Prediction and verification of the AD-FTLD common pathomechanism based on dynamic molecular network analysis

Multiple gene mutations cause familial frontotemporal lobar degeneration (FTLD) while no single gene mutations exists in sporadic FTLD. Various proteins aggregate in variable regions of the brain, leading to multiple pathological and clinical prototypes. The heterogeneity of FTLD could be one of the reasons preventing development of disease-modifying therapy. We newly develop a mathematical method to analyze chronological changes of PPI networks with sequential big data from comprehensive phosphoproteome of four FTLD knock-in (KI) mouse models (PGRNR504X-KI, TDP43N267S-KI, VCPT262A-KI and CHMP2BQ165X-KI mice) together with four transgenic mouse models of Alzheimer’s disease (AD) and with APPKM670/671NL-KI mice at multiple time points. The new method reveals the common core pathological network across FTLD and AD, which is shared by mouse models and human postmortem brains. Based on the prediction, we performed therapeutic intervention of the FTLD models, and confirmed amelioration of pathologies and symptoms of four FTLD mouse models by interruption of the core molecule HMGB1, verifying the new mathematical method to predict dynamic molecular networks.

Normal distribution of Log-transformed peptide signal ratios, which were used for comparison between a disease mouse model and its control, was examined as the basis of statistical analyses in this study. a) Distribution of raw values of the peptide signal ratio between disease model and control mice observed in mass analyses. b) Distribution of Log-transformed peptide signal ratios matched well with normal distribution (red curve) in all mouse models. c) Q-Q plot of Log-transformed ratios supported their normal distributions.

Supplementary Figure 2
Pathological networks of four FTLD and four AD mouse models at each time point a, b) Based on the integrated protein-protein interaction (PPI) database supplied by the Genome Network Platform of the National Institute of Genetics (https://cellinnovation.nig.ac.jp/GNP/index_e.html) including experimentally supported PPI database of the Human Genome Project (GNP) and databases from BIND (http://www.bind.ca/), BioGrid (http://www.thebiogrid.org/), HPRD (http://www.hprd.org/), IntAct (http://www.ebi.ac.uk/intact/site/index.jsf), and MINT (https://mint.bio.uniroma2.it/)., pathological protein networks of each FTLD mouse model (a) or each AD mouse model (b) at each time point were generated. Proteins including significantly changed phosphopeptide(s) in more than two models were used further for selecting core nodes with high betweenness scores.

Supplementary Figure 3
Venn's analyses of common changes of protein phosphorylation across multiple mouse models a) Comparison of phosphoproteins changed from four FTLD mouse models. b) Comparison of phosphoproteins changed from four AD mouse models. c) Integration of common proteins across FTLD models and common proteins across AD models. A1: phosphoproteins changed in 1 or more AD model(s). F1: phosphoproteins changed in 1 or more FTLD model(s). A2: phosphoproteins changed in 2 or more AD models. F2: phosphoproteins changed in 2 or more FTLD models. A3: phosphoproteins changed in 3 or more AD models. F3: phosphoproteins changed in 2 or more FTLD models. A4: phosphoproteins changed in 4 AD models. F4: phosphoproteins changed in 4 FTLD models.

Supplementary Figure 4 Permutation test of %commonness of AD core network and FTLD core network
The histogram shows a truncated normal distribution of %commonness calculated 10,000 times from 73 nodes (equal to the number of AD core nodes) and 62 nodes (equal to the number of FTLD core nodes) that were selected by bootstrap procedure from 1,965 nodes (equal to the total number of proteins detected by all the phosphoproteome analyses). 46.7% of %commonness between AD core nodes and FTLD core nodes was at mean + 44.6 SD in the truncated normal distribution.

Supplementary Figure 5
Verification of the AD-FTLD common core nodes and edges by comparison with the result from external data a) Five datasets of comparison between one mouse model as external data and the other four mouse models as internal data. The first dataset used APP-KI mice, which was not used for generating AD core network as external data. The second to fifth datasets are virtual datasets to confirm reliability our conclusion. b) Comparison of the first dataset of internal and external data indicated AD core nodes were precisely predicted at 79.2% and AD core edges were precisely predicted at 59.8%. c) Similarly performed comparisons with the other four datasets indicated mean precision rates for core nodes and for core edges were 81.8% and 64.6%.

Supplementary Figure 6 Cross-validation of three datasets of correlation values in each mouse model
Correlation values were calculated from two mice and one mouse in each dataset, and absolute error and root mean square error were calculated. a) Central dots indicate the values of mean absolute error and bar means standard deviation of absolute errors. b) Central dots indicate the values of root mean square error and bar means standard deviation of square errors.

Supplementary Figure 7
Generation of the AD-FTLD common network under a severe condition for selecting core nodes Core nodes were selected at the high threshold (betweenness > 2SD) and used for generation of AD core network, FTLD core network and AD-FTLD core network.