Data collection and sequence curation

Class I T-box riboswitch (transcriptional) leader sequences used to generate the database were obtained from a variety of primary sources including the RFAM 14.0 database, RibEX database, GeConT3 database and others (21,31–34). For class I transcriptional T-box riboswitch leader sequences that did not necessarily contain the terminator (RFAM and GeConT3 sequences) Entrez (NCBI) was used to extend sequences by 50 nt. Extended sequences were then subsequently trimmed to end in penta poly-U (3′-UUUUU). Sequences from these databases that were too short (<100 nt), too long (>500 nt), or redundant were also removed. Sequences of translational T-box riboswitches (class II) were found by using our custom covariance model (Supplementary Figure S2) to perform an INFERNAL search on all NCBI reference genomes in the Actinobacteria phylum (TaxID:1760) (35). The class II covariance model was additionally applied on T-box riboswitch sequences detected by the class I model but predicted to have a truncated Stem I (Supplementary Figure S3).

Structural and thermodynamic prediction of T-box riboswitch leaders

INFERNAL was used to predict the secondary structure of input sequences using either the RFAM 14.0 class I T-box riboswitch covariance model (RF00230.cm) (34,35), or our class II T-box riboswitch covariance model (Supplementary Figure S4). The INFERNAL output structure, corresponding to the antiterminator/antisequestrator fold, was then used to predict T-box riboswitch features according to the conserved patterns of stems and loops. Next, the MFE (Supplementary Figure S5) of the structure was evaluated using RNAfold (ViennaRNA) (36). For putative transcriptional T-box riboswitches, the terminator structure was determined by searching for a suitable terminator hairpin using RNALfold. The MFEs of both folds were calculated using RNAeval. Structures for sequestrator stems from translational T-box riboswitches are not currently predicted. Summary of MFE for T-box riboswitches represented in the TBDB are shown in Supplementary Table 2. Detailed descriptions of the feature extraction and thermodynamic prediction are available in the Supporting Information.

Pairing T-box riboswitches with putative cognate tRNAs

Top specifier sequence predictions were used to identify a putative tRNA family pair for each T-box riboswitch. To find the sequence of cognate tRNA, Entrez (NCBI) was used to query for all genomic records of the T-box riboswitch host organism. Genome sequences, either partial or full, were downloaded from RefSeq or GenBank (37,38). tRNAscan-SE (Lowe Lab, UCSC) was used to identify tRNAs in each host organism (39). tRNAs with predicted anticodons matching the specifier were considered paired. For cases where more than one possible tRNA gene was possible, a single tRNA was chosen from among the matching tRNAs for display. tRNA visualization was generated using VARNA (40).

Prediction of T-box riboswitch specifier sequence

The T-box riboswitch specifier region was assigned as the 1-5 bp (inclusive) 5′- from the end of the Stem I specifier bulge. In these five bases, three possible specifier sequence frames (‘−1’, ‘+0’, ‘+1’) were examined for meeting specifier-match criteria. For each possible specifier, we identified the putative tRNA family (by matching anticodon) that would bind. We then checked to see if (i) the predicted tRNA family had a discriminator base that could base pair with the T-box 5′-UGGN-3′ sequence, with wobble allowed, and (ii) if the predicted tRNA amino acid family matched the downstream gene ontology (where available). In the case of predicted His T-box riboswitches, discriminator matching was not used as a criterion as mature tRNA^His transcripts can have a paired discriminator base (41).

For matching T-box 5′-UGGN-3′ with tRNA acceptor end 5′-NCCA-3′ sequence, we first searched the host for all tRNAs of a given tRNA family and identified which discriminator base that specific host used for a particular tRNA. In cases where we could not identify matching tRNAs in the host organism, the bacterial discriminator base frequency information was extracted from tRNAviz and utilized in the specifier prediction model (42). The top specifier was then assigned as the specifier that met most of these conditions, equally weighted, with preference given in the following order: ‘+0’ > ‘−1’ > ‘+1’ specifier sequence frames. In cases where more than one specifier was possible, the top specifier was assigned as mentioned but alternative specifier sequence frames are also provided in TBDB.

Accessing TBDB content

The TBDB aims to be a comprehensive and approachable hub for predictions of T-box riboswitch structure and function. Database entries are provided in a searchable, tabulated format. Users can query entries based on fields that include sequence, host organism, specifier sequence, T-box 5′-UGGN-3′ sequence, or predicted tRNA family (Supplementary Figure S6).

Detailed information on each TBDB entry can be obtained by accessing the unique ID in the database table. Doing so brings users to a T-box riboswitch entry page that contains source, downstream protein annotation, structural, functional, and sequence information. The title of the entry page provides the T-box riboswitch ID, a unique identifier generated by TBDB, as well as the predicted tRNA family the T-box riboswitch interacts with in the host organism (Supplementary Figure S7). The source information panel gives a high-level summary of the T-box riboswitch entry and includes information regarding genetic locus and feature predictions. The following panel provides an interactive genome browser (NCBI) starting at the T-box leader sequence locus and ending 5,000 bp downstream. The genome browser allows users to visualize the genomic context of T-box riboswitch leader sequences and provides a quick method to assess the validity of T-box riboswitch specifier predictions. For example, a Trp T-box riboswitch is observed as a 5′-UTR of an operon involved in tryptophan biosynthesis (Supplementary Figure S8).

Towards the goal of making a T-box riboswitch registry-of-parts, in the subsequent panel we provide a T-box riboswitch sequence that stretches from the Stem I to the terminator poly-U region (for transcriptional T-box riboswitches, Supplementary Figure S9). Visualizations (VARNA) of the predicted secondary structures of the T-box riboswitch are given in the next panel (40). These 2D representations highlight important features of the T-box riboswitch entry including Stem I (light-yellow), possible specifier sequence (dark-yellow), antiterminator/antisequestrator (light-blue), the four nucleotide 5′-UGGN-3′ in the T-box bulge (dark blue), and terminator stem (red). The dot-bracket representations of the 2D structures are also provided (Supplementary Figure S10). Results for tRNA matching, generated using tRNAscan-SE, can be found in the following panel (Supplementary Figure S11) (39). Here, we show the sequence and description for the highest scoring predicted tRNA, with matching anticodon, in T-box riboswitch host organism (if available). tRNAs for alternative specifier sequence frames are also generated if our model could not identify a consensus specifier. Minimum free energy (MFE) predictions for each of the folds are given in the thermodynamics section, and are the result of structure refinement performed using ViennaRNA (Supplementary Figure S12) (36).

Finally, the INFERNAL panel has output information from structural searches, which can be used as a reference by users interested in structure prediction information and quality (Supplementary Figure S13). Here, users will also find information about predicted boundaries of important structural features. We believe the TBDB sequence demarcation features of the important Stem I, Stem II, Stem III, antiterminator/antisequestrator and terminator/sequestrator regions will spur future efforts to explore common themes and diversity that T-box riboswitches have accrued and help aid in their classification.

Identification of T-box riboswitch:tRNA pairs

T-box riboswitches tend to have a strict preference for canonical Watson–Crick base pairing between the specifier sequence and the anticodon of the cognate tRNA (9). Depending on the length of the specifier bulge, it is also likely that there are alternative specifier sequences, allowing for the possibility of multi-tRNA specificity in gene regulation (43). However, experimental work uncovering the determinants of multi-specificity in T-box riboswitches remains sparse. Our specifier prediction assignment takes into consideration downstream gene ontology and base pairing between the T-box bulge 5′-UGGN-3′ sequence and the tRNA acceptor end found in the host organism. In assigning T-box riboswitch specifier, priority was given to as the region 2-4 bp (inclusive) 5′- from the end of the Stem I specifier bulge, though ‘–1’ and ‘+1’ specifier sequence frames were also considered (Supplementary Figure S1). In a majority of the cases, the variable position (5′-UGGN-3′) on the T-box bulge shows Watson-Crick base pairing with the discriminator base (5′-NCCA-3′) of the cognate tRNA species. Exceptions were noted in the Trp family, where 46% of T-box riboswitches in our collection have a ‘U’ at the degenerate position while their cognate tRNAs have a ‘G’ at the discriminator position, suggesting a G:U wobble pair, as has been previously noted (41). Based on this observation we allowed for wobble base pairs between the tRNA discriminator and the degenerate nucleotide in the 5′-UGGN-3′ region of the T-box bulge sequence in our final model. In the case of putative His T-box riboswitches, discriminator base matching was not considered as tRNA^His transcripts can have an internally paired discriminator base (41).

In practice, the TBDB identifies the cognate tRNA pairs from T-box riboswitch hosts by first predicting the specifier sequence, then searching genome records of respective hosts for tRNAs that have a matching anticodon (Watson–Crick base pairing, no wobble allowed) and discriminator base pairing (both Watson-Crick and wobble pairing allowed). Our model gave a single specifier sequence frame prediction for 16,258 T-box riboswitch leader sequences, two possible specifier sequence frames for 2,884 sequences, and three specifier sequence frames for 3,551 sequences. For 48 sequences, we were able to predict a specifier but were unable to find the canonical 5′-UGGN-3′ sequence in the T-box bulge. In cases where more than one specifier is possible, preference is given to the ‘+0’ specifier sequence frame (Supplementary Figure S1C). In all, T-box riboswitch leaders containing Trp-, Leu- and Ile- tRNA matching specifiers were most commonly observed in our collection while Lys-, Glu- and Gln- matching specifiers were the least common, well in agreement with previous findings (20). Supplementary Table 1A and B show composition of the T-box Riboswitch Annotation Database amino acid family and specifier usage.

Through our tRNA search, we were able to match 79.4% of T-box riboswitches for which we predicted a specifier with a tRNA of the native host. Grouping T-box riboswitches by specifier sequence, we found that tRNA-matching was >80% for most specifiers (Supplementary Table 1B). Interestingly, we only identified a tRNA pair for 8.6% of T-box riboswitches with 3′-U specifiers, consistent with observations that 5′-A starting anticodons in bacterial tRNAs are rare (Supplementary Table 1C). In these cases, it is possible that these T-box riboswitches are controlled by tRNAs without matching anticodons (e.g. relying on wobble base pairing) or are using an alternative specifier sequence frame when binding anticodons (44). Alternative specifier sequence frames for T-box riboswitch sequences have previously been observed in at least one experimentally studied system (43).

T-box riboswitch specifier usage

Identifying specifier sequences for T-box riboswitches allowed us to interrogate the choice of tRNA anticodon, and therefore the tRNA, that is used for regulation. Analogous to ‘codon usage tables’, which summarize an organism’s codon preference for translating particular amino acid, Table 1 depicts a ‘specifier usage table’ generated from 23,535 sequences in the TBDB. For T-box riboswitches of amino acid families that only have two codons (Lys, Asp, Asn, Glu, Gln, His, Tyr, Cys, Phe), a single specifier is preferred in over 85% of corresponding T-box riboswitch sequences. The choice of specifier is also consistent, with 3′-A and 3′-C always favored over 3′-G and 3′-U respectively. T-box riboswitches of amino acid families decoded by four codons (Gly, Ala, Pro, Val, Thr) have more diversity in specifier usage. Much like the two-codon sets, there is a preference for Val, Thr and Gly T-box riboswitches to use 3′-A and 3′-C specifiers. Interestingly, Ala and Pro T-box riboswitches display a preference for 3′-U specifiers. T-box riboswitch families for amino acids with 6 codons (Leu, Ser, Arg) show a 3′-A and 3′-C preference, with an even stronger preference for 3′-C specifiers. In the particular case of Leu family T-box riboswitches, the CUC specifier is observed in 74% of sequences. For the special case of Ile, 3′-C (AUC) specifier is preferred. The large collection of sequences has allowed us to reinforce previous observations that the ‘C-rule’ (3′-C in specifiers) is prevalent, while additionally discovering that 3′-A usage is also largely preferred for specific amino acid classes (Supplementary Table 1B and C) (20,21,32). Indeed, recent structural analysis reveals that a non-canonical A-minor motif in the Stem II S-turn places local constraints directly to prefer a guanine on the tRNA anticodon, and therefore a cytosine in the third position of the specifier (15,30).

Table 1.

Specifier usage table based on ‘Top’ specifier predictions for T-box riboswitch leader sequences in the TBDB. Frequency given on an amino acid basis for the entire sequence collection in the TBDB. Specifier usage frequencies were rounded to two decimal places.

SPEC	AA	FREQ	SPEC	AA	FREQ	SPEC	AA	FREQ	SPEC	AA	FREQ
AAA	K	0.87	AGA	R	0.24	ACA	T	0.13	AUA	I	0.01
AAG	K	0.13	AGG	R	0.04	ACG	T	0.01	AUG	M	1.00
AAC	N	0.98	AGC	S	0.09	ACC	T	0.60	AUC	I	0.99
AAU	N	0.02	AGU	S	0.01	ACU	T	0.25	AUU	I	0.01
GAA	E	0.88	GGA	G	0.17	GCA	A	0.18	GUA	V	0.61
GAG	E	0.12	GGG	G	0.04	GCG	A	0.01	GUG	V	0.02
GAC	D	0.94	GGC	G	0.77	GCC	A	0.09	GUC	V	0.29
GAU	D	0.06	GGU	G	0.02	GCU	A	0.72	GUU	V	0.08
CAA	Q	0.95	CGA	R	0.01	CCA	P	0.16	CUA	L	0.02
CAG	Q	0.05	CGG	R	0.08	CCG	P	0.10	CUG	L	0.07
CAC	H	0.95	CGC	R	0.52	CCC	P	0.14	CUC	L	0.77
CAU	H	0.05	CGU	R	0.12	CCU	P	0.60	CUU	L	0.11
UAA	*	0.00	UGA	*	1.00	UCA	S	0.06	UUA	L	0.01
UAG	*	0.00	UGG	W	1.00	UCG	S	0.03	UUG	L	0.02
UAC	Y	0.99	UGC	C	1.00	UCC	S	0.65	UUC	F	0.99
UAU	Y	0.01	UGU	C	0.00	UCU	S	0.16	UUU	F	0.01

There are possible explanations for the source of specifier-usage bias in T-box riboswitches. First, specifier usage does not follow the same observed patterns of codon usage. In most cases, the preferred specifier is the least preferred codon for the amino acid family. For example in the taxonomic order Bacillales, the Phe UUU codon is used in approximately 70% of cases for translation (45), but is present in only 1% of Phe T-box riboswitch specifiers. One possible hypothesis for the specifier-use bias could be attributed to T-box riboswitches favoring interaction with a single tRNA species and disfavoring wobble base pairing. In the absence of tRNA with 5′-I (Inosine) anticodons, 3′-A and 3′-C codons are only decoded by a single tRNA species (5′-U and 5′-C anticodons), whereas 3′-G and 3′-U codons can be decoded by multiple tRNAs (5′-U/C and 5′-A/G anticodons). T-box riboswitches likely co-evolved to be highly specific in their response towards a single tRNA species, which would have been made more difficult if specifier binding is made competitive with two (or more) tRNA species (Watson-Crick basepair versus wobble). Additionally, tRNAs with 5′-A anticodons are not prevalent in bacteria, as the U:G wobble-pair is the preferred mechanism for decoding 3′-U codons (46). The consensus sequence of the 23,535 T-box riboswitches (Supplementary Figure S1) revealed that 5′-NNC-3′ specifiers were preferred overall, being represented at 59.9% of T-box riboswitch sequences with predicted specifiers.

Tool for T-box riboswitch scanning and feature extraction

In order to increase accessibility and reproducibility for detection and annotation of T-box riboswitches for arbitrary DNA sequences, we have also released a standalone CLI tool (tbox-scan) for finding and extracting features of T-box riboswitches. Tbox-scan uses INFERNAL to find T-box riboswitches in a FASTA sequence input, and then performs the same feature extraction used to build the TBDB. As outputs, the tool displays T-box riboswitch location, specifier sequence, T-box bulge 5′-UGGN-3′ sequence, and secondary structure prediction of antiterminator/antisequestrator and terminator folds. Covariance models for putative transcriptional (class I) and putative translational (class II) T-box riboswitches are also provided. Users can download tbox-scan from https://tbdb.io/tools/tbox-scan.html.