Fmrp targets or not: long, highly brain-expressed genes tend to be implicated in autism and brain disorders

There are two key facets of the reported Fmrp targets that motivated this analysis. First, as highlighted in the 2011 work [1], the Fmrp protein binds mRNA rather promiscuously, not being strongly restricted to RNAs with one particular motif in the CNS. Second, Fmrp itself is highly expressed throughout the nervous system, particularly, though not uniquely, in neurons [2]. Thus, as the confidence of Fmrp binding was dependent on read-depth, and the protein is both fairly promiscuous in sequence specificity and ubiquitously expressed in the brain, the most readily detected transcripts might be those that present the most opportunity for binding - those with the longest coding sequence and the highest expression. Thus, we tested the hypotheses that the reported Fmrp targets disproportionately represent the most abundant mRNAs in the brain (Figure 1A) and those with the longest coding sequence (Figure 2A). In support of this, using a list of genes with the highest expression in the human brain [3], one can demonstrate that the reported Fmrp targets disproportionately overlap with the most highly expressed neural genes in humans (P < 3e⁻¹⁶, Fisher’s exact test).

To test whether this intersection might sometimes lead to a statistical overlap between Fmrp targets and trait-associated genes in the brain, we conducted a simple experiment examining the overlap between a set of genes involved in a neurogenetic trait unrelated to Fragile X Syndrome or psychiatric disorder. It has been recognized that, since body weight tracks with consumptive behaviors, obesity is strongly influenced by genes that are expressed in the brain [4]. Thus, we tested the statistical overlap between reported genes for “obesity-related traits” and the reported Fmrp targets and see a statistical enrichment of a magnitude not too different from that sometimes seen in the literature for psychiatric disorder (P < 0.005,). Thus, the results of our experiment are consistent with the explanation that genes mediating any neurogenetic trait may show overlap with the reported Fmrp targets simply because both sets overlap with the set of genes highly expressed in neural tissue. Indeed, genes reported for several other neurocognitive traits also overlap Fmrp targets with nominal (P < 0.05) significance, for example, “hippocampal atrophy” (P < 0.009), “Alzheimer’s disease” (P < 0.022), and “cognitive performance” (P < 0.012).

Darnell et al. suspected a bias towards highly expressed or longer transcripts in their analysis and reported a very modest correlation between transcript abundance (cor = 0.1) and number of cross-linking and immunoprecipitation (CLIP)-seq tags ([1]; Additional file 1: Figure S1). However, their measures of abundance were based on microarray signal, which can be strongly biased by features unrelated to RNA abundance (for example, probe GC content), and lack the dynamic range of RNAseq analysis. Analyzing instead with RNAseq data from postnatal day 21, mouse cortex shows a correlation of 0.64 (Figure 1B) with expression and 0.45 with coding sequence (Cds) length (Figure 2B). Indeed, we found that a linear regression model incorporating expression, Cds length, and transcript length could account for >60% of the variance (r ²) in CLIP tag number (Table 1). We would emphasize that the presence of detectable, albeit weak, motifs in the targets [5] indicates they are non-random and thus the initial findings were not exclusively driven by these features. The more likely explanation was that the sensitivity was limited to the highly expressed Fmrp targets and that many less abundant (but perhaps equally high affinity) Fmrp targets might simply have been below the threshold of detection. In the future, the comprehensive identification of Fmrp targets in the brain could be revisited with approaches to allow greater sensitivity and perhaps cell-specific normalization for transcript abundance. However, in the meantime, we tested two candidate gene lists, the Simons Foundation Autism Research Initiative (SFARI) database of curated autism candidates (SFARIdb) and the recently identified rare de novo variants in autism probands (rDNV) [6], to see if a statistically significant overlap with Fmrp targets could be reproduced just using equally sized sets of random genes sampled to match as best possible the transcript abundance (Additional file 1: Figure S1) or coding sequence length (Additional file 2: Figure S2) of the Fmrp target genes.

We found in either case that random samples of abundant transcripts (Figure 1C) or transcripts with long Cds (Figure 2C) were significantly overlapped with the SFARIdb and rDNV genes, though not to the extent of that of the Fmrp targets. To make sure our results were not particular to a single sampling, we sampled 1,000 such gene lists. Most overlapped significantly with the autism candidates (Figures 1D and 2D), though again not as significantly as the reported Fmrp targets. Very similar results can be seen by comparing a contingency table overlapping the Fmrp targets and the rDNV genes relative to all brain-expressed genes (P < 6.02e − 12) or instead calculating the contingency table for just the genes with expression in the top quantile (P < 0.0002): gene expression level alone accounts for some of the overlap between rDNV genes and Fmrp targets, but not all of it.

However, when instead sampling random gene sets to match the Fmrp targets simultaneously on Cds length and expression level (Figure 3), we found that a substantial fraction of the sampled lists showed an equivalent or greater statistical overlap than the original target Fmrp list (Figure 3C). Finally, we repeated these analyses but excluded the Fmrp genes from the sampling pool. Because the Fmrp genes so strongly monopolize the long and highly expressed gene space (Figure 3A), it was impossible to sample a set of genes that perfectly matched the two-dimensional distribution of the Fmrp targets after excluding those genes (Additional file 3: Figure S3A, B). Nonetheless, a random set of non-Fmrp target genes, or simply a list of the longest and highest expressed non-Fmrp target genes, also significantly overlapped both of the autism candidate lists (Additional file 3: Figure S3C). This demonstrates that long, highly expressed genes in the brain tend to overlap with disease-risk genes, whether they are on the Fmrp target list or not. It is worth noting that a disproportionate amount of long and highly brain-expressed genes is not just a feature of the Fmrp targets, but of course many other sets of genes important for the functioning of the nervous system (for example, Gene Ontologies terms for ‘Synapse’ or proteomics studies of synaptic proteins), so some of the conclusions here extend beyond consideration of just the Fmrp targets. Likewise, genes with enriched expression in the brain are on average longer than genes expressed in other tissues [7], and thus any analysis that identifies long genes may tend to overlap statistically with brain-expressed genes, Fmrp targets, autism candidates, synaptic proteins, etc.

We have shown that the overlap between the reported Fmrp targets and at least two autism candidate gene lists can be reproduced by simply selecting for similarly long and highly expressed genes in the brain. This is consistent with long and highly brain-expressed genes also being more likely to be under selective constraint [8] or containing critical exons [9] and provides a straightforward explanation for why the Fmrp target list overlaps so frequently with sets of genes implicated in psychiatric disease by genetic studies. It is imaginable that similar features may explain why the Fmrp target list also frequently overlaps results of brain transcriptomic studies as well. This parsimonious explanation thus obviates complex hypotheses which require the Fmrp protein itself to be involved in the mechanism for many diverse disorders or different forms of ASD. Of course, mutations in Fmrp clearly do still cause Fragile X Syndrome, the most common form of monogenic ASD, and thus continued research into this protein remains important for that reason alone.

This model also provides reasonable explanations for two other puzzles about the Fmrp targets. First, it could explain why studies of the Fmrp targets in HEK cells [10] are less concordant with other studies [5] and why HEK cell data overlaps marginally if at all with psychiatric disorder candidate gene lists [11]. The HEK cell data should be biased towards long, highly expressed genes in HEK cells, which will likely contain few neural-specific transcripts. Second, this model might explain why identifying strong cis motifs or other features in the RNA that might mediate Fmrp binding has proven challenging [5]. Efforts to model the affinity of Fmrp for particular mRNAs will likely be aided by first removing the variance in the Fmrp CLIP data that can be explained by transcript length, Cds length, and transcript abundance. The authors of [1] suspected a bias towards highly expressed genes, but recognized the data were not available at that time to adjust for it, particularly if the level of Fmrp varies substantially across cell types in the brain. Thus, the definition of the Fmrp targets can probably now be revisited both with greater sensitivity and by models incorporating these covarying factors to identify additional features of the transcripts that account for the remaining variance in Fmrp binding.

In the end, it may well be that these studies find that Fmrp does bind preferentially to those transcripts whose protein levels most require precise regulation for normal CNS function. It is not unreasonable that this set of genes would also be vulnerable to haploinsufficiency [8,9] and of course be expressed highly in the brain. And a set of genes needing more precise regulation may indeed be selected by evolution to be longer (that is, allowing more potential sites for regulatory motifs). Thus, Fmrp binding may have been serving as a useful proxy for these other features. However, in the interim, we have provided a table (Additional file 4: Table S1) with precalculated weightings for length, expression, or length and expression for measurably brain-expressed genes. This can be used for drawing random samples for comparison to candidate gene lists, to help determine whether the candidate list is enriched in Fmrp targets specifically and/or long, highly brain-expressed transcripts generally.