NEWS

---

HiCPotts 1.1.1 (2026-04-21)

All changes involve improving, tidying and simplifying the codes and removing bugs

It also makes it easier for non-statisticians to understand

• run_metropolis_MCMC_betas() is now defined ONLY in R/RcppExports.R (plus its C++ backend). The duplicate R version in R/MCMC_chain.R has been deleted; it referenced an undefined mc_cores variable and used the wrong PACKAGE = "HMRFHiC" tag. Same for pz_123() in the now-deleted R/potts_posterior.R.
• likelihood_combined() for components 2 and 3 more simplified and efficient to evaluates the density on EVERY observation in the component. This makes it easier for non-statisticians to understand
• ZIP/ZINB zero-probability is now log(theta + (1-theta) * exp(-mu)) where mu = exp(eta) is the Poisson/NB mean. 1.0.0 used exp(-eta). Changed in both R/likelihood.R and src/potts_posterior.cpp.
• Neighbours_combined() (C++) rewritten to simplify the comparing of the reference configuration (current state or proposed state) to each of the four shifted copies of the current state.
• run_metropolis_MCMC_betas() ABC step for gamma uses the PROPOSED gamma to simulate synthetic data. 1.0.0 used chain_gamma[iter]. This helps to improve the accept/reject step.
• The ABC tolerance epsilon and the distance statistic are now on the same scale: mean |y - synthetic| (or RMS for distance_metric = "euclidean"), and epsilon defaults to the 10% quantile of per-cell |y - synthetic|. 1.0.0 compared a mean-difference distance against the 20% quantile of exp(|y - synthetic|).
• The Metropolis-Hastings ratio for the regression coefficients no longer includes a proposaldensity_combined() term. The proposal is a symmetric Gaussian random walk, for which the proposal-density ratio cancels.
• prior_combined() is now more DETERMINISTIC given the same parameters, data, and component assignments. 1.0.0 called rnorm() and rgamma() inside the prior, so pi(theta) changed between calls and the MH acceptance ratio was stochastic.
• compute_HMRFHiC_probabilities() no longer caps interactions at 500. The previous behaviour is available via max_interactions = 500L. burnin is now an integer that includes the last MCMC sample (fixes fractional-index warnings when iterations is odd).
• A new consistent_dist = TRUE default in compute_HMRFHiC_probabilities() uses the same emission distribution for all three components. Pass consistent_dist = FALSE for the previous 1.0.0 behaviour (ZIP/ZINB only for component 1, others reduced to Poisson/NB).
• compute_HMRFHiC_probabilities() now uses log-sum-exp normalisation, avoiding underflow for very small component densities.
• On Windows, mc_cores > 1 now emits a warning and falls back to 1 (parallel::mclapply does not fork on Windows).
• get_data() now raises an informative error when passed a .hic file (these are not HDF5 and were failing silently inside h5ls).
• The slice-rebin path in get_data() now aggregates duplicate entries produced by splitting a larger native bin into smaller target bins, instead of silently keeping only the last.
• posterior_combined() now calls likelihood_combined() with named arguments, eliminating positional fragility.
• process_data() now defaults to standardization_y = FALSE. When scaling is requested, it emits a warning because min-max rescaling of count data breaks the count-regression assumption.

PACKAGE METADATA

• BSgenome.Dmelanogaster.UCSC.dm6 moved from Imports to Suggests. Installing HiCPotts no longer forces users of other species to download the Drosophila genome.
• strawr removed from Imports (it was never called). Added to Suggests for users wanting to handle raw .hic files themselves.
• DESCRIPTION license tag (GPL-3), README badge (MIT), and documented R version (>= 4.2) reconciled: GPL-3 and R (>= 4.5) are now the package's authoritative values.
• Vignette updated: correct function name (get_data, not prepare_data); user_fixed_priors example restructured with correct field names; ZIP/ZINB chunk consistency fixed.

TESTING

• tests/testthat replaced with a suite that verifies numerical correctness (closed-form Poisson/NB/ZIP densities, hand-computed neighbour counts, and regression tests for every fix above).