rename_motif_models() now also renames interaction features. Previously, renaming a model that had interactions set the interaction columns of @model_features to NA and left @interactions on the old motif names, silently corrupting the model. (iq_regression() was unaffected - it renames before adding interactions.)
filter_traj_model() no longer crashes on small or aggressively-filtered models: it is a no-op on models with fewer than 2 motifs, never removes every motif (it keeps at least one so the model stays fittable), and its bits-based removal is now applied even when no feature also fails the R^2 threshold (previously it was silently skipped in that case).
The manifold/multi-trajectory beta filter (used by distill_traj_model_multi()) no longer crashes when unifying models that have no additional features, and is robust to single-motif or fully-filtered trajectories (it always keeps at least one motif).
add_interactions() is now a no-op (instead of crashing) on models with fewer than two features to interact, such as a single-motif model. This also fixes iq_regression() when filtering leaves a single motif.
regress_trajectory_motifs() no longer crashes or silently drops the motif name when the model is reduced to a single motif (e.g. max_motif_num = 1), including with include_interactions = TRUE.
add_interactions() now rescales predicted_diff_score to the trajectory’s accessibility-difference range (rescale_pred = TRUE), matching regress_trajectory_motifs(). Previously the default left predictions on the raw [0, 1] logistic scale, so adding interactions silently changed the units of predicted_diff_score (reported R^2 was unaffected, as it is correlation-based).
pbm_list.gextract(), pbm.gextract(), and pbm_list.multi_traj.gextract_energy() now extract sequences at the model’s trained size rather than the input intervals’ width. Previously, passing intervals whose width differed from the training peaks_size (e.g. 300bp peaks for a model trained at 500bp) produced incorrect energies in the documented inference path.
merge_trajectory_motifs() now removes the merged-away motifs from the model entirely (previously they were dropped from the model features but left behind in @motif_models and @normalized_energies).
add_interactions(min_signal_correlation = ...) no longer drops all interactions when none of them correlate with the signal.
compute_motif_energies(db_quantiles = ...) is documented more precisely: it normalizes against a stable precomputed reference instead of the observed quantile of the normalization intervals. This can change which motifs are selected during regression, but does not affect a trained model’s internal consistency or its predictions (energies are recomputed consistently at inference time).
Improvements
Faster model reports and add_interactions() on large motif sets; results are unchanged.
Breaking changes
Removed three exported functions that had no internal callers, no tests, and were never referenced from any vignette or public workflow: add_motif_models_to_traj(), adjust_energies(), adjust_motif_seq_lengths().
iq_regression(include_interactions = TRUE) now uses Akhiad-inspired tight single-pass defaults instead of the previous loose defaults:
interaction_threshold: 0.01 (was 0.001).
interaction_only_sig_motifs: TRUE (new argument, previously hard-coded to the add_interactions() default of FALSE). For exact numeric parity with 0.0.6 / paper / pycqream Phase 2 results, pass interaction_threshold = 0.001, interaction_only_sig_motifs = FALSE explicitly. Benchmarked on the gastrulation vignette data (5000 peaks, 21862-motif db) under two configurations:
At the default max_n_interactions cap (both configs bounded at 260 interactions): R^2 train/test 0.531/0.351 (0.0.7) vs 0.529/0.352 (0.0.6). Parity within measurement noise — expected, because the top-N correlation selection dominates when both configurations hit the same cap.
With max_n_interactions = 5000 (unbounded, probes the selection criteria directly): 0.0.7 produces 492 interactions at R^2 test 0.367; 0.0.6 produces 517 interactions at R^2 test 0.360. The tight default is more parsimonious (25 fewer interactions) and generalizes slightly better (+0.007 test R^2) — this is the evidence that motivates the default change.
iq_regression(strategy = "progressive") now errors at call time instead of warning. The default progressive builder (default_score_split_features()) causes silent test-R^2 collapse (~0.27 on gastrulation) because its helper-model predictions are defined only for training peaks and imputed to 0 at test. Errors are louder than warnings. The strategy = "progressive" path inside iq_regression is currently disabled — it will error unconditionally. Power users who want the two-pass workflow should call add_interactions_progressive() directly on a model that already spans train + test peaks and supply additional_features (with base_pred / end_pred) for both splits at inference time.
New features
Exported add_interactions_progressive() — N-pass interaction selection with a pluggable between-pass feature-injection hook. Mirrors the manual workflow used in Akhiad’s analysis notebooks.
Exported default_score_split_features() — helper that builds base_pred, end_pred, pred_diff_e_b engineered additional features from start- and end-bin ATAC scores by relearning the trajectory model twice. Using this as an additional_features_builder requires propagating the helper-model predictions to test peaks yourself — see ?default_score_split_features. Not safe to use inside iq_regression until test-time propagation lands (see the breaking-change note above — iq_regression(strategy = "progressive") errors unconditionally).
iq_regression() gains a strategy = c("single", "progressive") argument. "single" is the default. "progressive" is currently disabled (errors at call time — see the breaking-change note above); reserved for a future release once test-time propagation of the engineered features lands. Power users who want the two-pass workflow today should call add_interactions_progressive() directly on a traj_model that already spans train + test peaks and supply the engineered additional features explicitly at inference time.
add_interactions() gains interaction_scale_factor (multiplier on the normalized interaction matrix) and min_signal_correlation (post-filter: drop interactions with |cor(col, diff_score[train])| < threshold × max(|cor|)). min_signal_correlation = 1/8 mirrors the manual post-filter in Akhiad’s notebook.
Bug fixes
Fix: remove_interactions() silently failed to strip the logist-expanded interaction columns from @model_features when logist_interactions = TRUE, leaving orphan columns behind after a subsequent add_interactions(force = TRUE).
Fix: infer_trajectory_motifs() produced a “not valid for @‘interactions’” S4 validity error on any interaction-augmented model because rbind(matrix, data.frame) was assigned back to the matrix-only slot. Latent until now — no prior test exercised iq_regression(include_interactions = TRUE) through inference on the final model.
Fix: escape_vars() used a regex that made the escaped dot optional (\\.?), causing name-prefix matches to incorrectly hit variables that shared a prefix with the intended target (e.g. motif.low-energy also matched motiflow-energy).
Fix: cli_warn in infer_trajectory_motifs printed the full additional- features data frame instead of just the feature names.
Fix: plot_traj_model_clusters_report() would silently unlink(dir, recursive = TRUE) any user-supplied path — now refuses to delete /, $HOME, getwd(), or any path shallower than 3 components.
Fix: regress_trajectory_motifs() and regress_trajectory_motifs_manifold() (and therefore iq_regression()) now clamp kmer_sequence_length to peaks_size when it would exceed the extracted sequence length. Previously the mismatch (default kmer_sequence_length = 300, user peaks_size < 300) surfaced deep inside distill_motifs()’s parallel plyr::llply loop as the opaque task 1 failed - "kmer_sequence_length cannot be greater than the length" originating from prego::regress_pwm.
Improvements
relearn_traj_model() with use_cv = TRUE no longer refits glmnet after cv.glmnet; the path fit stored inside cv_model$glmnet.fit is reused. Saves one full fit per CV-enabled relearn.
Finished the partial tidyr::gather → pivot_longer and plyr::* → purrr::* migrations started in 0.0.6. plyr::llply(..., .parallel = TRUE) is retained in the hot paths where parallelism is active (distill-motifs.R, distill-multi-traj.R, inference.R, PBM.R).
Deferred
Integration with misha’s feat/glm-pred branch (glm_batch_quantiles(), glm_extract_features(), glm_pred.create() virtual tracks): deferred until the branch merges to misha master. Expected wins: 5–10× on per-motif quantile normalization, faster feature-matrix construction, and post-fit genome-wide scoring without per-position R round-trip. See .a5c/runs/deep-code-review-2026-04-19.md §3.
cv.glmnet parallelization via doParallel: deferred due to the OpenMP-oversubscription / deadlock risk when glmnet’s internal parallelism stacks with R-level foreach. Safe rollout requires forcing OMP_NUM_THREADS = 1 inside each worker and exposing it as opt-in via a package option. See .a5c/runs/deep-code-review-2026-04-19.md §5.2.
Test-time base_pred / end_pred propagation so that strategy = "progressive" with default_score_split_features doesn’t overfit on test. Would require inferring the base-only / end-only helper models on test peaks inside infer_trajectory_motifs.
iceqream 0.0.6
Bug fixes
Fix: prob1_thresh parameter in preprocess_data was silently ignored due to being passed as NULL instead of the computed value.
Fix: select_motifs_by_correlation fallback path always returned zero motifs.
Fix: distill_traj_model selected the weakest motif per cluster instead of the strongest (wrong sort order).
Fix: regress_trajectory_motifs_manifold modified the wrong object (mm instead of mm_new), discarding motif model updates.
Fix: iq_regression applied redundant TSS distance filtering twice (once in outer and inner call).
Fix: divide-by-zero in create_specific_terms and create_features_terms when interaction normalization factors are zero.
Improvements
New exported functions: norm_energy_dataset and strip_traj_model.
Migrated deprecated ggplot2 APIs: size aesthetic in line geoms replaced with linewidth; numeric legend.position replaced with legend.position.inside.
Virtual tracks created during preprocess_data, normalize_regional, and plotting functions are now cleaned up automatically via on.exit().
Graphics devices opened during report generation are now closed reliably via on.exit().
S4 validity checks added for IQFeature, PBM, and TrajectoryModel classes.
Fixed typo in function name: create_specifc_terms renamed to create_specific_terms.
Tests
Added 127 validation tests covering utility functions, S4 validity, refactored helpers, and API migration correctness (total: 575 tests).
iceqream 0.0.5
Bug fixes
Fix: plot_normalization_scatters no longer crashes when anchor_cell_type is NULL.
Fix: preprocess_data now correctly excludes the marginal track from cell type names, preventing size mismatch errors.
Fix: names(new_models_full) was incorrectly assigned in infer_trajectory_motifs_multi, leaving new_models_full unnamed.
Fix: S4 copy-on-modify bug in regress_trajectory_motifs_manifold where loop modifications to model params were silently discarded.
Fix: undefined variable symmetrize_spat in iq_regression when learning prego motifs de-novo.
Fix: add_type partial argument match corrected to add_types in infer_trajectory_motifs and get_model_coefs.
Improvements
preprocess_data: renamed peaks parameter to peak_intervals for API consistency.
regress_trajectory_motifs: bin_start and bin_end now accept column names (character) in addition to integer indices.
validate_atac_scores: relaxed validation to allow bin_start == bin_end only when they refer to different columns (removed bin_start < bin_end constraint).
Improved documentation linking preprocess_data()$atac_norm_prob to the atac_scores parameter in iq_regression and regress_trajectory_motifs.
Added input validation for norm_energy_matrix (matrix coercion, column name alignment).
Added input validation for interaction terms in infer_trajectory_motifs.
Tests
Added 423 unit tests covering energy utilities, S4 classes (PBM, TrajectoryModel, IQFeature family), validation functions, and preprocessing helpers.
iceqream 0.0.4
Fix: iq_regression now filters also test set peaks that are too close to TSS.
