Aggregation in gut: on the link between neurodegeneration and bacterial functional amyloids

Abstract

Amyloids are insoluble protein aggregates with a cross-beta structure, which are traditionally associated with neurodegeneration. Similar structures, named functional amyloids, expressed mostly by microorganisms, play important physiological roles, e.g., bacterial biofilm stabilization. Using a bioinformatics approach, we identify gut microbiome functional amyloids and analyze their potential impact on human health via the gut-brain axis. The results point to taxonomically diverse sources of functional amyloids and their frequent presence in the extracellular space. The retrieved interactions between gut microbiome functional amyloids and human proteins indicate their potential to trigger inflammation, affect transport and signaling processes; pathways typically affected by host-microbiome interactions. We also find a greater relative abundance of bacterial functional amyloids in patients diagnosed with Parkinson’s disease in two out of three analyzed datasets. Our results generate hypotheses on a tentative link between neurodegeneration and gut bacterial functional amyloids, which require further experimental validation.

Introduction

The influence of the human gut microbiome on health, including the occurrence of neurodegenerative disorders, has been well established by now^1–3. The microbiome is responsible for the supply of multiple substances, such as folate, riboflavin (vitamin B2), vitamin B12, and short-chain fatty acids (e.g., butyrate) that have anti-inflammatory properties and serve as an energy source for a host^4,5. Furthermore, it shapes immunological responses by modulating cytokine production and lymphocyte activation^6,7.

Recent studies showed that many bacterial strains are capable of producing proteins that are known as bacterial functional amyloids⁸. Their structures are strikingly similar to human misfolded proteins forming plaques in the brains of patients suffering from neurodegenerative diseases. However, unlike human pathological amyloids, bacterial functional amyloids support bacteria by performing vital functions, including biofilm formation, adhesion, or signaling⁹.

It has been proposed that bacterial functional amyloids could indirectly affect the onset and progression of neurodegenerative diseases, particularly Parkinson’s disease (PD). The interactions between bacterial functional amyloids and the human protein alpha-synuclein, whose aggregation contributes to the disease onset, increased the disease incidence¹⁰. Due to the structural similarity of human misfolded proteins and bacterial functional amyloids, they are capable of affecting the aggregation rates of their pathological human counterparts^11,12. Bacterial functional amyloids may trigger misfolding of human amyloids and their aggregation¹³. In this case, pathological aggregation could commence in the enteric nervous system and then propagate through the vagus nerve to the brain¹⁴. Several studies support this hypothesis. Alpha-synuclein aggregates appear early in the PD course in the enteric nervous system, and they are correlated with disease severity, suggesting that disease may start in the gut¹⁵. Moreover, application of inhibitors of the curli CsgA protein (gut bacterial functional amyloid, expressed e.g., by Escherichia coli) reduces aggregation of alpha-synuclein in the mouse brain model¹⁶. Biofilm-related bacterial functional amyloids are more abundant in PD patients than in healthy controls and can co-localize with alpha-synuclein in neurons and increase its aggregation¹⁷.

The potential role of bacterial functional amyloids in neurodegeneration is, however, not limited to triggering aggregation. Neurodegenerative diseases are characterized by gut dysbiosis, dysfunction and inflammation^18–20. Bacterial functional amyloids, due to their structural similarity to human pathological amyloids, could have similar cytotoxic activity in humans. Moreover, their presence, especially during dysbiosis, could be another pro-inflammatory factor leading to abnormal gut permeability²¹.

Although many studies have shed light on the potential effects of bacterial functional amyloids on neurodegeneration, multiple questions remain unaddressed. Are bacterial functional amyloids present on a large scale in the human gut, and if so, which bacteria produce them? Are they more abundant in the microbiomes of patients with neurodegenerative diseases than in those of controls? And, finally, which molecular pathways could be affected by gut bacterial functional amyloids? Answers to these questions could broaden our understanding of the onset of neurodegeneration and related molecular mechanisms governing the role of the microbiome, potentially enabling the development of novel therapeutic strategies or biomarkers for early diagnosis.

We develop an atlas of predicted bacterial functional amyloids expressed by the human microbiome through a detailed computational analysis of available datasets. Using the atlas, we investigate the abundance of bacterial functional amyloids in the microbiome metagenomic samples of patients with PD (the most extensively studied condition in the context of the potential influence of microbiome), Alzheimer’s disease (AD), and fungal brain infection, Cryptococcal meningitis (CM, negative control of the brain disease). Finally, we predict human interactors of gut bacterial functional amyloids to generate hypotheses on their potential effects on human molecular pathways.

Results

A vast array of bacterial functional amyloids may be present in the human gut microbiome

As testing the aggregation propensity of all bacterial proteins in the human gut microbiome proteome is infeasible, an indirect approach that narrows the pool of potential amyloids is needed. Aggregation propensity of bacterial functional amyloids should be evolutionarily conserved^22–24. Therefore, we performed a homology search of known bacterial functional amyloids (BFA dataset) in the human gut microbiome proteome (Unified Human Gastrontiestinal Proteome, denoted as UHGP dataset)^25,26. The homology search was intentionally performed with quite permissive parameters, in order to obtain as large and diverse a pool of reasonable initial candidates as possible. The permissive search, though, comes at the price of identifying very distant non-aggregating homologs, which were designated for a potential subsequent removal with aggregation propensity predictors, in the analyses of amyloidogenicity (see Fig. 1). Details on our amyloid identification pipeline and the benchmark of the amyloid predictors are given in the Methods and Supplementary Fig. S1.

Fig. 1. Bacterial functional amyloids in the UHGP dataset.

A Bioinformatics pipeline for the identification of potential bacterial functional amyloids in the human gut microbiome proteome. B Number of homologs found per BFA protein.

The homology search yielded 10,249 sequence matches and 9541 unique sequences in total. Longer proteins, such as biofilm-related or DNA/RNA-related, provided us with multiple homologous sequences due to different domain matching (WapA: 730, HelD: 714, Aap: 714, YghJ: 714, SuhB: 714, Bap: 713, Smu_63c: 711, PAc: 702, Tuf: 714). For three proteins (alpha phenol-soluble modulins and RdlB) from the BFA dataset, we identified no homologs in UHGP. The mean sequence identity was 45% with a standard deviation of 22%. The mean E-value was 5 × 10⁻⁶ (see also Supplementary Fig. S2).

The pool of initial candidates, identified with the homology search, was tested for their aggregation propensity using the full-length sequence predictor AMYPred-FRL²⁷, which discarded 90% of the sequences, leaving 855 proteins. In general, the percentage of predicted amyloids varied for different proteins. We observed no correlation between AMYPred-FRL score and any of the homology search parameters (Pearson’s |R| < 0.2). We identified beta-arch motifs with ArchCandy²⁸, and 50 (6%) sequences were discarded due to a lack of probable beta-arches. The resulting final dataset of predicted amyloids was labeled as UHGPAmyloids. The rate of false positives in UHGPAmyloids should be low, as all proteins are at least distant homologs of known bacterial functional amyloids (BFA), they passed a full-length sequence amyloid predictor (AMYPred-FRL) and a beta-arch detector ArchCandy. Nevertheless, the false negative rate is likely high and hence the size of UHGPAmyloids is underestimated, as only 14 out of 38 proteins in the BFA passed both AMYPred-FRL and ArchCandy filters (the full list is provided in the Supplementary Table S1) and non-homologs of the BFA proteins are not considered in our work.

Two additional datasets were defined that served as negative controls in our analyses of UHGPAmyloids. The first one is a dataset of shuffled sequences of UHGPAmyloids, denoted as Shuffled UHGPAmyloids. The second one is a random sample of 1000 sequences from the UHGP dataset, denoted as Random.

Bacterial functional amyloids have diverse taxonomic origins

The taxonomy of the UHGPAmyloids was analyzed to identify bacterial groups with potentially the highest capacity for producing bacterial functional amyloids. The taxonomy of the UHGPAmyloids can be naturally biased by the taxonomy of the source proteins in the BFA dataset, but the permissive homology search could have minimized this effect. Although 80% of proteins in UHGPAmyloids share their phylum with a BFA source protein (see Supplementary Figs. S3 and S4), for 20% we managed to reach beyond that. In consequence, the overall taxonomic diversity of UHGPAmyloids is statistically significantly different from that of BFA proteins (chi-square test p-value = 0.0001).

The UHGP proteins are mostly produced by Firmicutes (Bacillota), Bacteroidota, Proteobacteria (Pseudomonadota), and Actinobacteria (Actinomycetota) (Supplementary Fig. S5). Predicted bacterial functional amyloids in UHGPAmyloids are not evenly distributed between these groups (Fig. 2, Supplementary Fig. S4) and demonstrate enrichment in Firmicutes, Proteobacteria, Fusobacteria, and depletion in Actinobacteriota and Bacteroidota in comparison to the UHGP proteins (Supplementary Table S2). No statistically significant difference was identified for Campylobacterota and Myxococcota.

Fig. 2. Taxonomic distribution of UHGPAmyloids.

A Visualization of the taxonomic origin of homologs of BFA’s proteins on the phylum and family level. The taxonomy is consistent with the taxonomy provided by the authors of UHGP. Note that some nomenclature may therefore be outdated (Firmicutes-Bacillota, Proteobacteria-Pseudomonadota, Fusobacteriota-Fusobacteria, Actinobacteriota-Actinomycetota). B Shannon entropy for different groups of homologs.

Taxonomic distribution of identified homologs in each of their BFA source protein groups was examined (Fig. 2). Proteins in UHGPAmyloids were widely spread across the bacterial tree of life, while others turned out to be specific to certain taxa. For example, homologs of chaplins and Mtp were found exclusively in Streptomycetes. These microorganisms are capable of producing numerous antibiotics, antiproliferatives, and immunosuppressants. They are hypothesized to have been an important part of the human microbiota in the past, but nowadays are present in low abundance there and can, more easily, be found in water and soil²⁹. On the other side of the spectrum, the highest taxonomic diversity was found for homologs of the RNA-binding protein Hfq and some biofilm-related proteins, which span across different phyla. However, even in the group of biofilm-related proteins, some variation could be observed. Long biofilm-related proteins WapA, Bap, PAc, Aap, and Esp were identified predominantly in Firmicutes with much smaller representation in Proteobacteria, while short biofilm-related proteins, like CsgA and CsgB, were found almost exclusively in Proteobacteria. Interestingly, the changes in the ratios between these two phyla (Proteobacteria-Firmictures) were observed in many human diseases³⁰.

Gut microbiome of Parkinson’s disease patients is enriched with bacterial functional amyloids in two out of three cohorts

We examined the abundance of identified bacterial functional amyloids (determined in the previous section—UHGPAmyloids) in the microbiomes of patients with neurodegenerative (Parkinson’s disease (PD) and Alzheimer’s (AD)) and microbiome-associated inflammatory disease (Cryptococcal meningitis (CM)). We used three Whole Genome Sequencing (WGS) metagenomic studies of PD patients and their respective control groups (two datasets collected in different parts of the USA by Wallen et al. and Boktor et al., and one in Germany by Bedarf et al.)^31–33. To supplement the analysis, we also included one metagenomic study of Alzheimer’s disease (AD) patients and patients suffering from Cryptococcal meningitis (CM), which is a fungal infectious brain disease involving inflammation but not amyloid deposition^34,35. The original publications describing these datasets analyzed taxonomic differences in the gut microbiome between patients and controls. Here, we approached this data from a very different protein-related perspective and focused on the abundance of bacterial functional amyloids and their potential link to the disease onset. For all datasets, we extracted the original reads, performed quality control, assembled them, and searched for homologs of UHGPAmyloids in the samples (see “Methods” for details).

In two out of three studies involving PD (Bedarf and Wallen but not Boktor), we found significantly more bacterial functional amyloids in samples from PD patients compared to the controls (Mann–Whitney U-test p-value < 0.05) (Fig. 3A). In neither the AD nor the CM datasets was there a significant difference in the general abundance of amyloid proteins between groups.

Fig. 3. Abundance of bacterial functional amyloids in metagenomic samples.

A Mean total normalized amyloid abundances per run across five cohorts. Bars show the mean of summed abundances per sample for control vs. disease groups (PD, AD, CM). Sample sizes (N) are shown below each bar. p-Values from Mann–Whitney U tests (two-sided) are shown in panel titles. B Forest plots of log fold-change in amyloid abundances (sex-adjusted where possible) across cohorts. Each panel shows one homolog. Individual cohort estimates (points with 95% CI) and the pooled PD estimate (diamond) are shown at the bottom for PD cohorts only. Red color indicates statistically significant results (p < 0.05). Results are adjusted for sex, where available (Wallen, Boktor, Laske) and unadjusted for cohorts without sex data (Bedarf, Li). The analysis uses the highest-identity approach (each target mapped to the homolog with the highest identity per person/sample). Vertical dashed line indicates no effect (log fold-change = 0).

For each protein of the BFA dataset and its homologs from UGPAmyloids, we analyzed their individual abundances in each sample in order to identify protein families contributing most to the observed amyloid enrichment in PD samples (Fig. 3B). While there are significant differences in the abundance of certain proteins, none of them are consistent across different studies. For example, CarD homologs are uniquely enriched in Wallen and Hfq in Bedarf PD cohorts, while PAc homologs were statistically more abundant in the AD dataset by Laske et al. Even well-studied proteins like CsgA show no statistically significant differences. This protein holds a special place in this discussion, as it is known to interact with PD-related alpha-synuclein and AD-related Abeta peptide^36,37. It has also been found to promote amyloid-related neurodegeneration in C. elegans models of AD and PD, Huntington’s disease, and amyotrophic lateral sclerosis³⁸. This may suggest that, while there is a general difference in the overall abundance of amyloids, the presence of specific proteins can strongly vary between different populations. Such results seem natural in metagenomic studies as they may reflect heterogeneity in the studied populations and differences in methodological design, e.g., the use of different sequencing strategies. Therefore, analyses focusing on only one group of proteins, e.g., CsgA, might be inherently biased.

Identified intestine bacterial functional amyloids are often extracellular and predicted to interact with human intestinal proteins

The observed increased abundance of UHGPAmyloids in the PD microbiome raises a question about their potential routes of interaction with human cells. The putative pro-inflammatory or toxic effect of bacterial functional amyloids on the human body is closely related to their cellular localization. Bacterial proteins located inside cells are much less likely to come in contact with human proteins and be involved in undesired protein-protein interactions.

Prokaryotic proteins tend to localize mostly in the cytoplasm (up to 64%), membrane (up to 20%), and only <2% are secreted³⁹. The cellular localization predictions of the source proteins from the BFA dataset were different, and 60% of the proteins were predicted as Extracellular and 40% as Cytoplasm (Supplementary Fig. S6A). The UHGPAmyloids dataset followed this trend, as expected, with 43.1% of proteins predicted as Extracellular 43.1%, 48.2% as Cytoplasm, and 8.7% as Plasma membrane. While the difference from the BFA dataset was not statistically significant (chi-square test’s p-value = 0.07), UHGPAmyloids exhibited a statistically significantly different distribution of the cellular localization from control datasets (chi-square test p-value = 1e−4 for both Shuffled UHGPAmyloids and Random). Specifically, UHGPAmyloids were characterized by a high ratio of Extracellular and Plasma membrane proteins and a low of Cytoplasm. The predicted Extracellular proteins in UHGPAmyloids are homologs of 25 different amyloids, although the dominant group includes biofilm-related proteins, such as WapA and Bap, 187 and 129 homologs, respectively. To generate a dataset of surface-accessible bacterial functional amyloids, we filtered the UHGPAmyloids for Extracellular and Plasma Membrane localizations. This resulted in a dataset of 417 proteins denoted as UHGPAmyloids_SurfaceProteins (Supplementary Fig. S6B). Surface-protein subsets were also generated for the Random and Shuffled UHGPAmyloids control datasets.

Protein-protein interactions are central to understanding the molecular mechanisms of stimulus response. The predicted cellular localization provides a basis for analyzing which human proteins might interact with bacterial functional amyloids. We extracted proteins that are experimentally confirmed to be expressed in the human intestine and are biologically related to the extracellular membrane or the junction space (HPAIntestine_SurfaceProteins, 2361 proteins) from The Human Protein Atlas⁴⁰, (Supplementary Fig. S6B). All possible interactions between proteins from HPAIntestine_SurfaceProteins and UHGPAmyloids_SurfaceProteins were predicted (Supplementary Fig. S7A). The same protein-protein interaction predictions with HPAIntestine_SurfaceProteins were performed for surface-protein subsets of control datasets.

Out of 984,537 possible pairings between HPAIntestine_SurfaceProteins and UHGPAmyloids_SurfaceProteins, we predicted 183,742 interactions involving 1098 human intestinal proteins. The estimated false positive rate for these predictions is 0.12⁴¹. As is characteristic of protein-protein interaction networks, the distribution of bacterial interactors per human protein was skewed (Supplementary Fig. S8). In the following analyses, only those human proteins that had at least five interacting partners in UHGPAmyloid_SurfaceProteins or control datasets were considered. We observed that although even up to 88% human interacting partners of UHGPAmyloid_SurfaceProteins were shared with the control datasets, the frequency of the interactions was significantly higher for UHGPAmyloid_SurfaceProteins than for the controls (see Supplementary Table S3, Supplementary Fig. S7B, C).

The highest number of human interactors was found for homologs of WapA (114,058 in total) and the lowest for homologs of TapA (135 in total). The identified human protein potentially interacting with bacterial functional amyloids most abundantly was N-myc-interactor (Nmi protein, UniprotID: Q13287). The analysis showed 681 interactions with homologs from 20 different protein groups, mostly with biofilm-related homologs of WapA (346), SasG (92), and CsgA (70). All of these homologs originate from Firmicutes (Bacillota) or Proteobacteria (Pseudomonadota), the latter often seen as a disease marker³⁰. Nmi protein is a part of the human immune system and interacts with interleukin-2 and STAT proteins⁴². Its level can be elevated in multiple cancers⁴³. Among proteins representing the 5 highest numbers of interactions, we also found other immune response-related proteins (C-C motif chemokine 5, Interferon-induced transmembrane protein 1 and 2), claudins - controlling the epithelial barrier and ubiquitin ligases (Supplementary Table S4)⁴⁴. Interestingly, in this group, we also identified proteins that become human pathological amyloids, mainly alpha-synuclein (417 different interactions), major prion protein (417 interactions), and APP cutting protein—presenilin-1 (417 interactions). Other human amyloids interacting with UHGPAmyloids_SurfaceProteins included APP (294 interactions) and tau (229 interactions).

Bacterial functional amyloids in the human intestine may affect signaling and transport

Protein overrepresentation analysis aims to provide an overview of molecular pathways affected by interactions between human intestinal proteins and bacterial functional amyloids. We, therefore, extracted proteins from HPAIntestine_SurfaceProteins that are predicted to potentially interact with at least five putative amyloids from UHGPAmyloids_SurfaceProteins. Then, an overrepresentation analysis (ORA) of GO and KEGG terms was performed⁴⁵ to draw the hypothesis on molecular mechanisms that can be affected by bacterial functional amyloids in the human gut microbiome.

The overrepresentation analysis revealed that bacterial functional amyloids are not neutral to proteins of the human gastrointestinal tract, similarly to other bacterial proteins (Fig. 4, Supplementary Table S5). We observed enrichment in GO Biological Process terms related to vesicle transport, protein localization, signaling, cell–cell, and cell–matrix adhesion. The predictions suggested that identified amyloids might interact with proteins forming cell-cell junctions, like claudins, which are the main components of tight junctions (TJ). This group of proteins is essential for maintaining the integrity of epithelium and is associated with numerous gastrointestinal diseases, including inflammatory bowel disease (IBD), ulcerative colitis (UC), Crohn’s disease (CD), or colorectal cancer (CRC)^46,47. Other fundamental human processes, such as signaling and transport, can also be hypothetically affected by bacterial functional amyloids of the gut microbiome. We observed the enrichment in cytokine-binding proteins, growth factors, and G-protein-coupled receptors, as well as tyrosine kinase receptors, which all mediate cell response to internal and external factors. Similarly, terms related to endocytosis, exocytosis, and integrin-binding, all regulating protein transport, are also overrepresented. Notably, human pathological amyloids, such as Abeta-42 and tau, hallmarks of Alzheimer’s disease, also bind to integrin and integrin binding receptors^48,49, which suggests that some interactors might be shared by bacterial functional and human pathological amyloids.

Fig. 4. The results of the overrepresentation analysis for A.

The results of the overrepresentation analysis for A gene ontology molecular function, B gene ontology biological process, C KEGG pathways.

Analogously, overrepresentation results concerning KEGG pathways (Fig. 4C) and Disease Ontology (Supplementary Fig. S9) were analyzed. Apart from pathways related to cell-to-cell adhesion, including focal adhesion, cell and adherens junctions, or transport, multiple cancer-related pathways were identified. The most significant signal was observed for the pathway “Proteoglycans in cancer”. This pathway concerns proteins in the microenvironment of the cancerous cells, affecting numerous processes, such as cell proliferation, migration, and adhesion. It is also linked to the MAPK signaling pathway, which plays an important role in cell proliferation, differentiation, and cell death. In addition, the disease ontology results also indicate a significant population of proteins related to cancer disorders. Interestingly, bacterial functional amyloids can potentially bind to human proteins involved in viral infections, as well as other immunity-related pathways such as the chemokine signaling pathway and leukocyte migration.

Note that the identified enriched pathways presented in this section could also be affected by other bacterial proteins, as 88% of the human interactors of UHGPAmyloids_SurfaceProteins were shared with Random_SurfaceProteins. The extent to which these pathways are affected, though, could differ, as we predicted significantly more protein-protein interactions of bacterial functional amyloids than for random bacterial proteins.

Discussion

Growing evidence on the impact of the human gut microbiome on the onset and progression of neurological diseases naturally raised a question about the mechanisms governing this relation. Bidirectional interplay between gut microbiota and the central nervous system can occur via the enteric nervous system or immunological response mechanisms, interactions between neuroactive molecules derived from enteric microbiota with neuronal, immune, or neuroendocrine gut-derived proteins^18,20,50,51.

Recent studies suggested a potential role of certain bacterial functional amyloids in this arena³⁰. These proteins can be expressed by bacteria inhabiting human intestines and show intriguing structural similarity to misfolded human pathological amyloids that mark neurodegeneration. Importantly, little homology is not an obstacle to forming stable structural contacts between such human and bacterial proteins⁵². Based on the studies on PD, a few potential modes of action of the bacterial functional amyloids and their contribution to neurodegeneration could be proposed. First, in light of Braak’s hypothesis that PD is initiated in the enteric nervous system and then propagates to the brain via the vagus nerve¹⁴, it has been proposed that bacterial functional amyloids could serve as triggers of the pathological aggregation of human alpha-synuclein in colon neurons via direct protein-protein interactions, such as cross-seeding¹³. Second, the seeds from bacterial amyloids have been recently shown to be able “to propagate across the endothelial barriers through opening cadherin-based junctions”⁵³ and translocate to the brain through systemic circulation, despite the blood-brain barrier⁵³. Notably, these junctions open in the transport of CsgA protein across the gastric wall, which may indicate a significant role of this protein in PD, as suggested by Otzen et al.⁵³ and references therein. Finally, their remarkable structural similarity to human pathological amyloids could activate similar faulty molecular pathways, in particular, including immune-related ones²¹. For example, both pathological and functional amyloids are capable of activating Toll-like receptors and signal inhibitory receptors on leukocytes-1 (SIRL-1), hence modulating the immune response⁵⁴.

Currently available theoretical and experimental studies on the role of bacterial functional amyloids are often limited to a single protein. We aimed to provide a broader overview of several proteins and their potential role in neurodegeneration, which could be obtained with computational methods. This comes at the price of relying on predictions alone, and therefore, our results have a rather descriptive and correlative character than a causative one. Our approach allowed us to provide the first analysis of the landscape of bacterial functional amyloids likely produced by the human gut microbiome. We identified a significantly diverse set of putative bacterial functional amyloids in the human gut proteome via homology search in seven bacterial phyla, highlighting that the known bacterial functional amyloids may be just the tip of an iceberg. As our results are based on the predictions, the identified catalog of bacterial functional amyloids in the human gut microbiome should be interpreted as a high-confidence candidate set rather than a definitive list.

We investigated the conjecture concerning the potential role of these proteins in the disease using publicly available metagenomic datasets. However, we approached this data from the unexplored amyloid protein-related perspective, as previous analyses of these datasets only focused on taxonomical differences between microbial species in the patient and control cohorts. For two out of three publicly available PD studies (Bedarf and Wallen studies, but not Boktor), we showed that microbiomes of PD patients contain more bacterial functional amyloids than their healthy controls. Notably, the total abundances of bacterial functional amyloids in AD and CM patients did not show significant differences from their respective control groups. Although the overall greater abundance of gut bacterial functional amyloids was shown for two out of three PD cohorts (Fig. 3A), no statistically significant differences in the abundance of any particular protein group were observed, which can be a consequence of differences in the methodological designs and heterogeneity of the human populations (Fig. 3B). Additionally, the discovery of the clear signal may be impaired by different factors, such as diet, treatment and disease stage, all of which are known to influence the microbiome in PD, but were uncontrolled in our analysis. At the same time, it is possible that the overall abundance of microbial amyloids may be more important than the presence of any specific protein. This also highlights the possible limitations of most studies investigating the role of a single functional amyloid, e.g., CsgA.

It has been widely established that bacterial proteins could interact with the host and influence the human molecular processes⁵⁵. Bacterial functional amyloids are likely no exception to this and, according to our analyses, could interact more frequently with human proteins than other bacterial proteins. This can be related to two factors. First, the amyloid proteins found in our study (UHGPAmyloids) were predicted to be more often extracellular than other bacterial proteins; still, an experimental validation would be needed to confirm this. This could potentially enable them to engage in more molecular interactions with human proteins. Second, previous studies have suggested that structural similarity of human pathological amyloids and bacterial functional amyloids could enable molecular mimicry, leading to enhanced aggregation of pathological amyloids or activation of the same molecular pathways^12,54.

The role of the microbiome in shaping and regulating immunity has already been widely noted^12,54. Our results suggest that bacterial functional amyloids may also influence the inflammatory responses. The predicted interaction with human immune proteins was involved in chemokine signaling and leukocyte migration. Identified bacterial functional amyloids produced by gut bacteria (UHGPAmyloids dataset), similar to human pathological amyloids, were also predicted to bind integrin and calmodulin proteins. Abeta and tau binding to integrin has been associated with neurotoxic effects, and calmodulin binding to e.g., alpha-synuclein, Abeta, or apolipoprotein may correlate with neurodegenerative processes^48,56,57. This suggests that experimental exploration of the pro-inflammatory potential of the bacterial functional amyloids could be a promising direction.

Many bacterial species have been shown to influence gut permeability. Our predictions suggest that bacterial functional amyloids could also affect cell junctions^58,59. Although other bacterial proteins are likely to interact with cell junctions as well, we hypothesize that the structural similarity of bacterial functional amyloids to pathological ones could matter in this case. The negative effect of the protein aggregation phenomena on epithelial cell integrity has already been observed and alpha-synuclein and Abeta were found to disrupt the expression of tight junctions and their downregulation^58,60–62. Further experiments could verify whether the abundance of bacterial functional amyloids in the gut may significantly influence gut permeability, similarly to how human pathological amyloids affect the blood-brain barrier.

The predicted potential of bacterial functional amyloids to interact with the immune system could also be related to the structural similarity between the bacterial functional amyloids and pathological ones. Many bacterial functional amyloids can be recognized by the human immune system in the same manner as the Abeta-42 peptide, due to their structural features and despite different sequences of amino acids. For example, CsgA, similarly to Abeta-42, is recognized by the TLR2/TLR1 immune sensor-receptor system, consisting of 13 different TLR-type receptors that control the same up-regulation of IL-17A- and IL-22-mediated pro-inflammatory-signaling pathways^63,64. It is likely that due to structural similarity, other bacterial functional amyloids might trigger similar responses. This opens the door to further experimental endeavor, as the current experiments have focused mostly on CsgA. We hypothesize that the observed greater abundance of bacterial amyloids may in part contribute to the onset of PD by activating inflammation and activity of cytokines, deteriorating vascular permeability and levels of free radicals, as well as enhancing the fibrillation of human amyloidogenic peptides. To further test these mechanistic hypotheses, larger and multiple cohort studies, combined with extensive experimental validation in vivo and in vitro, are required.

Our results, although purely computational, provide the theoretical framework for discussion on the role of bacterial functional amyloids in the gut-brain axis in neurodegeneration. The specific microbiome composition observed in PD has altered functionality, which might be related to the possible increase of bacterial species expressing functional amyloids. This could lead to a pro-inflammatory positive feedback loop, in which bacterial functional amyloids could affect immune response and interact with tight junctions, inciting a greater gut permeability. This, in turn, could enable them to penetrate the epithelial barrier and promote further inflammation. Finally, the predicted potential of bacterial functional amyloids to interact with similar proteins as human pathological amyloids, as well as their structural resemblance, could induce alpha-synuclein aggregation in the enteric nervous system. This aggregation could then propagate via the vagus nerve to the brain, in accordance with the Braak hypothesis¹⁴. While each of these effects alone likely contributes to the onset of disease, their combination may have the potential to create “perfect storm” conditions. It should be noted that the hypotheses, derived from our predictions, still do not directly elucidate molecular mechanisms and instead highlight potential directions for future experimental studies. Nevertheless, we believe that the presented protein-level framework provides a unique perspective on putative molecular mechanisms of neurodegeneration and opens new avenues for novel research and therapeutic strategies.

Methods

Identification of amyloids in the human gut

The dataset of known bacterial proteins was represented by the Unified Human Gut Proteome UHGP from²⁶. The Unified Human Gut Proteome project provides multiple datasets. We used a version of UHGP clustered on 95% of sequence identity in a v1.0 edition (file name: uhgp-95.faa). The dataset was downloaded from the website http://ftp.ebi.ac.uk/pub/databases/metagenomics/mgnify_genomes/human-gut/v1.0/.

It was infeasible to test all UHGP proteins with the amyloid predictors. Classical sliding-window predictors, like e.g., Amylogram/FoldAmyloid^65,66 and many others, would find an aggregation-prone region in almost every protein sequence sufficiently long (see Supplementary Fig. S1). AmyPred-FRL possibly reduces the rate of false positive sequences, but it cannot be used on a large scale due to web server limitations. Therefore, we narrowed the pool of potential bacterial functional amyloids via permissive homology search of known bacterial functional amyloid proteins in UHGP. The BFA dataset used in this study was obtained and published in ref. ²⁵. The summary of the BFA dataset is given in Supplementary Table S6.

We searched for homologs of BFA in UHGP using mmseqs with the command: mmseqs search bfa uhgp-95 results tmp --comp-bias-corr 0 --mask 0⁶⁷. Multiple functional amyloids contain tandem repeats and low complexity regions, which are responsible for the aggregation²⁵, we did not mask such regions during homology search.

The amyloids were filtered using AMYPred-FRL software, available as a web server, with the cutoff value of 0.8²⁷. AMYPred-FRL is an entire sequence-based predictor. To further examine the aggregation potential of our sequences, we used ArchCandy²⁸. For each sequence, we used a cutoff value of 0.5 and ran the following command: java -jar ArchCandyV2_CLI.jar –TMfilter -t = 0.5 -i=Seq_ID Sequence.

The dataset of homologs of BFA in UHGP with positive aggregation predictions was further denoted as UHGPAmyloids.

Clusterization of UHGPAmyloids sequences was performed with CD-HIT using the following command: cd-hit -i input_file.fasta -o output_file.fasta -c 0.9/0.8/0.7/⁶⁸.

Shuffled and random sequences

Results for UHGPAmyloids were compared to the results for a random set of 1000 sequences from UHGP (Random dataset) and to the shuffled sequences of UHGPAmyloids (Shuffled UHGPAmyloids dataset). Random sequences were extracted with seqkit sample⁶⁹. Shuffling was performed using ushuffle with parameter -k = 2 to obtain shuffled sequences with the same distribution of k-mers of size 2 as the UHGPAmyloids proteins⁷⁰.

Taxonomic analysis

Taxonomic assignments of UHGPAmyloids sequences were already provided in the UHGP dataset under the file genomes-all_metadata.tsv. The taxonomic nomenclature throughout the text is consistent with the one provided in this file. Shannon entropy calculations were performed with the R phyloseq package designed for taxonomic analyses⁷¹.

Metagenomic analysis

To study the occurrence of functional amyloids in patients suffering from amyloid-related neurodegenerative diseases, we analyzed data from three Parkinson’s disease studies, Bedarf et al. (2017)³¹ (108 cases, 74 controls), Wallen et al. (2022)³² (319 cases and 156 controls), Boktor et al. (2023)³³ (47 cases and 49 controls), one Alzheimer’s disease study, Laske et al. (2022)³⁵ (75 cases and 100 controls), and Cryptococcal Meningitis study Li et al. (2022)³⁴ (15 cases and 16 controls). For the Bedarf dataset, some patients were represented with more than one sample. For some samples in the study by Wallen, metadata were missing, and such samples were excluded from our analysis. Raw reads for the metagenomics section of the analysis were downloaded from the Sequence Read Archive (SRA). All samples were quality-controlled using Trim Galore 0.6.10⁷² with default parameters. Quality-controlled reads were then assembled using MetaSpades 3.15.5 with default parameters⁷³. Genes were predicted using Prodigal 2.6.3⁷⁴. Then, to estimate coverage, reads were mapped into predicted genes using KMA 1.4.14⁷⁵. To estimate relative abundance while correcting for gene length bias, the average coverage depth (Depth) of each gene was used rather than raw read counts. The relative abundance was calculated by dividing the depth of each gene by the sum of all depth values within the sample. Previously identified amyloids from UHGPAmyloids were then searched in all samples using MMseqs2 15.6f452 with default parameters⁷⁶. Obtained hits were then filtered by percentage of identical residues (>90) to ensure that only highly similar sequences were analyzed (see Fig. 5).

Fig. 5.

The pipeline used for the identification of potential amyloids in metagenomic samples.

To test the difference in abundances of amyloids between case and control groups, Mann–Whitney U-test was applied. For comparing abundances of specific proteins between case and control groups, Mann–Whitney U-test with Benjamini-Hochberg correction was applied⁷⁷.

For each amyloid homolog, cohort-specific effect sizes were estimated using generalized linear models (GLM) applied to log-transformed normalized abundance values, with adjustment for sex, where available in cohort metadata. Effect sizes were reported as log fold-change comparing case versus control groups, with standard errors and 95% confidence intervals. For PD cohorts (Bedarf, Boktor, Wallen), when data were available from two or more cohorts, a random-effects meta-analysis was performed using the DerSimonian–Laird method to pool effect estimates. Heterogeneity across PD cohorts was assessed using Cochran’s Q statistic, and between-study variance (τ²) was estimated.

Prediction of cellular localization

We used BUSCA to predict cellular localization of proteins⁷⁸, BUSCA predictions are based on the identification of different peptides and domains. All predictions were run on the web server with a taxonomic group: Prokarya—Other—3 compartments. The distributions of cellular localization were compared using the chi2_contingency scipy function under the null hypothesis that samples follow the same distribution. The method parameter was set to PermutationMethod(), and the correction was set to False.

Protein–protein interaction prediction

For protein-protein interaction prediction, we only considered proteins from UHGPAmyloids and Extracellular or Plasma membrane localization according to BUSCA. The human proteins were extracted from The Human Protein Atlas⁴⁰, file normal_tissue.tsv (version: 23.0) is available in the Downloadable data section on the website. normal_tissue.tsv was filtered to contain only proteins expressed in the intestine. To do so, we considered proteins that contained the colon, small intestine, duodenum, or rectum in the Tissue column. Next, we filtered the dataset with respect to the cellular localization. For each intestine protein, we extracted its Subcellular location [CC] and checked if it contained an expression from the following set: {cell membrane, cell membrane, secreted, Secreted, extracellular, Extracellular, cell surface, Cell surface, junction, Junction, secretory, Secretory, cell wall, Cell wall}. This way, we identified 2361 proteins that are potentially at the first line of contact with substances secreted by the microbiome, which were further denoted as HPAIntestine_SurfaceProteins.

We performed predictions of protein-protein interactions with ProteinPrompt⁴¹, (using default parameters) between proteins from UHGPAmyloids with Extracellular or Membrane cellular localization according to BUSCA (UHGPAmyloids_SurfaceProteins) and HPAIntestine_SurfaceProteins, which can be defined as:

$$\begin{array}{c}\{\text{A}:BwhereA\in UHGPAmyloids\mathrm{\_}SurfaceProteins,\hfill \\ B\in HPAIntestine\mathrm{\_}SurfaceProteins\}\hfill \end{array}$$

ProteinPrompt assigns each protein-protein interaction a score from 0 to 1. Authors of ProteinPrompt recommend using a 0.5 threshold. The estimated specificity of ProteinPrompt is then 0.88 with a false positive ratio of 0.12, according to the authors.

The overrepresentation analysis was performed for those proteins from HPAIntestine_SurfaceProteins that had at least 5 interacting partners in UHGPAmyloids_SurfaceProteins. The identification of enriched terms was done with the ClusterProfiler package available in R⁴⁵ (version: 4.10). The specific functions used there were: enrichGO, enrichKEGG, and enrichDO.

The F-test was performed for top 20 terms for Biological Process, Molecular Function and KEGG between proteins HPAIntestine_SurfaceProteins which had at least 5 interacting partners in UHGPAmyloids and the entire set HPAIntestine_SurfaceProteins with f.test function in R programming language. Each p-value was multiplied by 60 in accordance with the Bonferroni correction.

General data analysis

All downstream analyses were performed using the Python programming language with NumPy 1.24.3⁷⁹, Pandas 2.0.3⁸⁰, SciPy 1.11.1⁸¹, and Matplotlib 3.7.2⁸² packages. Additional visualizations were performed in R using ggplot2 3.5.1⁸³.

Author contributions

J.W.W., A.W.W., J.S., T.K., and M.K.: conceptualization; J.W.W., A.W.W., K.Z., J.S., T.K., and M.K.: methodology and writing—review & editing; J.W.W., A.W.W., and K.Z.: investigation, formal analysis and visualization; A.W.W. and J.W.W.: writing—original draft; J.S., T.K., and M.K.: supervision; T.K. and M.K.: funding acquisition.

Data availability

The datasets produced in this paper (including UHGPAmyloids and predicted protein-protein interactions) are available on a Zenodo repository: https://zenodo.org/records/14016809.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41522-026-01005-8.