1.9.0

1.9.0

Features

• STEAM models (STEAM_v1, STEAM_evolution_aware, STEAM_evolution_naive) from the evolutionary transfer learning paper are now available in the model repository via {func}crested.get_model. (#217)
• {func}crested.tl.contribution_scores and {func}crested.tl.contribution_scores_specific now take a simplex_correction argument. When enabled, gradient-based scores are projected back onto the probability simplex by subtracting the per-position mean across the four nucleotide channels, reducing off-simplex noise as proposed in Majdandzic et al. (2023). (#218)
• {func}crested.tl.modisco.process_patterns now defaults to deterministic, order-independent agglomerative clustering (full pairwise TOMTOM + single cut at sim_threshold) instead of greedy leader clustering. New clustering, linkage_method, sort_by, and representative parameters; pass clustering="greedy" to reproduce earlier analyses. (#216)
• {func}crested.tl.modisco.create_tf_ct_matrix now defaults to NNLS-based TF selection (selection="nnls"): a non-negative ridge deconvolution that competes down broadly-expressed binders, followed by an expression-relevance gate (rel_keep_frac). Replaces the name-based paralog collapse; pass selection="threshold" for the original per-column gates. (#216)
• {func}crested.pl.modisco.clustermap_tf_motif gains a categorical row-class color strip and a per-column species stack. (#216)
• calculate_mean_expression_per_cell_type computes per-cell-type means on sparse X directly, avoiding densification on atlas-scale inputs. (#216)
• {func}crested.tl.modisco.tfmodisco can now run classes in parallel via the n_jobs argument (each class is independent), with workers pinned to a single thread to avoid oversubscription. joblib is now an explicit dependency. (#215)
• {func}crested.tl.contribution_scores_specific now takes a skip_existing argument, which skips classes already saved to output_dir so a long run can be resumed without recalculating finished classes. (#213)
• {func}crested.tl.zoo.borzoi now takes an absolute_positions argument to use absolute positional encoding in the multi-head attention instead of the default relative positions. (#211)

Bugfixes

• {func}crested.pp.change_regions_width now checks the resized coordinates (not the original ones) in its boundary filter, so regions whose widened window falls off a contig edge are correctly dropped instead of producing negative-start region names. Most apparent on assemblies with many short scaffolds (e.g. rheMac10). (#214)
• Agglomerative cluster IC is now the best member's IC (not the representative's), fixing discard/sort under a non-IC representative. (#216)
• {func}crested.tl.modisco.process_patterns skips classes with no matched pattern file instead of raising. (#216)
• {func}crested.tl.modisco.create_tf_ct_matrix skips constant-expression columns, avoiding divide-by-zero. (#216)
• _trim_pattern_by_ic no longer over-truncates/crashes on flat patterns. (#216)

Performance

• {obj}crested.Genome now reads the FASTA file first into the file cache before using fetch, speeding up region fetching on slow remote file systems. (#207)

Documentation

• Fixed the pinned pandas version used for intersphinx links. (#209)
• Enhancer code analysis notebook updated for the agglomerative + NNLS pipeline. (#216)
• Clarified that min_tf_gex is applied to the (optionally log-transformed) expression. (#216)