Automated Microscopic Analysis of Metal Sulfide Colonization by Acidophilic Microorganisms

The presented method for the assessment of mineral colonization allows accurate relative comparisons of the microbial colonization of metal sulfide concentrate particles in a time-resolved manner. Quantitative assessment of the mineral colonization development is important for the compilation of improved mathematical models for metal sulfide dissolution. In addition, deep-learning algorithms proved that axenic or mixed cultures of the three species exhibited characteristic biofilm patterns and predicted the biofilm species composition. The method may be extended to the assessment of microbial colonization on other solid particles and may serve in the optimization of bioleaching processes in laboratory scale experiments with industrially relevant metal sulfide concentrates. Furthermore, the method was used to demonstrate that DSF quorum sensing signals directly influence colonization and dissolution of metal sulfides by mineral-oxidizing bacteria, such as L. ferriphilum and S. thermosulfidooxidans.

ABSTRACT

Industrial biomining processes are currently focused on metal sulfides and their dissolution, which is catalyzed by acidophilic iron(II)- and/or sulfur-oxidizing microorganisms. Cell attachment on metal sulfides is important for this process. Biofilm formation is necessary for seeding and persistence of the active microbial community in industrial biomining heaps and tank reactors, and it enhances metal release. In this study, we used a method for direct quantification of the mineral-attached cell population on pyrite or chalcopyrite particles in bioleaching experiments by coupling high-throughput, automated epifluorescence microscopy imaging of mineral particles with algorithms for image analysis and cell quantification, thus avoiding human bias in cell counting. The method was validated by quantifying cell attachment on pyrite and chalcopyrite surfaces with axenic cultures of Acidithiobacillus caldus, Leptospirillum ferriphilum, and Sulfobacillus thermosulfidooxidans. The method confirmed the high affinity of L. ferriphilum cells to colonize pyrite and chalcopyrite surfaces and indicated that biofilm dispersal occurs in mature pyrite batch cultures of this species. Deep neural networks were also applied to analyze biofilms of different microbial consortia. Recent analysis of the L. ferriphilum genome revealed the presence of a diffusible soluble factor (DSF) family quorum sensing system. The respective signal compounds are known as biofilm dispersal agents. Biofilm dispersal was confirmed to occur in batch cultures of L. ferriphilum and S. thermosulfidooxidans upon the addition of DSF family signal compounds.

IMPORTANCE The presented method for the assessment of mineral colonization allows accurate relative comparisons of the microbial colonization of metal sulfide concentrate particles in a time-resolved manner. Quantitative assessment of the mineral colonization development is important for the compilation of improved mathematical models for metal sulfide dissolution. In addition, deep-learning algorithms proved that axenic or mixed cultures of the three species exhibited characteristic biofilm patterns and predicted the biofilm species composition. The method may be extended to the assessment of microbial colonization on other solid particles and may serve in the optimization of bioleaching processes in laboratory scale experiments with industrially relevant metal sulfide concentrates. Furthermore, the method was used to demonstrate that DSF quorum sensing signals directly influence colonization and dissolution of metal sulfides by mineral-oxidizing bacteria, such as L. ferriphilum and S. thermosulfidooxidans.

INTRODUCTION

The dissolution of metal sulfides is a chemical process catalyzed by the microbial oxidation of iron(II) ions and inorganic sulfur compounds (ISCs). It leads to the generation of acidic, sulfate, and heavy-metal laden acid mine drainage (AMD) waters. Mineral-attached microorganisms are crucial for the mineral breakdown (1) and are industrially exploited for the recovery of valuable metals from sulfide ores in biomining processes (2, 3). Although the mechanism of metal sulfide oxidation is an indirect chemical process (4, 5), contact of mineral-oxidizing microbes with metal sulfides may significantly increase dissolution kinetics. This is at least partially due to glucuronic acid residues in the extracellular polymeric substances (EPS) of Acidithiobacillus ferrooxidans and Leptospirillum ferrooxidans that accumulate the oxidative agent iron(III) ions (6, 7). The presence of biofilms is especially important for the persistence of active bioleaching microorganisms in commercial heap bioleaching operations (2, 8, 9). In addition, mineral-attached cells are particularly important for initiation of the metal sulfide dissolution. For instance, at dissolved iron ion concentrations of <200 mg/liter, mineral-attached cells of iron(II)- and ISC-oxidizing Acidithiobacillus ferrooxidans or Acidithiobacillus ferrivorans on pyrite surfaces are exclusively responsible for catalyzing its dissolution (10). Consequently, cell attachment to metal sulfides has been extensively studied (10–16). We compare metal sulfide colonization by the ISC oxidizer Acidithiobacillus caldus, the iron(II) oxidizer Leptospirillum ferriphilum, and the ISC- and iron(II)-oxidizing species Sulfobacillus thermosulfidooxidans.

Several methods for the assessment of mineral-attached cells on pyrite or chalcopyrite have been developed. Indirect microscopic cell counts rely on the decrease of planktonic cells during the initial contact with metal sulfides. However, this method cannot assess the temporal development of the mineral-attached cell population for prolonged cultivation periods. Other methods involve a cell detachment step, although quantitative separation of cells from the mineral is not possible and is prone to biases due to the release of fine mineral particles when samples are rigorously mixed. Molecular methods and microcalorimetry are alternative options for quantification and characterization of cells on mineral surfaces and were compared in a recent study (17). Quantitative PCR assays are currently the most reliable and common method for absolute quantification of attached cells in a species-specific manner, although DNA extraction from mineral samples has specific biases, such as differential susceptibility of different cell types to cell lysis, as well as interferences of iron ions with the remaining nucleic acids. In addition, intact cells have been found on chalcopyrite and pyrite mineral grains after aggressive chemical extraction methods, such as hot cell lysis and phenol treatment (see Fig. S1 in the supplemental material) (18). Epifluorescence microscopy (EFM) can be used to study the number of attached cells, as well as the structure of the biofilm (19). However, model systems that employ polished mineral coupons are not comparable with fine-ground mineral particles. Microbial metal sulfide colonization is generally heterogeneous, as particles devoid of bacterial colonization coexist with well-colonized surfaces. This requires the analysis of a sufficiently large number of particles to take into account random variation in mineral grain colonization.

In addition to quantitative information on mineral colonization, biofilm structures can also be investigated by using computational methods. Deep neural networks are the algorithms underlying “deep learning,” a method broadly used in areas of computer vision, for instance, to analyze and classify images. Popular examples are object recognition with smartphones and self-driving cars. Several tools exist for processing microscopy images and extracting relevant biological features that include cell or nucleus counting and eukaryotic phenotype analysis. These tools are based on open-source (R, EBImage [20]; and Java, ImageJ [21, 22]) or proprietary (MatLab, CellProfiler [23] and CellClassifier [24]) programming languages. The automated image analysis allows processing of many images with large numbers of mineral grains to be analyzed, reducing time for replication and therefore allowing testing and comparisons of multiple experimental conditions in a relative manner.

Automated image analyses could provide insights into aspects of biofilm development that are not yet fully understood. As such, the temporal dynamics of mineral colonization in acidophilic bacteria are largely unknown. However, it is known that the initial colonization of metal sulfide surfaces by A. ferrooxidans, A. ferrivorans, L. ferrooxidans, and Acidiferrobacter sp. strain SPIII/3 is influenced by quorum sensing (QS) signal compounds, such as N-acyl-homoserine lactones (25, 26). Those compounds are not produced by the three strains used in this study. However, genes encoding a diffusible soluble factor (DSF) QS system have recently been described for the L. ferriphilum^T (27). DSF family QS signal compounds synchronize virulence and biofilm dispersal in Xanthomonas campestris (cis-11-methyl-dodecenoic acid, termed DSF) and Burkholderia cenocepacia (cis-2-dodecenoic acid, termed BDSF). These compounds are also known to disperse biofilms. Pronounced interspecies biofilm dispersal effects are associated with DSF family signaling (28, 29). DSF QS systems are encoded by the rpfCFGR genes in those species where RpfF is the signaling compound synthase, while the corresponding two-component signal recognition system consists of the sensor kinase RpfC and the response regulator RpfG, which act directly on cyclic diguanylate (c-di-GMP) metabolism. In addition, the DSF signal receptor proteins homologous to the RpfR protein are known to become active c-di-GMP-hydrolyzing phosphodiesterases upon binding of DSF family signals. Lowered levels of c-di-GMP are typically associated with enhanced motility and decreased expression of biofilm-related genes (30).

In this study, we used a motorized EFM for automated image acquisition (Fig. 1) coupled to automated image analysis using algorithms that allowed quantification of mineral-attached cells (Fig. 2). In addition, we used deep neural network algorithms for classification of images based on species-specific biofilm patterns in samples with low microbial diversity. This methodology provides the possibility to assess directly microbial mineral colonization laboratory bioleaching assays of metal sulfide concentrate ores. We demonstrate that the method is suitable to follow the temporal development of biofilms in model cultures of A. caldus, L. ferriphilum, and S. thermosulfidooxidans in chalcopyrite and pyrite bioleaching assays. Furthermore, biofilm dispersal upon the addition of DSF molecules to biofilms formed by L. ferriphilum and S. thermosulfidooxidans is suggested to occur.

FIG 1.

Experimental setup for automated imaging and mounting of mineral grain samples. (A) Ten-well diagnostic glass slides were used for spotting mineral samples in mounting medium. (B) Stack images were recorded using a motorized epifluorescence microscope, for calculation of extended depth of focus image projections. (C) Determination of the mineral grain area (left) and cell counting (right) is illustrated. Detected cell counts are indicated by a yellow circle for generation of a report file.

FIG 2.

Illustration of the Python image analysis algorithm for quantification of attached cells on mineral grains.

RESULTS

Automated image analysis for monitoring biofilm on mineral grains.

Specimens with mineral grains for microscopy were prepared in a manner to achieve images with a minimum mineral coverage of 70%. For the assessment of the mineral grain colonization, images were grouped into four arbitrarily chosen equally large groups of one, two, nine, 18, 36, or 72 images. This was done in order to average the naturally nonhomogeneous mineral colonization in single microscopy images over a more representative mineral surface area in multiple images. The variation of the amount of images per group showed that the coefficient of variation of the mean values of each of the four groups decreased in a linear manner from 25% ± 10% to 8% ± 2.5% when two images per group (eight images in total) or 18 images per group (72 images in total) were considered from the same biological sample (Fig. 3).

FIG 3.

Development of the coefficient of variation with the amount of analyzed images. The coefficient of variation was calculated using mean values of mineral colonization data. Colonization data of the individual images were randomly binned into four arbitrarily chosen groups. The group size was varied from colonization values derived from 1, 2, 9, 18, 36, or 72 images from a data set of 300 images from the same mineral sample condition (mixed culture of A. caldus and S. thermosulfidooxidans after 12 days of cultivation on chalcopyrite). The colonized mineral stems from a biological triplicate experiment. The coefficient of variation among the groups was calculated repetitively with randomized selection of the colonization data 25 times in order to calculate the standard deviation of the coefficient of variation.

Mineral colonization data of every sample of mineral grains were derived from analysis of at least 36 images, corresponding to a coefficient of variation not larger than 16% ± 8%. For a hypothetical average mineral coverage of 75% of each image, this consideration corresponded to an analyzed top-view mineral surface area of 4.6 mm². This can be deduced to be the minimum mineral surface area that should be analyzed for colonization assessment of pyrite or chalcopyrite concentrate particles in the size range of 50 to 100 μm in order to achieve a coefficient of variation not larger than 16% ± 8%. Figure 3 shows that a higher accuracy of the method was achieved with more analyzed images, as the coefficient of variation fell below 10% with >80 analyzed images per sample.

L. ferriphilum efficiently colonizes pyrite and chalcopyrite surfaces.

Axenic cultures of A. caldus, L. ferriphilum, and S. thermosulfidooxidans were compared regarding their ability to colonize pyrite and chalcopyrite (Fig. 4). The inocula were not previously adapted by growth on pyrite or chalcopyrite or to the presence of copper ions. L. ferriphilum significantly outperformed A. caldus and S. thermosulfidooxidans in its capacity to attach on the minerals and was estimated to have 1.5 × 10⁻⁹ ± 6.2 × 10⁻⁷ or 1.2 × 10⁻⁹ ± 5.0 × 10⁻⁷ cells · g⁻¹ in chalcopyrite or pyrite cultures, respectively, averaging the highest levels of mineral colonization on days 14 and 21. The corresponding values for A. caldus and S. thermosulfidooxidans were 4.4 × 10⁻⁸ ± 7.3 × 10⁻⁷ or 4.8 × 10⁻⁸ ± 10⁻⁸ and 3.1 × 10⁻⁸ ± 4.8 × 10⁻⁷ or 3.1 × 10⁻⁸ ± 4.5 × 10⁻⁷ cells · g⁻¹ in chalcopyrite or pyrite cultures, respectively. Student's t tests showed that the difference is statistically significant (P < 10⁻⁴) between groups made of colonization data from 72 individual images (36 images from day 14 and 36 images from day 21 samples) of each of the mineral cultures of L. ferriphilum, A. caldus, and S. thermosulfidooxidans. In addition, L. ferriphilum was most effective in dissolution of pyrite or chalcopyrite in axenic batch experiments. This was reflected by the release of iron and for chalcopyrite copper ions (Fig. 5). In the case of chalcopyrite cultures, this difference was also represented by the development of the total cell numbers (Fig. 4A). For the pyrite cultures, S. thermosulfidooxidans showed the highest total cell numbers, likely due to its ability to utilize ISCs and iron(II) ions (31). On the one hand, ISCs are not used by the obligate iron(II) oxidizer L. ferriphilum, and due to its inability to oxidize iron(II) ions, A. caldus was unable to grow on pyrite. A. caldus and S. thermosulfidooxidans formed a lesser but detectable biofilm on both minerals (Fig. 4B). However, the initial colonization of chalcopyrite by cells of S. thermosulfidooxidans was significantly lower than that by A. caldus (Fig. 4B1).

FIG 4.

L. ferriphilum efficiently colonizes chalcopyrite and pyrite surfaces. (A to C) The temporal development of total cell numbers (A), mineral-attached cell per gram of metal sulfide mineral (B), and the fraction of the mineral-attached cell population of the total cell population (C) were compared in 150-ml cultures of axenic cultures of A. caldus (white diamonds), L. ferriphilum (black triangles), or S. thermosulfidooxidans (gray circles) containing 2% chalcopyrite grains (1) or 2% pyrite grains (2) of 50- to 100-μm grain size.

FIG 5.

Dissolution of pyrite or chalcopyrite indicates microbial growth in bioleaching assays. Axenic cultures of A. caldus (A, white diamonds), L. ferriphilum (L, black triangles), and S. thermosulfidooxidans (S, gray circles) were cultivated with pyrite (1) or chalcopyrite (2 and 3), as described. The development of the total iron and copper ion concentrations is shown.

In general, the development of the mineral-attached cell fractions in axenic cultures of all three strains clearly showed mineral-dependent differences. In chalcopyrite cultures, an initial peak of 45 to 78% attached cells was followed by a rapid decline within the first 10 days of cultivation to a level of 25 to 40% for all strains (Fig. 4C1). Interestingly, this peak in the percentage fraction of mineral-attached cells was the result of the initial mineral colonization, followed by growth of the planktonic cell population rather than detachment of biofilm cells. This finding is supported by the fact that the amount of mineral-attached cells did not decrease significantly during the respective time period (Fig. 4B1). Also, the total cell numbers per assay kept rising steadily from 10⁹ cells to at least 5 × 10⁻⁹ cells during the duration of the experiment for cultures of all three strains (Fig. 4A1). In the case of pyrite, the initial peak in the fraction of attached cells was less pronounced.

The fraction of attached L. ferriphilum cells in chalcopyrite cultures remained stable at 25 to 35% after the first 5 days of incubation, even though the amount of mineral-attached cells increased from 8.3 × 10⁻⁸ ± 8.1 × 10⁻⁷ on day seven to 1.6 × 10⁻⁹ ± 7.2 × 10⁻⁷ cells · g⁻¹ chalcopyrite on day 21. A. caldus and S. thermosulfidooxidans showed a different behavior, as after the first 10 days of incubation, the amount of attached cells decreased slightly to 4.4 × 10⁻⁸ ± 1.2 × 10⁻⁸ and 2.4 × 10⁻⁸ ± 3.3 × 10⁻⁷ cells · g⁻¹ on day 21, respectively. During this time, their percent fractions of attached cells gradually decreased to circa 10% (Fig. 4C1).

In the case of pyrite bioleaching, the fraction of mineral-attached cells in cultures of L. ferriphilum averaged over the time from day five until the end of the experiment on day 21 (Fig. 4C2) was ∼40 to 60% enhanced in comparison to the levels observed in chalcopyrite assays (25 to 35%). A similar observation was made for A. caldus (60 to 70% attached cells in pyrite assays compared to circa 10 to 30% in chalcopyrite assays), while S. thermosulfidooxidans showed the lowest pyrite colonization efficiency with a fraction of 20 to 30% attached cells (10 to 30% in chalcopyrite assays).

Deep neural networks can identify characteristic biofilm patterns on chalcopyrite in axenic and mixed cultures.

Deep neural networks trained on 600 microscopy images per experimental category were used to test their performance in recognizing cell attachment patterns on chalcopyrite grains. Samples from cultures with different inoculum compositions of A. caldus (A), L. ferriphilum (L), and S. thermosulfidooxidans (S) were used as pure or mixed cultures, resulting in the following categories: A, L, S, AS, LS, and ASL. These categories represent the biofilms formed on chalcopyrite grains after 5 days of incubation. A set of 100 test images per category not included in the training set were used to test the ability of the deep neural network to assign test images to one of the training set categories. Under the restrictions that only low-species-abundance samples are considered and individual training sets are available for each of the three species in axenic and mixed cultures, the technique allows the prediction of the microbial species present within a mixed-species biofilm on chalcopyrite samples (Table 1).

TABLE 1.

Deep learning prediction of species composition of mineral-attached cell populations

Actual class	Probability (%) by predicted classa
	A	L	S	AS	LS	ALS
A	96	0	3	1	0	0
L	0	94	0	1	0	5
S	2	0	93	3	0	2
AS	0	1	2	78	14	5
LS	1	0	0	11	84	4
ALS	0	0	3	0	1	96

^aProbabilities (%) were assigned by the deep learning analysis for the similarity of the 100 test set images to the convolutional neural network (CNN) class prediction. CNNs were trained with 600 images from five-day-old mineral cultures with different inoculum compositions of A. caldus (A), L. ferriphilum (L), and S. thermosulfidooxidans (S) that were used as pure or mixed cultures, resulting in the following categories: A, L, S, AS, LS, and ASL.

Expression of the DSF family quorum sensing system in L. ferriphilum.

A DSF synthase was found encoded in the L. ferriphilum genome (Table 2) (27). Genes likely encoding DSF family signal-specific two-component systems or response regulators, suitable for DSF signal perception, were identified in the genomes of A. caldus, L. ferriphilum, and S. thermosulfidooxidans (Table 2). The genes of the L. ferriphilum DSF QS system were found to be expressed in transcriptome analyses of cells grown in continuous cultures, as well as in chalcopyrite batch cultures. The DSF synthase LFTS_0514 was especially found to have high expression levels in the planktonic cell subpopulations. Those levels strongly exceeded the average expression of gene transcripts of this species in axenic, but also in mixed, cultures with S. thermosulfidooxidans (Fig. S2).

TABLE 2.

Presence of DSF family QS system-encoding genes in A. caldus, L. ferriphilum, and S. thermosulfidooxidans genomes identified using BLASTP^a

Species (reference and/or accession no.)	DSF quorum sensing system genes
	rpfF	rpfR	rpfC	rpfG
Acidithiobacillus caldusT (59) (GCA_000175575.2)		ACAty_RS14920, ACAty_RS14615, ACAty_RS02860	ACAty_RS07245, ACAty_RS04080
Leptospirillum ferriphilumT (27) (GCA_900198525.1)	LFTS_00514	LFTS_00511	LFTS_00515, LFTS_00516	LFTS_00517
Sulfobacillus thermosulfidooxidansT (GCA_900176145.1)		Sulth_1253, Sulth_1788, Sulth_2384	Sulth_1793	Sulth_2102

^aE value (<10⁻³⁰) (48).

DSF and BDSF signal compounds inhibit iron(II) oxidation and chalcopyrite dissolution.

A strong inhibitory effect on the metabolic activity of bioleaching bacteria was observed after the external addition of DSF or BDSF. These compounds prevented oxidation of the soluble substrates iron(II) ions and tetrathionate (Fig. S3) or the insoluble substrate chalcopyrite during a cultivation period of 32 days (Fig. S4), when 5 μM DSF or BDSF signal molecules were added simultaneously with the inoculum into cultures of L. ferriphilum and S. thermosulfidooxidans (Table 3). No effect of DSF or BDSF addition on soluble substrate oxidation was observed in tetrathionate cultures of A. caldus, while growth with chalcopyrite and its dissolution were inhibited by the addition of 5 μM DSF (Table 3).

TABLE 3.

Inhibitory effect of 5 μM DSF or BDSF addition on oxidation of soluble and insoluble energy sources in cultures of A. caldus, L. ferriphilum, and S. thermosulfidooxidans

Energy source (reference figure)	Inhibition by speciesa
	A. caldus	L. ferriphilum	S. thermosulfidooxidans
Soluble [tetrathionate/iron(II) ions] (Fig. S3)	−	+	+
Insoluble (chalcopyrite) (Fig. S4)	+	+	+

^a+, no biological oxidation of soluble substrates occurred within 32 days of incubation; chalcopyrite dissolution in assays with DSF or BDSF were significantly lower than in the control assays without DSF or BDSF addition; −, no inhibition, and substrate oxidation similar to assays without DSF or BDSF addition. Tetrathionate was used for A. caldus, iron(II) ions were used for L. ferriphilum, and tetrathionate or iron(II) ions were used for S. thermosulfidooxidans. Fig. S3 and S4 substantiate the summary represented by the indicators (±) shown here by providing quantitative measurements of iron(II) ions, pH values, and planktonic cell counts [Fig. S3, soluble energy sources of iron(II)ions or tetrathionate] and total copper ions (Fig. S4, insoluble energy source of chalcopyrite).

Computational image analysis detects biofilm dispersal upon addition of DSF family signaling compounds.

Biofilm dispersal was observed in cultures of L. ferriphilum, S. thermosulfidooxidans, and their combination in mixed cultures when 5 μM DSF was added after 5 days of incubation (Fig. 6). A similar effect was noted in mixed cultures of all three species (Fig. S5). In contrast, no biofilm dispersal was observed in cultures of A. caldus (Fig. S5). However, biofilm dispersal effects were short-lived, and recolonization of the chalcopyrite occurred in the batch experiment assays within 24 h after DSF addition. The addition of DSF to mixed cultures of L. ferriphilum and S. thermosulfidooxidans (Fig. 6C) caused a marked difference in the development of the sessile cell population, which was similar to the one observed in pure cultures of L. ferriphilum (Fig. 6A). Deep-learning analysis of this mixed-species biofilm under the influence of the DSF molecule (Table 4) confirmed a relatively high similarity with the biofilm pattern of axenic L. ferriphilum cultures (33%) and the one of mixed cultures of L. ferriphilum and S. thermosulfidooxidans (38%). However, DSF molecules had no influence on the biofilm pattern classification in all the other mixed or axenic cultures. In general, biofilm patterns on chalcopyrite grains after 12 days of incubation matched well to the true species composition in axenic or mixed cultures and are therefore similar to those observed in the training set images from day five of the experiment (Table 4).

FIG 6.

DSF molecules stimulate biofilm dispersal in L. ferriphilum and S. thermosulfidooxidans. (A to C) Axenic cultures of L. ferriphilum (A), S. thermosulfidooxidans (B), and mixed cultures of L. ferriphilum and S. thermosulfidooxidans (C) were cultivated with 2% chalcopyrite. DSF (5 μM) was added after 5 days of incubation (gray triangles), and the mineral-attached cell population was compared to control experiments without DSF (white diamonds).

TABLE 4.

Deep-learning classification of biofilm patterns on chalcopyrite after 12 days of incubation and addition of 5 μM DSF

Actual class	Predicted class (+ DSF/control)a
	A	L	S	AS	LS	ALS
A	89/89	6/0	3/3	NA	NA	8/3
L	6/3	86/92	3/0	NA	6/NA	3/3
S	6/3	0/3	83/89	NA	6/NA	6/6
LS	10/8	33/3	4/0	11/NA	38/83	4/6
ALS	3/3	3/3	6/3	NA	NA/3	89/89

^aProbabilities (%) were assigned by the deep-learning analysis for the similarity of 36 images to the CNN class prediction. CNNs were trained with 600 images from five-day-old mineral cultures with different inoculum compositions of A. caldus (A), L. ferriphilum (L), and S. thermosulfidooxidans (S) that were used as pure or mixed cultures, resulting in the following categories: A, L, S, AS, LS, and ASL. NA, not analyzed.

DISCUSSION

The presented method allows the direct assessment of the relative amount of mineral-attached bacterial cells in laboratory bioleaching cultures. It avoids laborious biochemical or molecular biology sample pretreatment procedures, such as nucleic acid extraction, and their biases. The method has its main strength in performing relative comparisons rather than accurate absolute quantification of the amounts of attached cells and was tested for mineral concentrates as a proof of concept. However, we propose that the method is extendable, with some specific ad hoc parameterization for the analysis of other industrially relevant concentrates and low-grade ore preparations. These requirements include, but are not restricted to, the mineral particle size of the ore sample, which has to be sufficiently small and homogenous for enabling the visible deliberation of metal sulfide phases and gangue mineral phases using standard microscopy equipment. Adapted image analysis algorithms may have to include manual or automated differentiation of mineral phases and exclusion of gangue and autofluorescent mineral phases. Consequently, we suggest that it will be possible to employ similar techniques for assessment of the microbial colonization of metal sulfides in complex and low-grade mineral samples.

For the species-specific attachment behavior on metal sulfides, similar findings have been published (32–34), supporting the validity of our approach. The reliable, relative, and quantitative evaluation of biofilm populations is an innovative and powerful avenue for industrial and academic efforts to improve biomining operations and devise inoculation strategies of bioleaching operations.

L. ferriphilum cells have a high capacity to form biofilms on chalcopyrite and pyrite ores, and our method proved this directly in time-series studies (Fig. 4). In contrast, A. caldus cells that are unable to oxidize pyrite exhibited a low affinity to its surface in short-duration studies (13, 33), which are based on an indirect assessment of the attached cells by counting planktonic cell numbers and following their decline during initial contact with metal sulfides. However, the ostensibly high affinity of this ISC-oxidizing strain to pyrite surfaces in longer-duration axenic culture experiments presented here (Fig. 4C2) may be explained since biofilm formation is a common microbial starvation survival strategy (35). S. thermosulfidooxidans showed fewer attached cells than A. caldus within the first week of cultivation in chalcopyrite cultures (Fig. 4B1), and this may explain difficulties encountered in RNA and protein extraction from biofilm cells of chalcopyrite cultures of this species. It may indicate that the attached S. thermosulfidooxidans population on chalcopyrite did not multiply. The poor initial attachment of S. thermosulfidooxidans alongside the slow increase of the number of attached A. caldus cells on chalcopyrite compared to pyrite grains (Fig. 4B1 and B2) is possibly related to the physiological effect of inhibitory levels of copper ions. Those reached concentrations of approximately 100 mg/liter after 5 days of incubation in cultures of S. thermosulfidooxidans (Fig. 5), and even lower copper concentrations are known to inhibit biofilm formation by iron-oxidizing acidithiobacilli (10). This is supported by the characteristic difference in the development of the fraction of biofilm cells in chalcopyrite (Fig. 4C1) and pyrite (Fig. 4C2) cultures.

The strong decrease in the mineral-attached cell population in L. ferriphilum pyrite cultures measured on day 21 (Fig. 4B2) may indicate a pronounced biofilm dispersal event, since A. caldus and S. thermosulfidooxidans cells exhibited a slower and more gradual decrease in attached cells than did L. ferriphilum. This dispersal may be related to multiple factors, including the toxicity of exudates, a lowered pH, and enhanced ionic strength that are known limiting and inhibitory factors for pyrite colonization (10, 19). However, an additional explanation for the dispersal may involve a QS-related effect. Christel and coworkers (27) revealed the presence of a DSF family QS system in L. ferriphilum. Even though DSF family signaling compounds of this species are not chemically identified, bioinformatic analyses suggest a possible function of these signal molecules in AMD and bioleaching microbial communities (Table 2). The high relative expression levels of the DSF synthase (LFTS_0514) support this suggestion (Fig. S2). Fatty acids were identified in extracts of pyrite cultures of L. ferriphilum and several other leptospirilli (36). Unknown compounds in those extracts inhibited iron oxidation in several acidophilic iron oxidizers, including S. thermosulfidooxidans and A. ferrooxidans^T (36). A similar observation was made in this study with DSF family signal compounds from Xanthomonas campestris (cis-11-methyl-dodecenoic acid, DSF) and B. cenocepacia (cis-2-dodecenoic acid [BDSF]) in S. thermosulfidooxidans and L. ferriphilum (Table 3 and Fig. S3 and S4). Furthermore, DSF family molecules are known biofilm dispersal agents with pronounced interspecies effects (28, 29, 37, 38). Consequently, it is not surprising that these compounds also caused biofilm dispersal in L. ferriphilum and S. thermosulfidooxidans (Fig. 6 and S4). Even though the biofilm dispersal effects were of short duration under batch culture conditions, the implications of this observation are of great importance under environmental conditions. Here, biofilm dispersal may be ensued by a succession in attachment by other microorganisms and detachment of bioleaching microorganisms may impact the performance of heap or stirred-tank bioleaching reactors. Furthermore, if DSF molecules are produced by mineral-oxidizing bacteria, cell-cell signaling mechanisms exerting strong inhibitory and presumably also biofilm dispersal effects on competing species may provide strategies to manipulate leaching activities in target strains.

Deep learning was used to classify biofilm images from experimental conditions that were not represented in the training image sets. Based on the visual features learned during the training, the deep learning correctly inferred the bacterial composition of the biofilms composed of combinations of the three species used in this study. The high accuracy achieved in classification of biofilm images after training with convolutional neural networks (CNNs) with a reduced number of images, compared to recent successful deep-learning applications (39–41), demonstrates that deep learning represents a valid imaging-based method for the analysis of low-diversity mixed-biofilm populations (Table 1). In combination with molecular validation, we anticipate that this method may be extended as an alternative to classical molecular methods for specific applications with characteristic and low-species-abundance microbial consortia.

Deep learning applied to images from chalcopyrite grains from mixed cultures of L. ferriphilum and S. thermosulfidooxidans after the addition of 5 μM DSF molecules suggested an intermediate situation between biofilms from axenic L. ferriphilum cultures (probability, 33%) and mixed L. ferriphilum and S. thermosulfidooxidans cultures (38%, Table 1). Further indications suggested a dominance of L. ferriphilum cells in those cultures after the addition of DSF. Phase-contrast microscopy indicated mainly small, curved, motile, and rod-shaped cells characteristic of L. ferriphilum in the planktonic cell population on day 12 of this experiment. Furthermore, a similar increase in the amount of biofilm cells, as shown in Fig. 6C, was observed in axenic cultures of L. ferriphilum with or without the addition of DSF molecules (Fig. 6A). Taken together, these results suggest that DSF molecules facilitated and accelerated L. ferriphilum to dominate the mixed culture with S. thermosulfidooxidans. The presence of DSF family QS genes in both species (Table 2) suggests that a complex signal molecule interaction of L. ferriphilum and S. thermosulfidooxidans may exist in mixed cultures. In general, competition for dissolved iron(II) ions and attachment sites on metal sulfides may be directly mediated by the DSF signal compounds, which trigger degradation of the second messenger c-di-GMP (42, 43). Low levels of c-di-GMP are primarily associated with upregulation of bacterial motility genes and downregulation of genes related to bacterial biofilm formation and exopolysaccharide (EPS) production (30, 44). However, the mechanism that explains the inhibition of iron(II) oxidation by DSF family signaling compounds is not yet understood. Likewise, it remains to be demonstrated if inhibition of iron(II) oxidation in L. ferriphilum is valid also for the DSF family compounds that are hypothesized to be produced by L. ferriphilum.

Conclusion.

The presented study is a proof of concept for a direct method for relative quantification of attached cells on metal sulfides using automated image acquisition and analysis. The results highlight the effects of DSF family signal compounds in cultures of L. ferriphilum and S. thermosulfidooxidans and suggest an important role of these signal compounds in colonization of metal sulfides, microbial interactions, and niche defense among chemolithotrophic mineral-oxidizing bacteria that compete for electron donors originating from interfacial processes that determine metal sulfide dissolution.

MATERIALS AND METHODS

Microorganisms, cultivation media, and mineral cultures.

The type strains Acidithiobacillus caldus DSM 8584 (45), Leptospirillum ferriphilum DSM 14647 (46), and Sulfobacillus thermosulfidooxidans DSM 9293 (31) were cultured with Mackintosh basal salt medium (MAC) (47). The medium was autoclaved at 121°C for 20 min. Cells were grown with soluble electron donors for inoculation of mineral cultures. This approach is a realistic scenario for the production of industrial bioleaching inoculum cells. In the case of L. ferriphilum, 4 g/liter iron(II) ions (provided as FeSO₄·7H₂O) was used. Precipitation of ferric salts was prevented by the addition of sulfuric acid to maintain the pH in the range 1.6 to 1.8. A. caldus and S. thermosulfidooxidans precultures were grown using 0.9 g/liter potassium tetrathionate (K₂S₄O₆), and for S. thermosulfidooxidans, the medium was amended with 0.02% yeast extract (YE) and 0.1 g/liter iron(II) ions. Cells were harvested by centrifugation at 11,270 × g for 10 min and washed with 100 ml MAC medium. Subsequently, cells were inoculated at an initial cell density of 10⁷ cells/ml to mineral cultures in 300-ml Erlenmeyer flasks with 150 ml MAC medium and 2% (wt/vol) pyrite or chalcopyrite grains (50- to 100-μm grain size). Equal proportions of cells of each species were used in mixed cultures. All strains were cultivated on a rotary shaker at 37°C and 150 rpm. For transcriptomic analyses, L. ferriphilum was additionally grown in continuous cultures, as described previously (27). Nucleic acid and protein extractions from free-swimming planktonic cells from batch mineral cultures, mineral-attached cells, and continuous-culture iron(II)-grown planktonic cells were done using a hot phenol protocol, as previously described (18, 27). Basic local alignment search tool (BLASTP) (48) was used to identify homologous proteins of known DSF family QS systems in the genome sequences of the three species.

For testing the effects of DSF family signal compounds, cis-11-methyl-dodecenoic acid (DSF; CAS 677354-23-3; Sigma) or cis-2-dodecenoic acid (BDSF; CAS 55928-65-9; Sigma) was used. A. caldus, L. ferriphilum, and S. thermosulfidooxidans were grown as described above, with the exception that YE was omitted in S. thermosulfidooxidans cultures. DSF family signal compounds were applied at 5 μM for testing their effects on cell growth and soluble substrate oxidation. Growth was evaluated by monitoring the planktonic cell number using a Thoma counting chamber and a phase-contrast microscope, spectrophotometric measurement of iron(II) ions (49), and following the development of pH for the tetrathionate cultures. DSF was also spiked into chalcopyrite cultures at a concentration of 5 μM for testing their effects on metal sulfide colonization and oxidation in axenic and mixed cultures of A. caldus, L. ferriphilum, and S. thermosulfidooxidans. Metal sulfide dissolution was monitored by measurement of the concentration of iron(II) ions, total iron ions, and total copper ions using the spectrophotometric phenanthroline and bicinchoninic acid assays, respectively (49, 50). All experiments were done in triplicate.

Mineral preparation.

Pure mineral samples were used in this study. Museum-grade pyrite grains (Navajun, Spain) used in leaching and attachment assays were from cube crystals crushed with a disc swing-mill (HSM 100M; Herzog). Chalcopyrite grains were obtained from a flotation concentrate provided by Boliden AB (Sweden). Mineral grains were wet sieved (Retsch, Germany) in order to use the particle fraction between 50 and 100 μm. Pyrite grains were boiled for 30 min in approximately 10 volumes of 6 M HCl, washed with deionized water until the pH was neutral, and stirred twice in approximately 5 volumes of acetone for 30 min in order to remove soluble sulfur compounds (51). Chalcopyrite grains were washed twice for 30 min in 10 volumes of washing solution (0.1 M EDTA, 0.4 M NaOH), followed by treatment with acetone, as described for pyrite grains. For sterilization of mineral preparations, aliquots were sealed under a nitrogen atmosphere and incubated for 10 h at 125°C.

Microscopy sample preparation.

Mineral grain particle samples were withdrawn from mineral cultures (∼25 mg) using a flame-sterilized spatula. These particles were incubated in 1 ml MAC medium (pH 1.8) with 4% formaldehyde at room temperature for 1 h for fixation of mineral-attached cells, followed by two washing steps with water and subsequently with 1 ml phosphate-buffered saline (PBS). Samples were stored at −20°C in 50% ethanol in PBS. Mineral particles were incubated for 10 min in 200 μl of an aqueous solution of 0.01% 4′,6-diamidine-2′-phenylindole dihydrochloride (DAPI) in 2% formaldehyde. Prior to and after staining of attached cells, mineral grains were washed with 1 ml PBS. Finally, mineral particles were mounted on 10-well diagnostic glass slides (10-well, 6.7 mm; Thermo Scientific) using a glycerol-based mounting medium (CitiFluor AF2) and 22- by 50-mm cover glasses (Fig. 1A).

High-throughput epifluorescence microscopy.

Automated image acquisition was performed as illustrated in Fig. 1A and B using an AxioImager M2m (Zeiss) fluorescence microscope equipped with a motorized microscopy stage (IM SCAN 130 × 85, DC 1 mm; Märzhäuser Wetzlar) and a AxioCam MRm camera. Image acquisition used a Zeiss filter set 09 for DAPI-stained samples or bright-field mode with background illumination for visualization of the localization of opaque mineral grains and transparent regions between. Images were recorded using a Zeiss Plan-Neofluar 20×/0.50 objective. Images were recorded as stack images with 2-μm step size, covering the entire maximum grain depth of 100 μm (50 layers). The extended-focus module of the Zen 2 software (blue edition, 2011; Carl Zeiss GmbH) was used to calculate projection images using the Wavelet option. Projections were exported as JPEG files. At least 36 images were analyzed for assessment of the amount of mineral-attached cells for every mineral sample and time point.

Image analysis. (i) Cell counting and mineral grain area determination.

Cell counting was carried out computationally as illustrated in Fig. 2 by first converting the EFM images into gray-scale images and subsequently using the “Determinant of the Hessian” method (“blob_doh” function of Python's scikit-image package) with the following parameters: min_sigma = 0.34, max_sigma = 1, num_sigma = 2, threshold = 2 × 10⁻⁵, overlap = 0.1, log_scale = false. The parameters were adjusted such that the analysis was accurate for a set of test images. A full description of the parameters is found on the “Determinant of the Hessian” Python scikit-image package. The analysis was subsequently applied to the entire image set. The mineral grain area was quantified from corresponding bright-field images with background illumination that were converted into gray-scale images and by setting a threshold in the color distribution using Otsu's method (“threshold_otsu” function of Python's scikit-image package assuming a bimodal pixel distribution in color intensity [nbins = 2]).

(ii) Calculation of mineral colonization and total cell numbers.

The evaluation of the method's statistical accuracy depended on the number of images considered. Cell counts were related to the two-dimensional mineral grain area depicted in microscopy images and expressed as cells per mm⁻². After manual removal of extreme values, representing the top and bottom deciles of images with extremely low or high cell counts, metal sulfide colonization values [cells per mm⁻²] of at least 36 images were normalized for the representation of 100% mineral grain area (i.e., the true percentage mineral area of each image and the corresponding cell count value were extrapolated to a theoretical image with 100% mineral coverage). Then, the values were randomly sorted using Microsoft Excel's random function and grouped in four arbitrarily chosen classes in order to calculate the mean of each class. These four classes can be understood as four sets of equal mineral areas used for averaging of the naturally nonhomogeneous mineral colonization over a larger area than that represented in a single microscopy image. The mean of the four mean values from each group and its coefficient of variation were calculated. For estimation of the metal sulfide colonization in cells per gram, the values in cells per mm⁻² were multiplied with the specific surface area in mm² · g⁻¹ of the mineral preparations (4.2 × 10⁴ and 4.8 × 10⁴ mm² · g⁻¹) for the used pyrite and chalcopyrite concentrates, as determined by gas adsorption according to the BET (Brunauer Emmet and Taylor) theory. In order to take into account the fact that the mineral grains were viewed only from the top, the resulting values were doubled in order to account for the unobserved bottom side, while no correction factor was used for extrapolation from two-dimensional areas to the true three-dimensional mineral objects. Total cell numbers were estimated by calculation from direct counts of planktonic cells using phase-contrast microscopy with a Thoma chamber in cells per milliliter multiplied by the medium volume in milliliters plus the estimated amount of mineral-attached cells, which were determined using the image analysis method presented in this study in cells per gram multiplied by the mass of mineral in the bacterial culture in grams.

Deep learning.

CNNs are a class of neural networks used in applications known as deep learning. They have shown high efficacy in areas of computer vision, such as image recognition and classification (52–54). The open-source program CAFFE was used to perform the deep-learning analysis (55). CNNs were used to perform deep-learning analysis of EFM images, where >600 images were used for model training and 100 images for model testing. In order to train our CNNs, images from mineral cultures with different inoculum compositions of A. caldus (A), L. ferriphilum (L), and S. thermosulfidooxidans (S) were used as pure or mixed cultures, resulting in the following categories: A, L, S, AS, LS, and ASL. These categories represent the biofilms formed on chalcopyrite grains after 5 days of incubation. Then, a network model for the CAFFE framework was defined and used along with the classified data to train the CNNs. Finally, the neural network analysis was validated by processing 100 images of each test category that were not used during the neural network training phase. It was also used to classify 36 images per species composition in chalcopyrite cultures after 12 days of incubation with or without addition of 5 μM DSF on day five.

RNA isolation, sequencing, and data analysis.

Leaching cultures were separated into mineral-attached and planktonic cell subpopulations. RNA was extracted from continuous culture samples and planktonic fractions according to Christel et al. (27), while RNA from mineral-attached cells was obtained as described previously (18). The RNA was purified with the RNeasy kit (Qiagen), including DNase treatment. RNA with sufficient quality was sequenced as described previously (27). Suitable RNA samples from chalcopyrite cultures of axenic L. ferriphilum (2 samples of mineral-attached cell subpopulation), mixed cultures of L. ferriphilum and S. thermosulfidooxidans (2 samples from the attached cell population and 4 samples from the planktonic cell subpopulation) were obtained. However, the success rate using this protocol was below 50% for chalcopyrite culture mineral samples. Raw reads for those samples are available under the accession no. PRJEB27815. Previously sequenced samples [3 L. ferriphilum continuous iron(II) culture samples and 2 samples from planktonic cells from chalcopyrite cultures] can be accessed under the accession no. PRJEB21842. Transcriptomic data were processed as described previously (27). In short, the resulting sequencing reads were mapped to the L. ferriphilum (27) reference genome with bowtie2 (56) after a quality filtering step. The resulting read counts for annotated coding sequences were normalized with DESeq2 (57) using a method introduced by Klingenberg and Meinicke (58).

Accession number(s).

Raw reads for the RNA samples are available under the accession no. PRJEB27815.