Introduction

Rare earth elements (REEs) are essential for the manufacturing of modern electronics1,2,3 and sustainable energy technologies, including electric motors and wind turbine generators4, solid-state lighting5, battery anodes6, high-temperature superconductors7, and high-strength lightweight alloys8,9. All of these applications place increasing demands on the global REE supply chain10. As the world demand for sustainable energy grows11, finding a reliable and sustainable source of REEs is critical.

Current methods for refining REEs often involve harsh chemicals, high temperatures, and high pressures and generate a considerable amount of toxic waste12. These processes give sustainable energy technologies reliant on REEs a high environmental and carbon footprint. As a consequence, due to its high environmental standards, the United States has no capacity to produce purified REEs13,14.

There is growing interest in biological methods to supplement, if not completely replace, traditional REE extraction and purification methods15,16,17,18,19. Biological extraction (bioleaching) is already used to extract 5% of the world’s gold20,21 and ≈15% of the world’s copper supply20,21 (in fact, Cu biomining in Chile alone accounts for 10% of the world’s Cu supply22,23).

The performance of REE bioleaching lags behind thermochemical processes. For example, while thermochemical methods have 89–98% REE extraction efficiency from monazite ore24,25, Aspergillus species can only achieve ≈3–5%16. The acid-producing microbe Gluconobacter oxydans B58 can recover ≈50% of REEs from FCC catalysts26. However, techno-economic analysis indicates that even this extraction efficiency is still not high enough for commercial application19.

Recent efforts to improve bioleaching have focused exclusively on process optimization27. To our knowledge, no genetic approaches have yet been taken for any bioleaching microbe28. With recent advances in tools for reading and writing genomes, genetic engineering is an attractive solution for enhancing bioleaching. However, applying these tools to non-model microorganisms such as G. oxydans can be a significant challenge9. While there have been some promising advances in editing the G. oxydans genome29,30,31,32,33,34, we do not yet know where to edit.

In the presence of glucose, G. oxydans secretes a biolixiviant rich in gluconic acid26. This is produced by periplasmic glucose oxidation by the pyrroloquinoline quinone (PQQ)-dependent membrane-bound glucose dehydrogenase (mGDH)35. The final pH of the biolixiviant is a major factor in REE bioleaching26. However, gluconic acid alone fails to explain bioleaching by G. oxydans: pure gluconic acid is far less effective at bioleaching than the biolixiviant produced by G. oxydans26. This means that even the most successful efforts to upregulate mGDH activity and gluconic acid production are unlikely to take full advantage of G. oxydans’ biolixiviant production capabilities.

In this work, we characterize the genome of G. oxydans and identify a comprehensive set of genes underlying its bioleaching capabilities through the construction of a carefully curated whole-genome knockout collection of single-gene transposon disruption mutants using Knockout Sudoku36,37 (Fig. 1). As the final pH of the biolixiviant is a good predictor of bioleaching efficiency26; we use acidification as a proxy for bioleaching potential and screen the collection for mutants that differ in their ability to produce acidic biolixiviant (Figs. 2 and 3). Finally, we demonstrate that a single gene disruption - only one of several potential enhancement strategies—can already significantly improve G. oxydans’ bioleaching capabilities (Fig. 4).

Fig. 1: Knockout Sudoku was used to curate a saturating coverage transposon insertion mutant collection for Gluconobacter oxydans B58.
figure 1

A The G. oxydans B58 genome contains 3283 genes. Two thousand five hundred and seventy genes were fully annotated with a BLAST hit, InterPro ID, and gene ontology (GO) group. An additional 163 genes had an annotation and GO group but lacked an InterPro ID; 399 retrieved only a BLAST hit, and 150 were unable to be assigned any annotation. B A Monte Carlo (MC) estimate (green curve) of the number of genes represented by at least one mutant as a function of the number of mutants collected demonstrated that picking 25,000 mutants would yield at least one disruption for 95% of genes, while picking 50,000 mutants would yield at least one disruption for 99% of genes. In total, we picked 49,256 single-gene disruption mutants and located at least one disruption for 2733 genes. A Monte Carlo simulation (blue curve) of picking with random drawing from the sequenced progenitor collection (PC) without replacements demonstrates that the genome coverage was truly saturated. The center of each curve is the mean value of the unique gene disruption count from 1000 simulations, while the upper and lower parts of each curve represent two standard deviations around this mean. C A one-sided Fisher’s exact test for gene ontology enrichment among the non-disrupted (putatively essential) genes revealed significant enrichment (p < 0.05, yellow line) of genes involved in translation and other ribosome-related functions. D The curated condensed collection (CC) contains 17,706 isolated colonies across 185 plates. High-throughput sequencing of the CC confirmed the location for 4419 unique disruption strains, representing disruptions in 2556 genes. Hundred and seventy-seven genes located in the PC were not located in the CC. No disruption mutant was detected in 550 genes.

Full size image
Fig. 2: High-throughput pH screens of the G. oxydans whole genome knockout collection were used to identify genes that control REE bioleaching.
figure 2

A Thymol blue (TB) was used to measure the endpoint acidity of biolixiviant produced by each well of the condensed collection. The ratio of TB absorbance (A) at 435 and 545 nm is linearly related to pH between 2 and 3.4 (Supplementary Fig. 1). CC plate 65 contains biolixiviant produced by δpstB strain in wells F7 and G7 (arrowhead), whose absorbance at 435 and 545 nm are shown (middle panel, red dots), along with the average absorbance of all occupied wells on the plate (n = 94) (white dot; error bars are SEM). The dashed line represents a typical absorbance spectrum for WT-produced biolixiviant. The A435/A545 ratio for these two wells (right panel, red bars and dots) compared with the mean ratio of the plate (orange bar; orange dots show individual data points) is well below the lower bound (LB) for the plate, indicating that δpstB produces a much more acidic biolixiviant than the average strain. B Bromophenol blue (BPB) was used to measure rate of change in pH at the onset of glucose conversion to organic acids. Rate was measured over a 6 min period within five minutes of adding bacteria to a glucose and BPB solution. Condensed collection (CC) plate 162 contains the δtldE strain in wells F11–C12 (arrowheads), whose changes in absorbance over time are graphed along with the average for that plate (middle panel, n = 94). A comparison of the normalized rate over OD for each well (right panel, dark blue bars and dots) versus the plate mean (light blue bar; dots show individual data points) shows how V/OD for these wells was below the lower bound for CC plate 162. C All 185 plates of the CC were screened for acidification using TB and BPB assays. Hits from both screens were verified in comparison with proxy WT strains. In total, 176 disruption strains were shown to significantly contribute to acidification. D, E Comparisons with proxy WT strains were made by a two-tailed t-test with a Bonferroni-corrected alpha (α = 0.05/N where N is the number of comparisons). Bars represent mean values and dots represent individual data points. Error bars represent standard deviation. D The 25 largest significant reductions in biolixiviant pH and 50 largest significant increases in biolixiviant pH. N = 120 or N = 242 for comparisons with pWT set A or set B, respectively. (The full data set and number of biological replicates for each strain can be found in Supplementary Data 6E). E All significant changes in acidification rate, N = 60 (The full data set and number of biological replicates for each strain can be found in Supplementary Data 6F).

Full size image
Fig. 3: Genes involved in phosphate signaling, carbohydrate metabolism and PQQ synthesis were significantly overrepresented in the significant hits from high-throughput screens of acidification by G. oxydans.
figure 3

A one-sided Fisher’s Exact Test was used to test for gene ontology enrichment (p < 0.05, yellow dashed line). Numbers at the base of bars are how many genes from the significant hits are from that gene ontology (GO), out of the total in the genome (in parentheses). Genes selected for further analysis of endpoint pH and bioleaching (Fig. 4) that contribute to an enriched GO are listed above the bars. A, B Enriched GO terms among genes that decreased and increased the end point pH. C, D Enriched GO terms among genes that increased and decreased the initial acidification rate. FBP fructose-bisphosphate, GDP-Man:DolP dolichyl-phosphate beta-d-mannosyltransferase, GGT glutathione hydrolase, G6P glucose 6-phosphate, HTA homoserine O-acetyltransferase, DD-transepeptidase D-Ala-D-Ala carboxypeptidase, HAG hydroxyacylglutathione, Membr membrane, Moco Mo-molybdopterin cofactor, MS monosaccharide, MT mannosyltransferase, M6P mannose-6-phosphate, Pi inorganic phosphate, PLP pyridoxal phosphate, PQQ pyrroloquinoline quinone, PSK phosphorelay sensor kinase, Q queuosine, RNase H DNA–RNA hybrid ribonuclease, SAM S-adenosyl-l-methionine, TPP thiamine pyrophosphate, TOP1 topoisomerase type 1, HK histidine kinase, UDP-G uracil-diphosphate glucose, 6-PGL 6-phosphogluconolactonase.

Full size image
Fig. 4: Increased acidification strains of G. oxydans B58 are able to increase rare earth extraction from retorted phosphor powder (RPP).
figure 4

A, B A subset of 22 disruption strains was tested for acidification with direct pH measurement. Gray circles represent the mean pH for each strain and comparisons were made with wild type by two-tailed t-test with Bonferonni correction (N = 22). pH measurements significantly different from pWT are labeled with asterisks: *p < 0.05/22; **p < 0.01/22; ***p < 0.001/22 (n = 5 biological replicates, df = 18). Error bars represent standard deviation. Individual data points are shown as black dots. C, D Ten disruption strains with the lowest final biolixiviant pH and four with the highest were tested for RPP bioleaching capabilities. Outer gray bars represent the mean total REE extracted for each biolixiviant. Inner multicolored bars represent fractional contributions of each REE. Error bars represent standard error around the mean total REE extracted. Percentages are total REE extraction efficiency (based on previously published REE amounts in the RPP26). Individual data points for total REE extracted are shown as white dots. C Comparison with a two-tailed t-test between each mutant and pWT demonstrated that eight strains were significantly better or worse at bioleaching total REEs (+p < 0.05; n = 5 biological replicates, df = 18). With a Bonferroni correction (N = 12), only one was significantly better (**p = 5.30 × 10−4 for δpstC), but two of the higher pH biolixiviants that extracted detectable REEs were significantly attenuated in bioleaching capability (***p = 1.71 × 10−11 and 6.32 × 10−14 for δtldE and δtldD, respectively). Light purple shading represents the mean and standard deviation for total REE bioleached by pWT biolixiviant. D Disruption mutants for mgdh and pqqC are only able to extract less than 1% of the REE that wild-type G. oxydans can but still extract significantly more REE than glucose alone when measured at a lesser dilution (two-tailed t-test with Bonferroni correction, N = 2: ***p = 9.71 × 10−10 and 2.47 × 10−8 for δpqqC and δmgdh, respectively, n = 5 biological replicates). E Total REE extraction is plotted versus pH for each replicate for each strain, demonstrating how the two are linearly related. Each data point is color coded by strain, as indicated in the plot legend.

Full size image

Results

Development of a knockout collection for G. oxydans

We built a saturating coverage transposon insertion mutant collection for G. oxydans B58 and cataloged and condensed it with the Knockout Sudoku combinatorial pooling method36,37 (Fig. 1). We sequenced the G. oxydans B58 genome and identified 3283 open reading frames (Fig. 1A and Supplementary Data 1; see “Methods” section). Following the recommendation of Monte Carlo simulations, we collected 49,256 transposon insertion mutants (the progenitor collection; PC) to ensure saturating coverage of the G. oxydans B58 genome (Fig. 1B; see “Methods” section).

Progenitor collection sequencing results indicate that we were able to generate at least one disruption mutant for almost every nonessential gene in the G. oxydans genome. In total, we identified disruption strains for 2733 genes out of the 3283 genes in the G. oxydans B58 genome. Since every predicted gene contains at least seven AT or TA transposon insertion sites (Supplementary Data 1), the remaining 550 non-disrupted genes are likely to be essential. A Fisher’s Exact Test for gene ontology (GO) enrichment representing 268 of the non-disrupted genes demonstrated significant enrichment (p < 0.05) in several essential ontologies, with the greatest enrichment in those relating to the ribosome and translation (Fig. 1C and Supplementary Data 2A).

The progenitor collection catalog was used to create a condensed G. oxydans disruption collection with at least one representative per nonessential gene. Forty-seven progenitor strains were verified by Sanger sequencing prior to condensing, of which 43 (92%) were confirmed to have the predicted transposon coordinate (Supplementary Data 3A). We selected one mutant for all 2733 disrupted genes, a second mutant for 2354 genes, and a third mutant for 50 genes where mutant location information was poor. All mutants were transferred for isolation of single colonies, and 2–10 colonies per mutant were picked, depending on the predicted number of cross-contaminating disruption strains in the originating well. This condensed collection contains 17,706 mutants in 185 96-well plates (Supplementary Data 4).

The condensed collection catalog was validated by a second round of combinatorial pooling and sequencing. Of the 17,706 wells in the condensed collection, we were able to confirm the identity of 15,257 (Supplementary Data 4). We confirmed 25 of these wells by Sanger sequencing, and 100% had the predicted transposon coordinates (Supplementary Data 3B). Among these wells, we were able to verify the identity and location of 4419 independent transposon insertion sites, representing a disruption mutant for 2556 unique genes (Fig. 1D). One thousand five hundred and eighty seven genes are represented by more than one disruption, and 3317 of all disruptions occur in the first half of the gene (Supplementary Data 4).

Genome-wide screening discovers genes linked to acid production

We screened the G. oxydans B58 whole genome knockout collection for disruption mutants with differential acidification capability (Fig. 2). We used the colorimetric pH-sensitive dye thymol blue (TB) to screen for changes in final biolixiviant pH (Fig. 2A and S1) and bromophenol blue (BPB) to screen for changes in the rate of acidification (Fig. 2B).

In total, we noted 304 genes that apparently controlled acidification (Fig. 2C and Supplementary Data 5). The TB screen discovered 282 genes whose disruption led to a differential change in biolixiviant acidity (Fig. 2C and Supplementary Data 5). Forty-seven mutants produced a more acidic biolixiviant, while 235 produced a less acidic biolixiviant (Fig. 2C and Supplementary Data 5). The BPB screen identified 82 gene disruptions with differential rates of acidification: 49 with a faster rate and 33 with a slower rate. Sixty mutants were identified by both screens (Fig. 2C and Supplementary Data 5).

Overall, we identified 165 genes that significantly (p < 0.05) changed the final biolixiviant pH, rate of acidification, or both but did not change the growth rate (Fig. 2C and Supplementary Data 6). We re-arrayed disruption strains with differential acidification into new 96-well plates alongside proxy wild-type (pWT) strains (see “Methods” section) that have a transposon insertion in an intergenic region and show nondifferential growth or biolixiviant production (Supplementary Fig. 2). The new collection was re-assayed with TB and BPB assays, and the strength and significance of each result were determined by comparison with pWT through a Bonferroni-corrected t-test (see “Methods” section). Mutants that cause the 25 largest reductions and 50 largest increases in endpoint acidity yet do not affect the growth rate are shown in Fig. 2D. A full set of mutants that cause significant changes in acidification are listed in Supplementary Data 6.

Thirty-one mutants that caused significant changes in the acidification rate without changing the growth rate are shown in Fig. 2E and Supplementary Data 6F. However, 14 of the faster strains, including δGO_868, a disruption of a LacI-type transcriptional repressor, which was the fastest strain, produced a less acidic biolixiviant than the wild-type, indicating that targeting these genes for engineering a faster acidifier would likely be at the expense of a more acidic biolixiviant. None of the strains with a faster rate of acidification also created a more acidic biolixiviant. This result suggests that multiple genetic engineering interventions will be needed to construct a strain of G. oxydans that simultaneously produces a more acidic biolixiviant than the wild-type at a faster initial rate.

Pi transport and PQQ synthesis are main controllers of acidification

We used Fisher’s exact test to determine which biological processes, metabolic functions, and cellular components are enriched among the gene disruptions that significantly changed biolixiviant acidification (Fig. 3). Among the disrupted genes that led to a stronger acidity, the most significant enrichment for all three GO categories involves the phosphate-specific transport system, represented by pstA, pstB, pstC, pstS, and phoR (Fig. 3A). Other enriched ontologies include those related to phosphate signaling and binding.

Among the disrupted genes that led to weaker acidity, several enriched GO terms are related to the synthesis or use of the redox cofactor PQQ, represented by pqqB, pqqC, pqqE, tldD, and mgdh (Fig. 3B). Other enriched ontologies include those related to carbohydrate metabolism, represented by five enzymes related to fructose metabolism: Fructokinase (GO_1072), GDP-mannose 4,6-dehydratase (GO_1441), Mannose-1-phosphate guanylyltransferase/Mannose-6-phosphate isomerase (GO_182), UTP--glucose-1-phosphate uridylyltransferase /Phosphomannomutase (GO_1923), and UDP-glucose 4-epimerase (GO_36) (Supplementary Data 2C).

Acidification rate is controlled by carbohydrate metabolism and respiration. Disruptions in the pentose phosphate pathway and related metabolic processes increased the acidification rate (Fig. 3C). Meanwhile, disruptions of the electron transport pathway components are the most significantly enriched group of mutants that decreased the acidification rate (Fig. 3D).

Dye assay results are validated by direct pH measurements

We selected 22 strains from those with the greatest increase or decrease in acidification (Fig. 2D, E) for further testing. Dye pH measurements were validated by direct pH measurements. Eleven of the 13 strains that produced significantly lower acidity biolixiviants in TB assays did the same in direct pH measurements (Fig. 4A). The most acidic biolixiviant was produced by a disruption in the phosphate transport gene δpstC at pH 2.09. In fact, 4 of the 11 mutant strains that produced a more acidic biolixiviant were disrupted in genes involved in the phosphate-specific transport system (δpstA, δpstB, δpstC, and δpstS).

Additional disruptions that led to a more acidic biolixiviant included those in a hypothetical protein with no similarity to anything previously characterized (δGO_1415); a gamma-glutamyltranspeptidase (δggt1); a periplasmic chaperone (δsurA); an HTH-type transcriptional regulator (δpetP); a two-component system sensor histidine kinase (δykoH); a pyridoxal 5-phosphate (PLP)-dependent ornithine decarboxylase (δspeC); and a TPR domain protein that is a putative component of the TonB iron uptake system (δtonB).

Nine of the tested strains produced biolixiviants significantly higher in pH than pWT (Fig. 4B). The most alkaline biolixiviant was produced by a disruption in the PQQ synthesis system, δpqqC, at pH 4.71. While δpqqC produced very little acid, this is below the pH of glucose in media alone, indicating that some bacterial acidification still occurred. In fact, three of the nine mutants that produce biolixiviants with reduced acidity either synthesize PQQ (δpqqC and δtldD) or use it as a cofactor (δmgdh).

Additional disruptions that led to a more alkaline biolixiviant than pWT include a fructose-bisphosphate aldolase class II (δGO_3252); a GTP and nucleic acid binding protein (δychF); a lipid A biosynthesis protein (δhtrB); a peptide chain release factor (δhemK); the LacI-type transcriptional repressor that increases the initial acidification rate (δGO_868); and the second component of a proteolytic complex with TldD (δtldE).

Disrupting the phosphate transport system increases bioleaching

We tested whether ten of the mutants that produced a more acidic biolixiviant could bioleach REE from retorted phosphor powder (RPP) from spent fluorescent lightbulbs more efficiently than pWT (Fig. 4C). For each mutant, the elemental composition of REE leachate was similar to that previously reported26. Six of these mutants significantly increased bioleaching. Two of the better bioleaching mutants disrupted the pst phosphate transport system (δpstC and δpstB). Overall, we found that bioleaching efficiency correlates with biolixiviant pH, as expected (Fig. 4E).

The δpstC mutant produced the most acidic biolixivant and extracted the most REE from RPP: 5.5% total extraction efficiency compared with pWT (4.7%). In other words, δpstC removed 18% more REEs from RPP than pWT. This increase in REE extraction remains significant even after adjusting the α to account for the possibility of a difference due to chance alone (Bonferroni correction, see “Methods” section). Without the adjustment, six of the better acidifiers were also better bioleachers than pWT (Fig. 4C). The remaining better bioleachers increased REE extraction by between 11% (δspeC) and 18% (δggt1) (Fig. 4C).

We speculate that disrupting phosphate transport and signaling derepresses acid production in G. oxydans. Six of the disruption strains that resulted in a lower biolixiviant pH (δpstC, δpstB, δggt1, δpstA, δpstS, and δykoH), including three that increased bioleaching (δpstC, δpstB, δggt1), along with many more identified by acidification high-throughput assays (e.g., δphoR, δenvZ), are involved in phosphate transport, sensing and signaling.

In its natural environment, G. oxydans produces biolixiviants to liberate phosphate from minerals, not metals38,39,40. Under phosphate-limiting conditions, the PstSCAB phosphate transporter will activate the histidine kinase PhoR, which in turn phosphorylates the transcription factor PhoB and activates the pho regulon, enabling phosphate solubilization and uptake41. Under sufficient phosphate conditions, PhoB is deactivated by PhoR, which in turn inhibits the expression of genes involved in the phosphate-starvation response. We speculate that by disrupting pstSCAB or PhoR, we prevent G. oxydans from sensing when there is adequate phosphate in its environment and when to stop producing biolixiviants.

Disrupting mgdh and PQQ synthesis genes decreases bioleaching

We also tested REE extraction by four mutants that produce a less acidic biolixiviant than pWT, and all were worse bioleachers than pWT (Fig. 4C, D). Unsurprisingly, the δmgdh mutant was the worst bioleacher of all tested mutants, considering its lack of gluconic acid production35. δmgdh reduced bioleaching by 97%. Disruption mutants that knocked out the synthesis of mGDH’s essential redox cofactor, PQQ, also produced significant reductions in biolixiviant acidity. δpqqC reduced bioleaching by ≈94%. While bioleaching by δmgdh and δpqqC was negligible compared to pWT, they were able to bioleach a statistically significant amount of REE compared to glucose alone. This indicates, as previously speculated26, that a bioleaching mechanism independent of mGDH exists in G. oxydans (Fig. 4D).

Disruption mutants in tldD and tldE were also much worse at bioleaching than pWT. δtldD reduces bioleaching by 92%, while δtldE reduces it by 63% (Fig. 4C). We speculate that TldD and TldE contribute to the supply of the PQQ cofactor to mGDH. δtldD strongly attenuates acid production (Fig. 4B), and the gene has already been implicated in PQQ synthesis in G. oxydans 621H42. In E. coli, TldD and TldE form a two-component protease for the final cleavage step in the processing of the peptide antibiotic microcin B1743. In a similar manner, PqqF and PqqG from Methylorubrum extorquens form a protease that rapidly cleaves PqqA, the peptide precursor to PQQ44. We speculate that TldD in G. oxydans plays the same role as PqqF from M. extorquens, while TldE plays the same role as PqqG. Deletion of pqqF in M. extorquens completely inhibits cleavage of PqqA, while we found that disruption of tldD in G. oxydans reduces REE bioleaching by 92%. Moreover, deletion of pqqG in M. extorquens only reduces PqqA cleavage by 50%44, while disruption of tldE only reduces REE bioleaching by 63%. These parallels strongly indicate a role for TldE in the biosynthesis of PQQ in G. oxydans.

Discussion

Bioleaching has the potential to revolutionize the environmental impact of REE production and dramatically increase access to these critical ingredients for sustainable energy technology. However, making REE bioleaching cost competitive with thermochemical methods will require increasing both the rate and completeness (overall efficiency) of REE extraction. This work provides a roadmap for improving bioleaching by genetic engineering.

By constructing a whole genome knockout collection for G. oxydans, one of the most promising organisms for REE bioleaching, we are able to characterize the genetics of this process with high sensitivity and high completeness. In total, we identified 165 gene disruption mutants that significantly changed the acidity of its biolixiviant, rate of production, or both. The regulatory elements of each of these genes represent dials that can be turned to improve bioleaching.

As well as producing gluconic acid, G. oxydans can be used in the production of industrially important products, including 2,5-diketogluconic acid, a precursor of vitamin C;45; 5-ketogluconic acid, a precursor of l(+)-tartaric acid used in the production of food and pharmaceuticals45, vinegar, sorbitol, and dihydroxyacetone45. The G. oxydans knockout collection will allow characterization of the production of these metabolites and improvement of their production.

REE bioleaching by G. oxydans is predominantly controlled by two well-characterized systems: phosphate signaling and incomplete glucose oxidation, which is supported by the production of the redox cofactor PQQ. Interrupting phosphate signaling control of biolixiviant production by disrupting a single gene (pstC) can increase REE extraction by 18%. Disrupting the supply of the PQQ cofactor to the membrane-bound glucose dehydrogenase reduces REE extraction by up to 94%.

Among the many genes whose disruption attenuated final biolixiviant acidity, several were related to fructose metabolism. G. oxydans converts mannitol (the carbon source used for growth in these studies) into fructose, which can be incorporated into the pentose phosphate and Enter-Douderoff pathways, the latter of which can result in accumulation of acetate in the growth media46. As these five disruptions did not affect growth on mannitol, the genes are likely dispensable for fructose catabolism, but may be important for acetate accumulation and some initial media acidification during growth on mannitol. Conversely, disruptions in other genes related to carbohydrate metabolism conferred a faster rate of acidification, most likely due to the disruption of glucose catabolism/incorporation into the pentose phosphate pathway and a preference for incomplete oxidation and gluconic acid production.

Comprehensive screening of the G. oxydans genome also revealed targets that could not have been predicted to contribute to REE bioleaching. For example, disrupting GO_1415, encoding a protein of unknown function, increases REE bioleaching by 15%. Additionally, our results highlight the potential for a previously uncharacterized role of TldE in PQQ synthesis. Several periplasmic dehydrogenases in G. oxydans depend on PQQ for their function47, and in the case of the d-sorbitol dehydrogenase, mSLDH, overexpression of the pqq synthase genes and tldD enhances conversion of N-2-hydroxyethyl-glucamine into 6-(N-hydroxyethyl)-amino-6-deoxy-l-sorbofuranose (6NSL), a precursor to the diabetes drug Miglitol48,49. The discovery of the potential contribution of TldE to PQQ biosynthesis may allow for exceptional enhancement of cofactor production through the additional overexpression of this gene and a consequent uptick in dehydrogenase activity. PQQ is an essential cofactor important for several other industrial applications of G. oxydans, including the production of l-sorbose50 and 5-keto-d-gluconate51. Furthermore, PQQ alone has many applications across many biological processes, from plant protection to neuron regeneration52.

Our results demonstrate the potential for improving bioleaching through genetic engineering. Furthermore, the creation of a whole-genome knockout collection in G. oxydans will facilitate its use as a model species for further studies in REE bioleaching and other industrially important applications of similar acetic acid bacteria. The findings of the two major systems contributing to acidification in G. oxydans suggest the first steps in the roadmap for greatly improving bioleaching: take the brakes off regulation of acid production by disabling the phosphate-specific transport system, while overexpressing mgdh along with the expanded synthesis pathway for its cofactor PQQ.

Methods

Gluconobacter oxydans B58 genome sequencing

Gluconobacter oxydans strain NRRL B-58 (GoB58) was obtained from the American Type Culture Collection (ATTC), Manassas, VA. In all experiments, GoB58 was cultured in yeast peptone mannitol media [YPM; 5 g L−1 yeast extract (C7341, Hardy Diagnostics, Santa Maria, CA), 3 g L−1 peptone (211677, BD, Franklin Lakes, NJ), 25 g L−1 mannitol (BDH9248, VWR Chemicals, Radnor, PA)] with or without antibiotic, as specified.

Genomic DNA was extracted from saturated culture using a Quick-DNA Miniprep kit from Zymo Research (Part number D3024, Irvine, CA). A genomic DNA library was prepared and sequenced using a TruSeq DNA PCR-Free Library Prep Kit (Illumina, San Diego, CA).

The prepared library was sequenced on a MiSeq Nano (Illumina, San Diego, CA, USA) with a 500 bp kit at the Cornell University Institute of Biotechnology (Ithaca, NY, USA). The resulting paired-end reads were trimmed using Trimmomatic53 and assembled with SPAdes using k-mer sizes 21, 33, 55, 77, 99, and 127 and an auto coverage cutoff54. Assembly quality was checked with QUAST55, and genome completeness was verified with BUSCO56 using the proteobacteria_odb9 database (available for download at https://datacommons.cyverse.org/browse/iplant/home/shared/iplantcollaborative/example_data/BUSCO.sample.data/proteobacteria_odb9/ancestral) for comparison. The resulting 62 contigs were annotated online using RAST (https://rast.nmpdr.org)57,58,59.

Gene ontology enrichment

DIAMOND60 was used to assign annotated protein models with the closest BLAST hit using the Uniref90 database (available for download at https://www.uniprot.org/uniref/), an E-value threshold of 10−10, and a block size of 10. InterProScan61 (version 5.50–84.0) was used to assign family and domain information to protein models.

Output from both of these searches was used to assign gene ontologies with BLAST2GO62. Gene set enrichment analysis was performed with the BioConductor topGO package63 using the default weight algorithm, the TopGO Fisher test, with a p-value threshold of 0.05.

Mating for transposon insertional mutagenesis

The transposon insertion plasmid pMiniHimarFRT36 was delivered to GoB58 by conjugation with E. coli WM3064. E. coli WM3064 transformed with pMiniHimarFRT was grown overnight to saturation in 50 mL LB (10 g L−1 tryptone, 5 g L−1 yeast extract, and 10 g L−1 NaCl) supplemented with 50 µg mL−1 kanamycin (kan) and 90 µM diaminopimelic acid (DAP; D1377, Sigma-Aldrich, St. Louis, MO), rinsed once with 50 mL LB, and then resuspended in 20 mL YPM.

GoB58 was grown for approximately 24 h in YPM, diluted to an optical density (OD) of 0.05 in 750 mL YPM and incubated at 30 °C for two doublings until the OD reached 0.2. GoB58 culture was distributed into 13 50 mL conical tubes, to which rinsed and resuspended WM3064 was added at a ratio of 1:1 by density (approximately 1 mL WM3064 to 50 mL B58). Bacteria were mixed by inversion and then centrifuged at 1900×g for 5 min. The supernatant was poured off, and the mixture was resuspended in the remaining liquid (≈0.5 mL), dispensed onto a YPM agar petri dish in five spots of 0.1 mL, and allowed to dry on the bench under a flame.

Mating plates were incubated at 30 °C for 24 h. Mating spots were collected by adding 4 mL YPM to a plate, scraping the spots into the liquid, and then suspending bacteria by drawing the liquid up and down several times with a pipette. Suspended cells were collected from each plate, and the suspension was distributed onto petri dishes with YPM agar supplemented with 100 µg mL−1 kanamycin at 100 µL per dish.

After 3 days of incubation at 30 °C, colonies were transferred into 96-well microplates using a CP7200 colony picking robot (Norgren Systems, Ronceverte WV, USA). Each well contained 150 µL YPM with 100 µg mL−1 kanamycin. For all high-throughput experiments, GoB58 was grown in polypropylene microplates sealed with a sterile porous membrane (Aeraseal, Catalog Number BS-25, Excel Scientific) and incubated at 30 °C with shaking at 800 rpm. Isolated disruption strains were grown for three days to allow cultures in all wells to reach saturated growth. Wells B2 and E7 of each plate were reserved as no-bacteria controls.

A Monte Carlo numerical simulation (collectionmc36) was used to approximate how many insertions would need to occur before a mutant is found representing a knockout of each gene in the genome, which demonstrated that approximately 55,000 mutants would need to be generated and selected to identify mutants in at least 99% of all GoB58 genes (Fig. 1B).

In total, 18 matings were required to recover and transfer a progenitor collection of 49,256 disruption strains into 525 microplates over the course of approximately two months. Microplates with saturated cultures were maintained at 4 °C for up to 3 weeks and incubated an additional night at 30 °C before pooling.

The progenitor collection was cataloged using the Knockout Sudoku combinatorial pooling method. Combinatorial pooling was done in three batches. The 525 plates were virtually arranged in a 20 by 27 grid, and creation of the progenitor collection catalog by combinatorial pooling, cryopreservation, pool amplicon library generation, and massively parallel sequencing were all performed using the protocol described by Anzai et al.36,37.

Curation of a whole-genome knockout collection

The condensed G. oxydans knockout collection was curated by manual curation of the progenitor collection catalog (as opposed to automatic curation in the Knockout Sudoku protocol36,37). To create a condensed collection, a disruption strain was chosen for each of the 2733 disrupted genes available in the progenitor collection, first prioritizing close proximity to the translation start and then the total probability of the proposed progenitor collection address. A second strain was chosen from the remaining strains for each gene that had another available. For 50 genes, the locations for both disruption strains selected were ambiguous, and thus, a third strain was selected from the remaining collection.

In total, 5137 disruption strains were transferred for isolation of single colonies. Many progenitor wells were predicted to have more than one possible strain per well, so for each strain, the number of colonies isolated was two times the predicted number of strains in the progenitor well, up to ten. The condensed collection, which amounted to 17,706 wells, was pooled, sequenced, and validated using the validation procedures detailed in the Knockout Sudoku protocol36,37. Unknown disruption strains significantly linked to acidification were identified with Sanger sequencing, also as previously described, with the exception of transposon-specific primers. For the first and second rounds of nested PCR, the transposon-specific primers were (5ʹ-GTATCGCCGCTCCCG-3ʹ, and (5ʹ-CATCGCCTTCTATCGCCTTC-3ʹ), respectively.

Thymol blue endpoint acidity assay

Endpoint acidity was measured using the pH indicator thymol blue (TB; SKU114545, Sigma-Aldrich, St. Louis, MO), which changes from red to yellow below a pH of 2.8 (https://www.sigmaaldrich.com/US/en/product/sial/114545). The lowest pH of biolixiviant generated by GoB58 was 2.326; thus, TB allows for distinguishing strains that lower the pH below that of the wild-type biolixiviant. To generate biolixiviants, the condensed collection was replicated by transferring on pins from cryopreserved culture plates into new growth plates containing 100 µL YPM with 100 µg mL−1 kanamycin per well. After two days of growth, an equal volume of 40% w/v glucose was added to saturated cultures for a final solution of 20% w/v glucose. All dispensing into microplates was done with a 96-channel pipette. The amount of glucose needed to lower the pH below 2.3 via the production of gluconic acid was estimated to be 13% w/v, but the higher concentration was used to account for any use of glucose as a carbon source and still maintain an excess amount. Viability tests demonstrated that the bacteria were still viable after two days of culture in such a solution.

Bacteria were incubated with glucose for 48 h to allow acid production to reach completion. Plates were then centrifuged for 3 min at 3200×g (top speed), and 90 µL of the biolixiviant supernatant was removed and added to 10 µL TB at a final concentration of 40 µg mL−1. After 1 min of mixing by vortexing, absorbance was measured for each well at 435 and 545 nm on a Synergy 2 plate reader (Biotek Instruments, Winooski, VT, USA). Because of the variation in background absorbance from well to well on each plate, absorbance was measured at these two wavelengths, and their ratio was used as a proxy for pH, which correlates linearly within the range of pH for the majority of biolixiviants produced by the collection (Supplementary Fig. 1).

Bromophenol blue acidification rate screen

Acidification rate was measured using the pH-indicating dye bromophenol blue (BPB; 161-0404, Bio-Rad, Richmond, CA). Knockout collection strains were grown for two days. OD was measured at 590 nm for each well, and then 5 µL of culture was transferred to a polystyrene assay plate containing 95 µL of 2% w/v glucose and 20 µg mL−1 BPB in deionized water per well. The initial pH of the culture is just above 5, and within moments of adding culture to glucose with BPB, the color begins to change rapidly. Assay plates were mixed by vortexing for one minute after the addition of bacterial culture and then immediately transferred to a plate reader, where the change in color was tracked by measuring absorbance at 600 nm every minute for 6 min, resulting in seven reads. Mean rate (V) and R-squared were calculated by the Gen5 microplate reader and imager software (Biotek Instruments). A plot of all V relative to OD demonstrated that the two are correlated; thus, V was normalized to OD for each well (Supplementary Fig. 3).

Hit identification in acidification end point and rate screens

Once every well had its assigned data point (A435/A545 for TB and V/OD for BPB), hits were determined by first identifying outliers for each plate. The interquartile range and upper and lower bounds were calculated in Microsoft Excel considering all wells with cultured disruption strains. Any data point that was more than 1.5× the interquartile range below or above the first and third quartile, respectively, was considered an outlier. A disruption strain was considered a hit if over half of the wells for that strain (or 1 of 2) were outliers.

End point acidity and rate quantification with colorimetric dyes

For each assay, knockout strains identified as hits were isolated from the knockout collection into new microplates, along with several blanks per plate, and proxy wild-type strains—GoB58 strains with an intergenic transposon insertion that should not affect the acidification phenotype (see next “Methods” section). Replicates for each knockout strain were distributed across several plates to control for any slight differences between assay plates. For purposes of comparison, all proxy WT strains were considered together; however, OD and acidification phenotypes were measured for each proxy WT strain separately to verify that growth and acidification were unaffected in these strains (Supplementary Fig. 4).

Acidification phenotypes for the disruption strains (for n, see Supplementary Data 6C and D) were compared to that of proxy WT (TB, n = 144; BPB, n = 31) with Student’s t-test in Microsoft Excel, two-tailed assuming equal variances. For the acidification rate assay, V/OD data was discarded if the R2 for V was below 0.5. A Bonferroni correction was used to determine significance to account for the possibility a comparison is significant by chance alone: a phenotype was considered significant if p < 0.05/N, where N is the number of comparisons being made (N = 120 or N = 242 for endpoint acidity comparisons with pWT set A or set B, respectively; N = 60 for rate of acidification comparisons with pWT).

Choice of proxy wild-type comparison

The biolixiviant end point pH and acidification rate of each G. oxydans mutant were compared against a proxy wild-type set of mutants for each phenotype. To account for the presence of a kanamycin cassette in the genome, the proxy wild-type set for each phenotype was constructed of several mutants with the transposon inserted in an intergenic region that had no growth defect and no apparent change in phenotype (Supplementary Fig. 2).

As the efficiency of E. coli WM3064 to G. oxydans mating was low, construction of the G. oxydans progenitor collection required 18 mating batches. As a result, there were slight variations in the wild-type background from batch to batch, notably two distinct saturated culture densities throughout the collection.

For the acidification rate, these variations did not affect the wild-type behavior across the collection due to normalization with culture density, and a single set of proxy wild-type strains could be used as a comparison with disruption strains in the quantification assays. For the end point pH measurement with TB, these two distinct growth phenotypes were found to affect the result, and thus a proxy wild-type set was used for each. Density measurements across the knockout collection revealed sets of plates resembling each phenotype. Proxy wild-type set A was used as a comparison for insertion mutants located in plates 1–76, 110–130, and 160–185 of the condensed collection (Supplementary Fig. 2A), and proxy wild-type set B, which had a lesser saturated culture density, was used for mutants in plates 77–109 and 130–159 (Supplementary Fig. 2B).

Optical density after 2 days of growth and endpoint acidity using the TB absorbance ratio (A435/A545) were compared for both wild-type sets (Supplementary Fig. 2). For wild-type set A, which was used for the BPB quantification assay, acidification rate of individual proxy WT strains was also compared. Pairwise comparisons were all made using the emmeans package in R with a Tukey p-value adjustment (https://github.com/rvlenth/emmeans).

Direct measurement of biolixiviant pH

Bacteria were grown for 48 h in culture tubes containing 4 mL YPM with 100 µg mL−1 kanamycin. One tube was left uninoculated as a no-bacteria control. OD was normalized to 1.9 and diluted in half with 40% glucose for a final 20% solution in 1.5 mL. Five replicates were created for each strain and control, and all mixtures were randomly distributed across two 96-well deep well plates. Seven hundred and fifty microliter of mixture was transferred from each well to a second set of deep-well plates for bioleaching experiments. All plates were incubated with shaking at 800 rpm at room temperature.

After two days, one set of deep-well plates was centrifuged for 10 min at 3200×g (top speed), and the pH of the supernatant was measured by insertion of a microprobe to the same depth in each well.

Four standards were used for meter calibration—pH 1, 2, 4, and 7—and the meter was recalibrated after every 12 measurements. pH measurements for each disruption strain (n = 5) were compared with those of proxy WT (n = 15) using Student’s t-test in Microsoft Excel, two-tailed, with equal variance. A biolixiviant pH was considered significantly different if p < 0.05/N, with N = 22.

Direct measurement of REE bioleaching

The second set of deep-well plates was centrifuged for 10 min at 3200×g (top speed), and 500 µL of biolixiviant was transferred from each well to a 1.7 mL Eppendorf tube. Twenty milligrams (4% w/v) of retorted phosphor powder (gift from Idaho National Lab26) was added to each tube for bioleaching. Tubes were shaken horizontally for 36 h at room temperature and then centrifuged to pellet the remaining solids. Supernatant with leached REE was filtered through a 0.45 µm AcroPrep Advance 96-well filter plate (8029, Pall Corporation, Show Low, AZ, USA) by centrifugation at 1500×g for 5 min.

All samples were diluted 1/200 in 2% trace metal grade nitric acid (JT9368, J.T. Baker, Radnor, PA) and analyzed by an Agilent 7800 ICP-MS for all REE concentrations (m/z: Sc, 45; Y, 89; La, 139; Ce, 140; Pr, 141; Nd, 146; Sm, 147; Eu, 153; Gd, 157; Tb, 159; Dy, 163; Ho, 165; Er, 166; Tm, 169; Yb, 172; and Lu, 175) using a rare earth element mix standard (67349, Sigma-Aldrich, St. Louis, MO) and a rhodium in-line internal standard (SKU04736, Sigma-Aldrich, St. Louis, MO, m/z = 103). Quality control was performed by periodic measurement of standards, blanks, and repeat samples. A pWT biolixiviant sample without bioleaching was spiked with 100 ppb REE standard and analyzed for all REE concentrations as a control. ICP-MS data were analyzed using the program MassHunter, version 4.5.

An additional 1/20 dilution in 2% nitric acid was analyzed for δmgdh and δpqqc disruption strains and the no-bacteria control (glucose).

Bioleaching measurements for each disruption strain (n = 5) were compared with those of proxy WT (n = 15) or glucose (n = 5) using Student’s t-test in Microsoft Excel, two-tailed, assuming equal variances. Total REE extracted was considered significantly different if p < 0.05/N, with N = 12 for those compared to pWT and N = 2 for those compared to glucose.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.