A novel method called CHASOS (CHromatin loop prediction with Anchor Score and OCR Score) to achieve accuracy chromatin loop predictions. CHASOS incorporates multi-receptive field large kernel convolutional modules and scale modules to extract features used for dealing with inconsistencies among multiple feature data. Using the extracted features, CHASOS constructs a gradient boosting tree model to accomplish the chromatin loop predictions without introducing extra sequencing.

• /experiments - the source code of comparison methods
• /figure - the source code of figures (Fig.2-4) in paper
• /source - the source code of CHASOS
  • /data_preprocess - source code of data preprocess
  • /pretrained_model - source code of anchor score model
  • /fine_tuned_ocr_model - source code of OCR score model
  • /loop_model - source code of loop prediction model
• /data - the data used in CHASOS
  • /ChIA-PET/CTCF/raw_data - raw data of detect chromatin loops used for training and testing
  • /ChIA-PET2_result - the result of chromatin loop detection by ChIA-PET2
  • /ChIP-seq - ChIP-seq data of CTCF used in chromatin loops filtering
  • /DNase - DNase data used in chromatin open region prediction in OCR score model
  • /FIMO_result - the result of motif scanning by FIMO
  • /negative_loop - the negative chromatin loops used in training and testing
  • /phastcons - sequence conservation score used in chromatin loop predictions
  • /positive_loop - the positive chromatin loops used in training and testing
  • /pretrained_data - the data used in training anchor score model
• /ref_block - some DL block tested in search space of model construction

Anchor score model dataset: /source/data_preprocess/pretrained_data_loader.py
OCR score model dataset: /source/data_preprocess/dnase_preprocessor.py
Loop prediction model dataset: /source/data_preprocess/loop_preprocessor.py, /source/data_preprocess/positive_loop_generator.py, /source/data_preprocess/negative_loop_generator.py, /source/data_preprocess/motif_strength_adder.py

• main training code: /source/pretrained_model/trainer.py
• model code: /source/pretrained_model/models.py - AnchorModel_v20230516_v1
• designed module in model: /source/pretrained_model/custom_layer.py
• training data preparation code: /source/pretrained_model/pretrained_data_loader.py

The anchor score model generates a single value between 0 and 1 as its output, which indicates the probability of the input sequence serving as an anchor for a loop. This output could be directly employed as a feature for predicting chromatin loops.

• main training code: /source/fine_tuned_ocr_model/ocr_trainer.py
• model code: /source/fine_tuned_ocr_model/ocr_models.py - OCRModel_v20230524_v1
• designed module in model: /source/pretrained_model/custom_layer.py
• training data preparation code: /source/fine_tuned_ocr_model/ocr_data_loader.py

The OCR score model produces a [1000,1] tensor representing 1000 positions of the input sequence. This tensor signifies the predicted signal values of DNase hypersensitive site at these 1000 positions. The average of these values is computed and employed as a feature for chromatin loop prediction.

• main training code: /source/loop_model/loop_ml_model_trainer.py
• training data preparation code: /source/loop_model/preprocess_data.py
• drawing K562 loop prediction example code (Figure 5 in paper): /source/loop_model/raw_predictor.py

In the loop prediction model section, the gradient boosting tree algo-rithm is employed. It combines the two simulated features ob-tained from the anchor score and OCR score models with the se-quence and functional genomics features commonly used for pre-diction. The gradient boosting tree outputs a single value as the prediction, which assesses the probability of constituting a chro-matin loop.

The anchor score model and OCR score model are trained efficiently on a NVIDIA GeForce RTX 3060 12G GPU.
The models are implemented in a Python 3.8 environment, utilizing PyTorch 1.12 as the backend framework.

model trained in experiment "Comparative evaluation on different chromosomes"

OCR model: /source/fine_tuned_ocr_model/model/OCRModel_val_loss_2023_06_30_23_34_50.pt
Anchor model: /source/pretrained_model/model/PretrainedModel_2023_06_29_20_59_01.pt
Loop model: /source/loop_model/model/RandomForestClassifier_mine_allfeats_8cellline_bychr.joblib (It's just named a random forest, but it's actually a gradient boosting tree)

model trained in experiment "Comparative evaluation in cancer cell lines"

OCR model: /source/fine_tuned_ocr_model/model/OCRModel_val_loss_2023_07_07_19_04_18.pt
Anchor model: /source/pretrained_model/model/PretrainedModel_2023_07_06_20_42_59_AnchorModel_v20230516_v1.pt
Loop model: /source/loop_model/model/RandomForestClassifier_mine_allfeats_8cellline_byhealth.joblib (It's just named a random forest, but it's actually a gradient boosting tree)