CASP9 protein, human

The new components aimed at IDR detection were trained with a concatenation of two datasets. The first was all entries (228 in total) from the Disprot v5.0 database (Sickmeier et al., 2007 (link)) flagged as being derived from either NMR or biophysical methods. The other dataset was a redundancy-reduced (percentage sequence identity <90%) subset of high resolution (≤2.2 Å) X-ray structure chains from PDB (Berman et al., 2000 (link)) derived from PISCES (Wang and Dunbrack, 2005 (link)) compiled in February 2010. Chains shorter than 25 amino acids were discarded. Missing residues, including those with occupancy equal to zero, were treated as disordered.
Position-specific scoring matrix (PSSM) scores were calculated for each residue using three iterations of PSI-BLAST (Altschul et al., 1997 (link)) running on the UniRef90 data bank (Suzek et al., 2007 (link)) with an inclusion E-value threshold of h = 0.001.
DISOPRED3 was registered as a server at CASP9 and CASP10—group ids 015 and 170, respectively—and made predictions for all assessed targets. The corresponding predictions are therefore available at the Prediction Center website; DISOPRED2 predictions for the same protein sets were generated as a prerequisite and stored locally.
The reference classification of the residues as ordered or disordered was taken from the Prediction Center website and was based on the structural data available before the final meetings. The amino acids were regarded as disordered if and only if either they were not assigned spatial coordinates, or the positions of their Cα atoms were more than 3.5 Å away across different chains or NMR models in the LGA (Zemla, 2003 (link)) structural alignment.

DisCo is derived from QMEANDist, a quasi-single model method that participated in the CASP9 experiment as a global quality predictor (Biasini, 2013 ; Kryshtafovych et al., 2011 (link)). We revisited the approach of assessing the agreement of pairwise residue–residue distances with ensembles of constraints extracted from experimentally determined protein structures that are homologous to the assessed model. Instead of generating global quality estimates, DisCo aims to predict local per-residue quality estimates. After extracting the target sequence of the model to be assessed, homologues are identified using HHblits (Remmert et al., 2011 (link), the used command line arguments are available in the Supplementary Materials). For each homologue k, all Cα positions are mapped onto the target sequence using the HHblits alignment. Gaussian distance constraints for residue pairs (i, j) are generated for all Cα–Cα distances μ_ijk below 15 Å:
$g_{\mathit{ijk}}\left(d_{\mathit{ij}}\right)\mathrm{ }=\mathrm{ }\hspace{0.17em}\text{exp}[-\frac{1}{2}(d_{\mathit{ij}}-\mathrm{ }{\mu }_{\mathit{ijk}}){}^2]$
The goal is to construct a pairwise scoring function s_ij(d_ij), that assesses the consistency of a particular pairwise Cα–Cα distance d_ij in the model with all corresponding constraints g_ijk(d_ij). In order to avoid biases towards overrepresented sequence families among all found homologues, they are clustered based on their pairwise sequence similarity as specified in the Supplementary Materials. Since the templates often do not cover the entire target sequence, some Cα–Cα pairs might not be represented in every template and consequently the number of templates n_ijc containing a Cα–Cα pair varies within a cluster for different (i, j). Only if a Cα–Cα pair is present in a cluster c, we construct a cluster scoring function h_ijc(d_ij):
$h_{\mathit{ijc}}\left(d_{\mathit{ij}}\right)=\frac{1}{n_{\mathit{ijc}}}\sum _{k\in c}{g}_{\mathit{ijk}}\left(d_{\mathit{ij}}\right)$
To get the desired pairwise scoring function s_ij(d_ij) we combine h_ijc(d_ij) from each cluster c in a weighted manner as exemplified in Figure 2. Clusters expected to be closely related to the target sequence contribute more than others:
$s_{\mathit{ij}}(d_{\mathit{ij}})=\sum _c{\mathrm{w}}_c{h}_{\mathit{ijc}}\left(d_{\mathit{ij}}\right)$ with weights w_c defined as exp[γSS_c] and normalized, so that the weights of all clusters in which the Cα–Cα pair is present, sum up to one. SS_c is the average normalized sequence similarity towards the target sequence of cluster c and γ is a constant that controls how fast the influence of a cluster vanishes as a function of SS_c. The default value for γ is 70 and the effect of varying γ is discussed in Supplementary Figure S3. The DisCo score of a single residue of the model at position i then is computed by averaging the outcome of all n pairwise scoring functions s_ij(d_ij) towards other residues j ≠ i with their Cα positions within 15 Å:
${\mathit{DisCo}}_i=\frac{1}{n}\sum _j{s}_{\mathit{ij}}\left(d_{\mathit{ij}}\right).$
As the accuracy of DisCo depends on the underlying templates, features describing its reliability are required to optimally weigh DisCo with the single model scores in a subsequent machine-learning step. For each residue i there are:

To lyse cells, 1 mL Trizol reagent (TaKaRa, Japan) was added to the sample tissue according to the actual manufacturer’s instructions and incubated on a shaker for 15 min at room temperature. Total RNA was subsequently extracted from target samples. One microgram of RNA was reverse transcribed into cDNA using the Revert Aid First Strand cDNA Synthesis Kit (Thermo, United States). Quantitative RT-PCR was then performed with Pro Taq HS Premix Probe qPCR Kit (Accurate, Hunan, China). The GAPDH gene was used as an endogenous control gene for normalizing the expression of target genes. Primers used in this study included CASP9 (forward 5′-CTG​TCT​ACG​GCA​CAC​AGA​TGG​AT-3′, reverse 5′- GGG​ACT​CGT​CTT​CAG​GGG​AA-3′), GAPDH (5′-CAG​GAG​GCA​TTG​CTG​AT-3′, 5′-GAA​GGC​TGG​GGC​TCA​TTT-3′).