RAS 1.1.2

This release adds a second region detector, a disk-backed engine in compiled code for large cohorts, and support for external GWAS summary statistics as regional weights. The compiled engine is now the implementation behind the familiar names; the 1.0.3 implementations remain available under _original names.

Function names

• The compiled, disk-backed implementations are now the default and carry the plain names: ras(), ras_scan(), ras_detect(), compute_gwas_weights(), screen_forward_max_region(), ras_scan_external(). The pure-R, in-memory implementations that 1.0.x shipped under those names are available, unchanged, as ras_original(), ras_scan_original(), ras_detect_original(), compute_gwas_weights_original(), screen_forward_max_region_original() and ras_scan_external_original().
• ras() and ras_scan() keep the 1.0.x calling convention: the first argument is still geno and still accepts an in-memory genotype matrix (converted to a temporary .rasbin file for the run); it now also accepts the path to a .rasbin file. For continuous traits the profile is identical to 1.0.x to machine precision. For binary traits the default scan is now the Rao score test (scan_test = "score"); the 1.0.x per-window logistic regression is still available as scan_test = "glm", which runs ras_original() and needs a genotype matrix.
• The _fast names of the development versions (ras_fast(), ras_scan_fast(), …) are kept as aliases.

Two region detectors

• ras() gains detector = c("box", "changepoint"), default "box", the new box-scan region detector; "changepoint" is the original two-pass detector, ras_detect() followed by ras_validate(), which reproduces the published method and the results of 1.0.3. ras_original() keeps "changepoint" as its default.
• ras() calibrates the box-scan threshold on the data: unless box_threshold or box_calibration is given, the scan is repeated on box_null = 100 permuted phenotypes (Freedman-Lane residual permutation for continuous traits) and ras_box_calibrate() turns their maxima into a threshold at family-wise level box_alpha = 0.05 (exact order statistic, or a Gumbel fit with box_method = "gumbel" when only 20 to 50 permutations are affordable). box_null_n runs the null scans on a random subset of individuals, which makes the calibration affordable at biobank scale. The calibration is returned as $box_calibration.
• ras_scan() gains rows, the subset of samples within which the train/hold-out splits are drawn.
• New ras_box_detect(): scores every window of a set of widths by how far its mean stands above both adjacent flanks and above the profile background, in the raw -log10(p) units of the profile, and reports the intervals [tau_L, tau_R] that clear a threshold. Because the window must stand above each flank, a window inside a plateau or on the shoulder of a peak scores about zero, so broad plateau-shaped regions are delimited at their edges and neighbouring peaks stay separate. No noise scale is estimated. Regions at the ends of a profile are reported (edge = "open").
• New ras_box_calibrate(): turns null scan profiles (the same scan on permuted phenotypes) into the threshold ras_box_detect() needs, as an order statistic of the null maxima, with an optional held-out error rate and a Gumbel extrapolation to longer profiles. New ras_box_stat() exposes the statistic itself.
• print() and plot() for "ras" objects understand both detectors: for a box-scan result the shaded areas are the detected intervals and the overlay plot shows the box-scan score in place of the Davies test significance.

Disk-backed C engine

• New .rasbin genotype format: a flat column-major binary file from which an arbitrary SNP range is one seek and one read. geno_to_rasbin() converts an in-memory matrix or an .rds/.csv/.tsv file; bed_to_rasbin() converts a PLINK 1 .bed/.bim/.fam fileset in fixed-size SNP chunks without ever holding the dense matrix (int8 storage, 1 byte per genotype); rasbin_header() and rasbin_read_chunk() read it back.
• ras_scan() and ras(): the Stage-1 scan streams SNP chunks from a .rasbin file through C (compute_gwas_weights(), screen_forward_max_region()), so peak memory is bounded by chunk_snps rather than by the chromosome. The statistics are identical to ras_scan_original() (exact FWL for continuous traits, Rao score test for binary traits), and the profiles agree to machine precision.
• ras_detect(): a C port of the first-pass changepoint detector (Muggeo’s segmented fit with bootstrap restarts and the Davies test). Resampling draws come from R’s RNG stream, so results are statistically, not bit-for-bit, equal to ras_detect_original().
• ras_scan()/ras() take cores (repetitions dispatched to a PSOCK cluster) and keep_reps (return the per-repetition profiles as well as their mean).
• Binary-trait GWAS weights on the fast route handle SNPs with missing dosages by downdating shared covariate cross-products in blocks instead of refitting each such SNP on its own, a 20-fold saving on post-QC sequencing data where almost every SNP carries at least one missing value.

External summary statistics as weights

• New ras_harmonize_sumstats() (and the building blocks ras_weights_from_sumstats(), ras_sumstats_report(), ras_read_variant_dictionary(), ras_map_ids_to_rsid(), ras_liftover_sumstats()): match an external GWAS to the target genotype map by rsID or by position and allele pair, check the genome build, align effect alleles, drop strand-ambiguous variants, and return aligned weights together with a one-row-per-stage QC table.
• New ras_scan_external() (and the in-memory ras_scan_external_original()): the Stage-1 scan with fixed external weights over every individual, with no train/holdout split and no repetitions.

Fixes

• The package now builds on Linux systems whose C library is not glibc (for example Alpine Linux / musl): malloc_trim() is a glibc extension and is now compiled only where glibc is present; release_memory() returns NA elsewhere, as it already did on Windows and macOS; the unit tests check 0/1 or NA according to what the build compiled in. (1.1.0 and 1.1.1 were submitted and withdrawn for this reason; neither was released.)
• ras_validate() no longer stops on a non-finite Davies p-value from a degenerate window, and no longer keeps a candidate whose profile value is NA.

RAS 1.0.3

Resubmission addressing the CRAN reviewer’s comments on the 1.0.0 submission, with a bug fix folded in.

• Removed the \dontrun{} examples that called the unexported internal functions plot_ras_scan() and plot_ras_zoom_regions() through :::. These functions are internal (@keywords internal) and are already exercised by the runnable example of the exported plot() method for "ras" objects.
• Made the ras_memory() “abort” example runnable by wrapping the intentional stop() in try(), instead of hiding it in \dontrun{}.
• The plotting functions now capture the caller’s graphics parameters and restore them via an immediate on.exit(par(oldpar)), so plotting a "ras" object no longer leaves the user’s par() settings modified.
• Fixed a crash in screen_forward_max_region() (and therefore ras()) when covariates contained missing values. Incomplete cases are now dropped before the model matrix is built, matching the documented “removes incomplete cases” behaviour, on all three scan paths (continuous exact-FWL, binary score, and the legacy glm path).