kundajelab/encode-chrombpnet-DNASE-sciatic-nerve-ENCSR672MOG-ENCSR903LXC

ENCODE ChromBPNet Atlas

As part of the ENCODE 4 Project, we trained ChromBPNet models on 1,512 ENCODE DNAse-seq and ATAC-seq across 408 biosamples. Here, we provide all models for open-source use.

For more information about the models, see:

• —Main ENCODE 4 Paper
• —A unified lexicon of predictive DNA sequence motifs from ENCODE transcription factor binding and chromatin accessibility assays (Yun, C. M. et al., Zenodo 2026)
• —ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants (Pampari, A. et al., bioRxiv 2024)

ChromBPNet model: DNASE in sciatic nerve (ENCSR672MOG)

• —Model: ChromBPNet
• —Assay: DNASE-seq
• —Experiment: ENCSR672MOG
• —Model annotation: ENCSR903LXC
• —Biosample: sciatic nerve (Full name: Homo sapiens sciatic nerve tissue female adult (41 years))
• —Cell slim(s): None
• —Organ slim(s): nerve
• —Developmental slim(s): ectoderm
• —System slim(s): peripheral-nervous-system
• —Assembly: hg38

Directory structure

• —fold_0: Model of 5-fold cross-validation: Fold 0
• —model.chrombpnet.fold_0.encid.h5: full chrombpnet model that combines both bias and corrected model in .h5 format
• —model.chrombpnet_nobias.fold_0.encid.h5: bias-corrected accessibility model in .h5 format (Use for all biological discovery)
• —model.bias_scaled.fold_0.encid.h5: bias model in .h5 format
• —model.chrombpnet.fold_0.encid.tar: full chrombpnet model that combines both bias and corrected model in SavedModel format. After being untarred, it results in a directory named "chrombpnet".
• —model.chrombpnet_nobias.fold_0.encid.tar: bias-corrected accessibility model in SavedModel format (Use for all biological discovery). After being untarred, it results in a directory named "chrombpnetwobias".
• —model.bias_scaled.fold_0.encid.tar: bias model in SavedModel format. After being untarred, it results in a directory named "biasmodelscaled".
• —logs.models.fold_0.encid: folder containing log files for training models
• —fold_1: Model of 5-fold coss-validation: Fold 1
• —fold_2: Model of 5-fold cross-validation: Fold 2
• —fold_3: Model of 5-fold cross-validation: Fold 3
• —fold_4: Model of 5-fold cross-validation: Fold 4

Instructions

1. Pseudocode for loading models in .h5 format

(1) Use the code in python after appropriately defining model_in_h5_format and inputs. \ (2) inputs is a one hot encoded sequence of shape (N,2114,4). Here N corresponds to the number of tested sequences, 2114 is the input sequence length and 4 corresponds to [A,C,G,T].

import tensorflow as tf
from tensorflow.keras.utils import get_custom_objects
from tensorflow.keras.models import load_model

custom_objects={"tf": tf}
get_custom_objects().update(custom_objects)

model=load_model(model_in_h5_format,compile=False)
outputs = model(inputs)

The list outputs consists of two elements. The first element has a shape of (N, 1000) and contains logit predictions for a 1000-base-pair output. The second element, with a shape of (N, 1), contains logcount predictions. To transform these predictions into per-base signals, follow the provided pseudo code lines below.

import numpy as np

def softmax(x, temp=1):
    norm_x = x - np.mean(x,axis=1, keepdims=True)
    return np.exp(temp*norm_x)/np.sum(np.exp(temp*norm_x), axis=1, keepdims=True)
    
predictions = softmax(outputs[0]) * (np.exp(outputs[1])-1)

2. Pseudocode for loading models in .tar format

(1) First untar the directory as follows tar -xvf model.tar. \ (2) Use the code below in python after appropriately defining model_dir_untared and inputs. \ (3) inputs is a one hot encoded sequence of shape (N,2114,4). Here N corresponds to the number of tested sequences, 2114 is the input sequence length and 4 corresponds to ACGT.

Reference: https://www.tensorflow.org/apidocs/python/tf/savedmodel/load

import tensorflow as tf

model = tf.saved_model.load('model_dir_untared')
outputs = model.signatures['serving_default'](**{'sequence':inputs.astype('float32')})

The variable outputs represents a dictionary containing two key-value pairs. The first key is logits_profile_predictions, holding a value with a shape of (N, 1000). This value corresponds to logit predictions for a 1000-base-pair output. The second key, named logcount_predictions`, is associated with a value of shape (N, 1), representing logcount predictions. To transform these predictions into per-base signals, utilize the provided pseudo code lines mentioned below.

import numpy as np
def softmax(x, temp=1):
    norm_x = x - np.mean(x,axis=1, keepdims=True)
    return np.exp(temp*norm_x)/np.sum(np.exp(temp*norm_x), axis=1, keepdims=True)
    
predictions = softmax(outputs["logits_profile_predictions"]) * (np.exp(outputs["logcount_predictions"])-1)

Docker image to load and use the models

• —https://hub.docker.com/r/kundajelab/chrombpnet-atlas/ (tag:v1)

Code for ChromBPNet

• —https://github.com/kundajelab/chrombpnet/

License & citation

External data users may freely download, analyze and publish results based on any ENCODE data without restrictions.

Released under the ENCODE data-use policy. Please cite the ENCODE Project Consortium and the model software: ChromBPNet (Pampari et al., bioRxiv 2024).