The Guinea grass M. maximus-associated epiphytic microbial community was the source of inoculum for selection of LAB and LB populations. Firstly, this grass was cut in 1cm² pieces. Then, in aseptic conditions, 10 fragments were inoculated in 10ml of Man, Rogosa, and Sharpe (MRS) broth (Merck, Darmstadt, Germany) and incubated at 39°C for 24h. After microbial growth, 1ml of the culture was inoculated in 9ml of MRS broth and incubated to the above conditions. This latter procedure was carried out to pre-enrich LAB populations by growth of the community in medium designed to enrich growth of LAB. Subsequently, 25μl of the pre-enriched culture were inoculated in 100ml-flasks with 25ml of MRS-modified broth (i.e., without glucose) containing 1% lignocellulosic substrate (M. maximus cv. Agrosavia Sabanera added as a carbon source; T1; Figure 1; Table 1), vitamin and trace elements solution (Jiménez et al., 2014b). Previously, the lignocellulosic substrate (supplied by Agrosavia) was knife-milled through a 1mm screen and washed twice with distilled water and ethanol 70% (v/v). The grass was collected at regrowth age of 30days in a sub-region of the Colombian dry Caribbean (Motilonia RC, Cesar) with an elevation of 103 masl. The culture flasks were incubated at 28°C under shaking conditions (130rpm). After 4days of incubation, aliquots (25μl) of microbial suspension were transferred to 25ml of fresh liquid medium containing 1% plant biomass (T2), following the dilution-to-stimulation approach (Jiménez et al., 2017). Two negative controls were also set up: (A) microbial inoculum in liquid modified MRS (without plant biomass and glucose) and (B) uninoculated liquid modified MRS (containing plant biomass and without glucose). During the growth-dilution transfers, liquid culture medium was sequentially modified (Table 1) to avoid the assimilation of labile carbon sources (e.g., ammonium citrate, yeast, and/or beef extract), increasing the selection for LB populations. Thus, these modifications were done until no growth in negative control A was observed (Figure 1; Table 1). At the end of each transfer cultures were filtered and the remain substrate was dried for 24h at 45°C to calculate the percentage of weight loss, following the formula proposed by de Lima Brossi et al. (2016). Bacterial growth was determined using optical density (OD) at 600nm and the number of viable cells (CFU/ml) was quantified by plate counting in R2A and MRS agar (Merck, Darmstadt, Germany). Moreover, samples from transfer 10 (T10) were conserved in 20% of glycerol at −20°C. ANOVA and post hoc Tukey-Kramer test were performed using the software R (R Core Team, 2008) to evaluate the significant difference in substrate consumption and bacterial growth along the transfers, with a CI of 99% (α=0.01).

To increase the bacterial biomass, the mixed culture after T10 was transferred to a final volume of 500ml. We carried out this procedure because this consortium was used as a starting bioinoculant in a silage of tropical forages in field experiments (data not shown). The scale-up growth process began with a pre-inoculum from T10 that was cultivated in 25ml of liquid medium (eight replicates), as previously described. Two of these replicates were mixed to produce an inoculum of 50ml. To ensure a high bacterial density, four 2L-flasks with 500ml of fresh medium (same composition as T10) were inoculated with the former 50ml and incubated at 28°C for 8days. Finally, four replicates were mixed and filtered to produce 2L of bacterial culture. Moreover, the consortium obtained at T10 was cultivated (i.e., perturbed) on the modified MRS medium adjusted to different pH values (4, 5, and unmodified pH 6.2) at 28°C for 4days, with shaking at 130rpm (in two sequential transfers). Number of viable cells (CFU/ml) was quantified by plate counting in R2A and MRS agar (Merck, Darmstadt, Germany). OD at 600nm and weight loss of the substrate were calculated with the parameters described above. After microbial growth in different pH values, the secreted fractions of the cultures were tested for the presence of plant polymers-degrading endo-enzymes using a set of six chromogenic polysaccharide hydrogels (CPH; i.e., 2-HE-cellulose, carboxymethylcellulose, arabinoxylan, xylan, xyloglucan, and glucomannan) and two insoluble biomass substrates (ICB; i.e., wheat straw and sugarcane bagasse; Kračun et al., 2015). The semi-quantification of the enzymatic activities was measured as reported Díaz-García et al. (2021) and following the instructions of the manufacturer (GlycoSpot IVS, Farum, Denmark). ANOVA and post hoc Tukey-Kramer test were performed using the software R (R Core Team, 2008) to evaluate significant differences in growth and enzyme activities, with a CI of 99% (α=0.01).

Total microbial DNA was extracted using the DNeasy UltraClean Microbial Kit (Qiagen, Hilden, Germany) according to the instructions of the manufacturer. For the 16S rRNA gene analysis, triplicate samples from the dilution-to-stimulation approach (T1, T2, T3, T4, T5, T6, and T10), after the scale-up process (S) and the pH perturbation (4, 5, and unmodified) were sequenced using the Illumina MiSeq technology (300bp pair-end reads). For each sample, a library of 16S rRNA gene amplicons (hypervariable region V3–V4) was prepared using the primers 341F and 785R (Klindworth et al., 2013). Bioinformatic analysis was done using the software Quantitative Insights Into Microbial Ecology (QIIME2; Bolyen et al., 2019). VSEARCH (Rognes et al., 2016) was used to join raw pair-end reads. Quality control and read correction to obtain amplicon sequence variants (ASVs) were done using Deblur pipeline (Amir et al., 2017). Rarefaction was performed (5,205 and 5,490 reads for the enrichment steps and the pH perturbation samples, respectively) per sample to calculate alpha and beta diversity (including Shannon’s index and weighted UniFrac distances), and for comparisons between samples (e.g., relative abundance of ASVs). MAFFT (Katoh and Standley, 2013) was used to construct multiple alignments of ASVs. A phylogenetic tree was built using FasTree (Price et al., 2009) based on MAFFT alignment. Taxonomic affiliation of ASVs was done using SILVA rRNA gene curated database (Quast et al., 2013).

Based on the 16S rRNA sequencing results, consortia T2, T10, and S (in duplicate) were subjected to Illumina MiSeq technology sequencing (300bp paired-end reads) in order to determine taxonomy profiles and metabolic potential. After whole-metagenome sequencing, the FastQ files were uploaded in the MG-RAST server (Meyer et al., 2008). Overlapping sequence pairs were matched, and non-overlapping reads retained as individual reads, after which, dereplication was performed. Duplicate read-based inferred SE estimation and quality trimming (phred score<20) used default settings. Gene predictions were done using the FragGeneScan software, and subsequently, the proteins were annotated based on BLASTX searches against the RefSeq and KEGG databases using an e-value cutoff of 1e-15, a minimum alignment length of 50 amino acids, and a minimum identity of 50% (Jiménez et al., 2016). Data from MG-RAST annotation were statistically analyzed using the STAMP package (Parks and Beiko, 2010). Moreover, to evaluate the relative abundance of reads per selected enzyme-encoding gene, the counts were normalized to hits, or unique matches, per million reads, thereby accounting for differences in metagenome sizes. Nine genes involved in heterolactic and/or homolactic fermentation of xylose (Tarraran and Mazzoli, 2018) and 10 genes involved in lignocellulose degradation (Jiménez et al., 2014a; Díaz-García et al., 2021) were analyzed. Heat maps were constructed in the web server Heatmapper using row-Z score for each enzyme (Babicki et al., 2016).

For the assembly of MAGs, raw sequence data (FastQ files) from T2, T10, and S samples were initially trimmed using the Sickle tool v1.33 with default parameters (available at https://github.com/najoshi/sickle). To avoid misleading results from subsequent binning analysis, carp artifacts were detected. For this, trimmed sequences were filtered against adapter and non-authentic primer sequences originating from the Illumina library preparation (Marter et al., 2021). Metagenome assembly was done using MEGAHIT v1.2.7 (Li et al., 2015) with default parameters. Concoct v1.1.0 (Alneberg et al., 2014), Bowtie v2.3.5 (Langmead and Salzberg, 2012), MetaBAT v2.12.1 (Kang et al., 2015), and MaxBin v2.2.6 (Wu et al., 2016) were used to bin the assembled sequences with default parameters. Binned contigs obtained were subsequently analyzed using DAS tool v1.1.2 (Sieber et al., 2018). To assess the completeness and contamination of the resulting bins, CheckM v1.0.13 was used with the lineage_wf option (Parks et al., 2015). The MAGs were structurally annotated and curated based on rRNAs genes (e.g., 5S, 16S, and 23S) and rpoB gene using DFAST v1.2.6 (Tanizawa et al., 2018). Taxonomic assignment was performed using GTDB v1.4.1 (Parks et al., 2020). Functional annotation was done using the RAST (Aziz et al., 2008) and dbCAN webservers (Yin et al., 2012). Moreover, 24 encoding-genes involved in hetero- or homo-fermentation of pentoses and hexoses (Hatti-Kaul et al., 2018; Tarraran and Mazzoli, 2018) were searched within the MAGs.

The selection of consortia was carried out by the dilution-to-stimulation approach using the epiphytic microbial community of M. maximus as starting inoculum (Figure 1). In the first stage of selection, we aimed to increase the populations of LAB by growing the phyllosphere-derived community in MRS broth. This pre-enrichment was used to inoculate the first transfer (T1), where glucose was removed and 1% (w/v) M. maximus was added as a carbon source. After this stage, increase of LB populations was the main goal. For this purpose, liquid MRS medium was sequentially modified from T2 to T6–T10, removing peptone, yeast extract, beef extract, triammonium citrate, and decreasing the concentration of sodium acetate, dipotassium phosphate, and tween 80 (Table 1). These modifications were carried out until microbial cell growth (measured by OD at 600nm) in the negative control (i.e., without plant biomass) was not observed, indicating that the consortium started to consume the plant polymers (Figure 2A). After T6, the selected consortium was able to consume ~14% of plant biomass and significant differences (value of p<0.01) in this capacity were evidenced compared to T5 (Figure 2B). Moreover, progressive reduction of bacterial growth rate was observed from T1 to T6, suggesting that labile carbon sources were depleted due to the liquid medium modification (Figures 2A,C). LAB populations were abundant at T1–T2. However, they decreased along the enrichment strategy (Figure 2D). At the end of the selective process (T10), around 10⁹ and 10⁵CFU/ml of total and LAB populations were obtained, respectively (Figures 2C,D).

The bacterial community changes during the selective process were evaluated by 16S rRNA gene sequencing. In total, this analysis generated 1,623 ASVs and ~1.2Gbp of high-quality reads. The results indicated that bacterial diversity slightly decreased from T1 to T5. However, the number of observed ASVs increased from T5 (44±3.0) to T6 (77±4.1) and T10 (73±4.5). Interestingly, after the scale-up process (to 500ml of culture volume) the number of ASVs increased to 99±4.0 (Figure 3A). The taxonomic affiliation of the 16S rRNA data revealed that Gammaproteobacteria and Bacillales species were mostly selected during the enrichment strategy. Regarding LAB populations, species belonging to the Enterococcaceae family were observed at T1–T2 (~10% of the total community; Figure 3B). They remained in low abundance in the final selected consortium (i.e., T10). Modifications in the MRS liquid medium did not greatly affect the bacterial structural composition between T2 and T10. Remarkably, the final consortium (i.e., T10) showed a significant change after its cultivation in a higher volume (i.e., 500ml; Figure 3C). This scale-up process (denominated S in Figure 3) mostly increased the relative abundance of Paenibacillaceae, Sphingomonadaceae, and Xanthobactereaceae, well known lignocellulolytic families (Figure 3B). At the end, the mixed bioinoculant (S), used in the field silage experiment, contained a higher proportion of LB populations and lower proportion of LAB (mostly species from Enterococcaceae family).

The mixed consortia obtained at T2, T10, and S was subjected to a whole-metagenomic sequencing and analysis. Approximately 1.3, 0.9, and 3.0Gbp were obtained in T2, T10, and S, respectively, with a read average size of 300bp. Based on the taxonomic affiliation of reads, four bacterial classes were highly abundant. In this regard, Enterobacteriales and Lactobacillales decreased from T2 to T10 and S, while Actinomycetales and Bacillales increased from T2 to T10 and S (Figure 4A). Enterobacteriales species were the predominant taxa (between 50 and 90% of total affiliated sequences) within consortia. At T10 and S, the relative abundance of Lactobacillales was lower than at T2, decreasing from about 4% to between 1 and 3% (Figure 4A). Regarding specific members of LAB, Enterococcus sp. was the most abundant genus (between 1 and 5% relative abundance) in T2, T10, and S. Other genera, like Lactobacillus sp., Leuconstoc sp., Pediococcus sp., and Weissella sp. were found in very low abundance (less than 0.15%) in all the stages of selection, although their populations increased in S with respect to T10 (Supplementary Figure S1). Regarding fungal communities, metagenome data revealed that they are found in very low relative abundance (less than 0.001%). Based on the function (KO level), consortia T2 and T10 showed a similar profile, compared with the consortium obtained after scale-up (S) process (Figure 4B). Moreover, the results showed that nine of 10 genes involved in lignocellulose transformation were enriched from T2 to T10 and S (Figure 4C). Regarding genes relevant to lactic acid production, we observed that genes encoding xylose dissimilation activities [e.g., transketolase (TK) and transaldolase (TA)] were highly abundant at T2, whereas others [e.g., 6-phosphofructokinase and L-lactate dehydrogenase (LDH)] were highly abundant at T10 and S (Figure 4C).

The ability of the selected consortium (i.e., T10) to grow in acidic silage conditions was evaluated by modifying the initial pH of the liquid medium. At pH 4, a significant (value of p<0.01) increase of LAB was evidenced, showing a viable population density of around 4×10⁸ CFU/ml. In contrast, total bacterial cells decreased at pH 4 and pH 5, compared to the unmodified liquid medium (pH around 6.2; Figure 5A). After the perturbation at different pH values, the metasecretome of the consortium was evaluated for the presence of plant polymer-degrading endo-enzymes. The highest and significant (value of p<0.01) enzymatic activities on a set of CPHs and ICBs were obtained in the metasecretome of the mixed consortium after growth at pH 5 (Figure 5B; Supplementary Figure S2). At this pH, the plant biomass consumption was 17.8±0.6%. Additionally, 16S rRNA gene sequencing analysis were carried out after the perturbation at different pH values. On average, 104 ASVs were generated from 50Mbp of high-quality reads. Alpha diversity results demonstrated a decrease of observed ASVs in the pH gradient, reaching a value of 31±1.0 at pH 4 (Figure 5C). Interestingly, a significant (value of p<0.001) increase in the relative abundance of Actinomycetales and Lactobacillales orders was observed at pH 5 (Figure 5D). However, the highest abundant order was Enterobacteriales (Supplementary Figure S3).

A total of 2, 3, and 9 high quality MAGs (>90% of completeness and <3.5% of contamination) were reconstructed from the T2, T10, and S metagenomes, respectively (Table 2; Supplementary Table S1). The genome lengths ranged from 2.28 to 5.93Mb. At T2 and T10, MAGs belong to Cellulosimicrobium funkei (Bin7_T2 and Bin8_T10) and Klebsiella aerogenes (Bin25_T2 and Bin2_T10) were obtained. Interestingly, the data revealed that three different MAGs from Enterococcus species were obtained. At T2, a single genome associated to Enterococcus faecalis, while in both T10 and S, two genomes associated to Enterococcus casseliflavus. Functional annotation of all reconstructed MAGs showed that Bin7_T2, Bin8_T10, Bin3_T10, Bin11_S, Bin6_S, Bin002_S, and Bin8_S contain more than nine genes-encoding glycosyl hydrolases (GHs) from CAZy families involved in degradation of (hemi)cellulose (Table 2). Moreover, Bin3_T10 and Bin6_S affiliated to E. casseliflavus harbor more than 20 GHs involved in transformation of oligosaccharides (e.g., GH1, GH2, and GH3). In addition, two LDHs encoding-genes were found in those two MAGs (Table 2; Supplementary Table S2).

A total of five MAGs (Bin6_S, Bin002_S, Bin2_S, Bin4_T10, and Bin44_S) were selected in order to analyze their potential to produce lactic acid from sugars. These MAGs contained at least one LDH encoding-gene (Table 2). Based on the functional annotation using the RAST server, encoding-genes involved in fermentation of pentoses and hexoses were detected within these five MAGs (Supplementary Table S2). The results showed that Bin6_S (E. casseliflavus), Bin002_S (C. funkei), and Bin2_S (K. aerogenes) encode complete enzymatic machinery to produce lactic acid. Specifically, Bin6_S and Bin002_S contain the genes necessary for heterolactic acid fermentation of pentoses (e.g., xylose and arabinose) through the phosphoketolase (PK) pathway (Figure 6). In addition, Bin002_S has the metabolic potential to carry out a homolactic fermentation of xylose through the pentose phosphate (PP) pathway (Figure 6) and heterolactic fermentation of maltose through the PK pathway. Moreover, Bin2_S contains the gene capability to produce D-lactate from glycerol through an Embden–Meyerhof (EM) glycolytic pathway (Supplementary Table S2).