Direct links to results¶

Motifs derived from filter-activating sequences

import os
import sys
sys.path.append(os.path.abspath("/users/amtseng/tfmodisco/src/"))
from motif.read_motifs import pfm_info_content, pfm_to_pwm
from util import figure_to_vdom_image
import plot.viz_sequence as viz_sequence
import h5py
import numpy as np
import pyfaidx
import matplotlib.pyplot as plt
import vdom.helpers as vdomh
from IPython.display import display
import tqdm
tqdm.tqdm_notebook()

/users/amtseng/miniconda3/envs/tfmodisco-mini/lib/python3.7/site-packages/ipykernel_launcher.py:14: TqdmDeprecationWarning: This function will be removed in tqdm==5.0.0
Please use `tqdm.notebook.tqdm` instead of `tqdm.tqdm_notebook`

0it [00:00, ?it/s]

Define constants and paths¶

# Define parameters/fetch arguments
filter_activations_path = os.environ["TFM_FILTER_ACTIVATIONS"]
filter_weights_path = os.environ["TFM_FILTER_WEIGHTS"]

if "TFM_MOTIF_CACHE" in os.environ:
    activation_motifs_cache_dir = os.environ["TFM_MOTIF_CACHE"]
else:
    activation_motifs_cache_dir = None

print("Path to filter activations: %s" % filter_activations_path)
print("Path to filter weights: %s" % filter_weights_path)
print("Saved activation-derived motifs cache: %s" % activation_motifs_cache_dir)

Path to filter activations: /users/amtseng/tfmodisco/results/filter_activations/singletask_profile_finetune/NR3C1-reddytime_singletask_profile_finetune_task14_fold5_filter_activations.h5
Path to filter weights: /users/amtseng/tfmodisco/results/filter_weights/singletask_profile_finetune/NR3C1-reddytime_singletask_profile_finetune_task14_fold5_filter_weights.npy
Saved activation-derived motifs cache: /users/amtseng/tfmodisco/results/reports/filter_derived_motifs/cache/NR3C1-reddytime_singletask_profile_finetune_task14_fold5_filter

# Constants/paths
input_length = 2114
filter_width = 21
reference_genome_path = "/users/amtseng/genomes/hg38.fasta"

if activation_motifs_cache_dir:
    os.makedirs(activation_motifs_cache_dir, exist_ok=True)

Helper functions¶

For extracting motifs

def dna_to_one_hot(seqs):
    """
    Converts a list of DNA ("ACGT") sequences to one-hot encodings, where the
    position of 1s is ordered alphabetically by "ACGT". `seqs` must be a list
    of N strings, where every string is the same length L. Returns an N x L x 4
    NumPy array of one-hot encodings, in the same order as the input sequences.
    All bases will be converted to upper-case prior to performing the encoding.
    Any bases that are not "ACGT" will be given an encoding of all 0s.
    """
    seq_len = len(seqs[0])
    assert np.all(np.array([len(s) for s in seqs]) == seq_len)

    # Join all sequences together into one long string, all uppercase
    seq_concat = "".join(seqs).upper()

    one_hot_map = np.identity(5)[:, :-1]

    # Convert string into array of ASCII character codes;
    base_vals = np.frombuffer(bytearray(seq_concat, "utf8"), dtype=np.int8)

    # Anything that's not an A, C, G, or T gets assigned a higher code
    base_vals[~np.isin(base_vals, np.array([65, 67, 71, 84]))] = 85

    # Convert the codes into indices in [0, 4], in ascending order by code
    _, base_inds = np.unique(base_vals, return_inverse=True)

    # Get the one-hot encoding for those indices, and reshape back to separate
    return one_hot_map[base_inds].reshape((len(seqs), seq_len, 4))

def extract_filter_activation_motifs(filter_activations_path, reference_genome_path):
    """
    Extracts the motifs that correspond to each filter. Returns an
    F x W x 4 array, where F is the number of filters and W is the width
    of each filter. The order of filters matches those in the saved HDF5/model.
    """
    reader = h5py.File(filter_activations_path, "r")
    activations_reader = reader["activations"]
    num_coords, two, num_windows, num_filters = activations_reader.shape
    
    assert two == 2
    assert num_windows == input_length - filter_width + 1
    
    print("Importing coordinates...")
    coords = np.empty((num_coords, 3), dtype=object)
    coords[:, 0] = reader["coords"]["coords_chrom"][:].astype(str)
    coords[:, 1] = reader["coords"]["coords_start"][:]
    coords[:, 2] = reader["coords"]["coords_end"][:]
    
    print("Fetching one-hot sequences...")
    genome_reader = pyfaidx.Fasta(reference_genome_path)
    one_hot_seqs = np.empty((num_coords, input_length, 4))
    batch_size = 128
    num_batches = int(np.ceil(num_coords / batch_size))
    for i in tqdm.notebook.trange(num_batches):
        batch_slice = slice(i * batch_size, (i + 1) * batch_size)
        one_hot_seqs[batch_slice] = dna_to_one_hot([
            genome_reader[chrom][start:end].seq for chrom, start, end in coords[batch_slice]
        ])
    
    pfms = np.empty((num_filters, filter_width, 4))
    for filter_index in range(num_filters):
        print("Extracting motif for filter %d..." % filter_index)
    
        print("\tComputing maximum activation...")
        acts = activations_reader[:, :, :, filter_index]
        max_act = np.max(acts)
        
        inds = np.where(acts >= 0.5 * max_act)
        
        windows, num_windows = np.zeros((filter_width, 4)), 0
        for coord_index, strand_index, pos_index in tqdm.notebook.tqdm(
            zip(*inds), total=len(inds[0]), desc="Extracting windows..."
        ):
            if strand_index == 0:
                window = one_hot_seqs[coord_index, pos_index : pos_index + filter_width]
            else:
                # Reverse complement; the positions are flipped
                window = np.flip(
                    one_hot_seqs[coord_index, input_length - filter_width - pos_index : input_length - pos_index],
                    axis=(0, 1)
                )
            windows = windows + window
            num_windows += 1
        
        pfms[filter_index] = windows / num_windows
    
    return pfms

def compute_filter_influence(filter_activations_path):
    """
    Extracts the influence of each filter by computing the difference
    in cross entropy when each filter is nullified.
    Returns an F-array, where F is the number of filters, containing the
    change in average cross entropy (after nullification - before
    nullification). The order of filters matches those in the saved
    HDF5/model.
    """
    reader = h5py.File(filter_activations_path, "r")
    print("Reading in cross entropies...")
    before_null_cross_ents = reader["predictions"]["cross_ents"][:]
    after_null_cross_ents = reader["nullified_predictions"]["cross_ents"][:]
    
    before_null = np.nanmean(before_null_cross_ents)
    
    num_filters = after_null_cross_ents.shape[1]
    
    influences = []
    for filter_index in tqdm.notebook.trange(num_filters):
        after_null = np.nanmean(after_null_cross_ents[:, filter_index])
        influences.append(after_null - before_null)
        
    return np.array(influences)

def save_activation_motifs(filter_pfms, filter_influences, path):
    """
    Saves the filter-activation-derived PFMs and influence values.
    """
    with h5py.File(path, "w") as f:
        f.create_dataset("pfms", data=filter_pfms, compression="gzip")
        f.create_dataset("influences", data=filter_influences, compression="gzip")

def load_activation_motifs(path):
    """
    Loads the filter-activation-derived PFMs and influence values.
    """
    with h5py.File(path, "r") as f:
        return f["pfms"][:], f["influences"][:]

Extract motifs from filter activations¶

Extract the motifs derived from each filter, ranked by filter influence.

Deriving a filter's motif:

Identify the top 10000 most well-predicted input sequences, ranked by cross entropy
For each window in each of these sequences, compute the filter activation for each 1st-layer filter
A filter's motif is the aggregation of sequence windows which activate that filter to at least half its maximum activation (over the top 10000 most well-predicted inputs)

Deriving a filter's influence:

Identify the top 10000 most well-predicted input sequences, ranked by cross entropy
Nullify each filter by setting it to the average activation over these 10000 most well-predicted inputs
A filter's influence is the average change in cross entropy before and after nullification

compute_motifs = True
if activation_motifs_cache_dir:
    # Import if it exists
    cache_path = os.path.join(activation_motifs_cache_dir, "filter_activation_motifs.h5")
    if os.path.exists(cache_path) and os.stat(cache_path).st_size:
        filter_pfms, filter_influences = load_activation_motifs(cache_path)
        compute_motifs = False

if compute_motifs:
    # Extract PFMs of highly-activating sequences
    filter_pfms = extract_filter_activation_motifs(filter_activations_path, reference_genome_path)

    # Compute influence of each filter
    filter_influences = compute_filter_influence(filter_activations_path)

    if activation_motifs_cache_dir:
        save_activation_motifs(filter_pfms, filter_influences, cache_path)

Importing coordinates...
Fetching one-hot sequences...

Extracting motif for filter 0...
	Computing maximum activation...

Extracting motif for filter 1...
	Computing maximum activation...

Extracting motif for filter 2...
	Computing maximum activation...

Extracting motif for filter 3...
	Computing maximum activation...

Extracting motif for filter 4...
	Computing maximum activation...

Extracting motif for filter 5...
	Computing maximum activation...

Extracting motif for filter 6...
	Computing maximum activation...

Extracting motif for filter 7...
	Computing maximum activation...

Extracting motif for filter 8...
	Computing maximum activation...

Extract motifs from filter weights¶

# Import the filter weights themselves
filter_weights = np.load(filter_weights_path)
assert len(filter_weights.shape) == 3
assert filter_weights.shape[:2] == (filter_width, 4)
filter_weights = np.transpose(filter_weights, axes=(2, 0, 1))  # Shape: F x W x 4

Motifs derived from filter-activating sequences¶

For each filter, its motif is constructed by averaging all of the sequences that activate it at least to half of its maximal activation. We show the PWMs. The filters are ranked by influence (i.e. the average difference in prediction cross entropy when the filter is nullified--that is, replaced with its average activation).

colgroup = vdomh.colgroup(
    vdomh.col(style={"width": "5%"}),
    vdomh.col(style={"width": "5%"}),
    vdomh.col(style={"width": "5%"}),
    vdomh.col(style={"width": "85%"})
)
header = vdomh.thead(
    vdomh.tr(
        vdomh.th("Rank", style={"text-align": "center"}),
        vdomh.th("Filter index", style={"text-align": "center"}),
        vdomh.th("Influence", style={"text-align": "center"}),
        vdomh.th("PWM", style={"text-align": "center"})
    )
)

body = []
for i, filter_index in enumerate(np.flip(np.argsort(filter_influences))):
    pwm = pfm_to_pwm(filter_pfms[filter_index])
    if np.sum(pwm[:, [0, 2]]) < 0.5 * np.sum(pwm):
        # Flip to purine-rich version
        pwm = np.flip(pwm, axis=(0, 1))
    fig = viz_sequence.plot_weights(pwm, figsize=(20, 4), return_fig=True)
    fig.tight_layout()
    
    body.append(
        vdomh.tr(
            vdomh.td(str(i + 1)),
            vdomh.td(str(filter_index)),
            vdomh.td("%.3f" % filter_influences[filter_index]),
            vdomh.td(figure_to_vdom_image(fig))
        )
    )
    
    if activation_motifs_cache_dir:
        # Save motif PWM
        fig.savefig(os.path.join(activation_motifs_cache_dir, "filter_activation_motif_%d.png" % filter_index))

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

/users/amtseng/tfmodisco/src/plot/viz_sequence.py:152: RuntimeWarning: More than 20 figures have been opened. Figures created through the pyplot interface (`matplotlib.pyplot.figure`) are retained until explicitly closed and may consume too much memory. (To control this warning, see the rcParam `figure.max_open_warning`).
  fig = plt.figure(figsize=figsize)

Motifs derived from filter weights¶

For each filter, we show its corresponding motif simply as the mean-normalized multiplicative weights in the filter. For consistency, we rank the filters by influence (as above).

colgroup = vdomh.colgroup(
    vdomh.col(style={"width": "5%"}),
    vdomh.col(style={"width": "5%"}),
    vdomh.col(style={"width": "5%"}),
    vdomh.col(style={"width": "85%"})
)
header = vdomh.thead(
    vdomh.tr(
        vdomh.th("Rank", style={"text-align": "center"}),
        vdomh.th("Filter index", style={"text-align": "center"}),
        vdomh.th("Influence", style={"text-align": "center"}),
        vdomh.th("Mean-normalized filter weights", style={"text-align": "center"})
    )
)

body = []
for i, filter_index in enumerate(np.flip(np.argsort(filter_influences))):
    weights = filter_weights[filter_index]
    weights = weights - np.mean(weights, axis=1, keepdims=True)
    if np.sum(weights[:, [0, 2]]) < 0.5 * np.sum(weights):
        # Flip to purine-rich version
        weights = np.flip(weights, axis=(0, 1))
    fig = viz_sequence.plot_weights(weights, figsize=(20, 4), return_fig=True)
    fig.tight_layout()
    
    body.append(
        vdomh.tr(
            vdomh.td(str(i + 1)),
            vdomh.td(str(filter_index)),
            vdomh.td("%.3f" % filter_influences[filter_index]),
            vdomh.td(figure_to_vdom_image(fig))
        )
    )
    
    if activation_motifs_cache_dir:
        # Save motif PWM
        fig.savefig(os.path.join(activation_motifs_cache_dir, "filter_weight_motif_%d.png" % filter_index))

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

Rank	Filter index	Influence
1	32	0.030
2	6	0.021
3	8	0.017
4	52	0.015
5	44	0.012
6	37	0.011
7	54	0.011
8	26	0.010
9	40	0.009
10	59	0.009
11	60	0.008
12	14	0.007
13	2	0.007
14	53	0.007
15	45	0.006
16	17	0.006
17	43	0.006
18	28	0.006
19	55	0.005
20	15	0.005
21	5	0.005
22	24	0.005
23	20	0.005
24	0	0.005
25	9	0.005
26	41	0.004
27	62	0.004
28	10	0.004
29	13	0.004
30	22	0.004
31	42	0.004
32	33	0.003
33	36	0.003
34	18	0.003
35	58	0.003
36	7	0.003
37	47	0.003
38	34	0.003
39	46	0.003
40	50	0.003
41	57	0.002
42	21	0.002
43	19	0.002
44	56	0.002
45	23	0.002
46	12	0.002
47	30	0.002
48	31	0.002
49	4	0.002
50	29	0.002
51	1	0.002
52	39	0.002
53	16	0.002
54	61	0.002
55	11	0.002
56	63	0.001
57	38	0.001
58	48	0.001
59	51	0.001
60	49	0.001
61	35	0.001
62	25	0.001
63	27	0.001
64	3	0.000

Rank	Filter index	Influence
1	32	0.030
2	6	0.021
3	8	0.017
4	52	0.015
5	44	0.012
6	37	0.011
7	54	0.011
8	26	0.010
9	40	0.009
10	59	0.009
11	60	0.008
12	14	0.007
13	2	0.007
14	53	0.007
15	45	0.006
16	17	0.006
17	43	0.006
18	28	0.006
19	55	0.005
20	15	0.005
21	5	0.005
22	24	0.005
23	20	0.005
24	0	0.005
25	9	0.005
26	41	0.004
27	62	0.004
28	10	0.004
29	13	0.004
30	22	0.004
31	42	0.004
32	33	0.003
33	36	0.003
34	18	0.003
35	58	0.003
36	7	0.003
37	47	0.003
38	34	0.003
39	46	0.003
40	50	0.003
41	57	0.002
42	21	0.002
43	19	0.002
44	56	0.002
45	23	0.002
46	12	0.002
47	30	0.002
48	31	0.002
49	4	0.002
50	29	0.002
51	1	0.002
52	39	0.002
53	16	0.002
54	61	0.002
55	11	0.002
56	63	0.001
57	38	0.001
58	48	0.001
59	51	0.001
60	49	0.001
61	35	0.001
62	25	0.001
63	27	0.001
64	3	0.000