotrophic for long-chain fatty acid (LCFA). Also, marine microbes often undergo membrane remodeling under nutrient limited conditions,⁵⁸ and seawater is a low nutrient environment with limited carbon sources.⁵⁹ It is therefore possible that bacteria containing TatT homologues require an efficient lipid uptake system to proliferate. We also observed that TatT homologues do not always coexist with other TRAP components, as approximately 25% of TatT’s exist with genomic partners that share no homology with the PQM proteins.² One example is BstC, which we revealed as part of a putative export system for sterols in M. caps (Figure 3B).¹

To test the hypothesis that TatTs are lipid binding proteins, we aimed first to identify their substrates using a lipid pulldown assay. For Es_TatT, the incubation of pure protein with the total lipid extract from E. salina followed by lipid extraction and untargeted lipidomics revealed an enrichment for long-chain fatty acids. The affinity for two abundant LCFAs was confirmed in MST binding assays.² We also obtained a 1.88 Å resolution crystallographic structure of Es_TatT bound to palmitic acid;² the structure confirmed that TatTs retained the overall fold seen in BstC and that the ligand site identified in BstC was present and was occupied by palmitic acid (Figure 5B).

While BstC and Es_TatT are both annotated as members of the TatT superfamily, their substrates are different. BstC binds to a variety of sterols, but Tp0956 and Es_TatT prefer LCFAs as substrates. The difference in substrates is not explained by the lack of sequence or structure similarity (Figure 6), as pairwise sequence similarity of the three proteins show ∼25–35% similarity and the root-mean-square deviation (rmsd) of their aligned Cα atoms in the structures is ∼1.3 Å. In summary, this work revealed the elusive substrates of some TatT proteins and implied that some TRAP systems could also be lipid transporters. Additionally, the cocrystal structure of Es_TatT with palmitic acid informed the structural basis for LCFAs engagement.

Figure 6.

Similar folds in TatT homologues that exhibit different substrate preferences. The structures are shown in cartoon representation with ligands (when present or modeled via molecular docking) are shown as cyan ball and stick.

Systematic Analysis of Lipid Binding Pockets

Thus far, our work on BstBC and TatT proteins demonstrated that lipid binding proteins can adopt a variety of structural folds. Furthermore, while lipids (including sterols and phospholipids) are structurally diverse, they do retain some chemical resemblance. For example, all lipids are amphipathic because of their characteristically hydrophobic tails and hydrophilic headgroups. At ligand sites, the lipid tails are stabilized through hydrophobic interactions while the polar headgroups engage in hydrogen bonds and Coulombic interactions. Additionally, these ligand sites are sufficiently large to accommodate the full tails. Given this chemical resemblance, we hypothesized that there are common physico-chemical features shared by all lipid binding sites, which would allow the detection of such sites within protein three-dimensional structures.

To identify such features, we performed the analyses on ligand-bound structures in the PDB (Figure 7).³ We restricted our analyses to proteins that bound their ligands within an enclosed cavity rather than on the surface. Also excluded were proteins whose ligand sites are formed upon oligomerization or in complex with other proteins. We chose cholesterol and LCFAs to comprise the lipid data set, and we assembled another data set of structures with adenosine, vitamin B12, β-glucose, and coenzyme A. fpocket,⁶⁰ an open-source pocket identification algorithm, was used to identify and extract physical descriptors of the ligand binding pockets. The descriptors cover a wide range of physiochemical properties including volume, surface area, hydrophobicity, solvent accessibility, charge, and more.

Figure 7.

Schematic description that summarizes the development of SLiPP. Initial PCA analyses of ligand-bound pockets revealed that hydrophobicity is a critical factor in lipid binding pockets; this information was leveraged to develop a ML model that detects pockets across entire proteomes and classifies them as lipid binding pockets.

Principal component analysis of the pockets revealed a clear distinction between lipid binding pockets and nonlipid binding pockets, which is best explained by hydrophobicity-related properties.³ A plot of the distribution of each individual property showed a pronounced distinction between the hydrophobicity-related properties but no clear separation in pocket volumes and surface areas. As such, the extracted features allow a segregation of lipid and nonlipid pockets but do not provide the resolution necessary to distinguish binding to the different classes of lipids. This is a feature that might be improved with a larger data set of many structurally distinct lipids or by extracting more descriptors of the ligand sites.

Predicting Lipid Binding Proteins

Given the clear separation of lipid binding pockets and nonlipid binding pockets in the PCA plots, we pursued training a machine learning model to predict pockets within protein structures that are consistent with binding lipids.³ We first identified the random forest as the best-performing algorithm among six commonly used algorithms in chemical and biological data. The multiple decision trees in random forest ensures generalizability and prevents overfitting. We identified that an imbalance in the data set due to the inclusion of more pseudopockets leads to suboptimal performance. To solve this problem, we rebalanced the data set by sampling different amounts of nonlipid binding pockets to include in the training data set. To maximize the sensitivity while retaining decent precision, we included five times more nonlipid pockets than lipid binding pockets. With the above-mentioned optimization of the algorithm, data set, and hyperparameters, we generated a lipid binding pocket classifier named SLiPP for structure-based lipid-interacting pocket predictor (Figure 7).³ Because our vision was to assess poorly annotated proteomes for potential lipid binding proteins, we leveraged AlphaFold’s ability to provide structural proteomes⁶¹ and aligned this with fpocket to enable the rapid detection of lipid binding pockets in proteomes. The structural proteome is first processed to remove signal peptides from the models (as they can introduce false pockets depending on how they fold) and filtered to avoid predictions that use poor AlphaFold models. The pocketomes are extracted from the models and submitted to SLiPP, which outputs a score for each protein model and provides pocket information for the most likely lipid binding site.

Validation of SLiPP was provided with the structures of known lipid binding proteins not included in the training data set. The predictor was used on well annotated proteomes such has E. coli, Saccharomyces cerevisiae, and Homo sapiens. SLiPP successfully identified lipid transporters in E. coli (MlaC, AsmA, YhdP⁶²), yeast, and humans. In the latter two proteomes, hits included proteins with gene ontology terms linked to lipids. Use of SLiPP with pathogenic bacteria such as Borrelia burgdorferi, Treponema pallidum, and Chlamydia trachomatis, produced hits of known lipid-interacting proteins as well as uncharacterized proteins.

Prior to our work, efforts to identify lipid-interacting proteins existed but differed principally in their search mechanisms. MBPpred⁶³ was developed by Nastou et al. and aimed to identify membrane binding proteins instead of proteins interacting with lipid molecules in enclosed pockets. DisoLipPred⁶⁴ was developed by Katuwawala et al. to predict the lipid binding capability of the intrinsically disordered regions in a given protein. LIBP-Pred⁶⁵ is similar to SLiPP in that machine learning was employed to learn features in lipid binding proteins and make predictions. However, LIBP-Pred does not have the ability to predict the location of potential lipid binding pockets. Additionally, LIBP-Pred performs poorly on proteins with poor pLDDT regions in corresponding AlphaFold models. Recent advances with AlphaFold3⁶⁶ have enabled predictions of ligand binding to proteins. While AlphaFold3 could be a useful tool for identifying lipid binding proteins, the available lipid ligands are currently limited to fatty acids. Furthermore, using AlphaFold3 for lipid binding protein discovery could be relatively low-throughput given that only one protein–ligand pair is assessed at a time and because the current iteration requires several hours to generate a model. Moreover, the extensive neural network used by AlphaFold3 is best run with parallel computing; this requires graphics processing units (GPU) for the structure prediction. SLiPP is fast and requires far less specialized computing capabilities. As an example, lipid binding proteins in the human proteome (∼22,000 proteins) were predicted by SLiPP within 24 h using a MacBook equipped with an Intel i5 processor.

Conclusions and Outlook

This Account summarizes several years of work that identified bacterial lipid transporters BstABC and TatT and inspired the development of SLiPP, an efficient discovery tool for detecting novel lipid binding proteins in proteomes. The BstABC transporters were discovered using traditional bioinformatics that focused on their genomic colocalization with sterol biosynthetic enzymes. The discovery of BstC led to TatTs, which were revealed a decade ago but were not functionally characterized. The fact that BstABC and TatT are unique in folds compared to other lipid binding domains prompted a re-evaluation of the approach wherein traditional bioinformatics focused on sequence and/or structure homology is used to identify functionally related proteins. This culminated with the development of a machine learning model that explores protein three-dimensional structures for pockets containing physicochemical features consistent with lipid binding pockets. Thus, the use of structural bioinformatics with SLiPP offers another “omics” approach to functionally characterize proteins. The tool is especially powerful when used to make predictions in poorly annotated genomes and is useful for organisms that are either genetically intractable or unculturable in a laboratory. Our hope is that it will be leveraged to accelerate the pace at which lipid-interacting proteins, which comprise viable drug targets, are discovered and modulated. Moreover, the approach is one that can be leveraged to reveal novel protein folds that engage with other classes of molecules and metabolites.

Biographies

Jonathan Chiu-Chun Chou is a Ph.D. candidate in Prof. Laura M. K. Dassama’s group at Stanford University. Born and raised in Taiwan, he received his B.S. degree in chemistry from National Taiwan University. While his interests cover multiple aspects of antimicrobial resistance, his current research aims to identify and characterize microbial transporters as an approach to mitigate antimicrobial resistance.

Laura M. K. Dassama is a chemical and structural biologist who leverages chemical and physical insights to understand complex biological phenomena. Her interests lie in modulating intractable disease-relevant human proteins and discovering novel strategies for infectious disease curtailment.

Author Contributions

CRediT: Jonathan Chiu-Chun Chou conceptualization, data curation, formal analysis, writing - original draft, writing - review & editing; Laura M. K. Dassama conceptualization, data curation, formal analysis, funding acquisition, writing - original draft, writing - review & editing.

The authors declare no competing financial interest.