6-propylchromone-2-carboxylic acid

We downloaded the RNA-seq data (RPKM value) across 32 tissues from GTEx V6 release (accessed on April 2016, https://gtexportal.org/). For each tissue, we regarded those genes with RPKM ≥ 1 in more than 80% samples as tissue-expressed genes. To measure the extent to which drug target-coding genes (a and b) associated with the drug-treated diseases are co-expressed, we calculated the Pearson’s correlation coefficient ( $PCC\left(a,b\right)$ ) and the corresponding P-value via F-statistics for each pair of drug target-coding genes a and b across 32 human tissues. In order to reduce the noise of co-expression analysis, we mapped PCC(a, b) into the human protein–protein interactome network (Supplementary Methods 2) to build a co-expressed protein–protein interactome network as described previously^{51 (link)}. The co-expression similarity of the drug target-coding genes associated with two drugs A and B is computed by averaging PCC(a,b) over all pairs of targets a and b with $a\in A$ and $b\in B$ as below: $⟨S_{\mathrm{co}}⟩=\frac{1}{n_{\mathrm{pairs}}}\sum _{\left\{a,b\right\}}\mid PCC\left(a,b\right)\mid$

Linear regression was performed to determine genetic associations with metabolites using KORA F4 CCA and imputed data and the results were compared with each other. For this analysis, we selected metabolite-SNP pairs for which (i) a genome-wide significant association could be identified in the meta-analysis of KORA F4 and TwinsUK cohorts in a previous GWAS (Shin et al. 2014 (link)) (summary statistics retrieved from http://www.gwas.eu); (ii) the proportion of each metabolite’s missing values in KORA F4 was between 10 and 70%; (iii) the metabolite was measured in the EPIC-Norfolk cohort, which we used to further benchmark the preservation of effect sizes; and (iv) a functional connection between the genetic locus of the SNP and the metabolite (e.g., metabolite is a known substrate of the enzyme/transporter) was evident according to manual curation of the GWAS results (Table S8). For each imputed dataset, 18 metabolite-SNP pairs were tested for genetic association using age- and sex-corrected linear regression models under the assumption of an additive genetic model $\text{(metabolite}\sim {\mathit{\beta }}_0+{\mathit{\beta }}_1\times \text{SNP}+{\mathit{\beta }}_2\times \text{age}+{\mathit{\beta }}_3\times \text{sex)}$ . To avoid spurious associations, metabolic data points greater than four SDs from the mean were removed prior to computing linear models. For MI approaches, the regression coefficients were pooled using Rubin’s rules as provided by the R package mice, version 2.25. For each metabolite-SNP pair, the variance of the regression coefficients and p-values were estimated using bootstrapping.
To explore which imputation approaches increased statistical power, p-values obtained for the effect sizes based on imputed data were compared with p-values obtained from CCA by calculating their ratio as $r_p=\frac{-{log}_{10}\left(\frac{p_{imp}}{p_{CCA}}\right)}{-{log}_{10}(p_{CCA})},$ where $p_{imp}$ was the p-value obtained for imputed data and $p_{CCA}$ was the p-value derived from CCA. A ratio less than or equal to zero indicated either no power gain or a power loss, whereas a ratio greater than zero indicated a drop in p-value, which suggested that statistical power increased when imputation was performed.
In addition to statistical power gain, the imputation approaches should be able to preserve effect sizes compared to CCA. Standardized effect sizes obtained from the imputed data $({\mathit{\beta }}_{imp})$ were compared with standardized effect sizes estimated for CCA $({\mathit{\beta }}_{CCA})$ based on the KORA F4 data (n = 1750) and the EPIC-Norfolk data (n = 10,634), assuming estimates from the EPIC-Norfolk data to be close to true effects. We calculated the ratio $r_{\mathit{\beta }}={log}_2\left(|\frac{{\mathit{\beta }}_{imp}}{{\mathit{\beta }}_{CCA}}|\right)$ , with a low ratio indicating a similar effect size between the imputed data and CCA. A highly negative or positive $r_{\mathit{\beta }}$ indicates an underestimation or overestimation of the effect sizes in imputed data, respectively. A well performing imputation method is assumed to obtain high $r_p$ and low absolute $r_{\mathit{\beta }}$ .

LC-MS/MS raw files from Lyso-APEX, Lyso-IP, and Lyso-BAR proteomic experiments were analyzed with Thermo Fisher Proteome Discoverer (2.4.1.15) software. For dynamic SILAC proteomic data, MaxQuant (1.6.17.0) software was used for data analysis. Swiss-Prot Homo sapiens database was used for i³Neuron data and Mus musculus database was used for mouse data with 1% false discovery rate (FDR) for protein identification. Custom-made contaminant protein libraries (https://github.com/HaoGroup-ProtContLib) were included in the data analysis pipeline to identify and remove contaminant proteins^{61 (link)}. Trypsin was selected as the enzyme with a maximum of two missed cleavages. Cysteine carbamidomethylation was included as fixed modification, and oxidation of methionine and acetylation of the protein N-terminus were selected as variable modifications.
Protein/peptide identification and peak intensities were output as excel files for further analysis using Python or R. Statistical analyses (t-test) and volcano plots for Lyso-APEX, Lyso-IP, and Lyso-BAR proteomics were conducted in Python. Lyso-APEX and Lyso-BAR data were normalized to the most abundant endogenously biotinylated protein (PCCA) before statistical analysis as described previously^{32 (link)}. For dynamic SILAC data, Maxquant output files were further processed with R to calculate heavy/light peptide ratios and construct the degradation and synthesis curves as well as curve-fitting to the first-order kinetic in multiple time point experiment. For single time point experiments, peptide level Maxquant output files were processed with Python to calculate the peptide half-lives using the equation: t_1/2 = t_s × [ln2 / ln (1+Ψ)], where t_s represents the sampling time after media switch, and Ψ represents the heavy-to-light abundance ratio of the peptide. Protein level half-lives were calculated by averaging the half-lives of unique peptides belonging to the specific protein. Statistical analysis was conducted with t-test, and multiple half-life datasets were merged by uniprot protein accession in Python. Protein GO enrichment analysis was conducted using ShinyGO^{62 (link)}. Protein network analysis was conducted with STRING^{63 (link)}.

PyEMMA v2.5.7 was used to build the MSM (Scherer et al., 2015 (link)) and carry out kinetic modeling of the tetramer. All 2090 trajectories were loaded into the software. The RMSD and backbone torsions of all residues in α3-β7 (G149-D171) and β12-α5 (A219-H239) loops from all four subunits were selected the input features. Next, the featurized trajectories were read in with a stride of 5 and were projected onto three independent components (ICs) using TICA. The produced projections can show the maximal autocorrelation for a given lag time (5 ns). The chosen ICs were then clustered into 200 clusters using k-means. In this way, each IC was assigned to the nearest cluster center. A lag time of 5 ns was chosen to build an MSM with seven metastable states according to the implied timescales (ITS) plot (Figure 2A). After passing the Chapman–Kolmogorov test within confidence intervals (Figure 2—figure supplement 1), the MSM was defined as good. This indicates the model highly agrees with the input data, and it is statistically significant for use. Bayesian MSM was used to build the final model in the system. The net flux pathways between macrostates, starting from state 1, were calculated using Transition Path Theory (TPT) function. These pathways all originate from state 1, as it shows the lowest stationary probability (the highest free energy) in the system. This is why state 1 is a reasonable starting point to illustrate all the relevant kinetic transitions through the full FE landscape. The structural results were selected from each PCCA distribution.

All trajectories were aligned to their crystal structure conformation using Moleculekit implemented in HTMD tools (Doerr et al., 2016 (link)). To visualize the structures representing each state, the structures collected from the PCCA distributions were loaded and superimposed in Pymol-mdanalysis (https://github.com/bieniekmateusz/pymol-mdanalysis; Bieniek, 2022 ). The structural analysis was performed using mdtraj (McGibbon et al., 2015 (link)) and MDAnalysis (Michaud-Agrawal et al., 2011 (link)). Figures capturing major conformational changes were generated using the Protein Imager (Tomasello et al., 2020 (link)). All plots were made using the matplotlib libraries (Hunter, 2007 (link)).