faster modisco for decima

[MuhammedHasan]

Speed up for the decima modisco-pattern function:
Old modisco with #22

s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
3948.6744       1:05:48 9562.70 18040.41        9521.04 9538.74 0.00    15611.66        261.27  10318.95

New modisco jmschrei/tfmodisco-lite#71

s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
1033.2864       0:17:13 15015.64        49914.56        12574.11        12719.28        0.00    3122.35 859.06  9723.68

Speed up for the decima modisco-reports function:
Old modisco

s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
736.9390        0:12:16 1065.41 6516.85 1018.21 1036.29 0.00    0.63    50.92   533.50

New modisco

s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
234.5583        0:03:54 6290.30 40260.50        4804.80 4889.07 0.00    0.70    189.41  548.03

Runtime of attribution prediction shared and takes on L40 for each replicate:

s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
336.5114        0:05:36 45098.63        229566.50       7871.31 12046.93        0.00    3842.39 98.06   334.85
s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
340.7673        0:05:40 46461.95        231611.04       8226.79 12526.48        0.00    3842.41 95.93   331.32
s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
343.4498        0:05:43 46129.14        231169.75       7997.02 12321.96        0.00    3842.39 95.39   331.99
s       h:m:s   max_rss max_vms max_uss max_pss io_in   io_out  mean_load       cpu_time
337.7918        0:05:37 46199.02        231601.22       8314.75 12585.25        0.00    3842.39 98.30   337.34

Code to call attributions:

from decima.interpret.modisco import predict_save_modisco_attributions


cell_types = [
    "Amygdala excitatory",
    "CGE interneuron",
    "Cerebellar inhibitory",
    "Deep-layer corticothalamic and 6b",
    "Deep-layer intratelencephalic",
    "Deep-layer near-projecting",
    "Eccentric medium spiny neuron",
    "Hippocampal CA1-3",
    "Hippocampal CA4",
    "Hippocampal dentate gyrus",
    "LAMP5-LHX6 and Chandelier",
    "Lower rhombic lip",
    "MGE interneuron",
    "Mammillary body",
    "Medium spiny neuron",
    "Midbrain-derived inhibitory",
    "Splatter",
    "Thalamic excitatory",
    "Upper rhombic lip",
    "Upper-layer intratelencephalic",
]
off_cell_types = [
    "Astrocyte",
    "Bergmann glia",
    "Choroid plexus",
    "Ependymal",
    "Microglia",
    "Oligodendrocyte",
]

predict_save_modisco_attributions(
    output_prefix='neuron',
    tasks=f'cell_type in {cell_types} and organ == "CNS"',
    off_tasks=f'cell_type in {off_cell_types} and organ == "CNS"',
    model=0,
    top_n_markers=250,
    num_workers=16,
)

The step to run modisco:

modisco_patterns(
    output_prefix='neuron',
    attributions=['neuron.attributions_0.h5', 'neuron.attributions_1.h5', 'neuron.attributions_2.h5', 'neuron.attributions_3.h5'],
    tasks=f'cell_type in {cell_types} and organ == "CNS"',
    off_tasks=f'cell_type in {off_cell_types} and organ == "CNS"',
    top_n_markers=250,
    tss_distance=10_000,
    max_seqlets_per_metacluster=10_000,
    num_workers=16,
)

and step to generate motifs:

modisco_reports(
    output_prefix='neuron',
    modisco_h5='neuron.modisco.h5',
    num_workers=16,
)

[Copilot]

Pull Request Overview

This PR significantly improves the performance of ModISCo (motif discovery) functions in Decima by upgrading to a faster ModISCo implementation and adding new functionality. The primary changes include updating to a faster ModISCo library version, parallelizing attribution loading, and adding seqlet bed file generation.

Key changes:

• Performance optimization through faster ModISCo library and parallel processing
• New motif utility functions for information content, trimming, and motif positioning
• Addition of seqlet bed file generation functionality

Reviewed Changes

Copilot reviewed 11 out of 12 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
setup.cfg	Updates to faster ModISCo library dependency
src/decima/utils/motifs.py	New utility functions for motif analysis and trimming
src/decima/interpret/modisco.py	Major refactoring with parallelization and new seqlet bed functionality
src/decima/core/attribution.py	Parallel processing support for attribution loading
src/decima/utils/io.py	BigWig writing improvements and gradient correction options
src/decima/hub/init.py	Better error handling for missing model/metadata files
src/decima/cli/*	CLI updates to expose new parameters and functionality
tests/*	New test coverage for motif utilities and updated existing tests

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}