4,4',5'-trimethylazapsoralen

Given an arbitrary data set as an input, TMAP encompasses four phases: (I) LSH forest indexing [37 , 38 ], (II) construction of a $c$ -approximate $k$ -nearest neighbor graph, (III) calculation of a minimum spanning tree (MST) of the $c$ -approximate $k$ -nearest neighbor graph [39 (link)], and (IV) generation of a layout for the resulting MST [40 ].
During phase I, the input data are indexed in an LSH forest data structure, enabling $c$ -approximate $k$ -nearest neighbor (k-NN) searches with a time complexity sub-linear in $n$ . Text and binary data are encoded using the MinHash algorithm, while integer and floating-point data are encoded using a weighted variation of the algorithm [41 –43 ]. The LSH Forest data structure for both MinHash and weighted MinHash data is initialized with the number of hash functions $d$ used in encoding the data, and the number of prefix trees $l$ . An increase in the values of both parameters led to an increase in main memory usage; however, higher values for $l$ also decrease query speed. The effect of parameters $d$ and $l$ on the final visualization is shown in Additional file 1: Fig. S1. The use of a combination of (weighted) MinHash and LSH Forest, which supports fast estimation of the Jaccard distance between two binary sets, has been shown to perform very well for molecules [44 (link)]. Note that other data structures and algorithms implementing a variety of different distance metrics may show better performance on other data and can be used as drop-in replacements of phase I.
In phase II, an undirected weighted $c$ -approximate $k$ -nearest neighbor graph ( $c$ – $k$ -NNG) is constructed from the data points indexed in the LSH forest, where an augmented variant of the LSH forest query algorithm we previously introduced for virtual screening tasks is used to increase efficiency [45 (link)]. The $c$ – $k$ -NNG construction phase takes two arguments, namely $k$ , the number of nearest-neighbors to be searched for, and $k_c$ , the factor used by the augmented query algorithm. The variant of the query algorithm increases the time complexity of a single query from $O\left(logn\right)$ to $O\left(k·k_c+logn\right)$ , resulting in an overall time complexity of $O\left(n\left(k·k_c+logn\right)\right)$ , where practically $k·k_c>logn$ , for the $c$ – $k$ -NNG construction. The edges of the $c$ – $k$ -NNG are assigned the Jaccard distance of their incident vertices as their weight. Depending on the distribution and the hashing of the data, the $c$ – $k$ -NNG can be disconnected (1) if outliers exist which have a Jaccard distance of 1.0 to all other data points and are therefore not connected to any other nodes or (2) if, due to highly connected clusters of size $\ge k$ in the Jaccard space, connected components are created. However, the following phases are agnostic to whether this phase yields a disconnected graph. The effect of parameters $k$ and $k_c$ on the final visualization is shown in Additional file 1: Fig. S2. Alternatively, an arbitrary undirected graph can be supplied to the algorithm as a (weighted) edge list.
During phase III, a minimum spanning tree (MST) is constructed on the weighted $c$ – $k$ -NNG using Kruskal’s algorithm, which represents the central and differentiating phase of the described algorithm. Whereas comparable algorithms such as UMAP or t-SNE attempt to embed pruned graphs, TMAP removes all cycles from the initial graph using the MST algorithm, significantly lowering the computational complexity of a low dimensional embedding. The algorithm reaches a globally optimal solution by applying a greedy approach of selecting locally optimal solutions at each stage—properties which are also desirable in data visualization. The time complexity of Kruskal’s algorithm is $O\left(E+logV\right)$ , rendering this phase negligible compared to phase II in terms of execution time. In the case of a disconnected $c$ – $k$ -NNG, a minimum spanning forest is created.
Phase IV lays out the tree on the Euclidean plane. As the MST is unrooted and to keep the drawing compact, the tree is not visualized by applying a tree but a graph layout algorithm. In order to draw MSTs of considerable size (millions of vertices), a spring-electrical model layout algorithm with multilevel multipole-based force approximation is applied. This algorithm is provided by the open graph drawing framework (OGDF), a modular C++ library [40 ]. In addition, the use of the OGDF allows for effortless adjustments to the graph layout algorithm in terms of both aesthetics and computational time requirements. Whereas several parameters can be configured for the layout phase, only parameter $p$ must be adjusted based on the size of the input data set (Additional file 1: Fig. S3). This phase constitutes the bottleneck regarding computational complexity.

After sequencing, reads were mapped to each genome reference sequence using the manufacturers’ alignment tools, tmap for PGM and blasr for PacBio (http://www.pacificbiosciences.com/products/software/algorithms). BWA [30 (link)] was used for mapping reads from the Illumina GAIIx, MiSeq and HiSeq. SAMtools [31 (link)] was then used to generate pileup and coverage information from the mapping output.

BLAST [16 (link)] was used to search the homology from the NCBI database. Amino acid sequences were deduced using ExPASy translate tool (http://web.expasy.org/translate/) [24 (link)]. The secondary structure of the COX1 in C. uncinata was determined by PSIPRED Protein Sequence Analysis Workbench (http://bioinf.cs.ucl.ac.uk/psipred) [25 (link)] to predict coil, strand, and helix chains of the COX1 protein. The transmembrane segment prediction was conducted and visualized by the EMBOSS packages tmap and pepwheel [26 (link)]. The 3D structure of COX1 protein was predicted by Swiss-Model online server (http://swissmodel.expasy.org/interactive) [27 (link)]. Swiss-PDB viewer [28 (link)] was used to check the stereochemical quality of the structure.
Aside from the COX1 gene sequence of C. uncinata obtained in this study, other sequences of COX1 were retrieved from the GenBank database using PhyloSuite v1.1.15 [29 (link)]. Sequences of the class Heterotrichea were selected as outgroups. Sequence alignments, best-fit model selection, and phylogenetic tree reconstruction were conducted using plug-in programs implemented in PhyloSuite: MAFFT v7.313 [30 (link)], ModelFinder [31 (link)], MrBayes v3.1.2 [32 (link)], and IQ-TREE v1.6.8 [33 (link)]. The GTR + I + F + G4 model was selected as the best model, which was used for both maximum likelihood (ML) and Bayesian inference (BI) methods. The ML analysis was performed by the IQ-TREE v1.6.8 [33 (link)] with 1000 bootstrap replicates. BI analysis was performed using MrBayes 3.2.6 [32 (link)] with 1,000,000 generations. Trees were sampled every 1000 generations with the initial 25% of sampled data discarded as burn-in. The ML and BI trees were visualized and edited using the Interactive Tree Of Life (iTOL v4) tool [34 (link)], with annotation files generated by PhyloSuite.

We tested the Spearman rank correlation between mean Ap-ha ERI in I. scapularis nymphs and adults and anaplasmosis incidence at the postal (ZIP) code tabulation area level gathered from the New York State Department of Health (NYSDOH) Communicable Disease Electronic Surveillance system as previously described (13 (link)). We assessed the correlation for each year, 2010–2018. We selected the Spearman rank test because of the underlying count data used to generate anaplasmosis incidence and Ap-ha ERI. We corrected the 18 correlation tests for multiple testing using the Bonferroni-Holm adjustment (24 (link)). We compared results of the Spearman rank tests with the results from a previous analysis using non–genotype-specific A. phagocytophilum ERI (13 (link)).
We tested for spatiotemporal interaction in the number of Ap-ha– and Ap-V1–infected I. scapularis nymphs and adults by year and across latitude and longitude categories using a generalized linear mixed model (GLMM) extension of zero-inflated negative binomial (ZINB) regression (25 ,26 (link)). ZINB regression accounted for overdispersion and excess zero-counts of Ap-ha–and Ap-V1–infected I. scapularis ticks, whereas the GLMM extension handled the repeated nature of our sampling scheme by allowing sampling sites to have varying intercepts. We binned tick collection data and corresponding PCR results by year, latitude, and longitude to increase the number of observations within each combination of covariates to fit the model. We binned tick data by year into 4 categories: 2008–2011, 2012–2014, 2015–2017, and 2018–2020. We binned tick collection sites by latitude into 3 categories: sites south of 42°N, at 42°N to 43°N, and north of 43°N. We binned collection sites by longitude into 3 categories: sites east of 74°W, from 74°W to 76°W, and west of 78°W. We built 4 models to analyze Ap-ha and Ap-V1 in I. scapularis nymphs and adults separately. We assessed interaction between year and latitude/longitude categories using the likelihood ratio test (LRT). We used the natural log of the total number of I. scapularis ticks of the target developmental stage (nymphs during sampling events in late May, June, July, and August; adult ticks during April, early May, October, November, and December) as an offset. We conducted data cleaning, generation of summary statistics, and data visualization using R version 4.0.3 (http://www.rstudio.com) and the dplyr (https://CRAN.R-project.org/package=dplyr), sf, ggplot2, and tmap R packages (27 (link)–29 ). We used the glmmTMB package in R (30 (link)) for modeling.

Low-quality and non-human RNA-seq reads were identified and removed from the analysis pipeline using the Kraken suite of quality control tools (25, 26 (link)). High-quality, trimmed, human RNA-seq reads were aligned to the human genome (GRCh37; hg19) using TMAP (v5.0.7) and gene counts were calculated using High-throughput sequence analysis in Python (HT-Seq) as described previously (24 (link)). Gene set enrichment analysis (GSEA, RRID:SCR_005724) comparing PDX samples, primary ovarian cancer ascites, and primary ovarian cancer solid tumors was analyzed using the GSEA software (http://www.gsea-msigdb.org/gsea/index.jsp; ref. 27 (link)). A total of 1,742 genes in the ovarian cancer dataset were compared with 7,871 gene sets from the Molecular Signature Database after filtering out for gene set size (minimum 15, maximum 500 genes/set). For comparison of PDX and primary ovarian cancer gene expression, nine solid tumor or ascites PDX samples were compared with 13 primary or omental ovarian tumors, and 10 primary ascites samples. Samples were divided into three groups: PDX, primary ascites, or primary solid tumor. Genes with significantly altered expression between groups were tabulated.
Deconvolution of normalized gene expression data was performed using the publicly available Carcinoma Ecotyper software (https://ecotyper.stanford.edu/carcinoma; ref. 28 (link)). Luca and colleagues analyzed 16 select tumor types, including ovarian serous cystadenocarcinoma, to identify the cellular composition and cell states based on gene expression clusters identified from single-cell RNA-seq datasets. The gene expression clusters can then be used to perform deconvolution on bulk RNA-seq tumor samples. Our ovarian cancer expression data from four PDX ascites samples, five PDX tumors, eight patient ascites samples, and 14 patient tumors were analyzed using this program. Output is given as the abundance by state for 12 cell types including immune cell types, cancer epithelial cells, endothelial cells, and fibroblasts. The average estimated abundance of monocytes/macrophages, CD4⁺ T cells, CD8⁺ T cells, and cancer epithelial cells by state is presented for the four sample types.