Clover

Thirty gene sets from FactorBook were selected for motif discovery tool comparison (Fig. 2D, Table S1). These gene sets have been selected because the motif of the ChIP'ped TF was detected as top enriched motif in the top 500 peaks in FactorBook. We extracted the top 200 genes having the highest peaks in their 20 kb region around the TSS. The comparison was performed on TF and motif recovery using the parameters indicated in Table S3. The parameters were left to default and when possible, we only adjusted the parameters to allow for larger upstream regions (when possible we choose TSS+−10 kb). iRegulon was compared to eight other publicly available motif enrichment tools, namely OPOSSUM [117] (link), DIRE [80] (link), [112] (link), PASTAA [32] (link), [113] (link), PSCAN [114] (link), Clover [16] (link), AME [118] (link), Allegro [115] (link) and HOMER2 [116] (link) (in the case of Homer2, de novo and known motif discovery are performed simultaneously but we consider them as different approaches and validate them separately). We selected these tools because they mostly take as input a set of human co-expressed genes, and they all return, at least to some extent, information on which TF could be regulating the input genes. For this reason, it not feasible to compare iRegulon with classical de novo motif discovery methods (e.g., MEME-like methods) because such methods are intractable on large human gene sets (e.g., 200 genes×20 kb×10 species represents a sequence set of 40 Mb), and they result in new motifs rather than candidate TFs. We also attempted to use SMART [119] (link) but we did not succeed in running the software. For tools that require regulatory sequences as input (AME and Clover) we used the same sequences as used by iRegulon. For some tools like Clover, it is theoretically possible to use a large search space but one run on one dataset takes too long (∼17 hours), and therefore we limited the analysis to 500 bp promoter sequences. In the case of AME, we found no positive results with a large search space (data not shown), so we show the results with the default search space. For comparison, we used the number of motifs/TFs found in top 1 and within top 5 positions. The total number of detected motifs was not reported for comparison, because some tools use more stringent thresholds than others. All these tools rely on the available motif annotation to identify the candidate TF such as Jaspar (J) or Transfac (T). However, we also manually re-associated the detected motifs to candidate TFs (mainly by comparison of the detected motif with the FactorBook motif) (see column “USING SIMILARITY” in the Table S3). For Homer2, 14 motifs that are derived from ENCODE ChIP-Seq data matching the actual Factorbook ChIP-Seq data were discarded from their in-house PWM collection to avoid over-fitting (indeed, iRegulon does not include FactorBook PWMs either, nor do any of the other tools). Note that for the other large-scale analysis (e.g. full ENCODE analysis), we use a command-line version of iRegulon.

The association function, F( , t_x, t_y) (Eqn. 4), computes the degree of association between the motif affinity score, X, and the biological signal, Y, which in some cases requires partitioning the (mapped) points, , in the X and/or Y dimensions using thresholds t_x and t_y, respectively. We design our association functions so that a larger value implies a stronger positive correlation between X and Y. For our first three association functions, the value of the function is the reciprocal of the p-value of a statistical test. The last two association functions compute non-statistical scores.
The first association function we study—the Fisher Exact Test—is perhaps the most frequently used association function in MEA approaches. When using the Fisher Exact Test, we create the 2-by-2 contingency table induced on the points by the thresholds t_x and t_y, and compute the p-value of the observed (or greater) number of points where X_g ≥ t_x and Y_g ≥ t_y. This is done using the hypergeometric distribution density function [14 (link)]. As noted above, we use the reciprocal of the p-value as the value of the association function.
Our second association function–the multi-Hypergeometric (mHG) Test—extends the Fisher Exact Test to multiple dimensions [15 (link)]. It requires that the affinity function have integral values in some fixed range [0,..., c]. We split the points in the Y dimension using the t_y threshold, and compute the p-value of the observed distribution (or more extreme) of X_g values in the points with Y_g ≥ t_y using the multi-hypergeometric distribution. The value of t_x is ignored by this association function.
Our third association function—the rank-sum test—also ignores the value of t_x. Instead, we sort the mapped points on X, and compute the sum of the ranks of points where Y_g ≥ t_y. We then compute the p-value that the sum of the ranks is as small or smaller than the observed value in the standard manner [16 (link)].
The fourth association function we study is the score computed by Clover (see [1 (link)], Eqn. 4). This association function can only be computed using the AMA motif affinity function, and its value is essentially the average of the motif affinity function over all possible subsets of points where Y_g ≥ t_y. As with the previous two association functions, the value of t_x is ignored by this function.
Our fifth association function is based on the mean-squared error of the linear, least-squares fit to the mapped data points. We assume a linear relationship between X and Y,
and perform least-squares regression on all mapped points. The value of the linear regression association function (LR) is
where E is the mean-squared error of the fit to the mapped data points, and sgn(m) is a function that returns -1 if m is less than 0, and 1 otherwise. (The dots in the function definition indicate that those arguments are not used.) This definition of F insures that its value is large and positive when there is a strong, positive correlation between X and Y. Note that to measure a negative association between X and Y, we could use -sgn(m) times the reciprocal of the mean-squared error in Eqn. 7. In the current study, however, we are only interested in positive correlations between the motif affinity score, X, and the biological signal, Y.
The final association function that we study is Spearman's rank correlation [17 (link)]. Like the linear regression association function, Spearman's rank correlation is threshold-free on both X and Y. Unlike linear regression, however, Spearman's rank correlation coefficient is a non-parametric measure of correlation. It does not assume a linear relationship between X and Y, rather, it assesses the degree to which an arbitrary monotonic function can describe the relationship between X and Y.

Experimentally verified transcription factor binding site regions were obtained from publicly available ChIP-seq experiments: a CTCF dataset (ENCSR000DLG) from ENCODE [20 (link)] and STAT3 dataset (GSM288353 [21 (link)]) from GEO [22 (link)]. These were selected because they were available in narrow peak format with peak max values, to give the highest probability of focusing on the true binding site, and because matching high-quality TRANSFAC TF models were available.
Sequences corresponding to 50 bases either side of the maximum signal of each ChIP-seq peak were obtained. This length was chosen to allow sufficient sequence to identify TFBSs, while minimising extraneous sequence. Backgrounds were produced by using 101 base genomic sequences 10,000 bases away from the peak, ensuring that none of the background sequences overlapped with surrounding peaks. CiiiDER scans were performed using deficits of 0.2. TRANSFAC scans were performed with equivalent core and matrix similarity score cut-offs of 0.8. Clover analyses were performed with default values. Prism 7 was used to generate ROC curves and associated area under curve (AUC) values for each program.

The cows mainly grazed pasture of perennial ryegrass (Lolium perenne) mixed with red clover (Trifolium pretense) and white clover (Trifolium repens). Besides pasture, cows grazed chicory (Cichorium intybus) in spring. To meet energy requirements and to cope with the seasonal changes in pasture quality and production (Machado et al., 2005 (link)), cows were also fed additional supplements including maize silage (Zea mays) and turnips (Brassica rapa) on various days during the summer and autumn seasons along with main feed (pasture). Supplementary feeds are used when quality pasture is less available, to fill the feed deficits and to support the cows to maintain energy intake and production (DairyNZ, 2022 ). The supplements were only used to provide energy when there was insufficient pasture available especially during summer and autumn. Moreover, the purpose of providing supplements to milking cows in autumn is also to achieve calving body condition score (BCS) targets, if the feeds are not supplemented, cows are more prone to lose as quality pasture is insufficient at that time of the year. Maize silage and turnip stems and leaves as such (in situ) were fed around midday in the paddock. The cows had ad libitum access to drinking water.

The European Council requires that pigs must have permanent access to manipulable material [Art. 3 (5 (link)), Annex 1 (4 (link))]. To comply with the legislation, two types of enrichment material were provided; either a rubber floor toy (weaner stage) or plank of wood (finisher stage) and a rack of fresh grass (Perennial Ryegrass and White Clover swards, both stages). Prior to the start of the experiment and at the end of each stage, the rubber floor toy and plank of wood were weighed to determine consumption by the pigs.
Metal racks (0.59 × 0.26 × 0.25 m) were fitted on the front wall of each pen adjacent to the corridor (0.6 m above ground and 0.8 m from the feeder). Racks were 27 cm in length in both the weaner and finisher stages (Figure 2). The grass was added two times a day to ensure that pigs had ad libitum access. The weight of the grass was recorded whenever it was renewed, and the total sum for each pen during each stage (weaner and finisher) was calculated.

U251‐MG, HeLa or KMWT1 cells were seeded in a 12‐well plate and the following day were transfected using Lipofectamine 3000 (Thermo Fisher Scientific) or Lipofectamine LTX (Thermo Fisher Scientific) (for KMWT1 cells) with 900 ng of pSBtet‐GP, pSBtet‐GP‐Regnase‐2, pSBtet‐GP‐Regnase‐2‐D196A, pSBtet‐Pur‐Clover‐Regnase‐2, pSBtet‐Pur‐Clover‐mRegnase‐2, pSBtet‐Pur‐Clover‐mReg‐2‐D195A, pSBbi‐Pur‐dCas9‐KRAB‐meCP2‐hU6‐sgRNA‐SapI or pSBbi‐Pur‐dCas9‐KRAB‐meCP2‐hU6‐sgRNA‐SapI‐Reg‐2 and 100 ng of the pCMV(CAT)T7‐SB100 vector. 24 h later, cells were trypsinized and one‐fifth was transferred into 6‐well plates. Cells transfected with transposon vectors were selected using puromycin (1 μg/mL, Invivogen) for seven days. pSBtet‐GP was a gift from Eric Kovarz (Addgene plasmid # 60495)⁴⁸ and pCMV (CAT)T7‐SB100 was a gift from Zsuzsanna Izsvak (Addgene plasmid #34879).⁴⁹