Reconsidering the causality of TIA1 mutations in ALS

T-cell-restricted intracellular antigen-1 (TIA1) has been recently reported as a novel amyotrophic lateral sclerosis (ALS) related gene, and has already been adopted in a resource frequently used in the clinic (Neuromuscular Disease Center) (1,2). The Project MinE Consortium has reviewed the data that support this inclusion, and suggest that the inclusion of this gene as causative is premature.

Although the biology of TIA1 is very appealing and the published functional experiments have been well-performed, it is also the case that current guidelines stipulate that experimental data must be interpreted with clinical caution, as they do not always prove causality (3). Here, we argue that data from the published pedigree that supports TIA1 is insufficient to prove causality. We hold that burden analysis lacks crucial methodological details, and is therefore potentially flawed. Because of this, without a replication effort, we hold that rare TIA1 mutations cannot be regarded as causal in ALS, and that it is premature to include this gene in diagnostic panels for ALS and FTD.

This caution is not to criticize the quality of the work that was performed in characterizing this gene, which is exemplary, and worth describing in detail. The purpose is to systematically evaluate the provided genetic evidence as an exercise in demonstrating the best possible methodologies to establish genetic causality.

In the TIA1 study, whole-exome sequencing was performed in a single pair of second-degree relatives. Variants not observed in the Exome Variant Server and observed less than three times in ExAC were extracted. After filtering non-synonymous variants present in both relatives, 17 genes containing variants remained. Five of the 17 genes were expressed in brain and contained variants that were predicted deleterious (CADD score >20). Out of these five genes, the variant P362L (rs757332023) in TIA1 was picked as the most interesting candidate. It affects a conserved residue and the variant was predicted to be possibly damaging or deleterious by several in silico prediction algorithms. However, it is also important to note that four more candidate genes were present in this pedigree derived from these filtering steps, with similar predictions for deleteriousness and CADD scores. In this context and without evidence for segregation, interesting observations, such as allele frequency in public databases, phylogenic conservation and pathogenicity in other diseases can be used to prioritize candidate genes, but do not prove causality.

To justify claims of causality, the candidate gene needs to pass the threshold for multiple testing in robust association analyses (in both a discovery and replication phase) and demonstrate phenotype-relevant downstream effects in functional follow-up.

The group correctly proceeded with a burden analysis in 1039 patients and 3036 controls, and generated a p-value for TIA1 variants in exon 11–13 (p = 8.7 ×10 ⁻⁶, SKAT-O adjusted for sex and age). However, this p-value remains difficult to interpret in the absence of specific methodological details (4). These should include: (1) a description of the criteria that were used to include variants in the burden analysis; (2) evidence of matching for ancestry or the use of covariates (e.g. principal components) used to correct for ancestry. This is because imperfect matching can severely inflate association statistics, especially in the analysis of rare variants (5); (3) the need to excluded related individuals from the analysis, or to account for relatedness methodologically; (4) the use of quality control measures; and (5) the use of uniform sequencing methodologies (Sanger sequencing vs. whole-exome sequencing), or to demonstrate that different methods are equally sensitive to detect all variants. Finally, the interpretation of association analyses requires evidence of a well-behaved teststatistic (i.e. through production of a QQ-plot and/or reporting LambdaGC) that ensures confounders have been addressed adequately.

For example, in this TIA1 study describing this gene as causative in ALS, it is not entirely clear how a p-value of 8.7 × 10⁻⁶ was obtained, given a test of six variant carriers among 1039 patients and 0 variants among 3036 controls. These numbers can be used to reconstruct their burden test using simulated data. Our simulation includes the same number of cases and controls, with identical allele counts. The SKAT-O test yields a p-value of ~1.37 × 10⁻⁴. Alternative aggregation tests such as SKAT, Firth logistic regression and Fisher’s exact, yield similar p-values (R-script available on https://bitbucket.org/ProjectMinE/tia1). This is, of course, an approximation because we do not have the exact same covariate data. Permuting the test 10,000 times with respect to the covariates does not yield a p-value <1.13 × 10⁻⁴. Without knowing the details of the burden test performed, the reported p-value of 8.7 × 10⁻⁶ might, therefore, be very well higher (i.e. less significant).

When authors focus on gene discovery, they must consider multiple testing burden. If we assume adequate methodology, we interpret the described p-value for TIA1 in the context of an exome-wide hypothesis-generating search, because TIA1 can only be regarded as a candidate gene at this point. Therefore, as with all potential newly discovered variants, stringent multiple testing correction of at least 1.7 × 10⁻⁶ (0.05, Bonferroni corrected for 21,000 protein-coding genes and 9000 long non-coding RNA genes) (3) is required. However, if a variety of tests and variant subsets have been considered, p<5 × 10⁻⁷ is the recommended threshold (5). This becomes an issue if test variation is performed solely within a specific domain, as was the case for TIA1. Depending on the definition of a domain the number of domains in the human genome far exceeds the number of genes (9). Therefore, it can be argued that, in a discovery phase, an even more stringent multiple testing correction is needed. Otherwise, the reported p-value does not meet any of these multiple testing criteria.

Finally, to avoid erroneous attribution of causation, replication is required. Association analyses, such as the SKAT-O burden test, benefit from the largest number of cases and controls available to decrease the chances of false-positive results. There are at least three available ALS-specific case-control cohorts (Table 1) including thousands of cases and controls.

Table 1.

Datasets available for replication.

	Method	Cases	Controls
Cirulli et al 2015. (6)	WES	2,869	6,405
Kenna et al. 2016 (7)	WES	1,022 (FALS)	7,315
Project MinE  Consortium et al. (8)	WGS	4,389	1,846

WES: Whole Exome Sequencing; WGS: Whole GenomeSequencing; FALS: Familial ALS cases.

To test the validity of TIA1 as a possible causative gene in ALS, we performed burden analysis on TIA1 exon 11–13 in the Project MinE Consortium dataset (4389 cases and 1846 controls). Data acquisition and quality control has been described previously (8). Genic SKAT-O tests in the whole dataset are well-behaved (lambdaGC 1.005). We assumed the same filtering was performed as in the pedigree exomes that prioritized TIA1 as the candidate gene. We extracted heterozygous variants that alter the amino acid sequence, that were seen two or fewer times in ExAC and were not observed in the Exome Variant Server. This did not yield a significant result (p40.32; SKAT-O adjusted for first 10 principal components, sex and platform). We acknowledge that this burden test could still be underpowered and/or may not explicitly test the variants tested in the initial report implicating TIA1. The absence of evidence does not eliminate TIA1 mutations as potentially associating with ALS. However, as is the case in all potentially causative genes, to increase power and provide a more definitive answer, we strongly encourage the combined analysis of all available data for the region of interest. Otherwise we are not providing an adequate service for our patients.

In conclusion, the extensive functional experiments performed by Mackenzie et al. shed light on the biological function of TIA1 and show that stress granule dynamics and phase separation may be a crucial aspect of ALS pathophysiology. However, there is currently insufficient genetic evidence to support a causal role for TIA1 mutations in ALS. We encourage research initiatives to provide robust genetic evidence before claiming causality. Therefore, we fully support the author’s remark that future studies in ALS and FTD patient cohorts and control populations will be critical to evaluate the contribution of TIA1 in ALS/FTD. Until then, we would advise against routine screening of TIA1 mutations in ALS/FTD patients in the clinic.