Python 3

The proteins identified by MaxQuant were excluded only in one sample. The quantitative data for each protein were normalized by dividing the LFQ values by the maximum of the LFQ value of the protein in all samples. A two-parameter linear model was developed for each protein:
where T is a period of time from sampling to therapy initiation and T₀ is a period of time from diagnostics to sampling (both in months). To estimate the quality of the models we used the following well-known metrics: the coefficient of determination R² as well as the F-values and p-values corresponding to the model coefficients.
In addition, a linear regression model T on T₀ and P₁ was built,

where P₁ is a projection of the vector of the LFQ values to the first principal component. In order to estimate the adequacy of this model, in addition to standard metrics, we used leave-three-out cross-validation [14 ]—the procedure of multiple calculations of the quality metrics using different splits of the data set to train/test sets.
All steps of data analysis were performed in Anaconda Python 3.

No statistical methods were used to predetermine sample size. Owing to the exploratory nature of our study, we did not use randomization or blinding. To compare weight amplitude distributions (Figures 5A and 5B) we used the paired Wilcoxon Rank Sum Test, taking paired components as the input (i.e., comparing red-light-transient versus green-light-transient, and so on). To assess weight correlations between cones (Figures 5C–5E and S2), we in each case list the Pearson correlation coefficient ρ and 95% confidence intervals (CI) based on the mean weights per cluster. Individual temporal weights were not considered in this analysis. All statistical analysis was performed in Python 3 (Anaconda) and/or Igor Pro 6 (Wavemetrics).

Data processing was conducted using the Anaconda distribution of Python 3.6 (Anaconda.org) and associated libraries—scikit-learn, pandas, numpy, visualization tools matplotlib, and seaborn and the statistical library scipy within the Jupyter lab environment. Python Jupyter Notebooks are available on github (at https://github.com/g-patlewicz/inhalation-ttc) and all the datasets are posted on the EPA FTP website (ftp://newftp.epa.gov/Computational_Toxicology_Data/CCTE_Publication_Data/CCED_Publication_Data/PatlewiczGrace/Frontiers-TTC/).

The CV and EIS data for each sample were transformed into a set of numerical features. Furtermire, we added class information “LC”/“H” (“lung cancer”/“healthy”). The relationship between features and class as a classification problem was studied. The whole analysis was performed using Anaconda Python 3.6. EIS models were simulated using ZView software (Scribner Associates Inc.; Southern Pines, NC, USA). The models obtained were fitted with the experimental data, and the parameters of equivalent circuits were calculated.

Data are publicly available from the Digital Database for Screening Mammography (DDSM) [30 ] and the Curated Breast Cancer Imaging Subset of DDSM [31 (link)] and provided by Google’s Kaggle.com, accessed on 5 January 2021 [32 ]. This analysis is a precursor to follow-on work supported by a National Cancer Institute Data Transfer Use Agreement (PLCOI-742). The data consisted of 71,249 images and labels, 55,885 pre-designated for training and 15,364 reserved for testing. The image data were sized 299 × 299 (single channel grayscale) but were augmented to three-color (‘RGB’) for use in models by replicating the channel. The label data included dichotomous classification (0 for true negative, 1 otherwise). All analyses were performed in Anaconda Python 3.7 and are available on Github [33 ]. An in-kind high performance computing grant from Advanced Micro Devices (250 teraflops computing power) provided the computational power for model training. The observations (pixels) used for the image data only were 299 pixels × 299 pixels × 3 channels × 71,249 images = 19.1 billion pixels.