Profiling electric signals of electrogenic probiotic bacteria using self-attention analysis

Abstract

Abstract

We fabricated a self-assembled electric circuit to detect the electrical signals produced by two electrogenic probiotic bacteria [Leuconostoc mesenteroides (L. mesenteroides) and Lactococcus lactis (L. lactis)] on chicken egg chorioallantoic membranes as well as in the intestine lumen of mice. Inoculation of L. mesenteroides or L. lactis plus glucose onto a ferrozine assay triggered the reduction of ferric ions to ferrous ions and the formation of ferrozine complexes, indicating the bacterial electron production. In the presence of glucose, L. lactis yielded higher electricity, measured by voltage changes, than L. mesenteroides in vitro. The spectra of the electrical signals generated by these two probiotic bacteria were highly distinguishable. We evaluated the importance of these differences with the application of a self-attention mechanism, a deep learning-based module, revealing several unique signals in the electrical spectra of L. mesenteroides as well as L. lactis bacteria. The specific electrical spectrum for each probiotic bacterium provided a dynamic signature for evaluation of the efficacy of various therapies using probiotics, antibiotics, and fecal microbiota transplantation in the future.

Key points

• The electrical signals produced by probiotic bacteria L. mesenteroides and L. lactis on chicken egg chorioallantoic membranes and in the mouse intestine lumen were detectable.

• In the presence of glucose, L. lactis yielded higher electricity than L. mesenteroides in vitro. Furthermore, the electrical spectra generated by these two bacteria were different.

• The importance of these differences with the application of a self-attention mechanism revealed several unique signals in the electrical spectra.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00253-025-13425-1.

Introduction

Human gut bacteria such as Enterococcus faecalis, Streptococcus agalactiae, and Staphylococcus aureus (S. aureus) exerted electrogenic activities, generating bioelectricity comparable to that of Shewanella oneidensis, a well-known soil bacterium (Tahernia et al. 2020a, b). Gram-positive bacteria yield bioelectricity via extracellular electron transfer (EET) (Verma et al. 2023). A flavin-based EET mechanism for gut bacterium, Listeria monocytogenes (Light et al. 2018), and fermentation-triggered EET for skin bacteria, Staphylococcus epidermidis (S. epidermidis), have been proposed. Fermentation-triggered EET has also been found in probiotic bacteria. Probiotic Lactococcus lactis (L. lactis) and Lactiplantibacillus plantarum produce electricity via EET to increase their growth during the fermentation of bovine milk (Marito et al. 2021). Probiotic Leuconostoc mesenteroides (L. mesenteroides) fermentatively metabolize linoleic acid to produce electrons which induce the reduction of ferric ions to ferrous ions (Pham et al. 2020). The ferrozine assay for detection of the iron reduction (Pham et al. 2020; Riemer et al. 2004) has previously been developed to screen electrogenic bacteria in the human fecal samples and skin swabs (Balasubramaniam et al. 2020a, b, c). L. mesenteroides was commonly found in fermented foods (Jung et al. 2012) and detectable in human breast milk (Ariute et al. 2023). The previous publication applied this assay to show that L. mesenteroides isolated from curd cheese yield electrons via fermentation of glucose (Yang et al. 2020) or linoleic acid (Pham et al. 2020). Measurement of the electricity from a single bacterial species has been mainly conducted in vitro mostly using a microbial fuel cell (MFC). In fact, it is difficult to detect electrical activity specifically produced by a single bacterial species in a mixed bacterial culture. Furthermore, the patterns of electrical signals elicited by electrogenic bacteria may be varied in the in vivo microbiome in a dynamic environment (Schwab et al. 2019).

Different bacterial species express different enzymes that ferment unique substrates (Wang et al. 2016). For example, S. aureus bacteria are known to utilize glucose and mannitol (Erasmus 1985), but not galactitol or saccharic acid (Safonova Tb Fau—Shcherbakova et al. 1978), as substrates for fermentation. The same carbohydrate substrates induced the synthesis of short-chain fatty acids (SCFAs) at varied amounts (Reichardt et al. 2018). Both carbohydrates and SCFAs (Frühauf-Wyllie et al. 2022) were favorable as electron donors for bacteria to produce electricity. Thus, different electrogenic bacteria may generate the distinguishable patterns of electrical signals via fermentation-triggered EET despite utilizing the same carbohydrate substrate or electron donor. Identification of a unique, species-specific pattern of electrical signals makes it possible to detect specific bacterial species in an in vitro mixed culture and in vivo human microbiome.

Electricity elicited by a single bacterial species in vivo can be distinguished against the background activity of other bacterial species and host physiological conditions including heartbeat and reduction–oxidation (redox) status (Tahernia et al. 2020a, b). The chicken embryo, an alternative to rodents, has been widely used as an experimental animal model for many studies including in the field of bioelectrochemistry (Haddad 2022). Notably, the microbiota in chicken embryo (Ding et al. 2021) is relatively simple compared to that in rodents which harbors a complex population of microorganisms. The chicken egg chorioallantoic membrane (Finotello et al. 2018), a highly vascularized extraembryonic membrane, was first exploited for bacterial culture (Morrow G Fau—Syverton et al. 1938; Ribatti et al. 1999), although it has not been extensively used for microbiome studies. The electrical properties of CAM have been characterized by the potential difference and circuit current provoked by the transport of sodium across the CAM (Pácha J Fau—Ujec et al. 1985).

Electrical signals provide insight into the physiological health of humans. The electrocardiogram (ECG) (Ansari et al. 2023), electroretinogram, electroneurogram, electroencephalogram (EEG) (Wang et al. 2024), electromyogram, and electrooculogram are typical electrical signals which can be collected from different tissues. More recently, the assessment of electrodermal activity (Kokubo et al.1999) from the skin surface is an emerging technique for monitoring the progression of various diseases (Nardelli et al. 2022; Singaram et al. 2023). Deep learning analysis of the spectra of electrical signals in large datasets allows the precise differentiation of various electrical signals in health and disease (Petmezas et al. 2022). Self-attention used in deep learning architecture provides focus into specific parts of datasets (Zhao et al. 2022). In this study, the self-attention mechanism was executed to interrogate electrical signals and distinguish specific patterns generated by different bacteria. Like other electrical signals, the unique patterns produced by specific bacteria may serve as microbiome-derived biomarkers for prediction of the effectiveness of various therapies using probiotics, prebiotics, antibiotics, and fecal microbiota transplantation (FMT) (Rees et al. 2022). The polymerase chain reaction (PCR) technologies and next-generation sequencing (NGS) have been currently used for microbiome profiling (Cason et al. 2022; Wensel et al. 2022). PCR technologies, being targeted to specific microorganisms and genes, cannot profile the whole microbiome. The NGS for simultaneously sequencing millions of bacterial DNA fragments has provided a thorough picture of the microbiome (Finotello et al. 2018). It can be applied to amplicon sequencing, whole metagenome shotgun analyses, and 16S ribosomal RNA (rRNA) sequencing (Wensel et al. 2022). Although PCR technology and NGS can accurately reveal the sequences of targeted genes or whole genome in the microbiome, they fail to detect the dynamic changes of the microbiome. Unlike PCR technology and NGS for microbiome profiling (Tozzo et al. 2022), detection of electrical signals of electrogenic bacteria provides a means to dynamically monitor the microbiome in real time.

Materials and methods

Ethics statement

This study was performed in accordance with the protocols (No. 2023–10-01) of the Institutional Animal Care and Use Committee (IACUC) of Dalian University Affiliated Xinhua Hospital.

Bacterial culture

L. mesenteroides, L. lactis bacteria isolated from curd cheese (Pham et al. 2020) or Escherichia coli (E. coli; ATCC35218) were cultured in tryptic soy broth (TSB) (Sigma, St. Louis, MO, USA). The 16S rRNA sequences of isolated L. mesenteroides and L. lactis shared 100% identity to L. mesenteroides strain LK-151 and L. lactis strain AN42 16S ribosomal RNA genes, partial sequences, respectively. Both isolated L. mesenteroides and L. lactis strains used in this study have been deposited at China Center for Type Culture Collection (CCTCC) with strain no. and M2024452 and M2024074, respectively. After 1:100 dilution, bacteria were cultured overnight at 37 °C to an optical density of 600 nm (OD₆₀₀) = 1.0. Bacterial pellets were obtained by centrifugation at 1957 × g for 10 min, washed with phosphate-buffered saline (PBS), and suspended in PBS for all experiments.

Ferrozine assays

For preparation of ferrozine assays in 96-well plate format, TSB media containing 0.8% agar, 4 mM ferrozine (Alfa Aesar Chemicals, Tewksbury, MA, USA), and 50 mM ferric ammonium citrate (Sigma) were added into a 96-well plate (ThermoFisher Scientific, Waltham, MA, USA). TSB alone, 2% glucose alone (100 µL), or bacteria [10⁷ colony forming unit (CFU)/100 µL in TSB] supplemented with or without 2% glucose were pipetted onto the 96-well assay agar plate. OD₅₆₂ measurements were made after 37 °C incubation for 24 h. The color change of agar plate was detected and quantified by a calibration curve.

Detection of bacterial electricity on chicken egg CAM

A circle with a diameter of 3 cm on the shell of a Hy-line brown egg (Hy-line International, Iowa, USA) at day 11 of embryo development was opened with a drill. An electric circuit with electrodes (cathode and anode) connected to a digital multimeter (Lutron, DM-9962SD, Sydney, Australia) was fabricated as previously described (Marito et al. 2021). In brief, a carbon filth (1 × 0.5 cm square) (Tmax Battery Equipment, Ximen, China) served as an anode. A copper wire wrapped with a nafion membrane N117 (Tmax Battery Equipment), a proton exchange membrane (PEM), served as a cathode. To detect the bacterial electricity on chicken egg CAM, the anode was laid on the top surface of the CAM, and cathode with PEM was inserted into the CAM and placed underneath the anode. Bacteria (10⁷ CFU/100 µL) in TSB media in the absence or presence of 2% glucose were applied onto the anode. Application of the same volume (100 µL) of TSB alone or glucose alone onto the anode acted as controls. To create a sterile environment, chick embryos were placed in a Class II Biosafety Cabinet at room temperature during electricity detection. The digital multimeter read the voltage change per second as an indicator of the electricity production of bacteria. The recorded voltage was plotted against time. The area under the curve (voltage versus time) was calculated to quantify the amount of electricity. To detect egg bacteria, areas (1 × 1 cm square) on the egg shell and CAM were gently swabbed with sterile cotton swabs before spreading onto TSB agar plates for the growth of bacterial colonies.

Detection of bacterial electricity in mouse intestine

BALB/c male mice (8–9 weeks old) purchased from the Xishan Biotechnology Co., Suzhou, China, were anesthetized by inhaling 4% isoflurane. A copper wire was inserted into carbon filth (0.2 cm square) to create an anode. The carbon filth inserted with a copper wire was placed into a high borosilicate glass tube composed of silicon dioxide (Jinan Munan Pharm Glass Co. Ltd., Shandong, China) before implantation into the mouse intestine. An electric copper sheet (1 × 0.5 cm square) wrapped with a PEM (nafion membrane N117) acted as a cathode. Electrodes (cathode and anode) were autoclaved for mouse experiments. An uncut segment (3 cm) of the mouse intestine was relocated onto a Pertri dish with (10 mL) PBS. Feces in the intestine segment were removed by washing three times with PBS. Antibiotics [Penicillin–Streptomycin (1 mM, 50 µL), ThermoFisher Scientific] and 1% sodium hypochlorite were injected into the intestine lumen for 20 min. After washing the lumen three times with PBS, the anode was inserted into the intestine lumen. The cathode was placed on the outer surface of the intestinal segment. TSB media (50 µL), 2% glucose, or bacteria (10⁷ CFU/50 µL) with or without glucose were injected into the intestinal segment that was 0.5 cm away from electrodes. Electrodes were a digital multimeter to record voltage changes for 3 min. All experiments using mice (n = 3 per group) were conducted according to institutional guidelines. At the end of the experiment, all mice were sacrificed under CO₂ anesthesia in a sealed chamber.

Pearson correlation coefficient

Four samples (TSB media alone, 2% glucose alone, bacteria alone, and bacteria plus 2% glucose) were prepared for detection of electricity production. Each sample was pipetted onto the anode. Electricity measured by voltage changes (Ariute et al. 2023) was detected by a digital multimeter every second for 3 min to generate 180 (n = 180) data points for each sample. The linear relationship of data from each sample was analyzed through the formula of Pearson correction coefficient (r).

$$r=\frac{{\sum }_{i=1}^{\text{n}}\left({\text{x}}_i,-,\overline{\text{x}}\right)\left({\text{y}}_i,-,\overline{\text{y}}\right)}{\sqrt{{\sum }_{\text{i}=1}^{\text{n}}{\left({\text{x}}_i,-,\overline{\text{x}}\right)}^2}\sqrt{{\sum }_{\text{i}=1}^{\text{n}}{\left({\text{y}}_i,-,\overline{\text{y}}\right)}^2}}$$

In this formula, n is the number of values, x is a set of data points collected from a selected sample, x̄ is the mean of the x data set, y is a set of data points obtained from a related sample, ȳ is the mean of the y data set, and Σ represents a summation of all values (Blyth 1994; Hu et al. 2022). The correction values, displayed in a 4 × 4 correlation matrix, were scaled from − 1 to 1, where 1, 0, and − 1 mean perfect positive correlation, no linear relationship, and perfect negative correlation, respectively.

Self-attention mechanism for electrical signal analysis

Voltage change was recorded every second for 3 min to obtain 180 data points with varied mV. A spectrum (mV versus s) of the electrical or voltage fluctuation was obtained for each sample. Self-attention mechanism was performed to extract the unique features of peak signals in each spectrum of electrical fluctuation signals. The self-attention mechanism contains three Query (Q), Key (K), and Value (V) vectors (Ashish Vaswani 2023). Using the data point (Ariute et al. 2023) of each peak signal in a spectrum of electrical fluctuation signals as inputs, the self-attention mechanism computed the linearly projected versions of the query, key, and value vectors through multiplication with the respective weight matrix (Ji et al. 2024). The weights were calculated to acquire the similarity/attention scores between the query and key vectors based on the formula of Attn (i, j) = Softmax [Q.K^T/square root (d_k)]. K^T denoted the transpose of K, and d represented the feature dimensionality at each time step (Ashish Vaswani 2023). The attention weights as Attn (i, j) were attained by computing a scalar value that signified the degree of association between the ith and jth peak signals in the electrical spectra. Softmax normalization was implemented to assure that all the numbers in the vector sum to 1 and reflect the importance of each peak signal in an electrical spectrum. The similarity/attention score was presented as an attention matrix. Finally, the attention weights were multiplied by the weighted sum of each value vector (Attention = Attn x V) (Jung et al. 2012; Yi Tay et al. 2022) to gain the final outputs. The process effectively enabled the self-attention mechanism to learn complex dependencies between different electrical spectra as inputs and gained a more accurate representation. The attention distribution with final outputs was superimposed on an electrical spectrum to pinpoint the peak signals with high attention scores.

Statistical analysis

Data were analyzed by unpaired t-test and two-way ANOVA using GraphPad Prism® 8.0 software. The P-values of < 0.05 (*), < 0.01 (**), and < 0.001 (***) from unpaired t-test and a P-value < 0.05 from two-way ANOVA denoted statistically significant results. The mean ± standard deviation (SD) for at least three separate experiments was calculated.

Results

Characterization of electrogenic probiotic bacteria in a ferrozine assay

In the current study, we characterize the electrogenicity of L. mesenteroides (Pham et al. 2020) and L. lactis (Plupjeen et al. 2020) a probiotic bacterium detectable in the human gut microbiome (Plupjeen et al. 2020), in the presence of glucose with the ferrozine assay. As shown in Fig. 1A and B, the cultures appeared light purple after the application of TSB alone or 2% glucose alone onto the 96-well ferrozine assay agar plates for 24 h. However, dark purple complexes developed following the application of probiotic bacteria [L. mesenteroides or L. lactis] alone or bacteria plus 2% glucose. Induction of the formation of dark purple agars by bacteria without exogenously adding 2% glucose may be caused by bacterial consumption of dextrin, a mixture of glucose polymers, in TSB media (Kokubo et al. 1999). Colorimetric quantitation demonstrates that the most robust ferrous-ferrozine complex formation occurred when agar plates were applied with L. mesentroides or L. lactis plus 2% glucose (Fig. 1C, D). This result indicates that probiotic L. mesenteroides and L. lactis bacteria in the presence of glucose may convert the ferric ions to ferrous-ferrozine complexes via electron production.

Fig. 1.

Conversion of ferric ions to ferrozine-chelatable ferrous ions by probiotic bacteria in the presence of glucose. L. mesenteroides (LM) or L. lactis (LL) bacteria (10⁷ CFU/100 µL) in the presence (LM + GLU; LL + GLU) or absence of 2% glucose (GLU) were added onto a 96-well assay. Addition of 100 µL of TSB or 2% glucose served as controls. After 37 °C incubation for 24 h, the dark brown complexes of ferrozine-chelatable ferrous ions were photographed (A, B) and quantified (C, D) using values of OD₅₆₂ absorbance. The P-values of < 0.001 (***) and < 0.05 (*) via unpaired t-test from three different experiments with mean ± SD were shown. The statistical significance was also confirmed by two-way ANOVA with a P-value < 0.05. ns, non-significant

Chicken egg CAM as a platform for electricity detection in ovo

Electric circuits with miniatured anode and cathode were fabricated to detect the electron production of probiotic bacteria. To evaluate whether this platform was able to discriminate bacterial activity in the presence of host factors in vivo, live chicken embryo CAM was chosen as a model to monitor the electricity of L. mesenteroides and L. lactis. In contrast to the large numbers of bacterial colonies harvested from the chicken shell surface, no culturable bacterial colonies formed after spreading swabs collected from the surfaces of the CAM on a TSB agar plate (Fig. S1). To monitor the electricity of probiotic bacteria in a dynamic in vivo environment, electric signals from probiotic bacteria were recorded per second continuously for up to 180 s immediately after inoculation of L. mesenteroides or L. lactis in the presence of 2% glucose on the anode which was placed on the CAM surface (Fig. 2A–D). Inoculation of TSB alone, glucose alone, or bacteria alone were included as control groups. A live chicken embryo with cardiac activity, a prominent head, and an extensive vascular network between the CAM and the embryonic body was observable during electrical signal recording (Fig. 2E).

Fig. 2.

An assembly of an electric circuit on chicken egg CAM for detection of electricity of probiotic bacteria. Copper wires with a carbon filth and PEM as anode and cathode (A), respectively, were connected to a digital multimeter (M) (B). The anode was positioned on the top surface of CAM (C) after cathode was placed underneath CAM (D) of an 11-day chicken embryo (CE). One hundred microliters of probiotic bacteria (10⁷ CFUs in TSB) with/without glucose, TSB alone, or bacteria alone were pipetted onto anode. E Live embryo (dash line) with a large head (arrow) and a network of blood vessels (V) were visible during electricity detection. Bars = 0.5 cm

Dynamically monitoring electricity of probiotic bacteria supplemented with glucose

The voltage changes on chicken egg CAM, as an indicator of bacterial electricity, were recorded. As shown in Fig. 3A and C, in the presence of 2% glucose, probiotic L. mesenteroides and L. lactis provoked extensive electrical fluctuations with an electric signal reaching its highest level at approximately 10 and 20 mV, respectively. TSB, glucose, or bacteria alone generated relatively stable electrical signals with values of voltage changes below 5 mV, although a peak voltage of about 18 mV was detected when the anode was inoculated with L. lactis alone. A Pearson correlation coefficient analysis demonstrated that most spectra had low similarities with correction values below 0.5 (Fig. S2 A, B), indicating that electrical spectra generated on chicken egg CAM were different from each other. Furthermore, the spectra of the electrical spectra generated by L. mesenteroides and L. lactis were highly distinguishable (Fig. S2 C). The levels of bacterial electricity were quantified by calculation of the area under the curve of electrical spectra. Data in Fig. 3B and D showed that, compared to other control groups, probiotic bacteria plus 2% glucose yielded the high levels of electricity. These results are consistent with the findings of the ferrozine assay in Fig. 1 and demonstrate that both probiotic L. mesenteroides and L. lactis bacteria can metabolize glucose to elicit electrons.

Fig. 3.

Spectra of electrical signals of electricity yielded by probiotic bacteria. Electricity detected by voltage changes (Ariute et al. 2023) was recorded for 180 s after addition of L. mesenteroides (LM) or L. lactis (LL) bacteria alone or bacteria plus 2% glucose (LM + GLU or LL + GLU) onto the surface of anodes. Addition of TSB alone or glucose (GLU) alone acted as controls. A spectrum of electrical signals of electricity was displayed by plotting a graph of voltage against time (A, C). The area (mV × s) under the curve of electrical signals, referred to the amount of electricity (B, D), was calculated. The P-values of < 0.01 (**) and < 0.05 (*) from three independent experiments with mean ± SD were denoted. ns, non-significant

Distinct spectra of electrical signals generated by two probiotic bacteria

Both L. mesenteroides and L. lactis bacteria in the presence of glucose produced measurable electricity on the CAM, with distinguishable spectra of electrical signals (Fig. 4A, C). Self-attention is a deep learning mechanism that allows an analysis of different peak signals within a spectrum of the electrical fluctuations by assigning a weight to each peak signal, determining its significance in an electrical spectrum. The self-attention mechanism was executed to identify unique peak signals across the electrical spectra generated by L. mesenteroides and L. lactis bacteria in the presence of 2% glucose. Five peak signals appearing at 36, 76, 109, 131, and 172 s showed high attention scores in the electrical spectrum of L. mesenteroides plus 2% glucose (Fig. 4A). By contrast, L. lactis plus 2% glucose produced five peak signals with high attention scores detected at 22, 60, 143, 161, and 175 s in the electrical spectrum (Fig. 4A, C). A correlation matrix highlighted a heatmap with the pairwise correlation values in a 2-dimensional matrix (Tang et al. 2020). As displayed in Fig. 4B and D, peak signals with high attention scores appeared at 120–180 and 0–60 periods of time (seconds) in the electrical spectrum of L. mesenteroides or L. lactis plus 2% glucose, respectively. Furthermore, five highest attention scores in the electrical spectrum from each bacterium can be located in a correlation matrix. To investigate if peak signals from each individual bacterium were still detectable when two bacteria grew together, a mixture of L. mesenteroides and L. lactis in the presence of 2% glucose was applied onto the anode placed on the CAM surface. Five peak signals with high attention scores were found at 2, 22, 36, 41, and 175 s in the electrical spectrum (Fig. S3). A peak signal appearing at 36 s was measurable in electrical spectra of L. mesenteroides and mixed bacteria. Peak signals showing up at 22 and 175 s were detectable in electrical spectra of L. lactis and mixed bacteria. This data suggested that electric signals from either L. mesenteroides or L. lactis were detectable when two bacteria were mixed. The results above emphasize the practicability of the self-attention mechanism to separate the different spectra of electrical signals generated by L. mesenteroides and L. lactis. To exam if electricity produced by L. mesenteroides or L. lactis bacteria was detectable in the mammalian intestine, TSB, 2% glucose or bacteria in the presence or absence glucose was injected into the intestine lumen of mice (Fig. 5). Nine peak signals (10, 13, 19, 31, 36, 76, 109, 132, and 177 s) and fourteen peak signals (4, 7, 22, 40, 60, 100, 104, 120, 143, 158, 161, 170, 171, and 174 s) with high attention scores were detected from electrical spectra of L. mesenteroides and L. lactis, respectively, when bacteria were exposed to intestine lumen in the presence of glucose. Peak signals at 36, 76, and 109 s from L. mesenteroides and 22, 60, 143, and 161 s from L. lactis appeared both on the CAM surface and in the intestine lumen of mice (Figs. 4 and 5 D–G). The results demonstrated that electrical spectra of two bacteria were discriminable in different in vivo environments.

Fig. 4.

Distinguishable electrical spectra of two probiotic bacteria with unique signal peaks via self-attention analysis. Four spectra (blue lines) of electrical signals collected from anodes applied with TSB, glucose (GLU), bacteria [L. mesenteroides (LM) or L. lactis (LL)], or bacteria plus glucose (LM + GLU or LL + GLU) are displayed in Fig. 3. The curve of self-attention distribution (pink) was placed over each electrical spectrum (A, C). After Softmax normalization, the peak signals with high attention scores in the spectra of electrical signals collected from anodes applied with bacteria plus glucose were pinpointed (arrows). The attention scores (− 1 to 1) were presented as attention matrixes (B, D). The areas (120–180 and 0–60 s) with high scores in attention matrixes were shown (dashed squares)

Fig. 5.

Detection of electricity of probiotic bacteria in mouse intestine for self-attention analysis. A A copper wire-inserted carbon filth served as an anode which was put into a high borosilicate glass tube for intestinal implantation. An electric copper sheet wrapped with PEM (arrow) acted as a cathode. B The anode was inserted into the lumen and the cathode was placed on the surface of an uncut segment of the mouse intestine (red line). C A schematic drawing showed the locations of implanted electrodes and injection (solid triangles) of bacteria. D, F Four spectra (blue lines) of electrical signals collected from anodes after injection of TSB, glucose (GLU), bacteria [L. mesenteroides (LM) or L. lactis (LL)], or bacteria plus glucose (LM + GLU or LL + GLU) into the intestine lumen, and the curves of self-attention distribution (pink areas) were shown. (E, G) The attention matrixes derived from the spectra of electrical signals from anodes after luminal injection of bacteria plus glucose were created. The signal peaks (arrows) and areas (dashed squares) with high scores in attention matrixes were presented. One representative (D–G) with similar results from three mice per group was shown per group Bars = 0.5 cm

Discussion

The electrogenicity of several bacteria in the human microbiome has been well characterized. Skin bacteria such as S. epidermidis and S. hominis undergo fermentation-triggered EET using glycerol or other substrates to produce electrons (Balasubramaniam et al. 2020a, b, c). Gut bacteria including Enterococcus faecalis, Streptococcus agalactiae, S. aureus (Balasubramaniam et al. 2020a, b, c; Balasubramaniam et al. 2020a, b, c; Tahernia et al. 2020a, b), and Lactobacillus rhamnosus (Ganzorig et al. 2023) yield electrons in the presence of carbohydrates in culture media as electron donors. Probiotic L. mesenteroides and L. lactis use cyclophilin A (Pham et al. 2020) and quinone (Freguia et al. 2009) as electron mediators, respectively, to perform EET for electron production. E. coli, a dominate microorganism in the gut, is not naturally electrogenic (Cialek 2024). As shown in Fig. S4, in the presence of glucose, electricity measured by voltage changes was not generated by E. coli. Our results in Fig. 3 demonstrated that L. lactis elicited higher electricity, measured by voltage changes on chicken egg CAM, than L. mesenteroides in the presence of 2% glucose. Both L. mesenteroides and L. lactis are facultative anaerobic lactic acid bacteria, and fermentation by L. lactis has been reported under anaerobic conditions (Gu et al. 2023). However, in the presence of heme, oxygen can be beneficial to L. lactis by enhancing bacterial growth and survival (Brooijmans et al. 2007). In this study, we measured bacterial electricity under aerobic conditions on chicken egg CAM which contained heme (Preuße et al. 2023). L. mesenteroides and L. lactis bacteria enter anaerobic intestinal tract after probiotic supplementation, although the mechanism responsible for maintaining the anaerobic environment of the intestine lumen remains elusive (Albenberg et al. 2014). Profiling the electrical signals of probiotic bacteria can be achieved by measuring electricity using human fecal samples (Pandit et al. 2022) or inserting an electrode-integrated colonoscopy into anaerobic intestine lumen (Nguyen et al. 2022). Furthermore, using ex vivo mouse caecum, the effects of electrical potentials of gut epithelia on bacterial locations have been studied (Sun et al. 2024). As shown in Fig. 5, both L. mesenteroides and L. lactis bacteria can mediate glucose to exert electricity in mouse intestine, although oxygen levels in the intestine lumen of mice were not measured. Notably, the electric spectra from L. mesenteroides and L. lactis bacteria were distinguishable either on the CAM surface or in the mouse intestine (Figs. 4 and 5). The in vivo electrical spectra of probiotic bacteria can serve as biomarkers for probiotic or antibiotic treatments of gastrointestinal diseases in humans. Thus, it is worth comparing the electrical spectra of probiotic L. mesenteroides and L. lactis bacteria when they are exposed to an anaerobic environment.

The chicken egg CAM with a rich vascular network for gas exchange is a homolog of the mammalian placenta. It has been widely used as a model for evaluating angiogenesis during tumor growth or wound healing (García-Gareta et al. 2020). The Animal Scientific Procedures Act (ASPA) stated that a chicken embryo that has not reached the 14th day of its gestation period can be used for experimentation without ethical restrictions (García-Gareta et al. 2020). The microbiota in chicken embryo gut during stages of developments have been identified by 16S rRNA sequencing (Akinyemi et al. 2020). Salmonella typhimurium via egg shell penetration has been isolated from chicken egg CAM (Padron 1990). In our study, no culturable CAM bacteria were detected on TSB agar plates (Fig. S1). Although we cannot rule out the possibility of the existence of endogenous microbes on chicken egg CAM which influences the electricity detection, two electrical spectra derived from L. mesenteroides and L. lactis bacteria were very different from each other (Figs. 3 and 4). Glucose, a digestible carbohydrate with a concentration of approximately 48 mmol/L in the human gut (Gromova et al. 2021) was used as an electron donor for electrogenic probiotic bacteria in this study. Other carbohydrates or molecules such as formate and acetate (Hughes et al. 2017) with properties of electron donors have been found in the human gut. As an electron donor, glucose triggered the L. mesenteroides and L. lactis bacteria to yield electrons. Five peak signals with the high attention scores in the electrical spectra of L. mesenteroides as well as L. lactis have been selected as unique electric signals (Fig. 4). Future work will determine if peak signals selected by self-attention mechanism in this study are inducible when different electron donors are given to L. mesenteroides or L. lactis. Some of selected peak signals from either L. mesenteroides (e.g., a peak at 36 s) or L. lactis (e.g., peaks at 22 and 175 s) were detected in the mixtures of L. mesenteroides and L. lactis (Fig. S3). The human gut harbors a complex community of over 100 trillion microbial cells with greater than one thousand bacterial species (Guinane et al. 2013). Five Gram-positive bacteria including E. faecalis, S. agalactiae, S. aureus, Lactobacillus rhamnosus, and Lactobacillus reuteri in the human gut have been identified as electroactive bacteria. The in vivo activities of bacterial EET have been detected in the mouse gut (Wang et al. 2019). By using antibiotics specific for Gram-positive and Gram-negative bacteria, scientists have found that Gram-positive bacteria in the mammalian gut exerted higher EET activities than Gram-negative bacteria (Wang et al. 2019). Although the distinguishable electrical spectra presented in the current study were only from two probiotic bacteria, future studies will include using specific antibiotics to differentiate the distinct electrical spectra in the gut microbiome.

Several signals in the spectra of electrical fluctuations generated by L. lactis without glucose had high mV values, but low attention scores (or vice versa) (Fig. 4), indicating that self-attention mechanism did not essentially select those high signals as unique features. The self-attention mechanism via deep learning enables weigh the importance of different electrical signals as input data and dynamically adjusts their influence on output data (Shi et al. 2023). Several techniques by using random split validation approaches (Chopannejad et al. 2024) have been developed to verify the models of deep learning including self-attention mechanism. However, the use of a biochemical approach to knockout the key components such as type II NADH dehydrogenase (Ndh2) in an EET route (Tolar et al. 2023) for inhibiting the electricity production of bacteria may further validate the specificity of electrical signals selected by self-attention analysis. NGS (Wensel et al. 2022) for construction of microbial genome in combination with Basic Local Alignment Search Tool (BLAST) searching on the NCBI database can identify microbes to species level. A limited number of electrical spectra of bacteria were used to differentiate the electric profiles of two probiotic bacteria in the current study. Databases stored the vast amounts of electric profiles of bacteria that may be needed to distinguish bacterial species in the microbiome via self-attention analysis. Bio-signals collected in vivo are frequently contaminated with artificial noise from host physiological activities including muscle contraction and heartbeat (Rodrigues et al. 2017). Although the in vivo electrical spectra of probiotic L. mesenteroides and L. lactis bacteria can be clearly distinguished in this study, future works will adjust the sampling rate for collection of electrical signals (Silva et al. 2023) and combine self-attention analysis with other machine learning algorithms (Rodrigues et al. 2017; Zhao et al. 2020) for denoising and enhancing the accuracy of bacteria-specific electrical spectra. Although many methods such as 16S rRNA sequencing, mass spectrometry-based metabolomics (Bauermeister et al. 2022), and next-generation microarrays (Thissen et al. 2019) can be used for profiling the microbiome, the electrical spectra with self-attention analysis can efficiently separate electrical signals from different bacteria and pave a way for a future study of monitoring the dynamic changes of microbiome in a real time.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

QC, JT, and MZ conducted experimental investigation, data analysis, and methodology. CJ, SC JT, QC, JC, and SS performed data analysis, methodology, and resources. DRH performed the writing — review and editing. QGZ and ZW were in charge of conceptualization and project administration. CMH played a role in conceptualization, methodology, resources, supervision, writing — original draft, writing — review and editing, and funding acquisition.

Funding

The study was supported by internal research funds of Dalian University and a grant from National Key Research and Development Program (2023YFC2508200).

Data availability

The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Declarations

Ethical approval

This article does not involve any ethical approval.

Conflict of interest

The authors declare no competing interests.