UFMP was supplied by the Center for Dairy Research (Madison, WI, USA) and stored frozen at −20°C until use. The microbial inoculant was collected from an acid-phase anaerobic digester at the Madison Metropolitan Sewerage District (Madison, WI, USA) and used right after collection for bioreactor inoculation.

A continuous stirred tank reactor with a 3-L working volume was constructed in-house using a 4-L glass Erlenmeyer flask and a butyl rubber stopper. This bioreactor was continuously stirred at 200 rpm and heated to 35°C with a stir plate and heat tape that was temperature controlled using a thermocouple/controller system. The pH was monitored and maintained at 5.5 with a pH probe, a pH controller, and automated addition of 5M NaOH, when needed. The reactor was seeded with equal volumes of feedstock and inoculum from an acid-phase digester operated in the Madison Metropolitan Sewerage District’s wastewater treatment plant (Madison, WI). The feedstock, UFMP amended with 400 mg N/L (as NH₄Cl), was continuously fed at a rate of 0.5 L/day throughout 282 days of operation resulting in solids and hydraulic retention times of 6 days. Temperature and pH were maintained constant at 35°C and 5.5, respectively. Care was taken to minimize input of air into the vessel.

Samples were collected roughly every 6 days from the bioreactor for chemical analyses and DNA extraction. Fully mixed reactor broth was collected, some of which was stored at 4°C, and the rest centrifuged at 10,000 g for 10 min. The supernatants of centrifugated samples were passed through a 0.22-µm polyethersulfone membrane syringe filter and stored at −20°C or 4°C. The microbial biomass pellets remaining after centrifugation were stored at −80°C.

Analytical tests were performed on UFMP samples and on samples collected from the bioreactor. Manufacturer’s protocols were used in each case. Chemical oxygen demand (COD) was measured on filtered and unfiltered samples using COD2 mercury-free high range (20–1,500 mg/L) digestion vials (2565115, Hach, Loveland, CO, United States). Total nitrogen was measured on filtered samples using a high range total nitrogen reagent set (TNT, 2714100, Hach). Orthophosphate assays were performed on filtered samples using PhosVer 3 phosphate reagent powder pillows (216069, HACH). Biomass was estimated by quantifying the volatile suspended solids using Method 2540G (American Water Works Association and Water Environment Federation, 2005).

The concentrations of carbohydrates (lactose, glucose, galactose), some organic acids (formic, acetic, lactic, and succinic) within filtered samples were quantified by high-performance liquid chromatography (HPLC), utilizing an Agilent 1260 Infinity HPLC system equipped with a refractive index detector (Agilent Technologies Inc., Palo Alto, CA, United States). Analyte separation was achieved using a Bio-Rad Aminex HPX-87H column (300 × 7.8 mm, 1250140, Bio-Rad Inc, Hercules, CA, United States) and a cation-H guard column (1250129, Bio-Rad), with a flow rate of 0.5 mL/min, mobile phase of 0.02 N H₂SO₄, and column and detection temperatures of 50°C. The concentration of ethanol and short and medium chain fatty acids (propionic, butyric, pentanoic, hexanoic, heptanoic, and octanoic) within the filtered samples were quantified with tandem gas chromatography-mass spectroscopy (GC-MS) utilizing an Agilent 7890A GC instrument (Agilent Technologies Inc.) equipped with an L-PAL3 autosampler system (LECO Corporation, St. Joseph, MI, United States), a solid phase microextraction (SPME) fiber (StableFlex fiber assembly 23 Ga, 57298-U, MilliporeSigma, Burlington MA, United States), and a Pegasus BT TOF-MS detector (LECO Corporation). For each run, headspace sampling with the SPME fiber occurred over 20 min with agitation at 95°C followed by a sample desorb time in the injection port of 20 min. The oven temperature ramp followed the following scheme: oven equilibration of 1 min, 50°C for 2 min, ramp of 8°C per min to 250°C, then 5.5 min at 250°C for a total run time of about 30 min.

For metagenomic analyses, DNA was extracted from microbial biomass pellets using a published phenol-chloroform extraction procedure (Scarborough et al., 2020), but omitting the bead-beating step. DNA aliquots were submitted to the Joint Genome Institute (JGI) (Berkeley, CA, USA) for sequencing, which was performed using either paired-end 2 × 150 bp Illumina NovaSeq S4 sequencing (Illumina, Inc., San Diego, CA, USA) or single-molecule real-time (SMRT), long-read sequencing on a Sequel II platform (Pacific Biosciences, Inc. [PacBio], Menlo Park, CA, USA). DNA processing, sequencing, quality checking, assembly, binning, annotation, and taxonomy classification was performed using published procedures (Walters et al., 2022). This analysis resulted in a total of 173 metagenome-assembled genomes (MAGs), which were reported elsewhere (Walters et al., 2022).

For microbial community analyses we used a dataset of 217 non-redundant MAGs (Supplementary Table S1), which represent the collective set of metagenomes we have assembled from several fermentation bioreactors seeded with the same inoculum but fed different carbohydrate-rich feedstocks (Myers et al., 2023). In the assembly of this non-redundant dataset we used the 173 MAGs obtained from the UFMP-fed bioreactor described in this study, 105 MAGs from a UFMP-fed upflow sludge blanket bioreactor (Walters et al., 2022), 10 MAGs from a cellulosic ethanol thin stillage bioreactor (Scarborough et al., 2018a), 8 MAGs from a xylose-rich synthetic cellulosic medium reactors (Scarborough et al., 2022b), 51 MAGs from bioreactors using starch-ethanol thin stillage as the feedstock (Fortney et al., 2022), and 48 MAGs from bioreactors fed dairy manure hydrolysate (Ingle et al., 2022). These MAGs were consolidated by dereplication (dRep, v3.2.2; “dereplicate” command with “–conW 0.5” and” –N50W 5″ custom parameters) (Olm et al., 2017), obtaining the final library of 217 non-redundant MAGs used for further analysis. This set of non-redundant MAGs has 151 MAGs assembled from Illumina and 66 MAGs assembled from PacBio sequencing and includes MAGs with greater than 75% completion and less than 7.5% contamination (Supplementary Table S1).

The presence and abundance of microorganisms in different bioreactor samples was assessed as follows (default settings used unless otherwise specified). First, sequence reads from the FASTQ sequencing files were mapped to the concatenated FASTA files of the non-redundant MAG library using Bowtie2 (v2.2.2) (Langmead and Salzberg, 2012) with the “bowtie2” command. Resulting SAM files were converted to BAM files and sorted using samtools (v1.15.1; “samtools view” and “samtools sort” commands) (Li et al., 2009). CoverM (v0.4.0) (https://github.com/wwood/CoverM) was used to generate coverage and relative abundance statistics of mapped reads using the “coverm genome” command on the sorted BAM files. For phylogenetic tree generation of MAGs and related organisms, GTDB-Tk (v1.5.1, database release 202) (Chaumeil et al., 2020) was used for alignment based on the concatenation of 120 bacterial single-copy marker genes (Bac120) (“gtdbtk identify” and “gtdbtk align” commands), RAxML-NG (v0.9.0) (Kozlov et al., 2019) for tree construction (“raxml-ng --parse” command with “--model LG+G8+F” model specification) using 1000 bootstraps, and TreeViewer (v2.0.1) for visualization.

A metabolic network containing known routes for fermentation of lactose to the seven major extracellular products observed during bioreactor operation was constructed from previously described pathways and the MetaCyc database (v.26.5) (DeMoss et al., 1951; Maxwell et al., 1962; Bissett and Anderson, 1973; Song and Lee, 2006; Pokusaeva et al., 2011; Buckel and Thauer, 2013; Cao et al., 2013; Caspi et al., 2014). This network was used as a basis for predicting gene and metabolic pathway presence within MAGs. Predictions of gene presence and metabolic pathways were done using enzyme commission numbers (EC numbers), KEGG orthology numbers (KO numbers), cluster of orthologous genes numbers (COG numbers), and gene product names described in gene annotations predicted by the JGI metagenome processing pipeline (Clum et al., 2021). The General Feature Format (gff) files containing these annotations are available on the GitHub repository https://github.com/GLBRC/UFMP-metagenomics. To supplement these annotations, the “Pathway Hole Filler” tool of the PathoLogic component of Pathway Tools (version 25.5 tier 1) (Karp et al., 2011; Karp et al., 2021) was used to search for genes predicted to be missing in reconstructed metabolic pathways. Additionally, using the Geneious Prime software (v2022.1.1), tblastn searches (default parameters; percent identity ≥25%; query coverage ≥75%) (Gerts et al., 2006) were used to identify electron confurcating lactate dehydrogenase (ecLDH) homologues by sequence comparison to the ecLDH gene of Acetobacterium woodii strain DSM 1030 (Weghoff et al., 2015) (Supplementary Table S2A). A similar tblastn approach was used to identify lactose permease homologues (Supplementary Table S2B), lactose-specific PTS permease components (Supplementary Table S2C), and lactate permease homologues (Supplementary Table S2D) (Saier et al., 2021). A summary of the predicted presence of gene homologues in the 10 most abundant MAGs is provided in Supplementary Table S3.

Phylogenetic and gene neighborhood approaches were used to infer whether identified electron transfer flavoproteins (EtfAB) were partaking in reverse β-oxidation or lactic acid utilization (Costas et al., 2017; Detman et al., 2019). First, a phylogenetic tree based on the EtfB subunit was constructed, including genes for EtfB proteins that were biochemically characterized as partaking in either reverse β-oxidation or lactic acid utilization. Inferences of the function of MAG EtfB were made for those that clustered with characterized EtfB. To corroborate these inferences, an assessment of the presence of relevant genes flanking the etfB genes was performed. The presence of acetyl-CoA acetotransferase (ACAT), 3-hydroxyacyl-CoA dehydrogenase (HAD), enoyl-CoA hydratase (EcoAH), and acyl-CoA dehydrogenase (ACD) genes flanking or nearby (within 5 kb) etfAB genes was an indication of the EtfAB pair being relevant to reverse β-oxidation, whereas presence of lactate permease (LacT) and ecLDH genes was indicative of relevance of the EtfAB pair in lactic acid utilization. For phylogenetic tree generation of EtfB amino acid sequences, MUSCLE (v3.8.31) (Edgar, 2004) was used for alignment using the “muscle” command and RAxML (v8.2.11) (Stamatakis, 2014) for tree construction (“raxmlHPC-SSE3” command with “-m PROTCATAUTO -f a” flags) using 500 bootstraps. The amino acid sequences used in this analysis are provided in Supplementary Table S4. Figure 5 was created using Biorender.com.

Bioreactor performance was monitored by measuring extracellular fermentation products and carbohydrates during the operational period (Figure 1A). Lactose was metabolized and converted to mostly organic acids and ethanol (Table 1), demonstrating efficient fermentation of this sugar during the operational period. Additionally, the measured levels of phosphate and total nitrogen in the bioreactor indicated that neither of these macronutrients became limiting during bioreactor operation (Table 1). After the initial dilution of the inoculant biomass (18 days), biomass concentration showed no major sustained directional trend (Supplementary Table S5) indicating relatively consistent growth of the microbial community.

The bioreactor media initially had a high concentration of lactose because at start up half of the reactor volume was composed of the UFMP feedstock (Figure 1A). After this initial lactose was depleted, lactose and one of its monomeric sugar constituents, galactose, were only rarely detectable in the bioreactor (e.g., day 132 in Figure 1A).

The accumulation of extracellular fermentation products was dynamic during the operational period (Figure 1A), even though flow rates, temperature, and pH were kept constant. After an initial period of acclimation of the microbial community to the UFMP feedstock, about 80% of the soluble COD in the bioreactor effluent could be accounted for by the accumulation of the seven fermentation products included in Figure 1A (ethanol and lactic, succinic, acetic, butyric, hexanoic, and octanoic acids). The sum of these extracellular fermentation products represented, on average, 75% of the COD in the feedstock (Table 1). To facilitate the analysis of the bioreactor operation, we define five distinct periods chiefly based on the identity of the two most abundant fermentation products (Figure 1A); reactor performance during these periods is summarized in Supplementary Table S7:

Period A, from day 0–36, corresponds to a period when acetic, butyric, and hexanoic acids are the predominant extracellular fermentation products.

Period B, from day 44–54, is characterized by a decrease in the extracellular concentration of butyric acid and the accumulation of ethanol, acetic, lactic, hexanoic, and octanoic acids in the media.

Period C, from day 60–84, is a time in which lactic acid is the predominant extracellular product.

Period D, from day 91–108, is characterized by a decreased concentration of extracellular lactic acid and increased accumulation of butyric and succinic acids as major fermentation products.

Period E, from day 114–282, is one in which butyric and acetic acids become the major extracellular fermentation products along with the sporadic accumulation of lactic acid.

The community structure throughout bioreactor operation was analyzed by mapping DNA sequence reads from 20 bioreactor samples to a set of 217 non-redundant MAGs and assessing coverage and relative abundance statistics for each MAG (Supplementary Table S6). Microbial community compositions across time points show notable differences (Figure 1B). Moreover, the operational periods A-E, which were defined based on accumulation of fermentation products (Figure 1A), differed in which community members were most abundant, showing distinct periods of microbial community structure (Figure 1B).

We focused further analysis on a set of 10 MAGs, which represented the highest relative abundances observed during the operational period of the bioreactor. We selected these 10 MAGs after sorting the maximum observed relative abundances of each MAG at all sample points (Table 2). This set of 10 MAGs includes one from the Spirochaetota phylum (SPH2), six from the Actinobacteriota phylum (ACT2, ATO3, ATO6, BIF2, BIF11, BIF18), and three from the Firmicutes phylum (ACID1, CLOS1, LCO1). The assembled genomes of these 10 MAGs have completeness greater than 94% and contamination less than 5%, and encompass MAGs with observed relative abundances as high as 61.5% and as low as 10.2% (Table 2). Furthermore, half of these 10 MAGs were obtained from PacBio sequencing with assembled genomes containing as few as 1 to 4 contigs, whereas the other half were obtained from Illumina sequencing and assembled into 17 to 63 contigs. Only five of these 10 high abundance MAGs had a close representative genome in the GTDB database, indicating that several of these MAGs may be derived from or represent yet uncharacterized microorganisms. To provide context, a phylogenetic analysis was conducted of these 10 MAGs compared to the most closely related isolates for which fermentation products have been reported (Figure 2).

When the relative abundance of the 10 MAGs is evaluated at each sampled time point (Figure 1B), it is evident that none of these MAGs were abundant in the inoculum, suggesting that these community members were enriched at one or more periods during operation of the bioreactor. For periods B to E, there is agreement between the enriched MAGs (Figure 1B), the observed pattern of extracellular fermentation products (Figure 1A), and inferences that can be made from the phylogenetic analysis (Figure 2). These observations are further explored in conjunction with the metabolic analysis described below to infer the functional role of the abundant most abundant organisms.

We queried the set of 10 abundant MAGs (Table 2) for the presence of genes predicted to encode the enzymes associated with the known routes for fermentation of lactose to the seven major extracellular products observed during bioreactor operation (Figure 3). This analysis, described below, allowed us to hypothesize the role of different microbial community members in the metabolism of lactose and accumulation of the major extracellular fermentation products that were identified in the bioreactor.

An emerging picture from analyzing lactic acid-producing communities at the metagenomic and metatranscriptomic level (Scarborough et al., 2018a; Crognale et al., 2021; Myers et al., 2023) is that the microorganisms enriched will likely depend on the composition of the feedstock. For instance, lactic acid producing bacteria in the UFMP-fed bioreactor described here were characterized by having the Leloir pathway for lactose metabolism and were mostly Actinobacteriota with a complete bifid shunt that suggested a heterofermentative process with lactic and acetic acids as the main products. In another lactose-fermenting microbial community, Actinobacteriota were also the most abundant community members inferred to produce lactic acid (Duber et al., 2018). In contrast, the lactic acid producing bacteria enriched in a xylose-rich lignocellulosic feedstock were primarily Firmicutes in the Lactobacillaceae family and encoded pathways for pentose utilization (Scarborough et al., 2018a). Likewise, members of the Lactobacillaceae family were enriched in a bioreactor fed thin stillage derived from a starch ethanol production (Fortney et al., 2021; Fortney et al., 2022). Within microbial communities fed food waste extract, the dominant lactic acid producing bacteria differed based on the organic loading rate, with Firmicutes in the Streptococcaceae family being dominant at low loading rates and Actinobacteriota within the Atopobiaceae family being dominant at high loading rates (Crognale et al., 2021). The implications of these observations are twofold. First, it suggests that a community producing lactic acid from one feedstock may not have the ability to efficiently produce lactic acid when fed another feedstock having a markedly different composition. On the other hand, our observations also suggest that a feedstock that contains a mixture of carbohydrates would enrich for microorganisms capable of fermenting one or more of those carbohydrates. Predicting which microorganisms could be enriched on a particular feedstock will require knowledge of their genetic potential to metabolize the specific carbohydrates in the feedstock.

The inferred connection between some Actinobacteriota fermenting lactose to lactic acid and some Firmicutes utilizing lactic acid for chain elongation derived from this study has been described or hypothesized by others for microbial communities fermenting other carbohydrate-rich feedstocks (Scarborough et al., 2018a; Duber et al., 2018; Carvajal-Arroyo et al., 2019; Crognale et al., 2021; Liu et al., 2022), including within human gut microbial communities (Duncan et al., 2004; Belenguer et al., 2006). For instance, Scarborough et al. analyzed the microbial community in a bioreactor producing acetic, butyric, hexanoic, and octanoic acids from a xylose-rich feedstock and proposed that Actinobacteriota of the Coriobacteriaceae family fermented xylose to lactic and acetic acid, while members of the Eubacteriaceae family of Firmicutes produced hexanoic and octanoic acid from chain elongation of lactic acid (Scarborough et al., 2018a). A similar association was proposed by Carvajal-Arroyo et al. for a granular bioreactor fermenting thin stillage, where Actinobacteriota members of the Coriobacteriaceae family were proposed to ferment the sugars in this feedstock to lactic acid and Firmicutes in the Ruminococcaceae family were proposed to elongate lactic acid to hexanoic acid (Carvajal-Arroyo et al., 2019). In a bioreactor converting food waste extract into hexanoic acid, Crognale et al. proposed that Pseudoramibacter alactolyticus utilized lactic acid for chain elongation, with the lactic acid produced by Actinobacteriota (Crognale et al., 2021). Furthermore, Belenguer et al. used human gut strains in an isotope tracer experiment to demonstrate the flow of carbon from Bifidobacterium adolescentis (Actinobacteriota) to butyrate-producing Firmicutes (Belenguer et al., 2006).

There were 3 Firmicutes among the 10 most abundant MAGs in the UFMP-fed bioreactor analyzed in this study. Based on the genomic information for each MAG (Figure 3B), their abundance (Figure 1B), and the bioreactor fermentation product profiles (Figure 1A), we proposed that all three Firmicutes performed chain elongation, but from different substrates. CLOS1 was hypothesized to use lactic acid, LCO1 was hypothesized to use lactose, and the substrate utilized by ACID1 for chain elongation was unclear but hypothesized to be something other than lactose or lactic acid and possibly ethanol (Figure 5). The following genomic features emerged as the key features that allowed us to make these inferences.

Although we have proposed genomic features that could be diagnostic for whether the substrate for chain elongation is lactic acid, ethanol, or a carbohydrate, we were not able to propose a genomic feature diagnostic of the length of the terminal product of chain elongation. Our proposal of butyric acid as the major chain elongation product of CLOS1 (Figure 5) was based on the observation that this MAG was abundant only when butyric acid was the main product of chain elongation. Our inferences regarding LCO1 and ACID1 producing longer chain fatty acids were made using similar criteria. Isolated organisms have shown a preference for chain elongation products of different lengths, (e.g., C. tyrobutyricum produces butyric acid while CPB6 produces butyric and hexanoic acids) (Wu and Yang, 2003; Zhu et al., 2017) and recent studies utilizing cell-free prototyping to screen enzyme homologs have indicated that thioloases (ACAT) and terminal enzymes of reverse β-oxidation are key determinants in chain length selection (Tarasava et al., 2022; Vögeli et al., 2022). These findings warrant further investigation regarding their implementation in elucidating genomic features for the determination of chain length specificity, possibly through a phylogenetic comparison of enzyme homologs. Furthermore, factors independent of genomic considerations have been implicated in the selection of chain length through thermodynamic-based effects, (e.g., hydrogen partial pressure and substrate type and concentration), exemplifying the need to consider these parameters when designing MCFA-producing bioreactors targeting specific chain lengths (Cavalcante et al., 2017; Scarborough et al., 2018a).

The ability to use microbial communities to generate valuable fermentation products from organic feedstocks will ultimately depend on having the knowledge to control the members of the microbial community and directing it to produce the product of interest while minimizing the accumulation of other fermentation products (Lawson et al., 2019). Our analysis adds to the growing number of approaches for analyzing, predicting, and eventually controlling fermenting microbial communities (Han et al., 2019). The following are genomic features that we propose as diagnostic features to identify the roles of key members of these microbial communities: 1) metabolic pathways for carbohydrate utilization, 2) the reverse β-oxidation pathway, including the ACD-complexing electron transfer flavoproteins (EtfAB), 3) the presence of LacT, ecLDH, and ecLDH-complexing EtfAB pair, 4) presence of ADA and ADH, and 5) predicted absence of genomic features that identify chain elongation from specific substrates. Notably, complications remain in evaluating the absence of a genomic feature due to either incomplete assembly of metagenomes or a missed gene prediction. As such, inferences should be made with caution, ideally with corroborating information such as phylogenetic inferences, correlation between MAG abundance and product accumulation, or metatranscriptomic data. Fortunately, as long-read sequencing becomes more prevalent and as gene prediction tools improve, this limitation will decrease in its practical impact. Finally, genomic features that could be used for predicting the terminal acid in chain elongation (e.g., butyric, hexanoic, or octanoic acid) or improving prediction of chain-elongation from ethanol remain knowledge gaps that require further investigation.