Campolina: A Deep Neural Framework for Accurate Segmentation of Nanopore Signals

This repository contains trained models for Campolina, a deep learning framework for event border detection in nanopore sequencing signals. We provide models trained for R10.4.1 (R10_model.pth) and R9.4.1 (R9_model.pth) nanopore versions.

Campolina Framework Overview:

• a) Feature Extraction Pipeline:
  Raw signal data is split into non-overlapping chunks of fixed length (L = 6000 samples). Each chunk is z-normalized and augmented with four point-wise statistical descriptors:

  1. Rolling mean

  2. Rolling standard deviation

  3. First-order difference

  4. Rolling-window t-statistic

  This results in a 5-channel input per chunk.

• b) Model Architecture:
  The 5-channel input is processed by a convolutional neural network comprising six convolutional blocks with increasing channel sizes (32, 64, 64, 128, 128). The first block uses a kernel size of 3, while all others use a kernel size of 31. All activations are GELU. A linear classification head outputs a logit (non-normalized probability) per sample point, indicating the likelihood of an event border. The model outputs a sequence of shape (6000 × 1).

Training Objective:

Campolina is trained to identify signal positions that correspond to event borders. Ground truth event borders are generated through a two-step process:

1. The signal is basecalled using Dorado’s “super accurate” basecaller and aligned to a reference using minimap2. This produces an initial segmentation.

2. The initial borders and k-mer model are refined using Remora, producing high-accuracy event border annotations.

Loss Function:

The model is trained using a composite loss function:

• Focal Loss: Classification loss that handles class imbalance due to the rarity of event borders.

• Huber Loss: Constrains the total number of predicted event borders.

• Consecutive Loss: Prevents over-prediction of consecutive borders, which can occur due to variable translocation speed in nanopore sequencing.