Lipoproteins on Neurosnap: from circulating particles to bacterial lipid anchors

Figure 1. Two meanings of lipoprotein: a generic circulating lipid–protein particle and a bacterial protein attached to a membrane by a covalent lipid anchor. The cutaway shows a schematic three-tail E. coli-type anchor; bacterial lipid chemistry varies. This illustration is neither a coordinate reconstruction nor a comparison of relative scale, and the bacterial protein is not a literal Lpp structure.

A lipoprotein can be a particle carrying lipids through the bloodstream, or a protein tethered to a bacterial membrane by a covalently attached lipid. The shared name describes a partnership between lipid and protein, but the objects are different. In one case, the protein helps organize a changing assembly of many molecules. In the other, lipid modification is part of the maturation of a particular protein chain.

That distinction changes the questions we can ask of a sequence or structure. We examine protein-only structure prediction for apoA-I, sequence-model scores for apoE variants, and the cleavage boundary that preserves a bacterial lipid-attachment site in E. coli Lpp. These are demonstrations using recognizable reference proteins, which may overlap model training data, not held-out benchmarks.

ApoA-I: a protein shaped by its lipid environment

Apolipoprotein A-I is a major protein component of high-density lipoprotein (HDL). Its amphipathic helices have surfaces compatible with both water and lipid, making them useful at the interface of a lipid-containing particle and its aqueous surroundings. HDL is not a single fixed structure: its composition and particle size change, and apoA-I can adopt different arrangements in different lipid contexts. Experiments on particles of different sizes reveal changes in apoA-I structure, while work on lipid-free apoA-I also supports a flexible, constrained structural model (Chetty et al.; Melchior et al.).

The sequence submitted to a model already makes a biological choice. Human apoA-I P02647 is annotated as a 267-residue precursor, with a signal peptide at positions 1–18 and a proapoA-I region at 19–267. The mature chain begins at precursor residue 25 and contains 243 residues. Thus mature-chain residue numbers are 24 lower than precursor numbers. Removing the signal alone would still leave the propeptide.

We submitted the mature chain to Boltz-2 as one protein chain, without lipids, partners or a template. The run used an internal sequence alignment, six recycles, 200 sampling steps and three diffusion samples. Figure 2 shows the top-ranked sample and its associated confidence outputs. This is a prediction for the supplied protein-only input, not a reconstruction of HDL.

Figure 2. Top-ranked Boltz-2 prediction for mature apoA-I supplied as a single protein chain without lipids or partners, with per-residue confidence and predicted aligned error (PAE). The exported pLDDT values on a 0–1 scale are shown here multiplied by 100; PAE is in Å and describes uncertainty in relative placement.

The result has modest confidence: mean per-residue pLDDT is 57.9, and 209 of 243 residues score below 70. PAE adds information about uncertainty in relative placement across the model. Neither metric measures a physical ensemble, and low confidence alone does not prove that a region is intrinsically disordered. Most importantly for this example, a protein-only prediction cannot tell us which surfaces would be occupied or reshaped by a particular HDL particle.

The same distinction matters for surface analysis. Neurosnap's SASA Calculator and Aggrescan3D can examine a supplied protein structure, but their hosted analyses strip lipid heteroatoms. An exposed surface in that input is therefore not automatically an exposed surface in a native particle. SASA's hydrophobicity metric is a fraction of exposed residues by count, not a measurement of lipid affinity. We do not infer native apoA-I aggregation hotspots from this example.

A different particle makes the scale of the problem clear. ApoB100 wraps around the much larger low-density lipoprotein (LDL) particle, and structural work resolves it in the context of its lipid cargo (experimental LDL structure). A single isolated chain and a particle containing protein, many lipid species and a large lipid core are different structural questions. Even a useful protein model needs its input context attached to the interpretation.

ApoE: variants, numbering and model preference

Apolipoprotein E participates in lipid transport and receptor interactions. Its common protein forms differ at two sites, often discussed using mature-protein numbering. The canonical apoE3 reference sequence is 317 residues as a precursor, including an 18-residue signal peptide; the mature chain has 299 residues. Mature positions 112 and 158 correspond to precursor positions 130 and 176.

ApoE3 has cysteine at mature position 112 and arginine at 158. Relative to it, apoE4 carries C112R, while apoE2 carries R158C. These are allelic protein forms, not splice products. Keeping the reference and numbering convention explicit avoids turning the right substitution into the wrong coordinate.

We submitted mature apoE3 to ESMC Mutation Scoring using the 600M model. For each substitution, the reported log-likelihood ratio (LLR) is the model's log probability of the mutant residue minus that of the reference residue at the masked position. C112R scored +3.0000 and R158C −2.3906. Each score here concerns one substitution in the same reference sequence.

Figure 3. ApoE variant sites and sequence-model scores. The molecular panel is the experimental N-terminal domain in PDB 1LPE, mature residues 23–166, with positions 112 and 158 marked. The ESMC scores are separate computations on mature apoE3; the structure is not a mutant prediction, receptor complex or complete lipoprotein particle.

The positive score for the apoE4 substitution is particularly instructive. It says the model assigned the substituted residue a higher conditional likelihood than the reference residue at that position. It does not mean that apoE4 is healthier, or measure receptor affinity, cholesterol efflux, clinical risk or biological benefit. Those outcomes depend on biological context and require evidence beyond a sequence-model score. Likewise, the negative score for R158C is not a quantitative prediction of any of those properties.

The experimental domain structure provides spatial context for the sites, not a readout of the model's reasoning. ApoE's domain organization and its interactions are subjects of structural study (1LPE; Wilson et al.). A marked residue in that domain helps locate a change; it does not show how an intact particle or receptor complex responds to it.

Lpp: a cleavage boundary with a lipid attachment site

Bacterial lipoproteins use the word differently. E. coli Lpp is synthesized as a precursor with an N-terminal signal sequence. Its mature protein begins with a cysteine that becomes lipid modified and anchors the protein to the membrane. The original biochemical work on Lpp established lipid attachment to the protein (Braun and colleagues).

In the E. coli pathway, Lgt transfers a diacylglyceryl group to the conserved cysteine, LspA (signal peptidase II) cleaves the signal peptide, and Lnt adds an N-acyl group to the resulting N terminus. Structural and biochemical work on Lnt illuminates this maturation step (Lnt study). This description is specific to the E. coli system; bacterial lipoprotein chemistry and enzyme requirements are not identical everywhere.

The E. coli K-12 Lpp reference has 78 residues. Its annotated signal occupies positions 1–20, so the 58-residue mature chain begins at Cys21, with CSSNAK.... That first cysteine is more than a convenient marker: it is the residue on which the lipid anchor is built. A boundary that removes it changes the biological object represented by the exported sequence.

We submitted the full precursor and, as a missing-signal control, the mature-chain sequence alone to three services. TSignal classified the precursor as Sec/SPase II and resolved cleavage after residue 20. Its exported tail matches the annotated mature chain and retains Cys21. The mature-chain-only control was classified No Signal Peptide, with no cleavage call. TSignal's first-label and cleavage confidences were 1.000 to the precision reported, but these are conditional decoder confidences, not calibrated experimental probabilities (TSignal paper).

USPNet-fast, with Gram-negative bacteria specified, also classified the precursor as Sec/SPase II, placed cleavage after residue 20 and exported the same complete mature tail. It classified the mature-chain control as No Signal Peptide, without a cleavage call. Its global class score of 0.416 is cosine-normalized, not a probability; the cleavage-label softmax confidence was 0.174. Agreement with the reference here should therefore not be described as a high-confidence prediction (USPNet paper). Both services analyze the first 70 residues but export the complete post-cleavage tail.

Figure 4. Annotated Lpp processing boundary compared with the precursor predictions from TSignal, USPNet-fast and DeepSig. Cys21 is highlighted: the reference, TSignal and USPNet-fast boundaries retain it, while DeepSig's predicted cut after residue 25 does not. The diagram shows sequence boundaries, not predicted lipid coordinates.

DeepSig classified the full precursor as Signal Peptide, with a raw signal probability of 0.884, but placed cleavage after residue 25. Its export begins KIDQL... at Lys26, omitting Cys21 and residues 22–25. It is therefore not the correctly processed mature Lpp sequence. The mature-chain control was called Other. DeepSig is a general signal-peptide classifier rather than a lipoprotein-specific SPase-II classifier; this one example does not establish why the boundary differs or rank the tools' general performance (DeepSig paper).

The control also shows why the submitted processing state matters. Once the signal sequence has been removed, the predictor receives a different N terminus and no longer sees the precursor signal. Even an export with the expected boundary is an amino-acid sequence, however: FASTA does not encode the covalent lipid anchor. Matching the annotation does not establish that lipidation or secretion occurred, or determine membrane localization, immune activation or virulence.

A circulating particle and a bacterial lipid-anchored protein invite different kinds of evidence. ApoA-I's conformation depends on the lipid assembly around it; apoE scores describe a model's sequence preferences; and Lpp's cleavage boundary determines whether the sequence retains the cysteine needed for its anchor. In each case, the molecular object, processing state and lipid context give the computational output its meaning.

Sources

Reference sequences: apoA-I P02647, apoE P02649, and E. coli Lpp P69776. See also the Boltz-2 paper.