src/02.1_assemble_datasets.R

#' Assemble a Rosetta models dataset
#'
#' @param data_path character directory .pdb.gz files are located
#' @param output_path character output .parquet path
#'
#' Write output_path .parquet file with columns
#'   <id> <pdb> [<scores>]
#' where [<scores>] are key-value entries following
#' the TER line in each .pdb.gz file
assemble_rosetta_models <- function(
    data_path,
    output_path) {

    cat(
        "data path: ", data_path, "\n",
        "output path: ", output_path, "\n",
        sep = "")

    file_index <- 1
    models <- list.files(
        path = data_path,
        full.names = TRUE,
        pattern = "*.pdb.gz",
        recursive = TRUE) |>
        purrr::map_dfr(.f = function(path) {
            file_handle <- path |>
                file(open = "rb") |>
                gzcon()

            if( file_index %% 1000 == 0) {
                cat("Reading  '", path, "' ", file_index, "\n", sep = "")
            }
            file_index <<- file_index + 1

            lines <- file_handle |> readLines()
            file_handle |> close()

            ter_line_index <- which(
                lines |> stringr::str_detect("^TER"),
                arr.ind = TRUE)

        lines[(ter_line_index + 1) : (length(lines) - 1)] |>
            paste0(collapse = "\n") |>
            readr::read_delim(
                delim = " ",
                col_names = c("key", "value"),
                show_col_types = FALSE) |>
            dplyr::mutate(
                id = path |>
                    basename() |>
                    stringr::str_replace_all(".pdb.gz", ""),
                .before = 1) |>
            dplyr::mutate(
                pdb = lines[1:ter_line_index] |> paste0(collapse = "\n"))

        })
    models <- arrow::arrow_table(models)
    models$pdb <- models$pdb$cast(arrow::string())
    models |> arrow::write_parquet(output_path)
}


# call assemble_rosetta_models for the high_quality and low_quality datasets
dataset_tag <- "rosetta_high_quality_models"
assemble_rosetta_models(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))

dataset_tag <- "rosetta_low_quality_models"
assemble_rosetta_models(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))

#' Assemble a DMP-Fold models dataset
#'
#' @param data_path character directory .pdb.gz files are located
#' @param output_path character output .parquet path
#'
#' Write output_path .parquet file with columns
#'   <id> <pdb>
#'
#' Note that dmpfold doesn't write out score lines like Rosetta
assemble_dmpfold_models <- function(
    data_path,
    output_path) {

    cat(
        "data path: ", data_path, "\n",
        "output path: ", output_path, "\n",
        sep = "")

    file_index <- 1
    models <- list.files(
        path = data_path,
        full.names = TRUE,
        pattern = "*.pdb.gz",
        recursive = TRUE) |>
        purrr::map_dfr(.f = function(path) {
            file_handle <- path |>
                file(open = "rb") |>
                gzcon()

            if (file_index %% 1000 == 0) {
                cat("Reading  '", path, "' ", file_index, "\n", sep = "")
            }
            file_index <<- file_index + 1

            lines <- file_handle |> readLines()
            file_handle |> close()

            ter_line_index <- which(
                lines |> stringr::str_detect("^TER"),
                arr.ind = TRUE)

            data.frame(
                id = path |> basename() |> stringr::str_replace_all(".pdb.gz", ""),
                pdb = lines[1:ter_line_index] |> paste0(collapse = "\n"))
        })
    models |>
        arrow::write_parquet(output_path)
}

# call assemble_rosetta_models for the high_quality and low_quality datasets
dataset_tag <- "dmpfold_high_quality_models"
assemble_dmpfold_models(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))

dataset_tag <- "dmpfold_low_quality_models"
assemble_dmpfold_models(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))


####################################
##                                ##
##  Assemble Function Predictions ##
##                                ##
####################################


#' Assemble DeepFRI Function Prediction dataset
#'
#' @param data_path character directory where *_pred_scores.json.gz files are located
#' @param output_path character output .parquet path
#'
#' Write output_path .parquet file with columns
#'   <id> <term_id> <term_name> <Y_hat>
#'
#' <id>: Structure identifier like `MIP_00004873`
#' <term_ontology>: term onlogy, one of [BP, CC, EC, or MF]
#' <term_id>: GO or EC term identifiers like `GO:0009225`
#' <term_name> is the description of the term
assemble_DeepFRI_function_predictions <- function(
    data_path,
    output_path) {

    cat(
        "data path: ", data_path, "\n",
        "output path: ", output_path, "\n",
        sep = "")

    file_index <- 1

    scores <- c("BP", "CC", "EC", "MF") |>
        purrr::map_dfr(.f = function(ontology) {
            cat("Reading predictions cores for ontology ", ontology, "\n", sep = "")
            list.files(
                path = data_path,
                full.names = TRUE,
                pattern = paste0("*_", ontology, "_pred_scores.json.gz"),
                recursive = TRUE) |>
                purrr::map_dfr(.f = function(path) {
                    cat("Reading  '", path, "' ", file_index, "\n", sep = "")
                    file_index <<- file_index + 1

                    data <- jsonlite::fromJSON(txt = path)

                    scores <- as.data.frame(data$Y_hat)
                    names(scores) <- data$goterms
                    scores <- scores |>
                        dplyr::mutate(
                            id = data$pdb_chains,
                            .before = 1) |>
                        tidyr::pivot_longer(
                            cols = -"id",
                            names_to = "term_id",
                            values_to = "Y_hat") |>
                        dplyr::left_join(
                            data.frame(
                                term_ontology = ontology,
                                term_id = data$goterms,
                                term_name = data$gonames),
                            by = "term_id") |>
                        dplyr::select(
                            id,
                            term_ontology,
                            term_id,
                            term_name,
                            Y_hat)
                })
        })

    scores |>
        arrow::write_parquet(output_path)

}

# call assemble_rosetta_models for all the datasets
dataset_tag <- "rosetta_high_quality_function_predictions"
assemble_DeepFRI_function_predictions(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))

dataset_tag <- "rosetta_low_quality_function_predictions"
assemble_DeepFRI_function_predictions(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))


dataset_tag <- "dmpfold_high_quality_function_predictions"
assemble_DeepFRI_function_predictions(
    data_path = paste0(
        "data/mxoicrobiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))

dataset_tag <- "dmpfold_low_quality_function_predictions"
assemble_DeepFRI_function_predictions(
    data_path = paste0(
        "data/microbiome_immunity_project_dataset/dataset/",
        dataset_tag),
    output_path = paste0("intermediate/", dataset_tag, ".parquet"))