Timeless and Armadillo: A link too far (Comment on Vodovar et al. Curr. Biol. (2002) 12, R610–R611)

Public availability of automated methods for prediction of protein structure from sequence greatly benefits the community of 'wet-laboratory' biologists. Their reliability is assessed regularly on behalf of newly solved crystal and NMR structures, most prominently through the so-called CAFASP-experiments [1], and seems to be increasing on the whole. Clearly, however, the degree of confidence accorded to predictions based on results by automated servers must be in measure with the corroborating evidence that is presented. In this light, we have examined the validity of associating the circadian clock protein Timeless with the Armadillo or HEAT families of proteins (referred to as Armadillo/HEAT below), based on the predictive analyses by Vodovar et al. [2]. Our investigation casts serious doubts on the legitimacy, and correctness, of their claims. The authors proposed two 'Arm domains' (AD) each in Drosophila melanogaster Timeless and Timeout (Figure 1A). These were identified, and their boundaries determined, through submissions to the fold prediction server 3D-PSSM [3]. The core of Armadillo/HEAT structures is α α-helical and can be expected to show considerable sequence similarity between homologs. Valid candidate regions in which to begin looking for Armadillo/HEAT repeats should fulfill these criteria. 3D-PSSM failed to return even sub-significant matches for any of these regions in either paralog (Figure 1; Supplemental Data). In contrast, significant E-values (E < 0.5) could only be obtained if the fragments used for the query are extended to Vodovar et al.'s ADs. Especially in the case of Timeless, the extension includes a substantial non-helical portion. This should immediately have raised suspicions that this could be a false positive prediction, as should have the comparatively weak threading alignment accompanying the prediction. Indeed, we found that 3D-PSSM regularly overpredicts Armadillo/HEAT in largely α α-helical fragments exceeding 400 residues. A control study revealed several clearly false positive predictions at comparable levels of significance (Supplemental Data). Submitting various amino-terminal fragments of Timeless and Timeout exposed a strong length dependence of the results (Supplemental Data), which explains why the slightly shorter fragments corresponding to AD1 in Timeless orthologs do not elicit false positive predictions below the significance threshold (Figure 1B). Regarding the existence of AD2, only extremely weak indications were reported in the first instance — a match only just at the significance threshold for Timeless, and an E-value well above threshold for Timeout (Figure 1B). Given these observations, adequate corroborating evidence becomes essential to the question whether even the slightest justification …