DiChIPMunk: on peaks and on multi-task seqlets and on single-task seqlets

HOMER: on peaks and on multi-task seqlets and on single-task seqlets

MEME: on peaks and on multitask seqlets and on single-task seqlets

In [1]:
import sys
import os
sys.path.append(os.path.abspath("/users/amtseng/tfmodisco/src/"))
import numpy as np
import h5py
import matplotlib.pyplot as plt
import motif.read_motifs as read_motifs
import motif.match_motifs as match_motifs
from util import figure_to_vdom_image
import plot.viz_sequence as viz_sequence
import vdom.helpers as vdomh
from IPython.display import display
In [2]:
# Define parameters/fetch arguments
tf_name = os.environ["TFM_TF_NAME"]
multitask_fold = int(os.environ["TFM_MULTITASK_FOLD"])
head = os.environ["TFM_HEAD"]
assert head in ("profile", "count")
if "TFM_TASK_INDEX" in os.environ:
    task_index = int(os.environ["TFM_TASK_INDEX"])
    singletask_fold = int(os.environ["TFM_SINGLETASK_FOLD"])
else:
    task_index = None
    singletask_fold = None
motif_file = os.environ["TFM_MOTIF_FILE"]  # TF-MoDISco motifs
    
print("TF name: %s" % tf_name)
print("Multi-task fold: %s" % multitask_fold)
print("Head: %s" % head)
print("Task index: %s" % task_index)
print("Single-task fold: %s" % singletask_fold)
print("TF-MoDISco motif file: %s" % motif_file)

TF name: NR3C1-reddytime
Multi-task fold: 5
Head: count
Task index: 0
Single-task fold: 5
TF-MoDISco motif file: /users/amtseng/tfmodisco/results/reports/tfmodisco_results/cache/multitask_profile_finetune/NR3C1-reddytime_multitask_profile_finetune_fold5/NR3C1-reddytime_multitask_profile_finetune_task0_fold5_count/all_motifs.h5
In [3]:
# Define paths and constants
base_path = "/users/amtseng/tfmodisco/results/classic_motifs/"

multitask_seqlets_dir = os.path.join(
    base_path, "seqlets", "multitask_profile_finetune",
    "%s_multitask_profile_finetune_fold%s" % (tf_name, multitask_fold)
)

if task_index is None:
    peaks_path = os.path.join(base_path, "peaks", tf_name, "%s_peaks_taskall" % tf_name)
    multitask_seqlets_path = os.path.join(
        multitask_seqlets_dir,
        "%s_seqlets_%s_taskall" % (tf_name, head)
    )
else:
    peaks_path = os.path.join(base_path, "peaks", tf_name, "%s_peaks_task%d" % (tf_name, task_index))
    multitask_seqlets_path = os.path.join(
        multitask_seqlets_dir,
        "%s_seqlets_%s_task%d" % (tf_name, head, task_index)
    )
    
    singletask_seqlets_dir = os.path.join(
        base_path, "seqlets", "singletask_profile_finetune",
        "%s_singletask_profile_finetune_fold%s" % (tf_name, singletask_fold),
        "task_%d" % task_index
    )
    singletask_seqlets_path = os.path.join(
        singletask_seqlets_dir,
        "%s_seqlets_%s_task%d" % (tf_name, head, task_index)
    )

Helper functions

In [4]:
def import_tfmodisco_motifs(motif_file):
    """
    Imports a set of motifs from the saved HDF5.
    Returns a list of motifs as L x 4 arrays, and a parallel list of
    motif names.
    """
    motifs, motif_names = [], []
    with h5py.File(motif_file, "r") as f:
        for key in f.keys():
            motif_names.append(key)
            motifs.append(f[key]["pfm_trimmed"][:])
    return motifs, motif_names
In [5]:
def pfm_to_forward_pwm(pfm):
    """
    Converts PFM to PWM and flips to purine-rich orientation.
    """
    pwm = read_motifs.pfm_to_pwm(pfm)
    if np.sum(pwm[:, [0, 2]]) < 0.5 * np.sum(pwm):
        # Flip to purine-rich version
        pwm = np.flip(pwm, axis=(0, 1))
    return pwm
In [6]:
def show_matched_motifs_table(
    tfm_motifs, tfm_names, bench_type, bench_motifs, bench_scores, matches
):
    """
    Shows a table of motifs from the given matched results.
    `bench_type` is either `dichipmunk`, `homer`, or `meme`.
    `matches` is NumPy array parallel to `bench_motifs`, where
    each entry is the index of the match in `tfm_motifs`, and -1
    if no match was found.
    """
    assert len(matches) == len(bench_motifs)
    assert bench_type in ("dichipmunk", "homer", "meme")
    if bench_type == "dichipmunk":
        score_name = "Supporting sequences"
        bench_name = "DiChIPMunk"
    elif bench_type == "homer":
        score_name = "Log enrichment"
        bench_name = "HOMER"
    else:
        score_name = "E-value"
        bench_name = "MEME"
        
    colgroup = vdomh.colgroup(
        vdomh.col(style={"width": "5%"}),
        vdomh.col(style={"width": "45%"}),
        vdomh.col(style={"width": "5%"}),
        vdomh.col(style={"width": "45%"})
    )
    header = vdomh.thead(
        vdomh.tr(
            vdomh.th("TF-MoDISco motif ID", style={"text-align": "center"}),
            vdomh.th("TF-MoDISco PWM", style={"text-align": "center"}),
            vdomh.th(score_name, style={"text-align": "center"}),
            vdomh.th("%s PWM" % bench_name, style={"text-align": "center"})
        )
    )
    
    # First, show a table of matches
    display(vdomh.p("Matched motifs"))

    body = []
    for bench_index, tfm_index in enumerate(matches):
        if tfm_index == -1:
            continue
        tfm_fig = viz_sequence.plot_weights(
            pfm_to_forward_pwm(tfm_motifs[tfm_index]), figsize=(20, 4), return_fig=True
        )
        tfm_fig.tight_layout()
        bench_fig = viz_sequence.plot_weights(
            pfm_to_forward_pwm(bench_motifs[bench_index]), figsize=(20, 4), return_fig=True
        )
        bench_fig.tight_layout()
        
        body.append(
            vdomh.tr(
                vdomh.td(str(tfm_names[tfm_index])),
                vdomh.td(figure_to_vdom_image(tfm_fig)),
                vdomh.td(str(bench_scores[bench_index])),
                vdomh.td(figure_to_vdom_image(bench_fig))
            )
        )

    display(vdomh.table(colgroup, header, vdomh.tbody(*body)))
    plt.close("all")
    
    # Now, show a table of leftovers
    display(vdomh.p("Motifs without matches"))
    
    tfm_lone_indices = np.setdiff1d(np.arange(len(tfm_motifs)), matches)
    bench_lone_indices = np.where(matches == -1)[0]

    rows = []
    for i in range(max(len(tfm_lone_indices), len(bench_lone_indices))):
        row = []
        if i < len(tfm_lone_indices):
            tfm_fig = viz_sequence.plot_weights(
                pfm_to_forward_pwm(tfm_motifs[tfm_lone_indices[i]]), figsize=(20, 4), return_fig=True
            )
            tfm_fig.tight_layout()
            row.extend([
                vdomh.td(str(tfm_names[tfm_lone_indices[i]])),
                vdomh.td(figure_to_vdom_image(tfm_fig))
            ])
        else:
            row.extend([vdomh.td(), vdomh.td()])
        if i < len(bench_lone_indices):
            bench_fig = viz_sequence.plot_weights(
                pfm_to_forward_pwm(bench_motifs[bench_lone_indices[i]]), figsize=(20, 4), return_fig=True
            )
            bench_fig.tight_layout()
            row.extend([
                vdomh.td(str(bench_scores[bench_lone_indices[i]])),
                vdomh.td(figure_to_vdom_image(bench_fig))
            ])
        else:
            row.extend([vdomh.td(), vdomh.td()])
        rows.append(row)
    body = [vdomh.tr(*row) for row in rows]

    display(vdomh.table(colgroup, header, vdomh.tbody(*body)))
    plt.close("all")

Import motifs

In [7]:
tfm_pfms, tfm_names = import_tfmodisco_motifs(motif_file)
In [8]:
dichipmunk_peak_pfms, dichipmunk_peak_num_seqs = read_motifs.import_dichipmunk_pfms(
    os.path.join(peaks_path, "dichipmunk")
)
dichipmunk_multitask_seqlet_pfms, dichipmunk_multitask_seqlet_num_seqs = read_motifs.import_dichipmunk_pfms(
    os.path.join(multitask_seqlets_path, "dichipmunk")
)

homer_peak_pfms, homer_peak_enrichments = read_motifs.import_homer_pfms(
    os.path.join(peaks_path, "homer")
)
homer_multitask_seqlet_pfms, homer_multitask_seqlet_enrichments = read_motifs.import_homer_pfms(
    os.path.join(multitask_seqlets_path, "homer")
)

meme_peak_pfms, meme_peak_evalues = read_motifs.import_meme_pfms(
    os.path.join(peaks_path, "meme")
)
meme_multitask_seqlet_pfms, meme_multitask_seqlet_evalues = read_motifs.import_meme_pfms(
    os.path.join(multitask_seqlets_path, "meme")
)

if task_index is not None:
    dichipmunk_singletask_seqlet_pfms, dichipmunk_singletask_seqlet_num_seqs = read_motifs.import_dichipmunk_pfms(
        os.path.join(singletask_seqlets_path, "dichipmunk")
    )
    homer_singletask_seqlet_pfms, homer_singletask_seqlet_enrichments = read_motifs.import_homer_pfms(
        os.path.join(singletask_seqlets_path, "homer")
    )
    meme_singletask_seqlet_pfms, meme_singletask_seqlet_evalues = read_motifs.import_meme_pfms(
        os.path.join(singletask_seqlets_path, "meme")
    )

Match benchmarked motifs to TF-MoDISco motifs

DiChIPMunk on peaks

In [9]:
dichipmunk_peak_matches = match_motifs.match_motifs_to_targets(dichipmunk_peak_pfms, tfm_pfms)
show_matched_motifs_table(
    tfm_pfms, tfm_names, "dichipmunk", dichipmunk_peak_pfms, dichipmunk_peak_num_seqs,
    dichipmunk_peak_matches
)

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWMSupporting sequencesDiChIPMunk PWM
0_0
568
0_1
228
0_3
50

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWMSupporting sequencesDiChIPMunk PWM
0_2
26
0_4
5
1

DiChIPMunk on multi-task seqlets

In [10]:
dichipmunk_multitask_seqlet_matches = match_motifs.match_motifs_to_targets(dichipmunk_multitask_seqlet_pfms, tfm_pfms)
show_matched_motifs_table(
    tfm_pfms, tfm_names, "dichipmunk", dichipmunk_multitask_seqlet_pfms, dichipmunk_multitask_seqlet_num_seqs,
    dichipmunk_multitask_seqlet_matches
)

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWMSupporting sequencesDiChIPMunk PWM
0_0
743
0_1
412
0_2
127
0_3
40

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWMSupporting sequencesDiChIPMunk PWM
0_4
1

DiChIPMunk on single-task seqlets

In [11]:
if task_index is not None:
    dichipmunk_singletask_seqlet_matches = match_motifs.match_motifs_to_targets(dichipmunk_singletask_seqlet_pfms, tfm_pfms)
    show_matched_motifs_table(
        tfm_pfms, tfm_names, "dichipmunk", dichipmunk_singletask_seqlet_pfms, dichipmunk_singletask_seqlet_num_seqs,
        dichipmunk_singletask_seqlet_matches
    )

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWMSupporting sequencesDiChIPMunk PWM

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWMSupporting sequencesDiChIPMunk PWM
0_0
1224
0_1
301
0_2
12
0_3
1
0_4

HOMER on peaks

In [12]:
homer_peak_matches = match_motifs.match_motifs_to_targets(homer_peak_pfms, tfm_pfms)
show_matched_motifs_table(
    tfm_pfms, tfm_names, "homer", homer_peak_pfms, homer_peak_enrichments,
    homer_peak_matches
)

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWMLog enrichmentHOMER PWM
0_0
-463.687679
0_1
-366.397049
0_3
-284.060976
0_4
-172.174005
0_3
-64.98932
0_0
-64.692095
0_1
-57.883603
0_2
-53.380344

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWMLog enrichmentHOMER PWM
-222.794162
-111.521251
-111.435294
-103.282832
-81.136063
-80.841664
-80.238425
-71.30052
-71.12152
-38.566296
-26.644662
-25.91339

HOMER on multi-task seqlets

In [13]:
homer_multitask_seqlet_matches = match_motifs.match_motifs_to_targets(homer_multitask_seqlet_pfms, tfm_pfms)
show_matched_motifs_table(
    tfm_pfms, tfm_names, "homer", homer_multitask_seqlet_pfms, homer_multitask_seqlet_enrichments,
    homer_multitask_seqlet_matches
)

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWMLog enrichmentHOMER PWM
0_0
-1595.283747
0_1
-852.126851
0_3
-368.637113
0_2
-362.92872
0_4
-214.597544
0_1
-163.138461

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWMLog enrichmentHOMER PWM
-148.788003

HOMER on single-task seqlets

In [14]:
if task_index is not None:
    homer_singletask_seqlet_matches = match_motifs.match_motifs_to_targets(homer_singletask_seqlet_pfms, tfm_pfms)
    show_matched_motifs_table(
        tfm_pfms, tfm_names, "homer", homer_singletask_seqlet_pfms, homer_singletask_seqlet_enrichments,
        homer_singletask_seqlet_matches
    )

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWMLog enrichmentHOMER PWM
0_0
-94.520167
0_2
-44.634734
0_4
-37.311805
0_4
-27.624514

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWMLog enrichmentHOMER PWM
0_1
-107.637503
0_3
-66.442118
-49.109607
-44.634734
-40.619424
-38.558637
-37.207106
-22.342533
-18.261804
-16.867176
-15.804633
-14.508918

MEME on peaks

In [15]:
meme_peak_matches = match_motifs.match_motifs_to_targets(meme_peak_pfms, tfm_pfms)
show_matched_motifs_table(
    tfm_pfms, tfm_names, "meme", meme_peak_pfms, meme_peak_evalues,
    meme_peak_matches
)

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWME-valueMEME PWM
0_1
3.7e-104
0_0
3.8e-091
0_3
1.7e-056
0_1
3.1e-002

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWME-valueMEME PWM
0_2
5.7e-064
0_4
8.7e-025
7.2e-012
1.3e-013
2.3e-004
2.5e-003

MEME on multi-task seqlets

In [16]:
meme_multitask_seqlet_matches = match_motifs.match_motifs_to_targets(meme_multitask_seqlet_pfms, tfm_pfms)
show_matched_motifs_table(
    tfm_pfms, tfm_names, "meme", meme_multitask_seqlet_pfms, meme_multitask_seqlet_evalues,
    meme_multitask_seqlet_matches
)

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWME-valueMEME PWM
0_0
1.7e-560
0_1
1.5e-386
0_3
2.5e-059
0_2
2.6e-037
0_3
6.7e-021
0_4
7.8e-018

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWME-valueMEME PWM
7.3e+003
2.0e+006
2.3e+006
1.1e+006

MEME on single-task seqlets

In [17]:
if task_index is not None:
    meme_singletask_seqlet_matches = match_motifs.match_motifs_to_targets(meme_singletask_seqlet_pfms, tfm_pfms)
    show_matched_motifs_table(
        tfm_pfms, tfm_names, "meme", meme_singletask_seqlet_pfms, meme_singletask_seqlet_evalues,
        meme_singletask_seqlet_matches
    )

Matched motifs

TF-MoDISco motif IDTF-MoDISco PWME-valueMEME PWM
0_4
3.1e+003
0_1
5.1e+004

Motifs without matches

TF-MoDISco motif IDTF-MoDISco PWME-valueMEME PWM
0_0
5.8e-154
0_2
7.4e-010
0_3
2.2e-001
7.2e+000
5.1e+003
2.2e+004
2.6e+004
1.6e+003