BPNet

Unlike window-based classifiers that output a single score per region, BPNet predicts a continuous binding probability profile at every nucleotide position in the input sequence. This is achieved through a profile head that outputs a vector matching the input length, trained against ChIP-nexus data which provides single-nucleotide resolution of protein-DNA contacts. The model learns to predict the exact shape of the binding event, including the central peak and surrounding accessibility pattern.

BPNet employs stacked dilated convolutions with exponentially increasing dilation rates (e.g., 1, 2, 4, 8, 16, 32) to capture regulatory patterns across multiple spatial scales without pooling. This design:

• Exponentially expands the receptive field to span hundreds of base pairs
• Preserves spatial resolution at every layer
• Enables detection of both local motif instances and long-range motif interactions
• Maintains parameter efficiency compared to large-kernel convolutions

BPNet simultaneously optimizes two complementary objectives through a shared representation:

• Profile head: Predicts the base-resolution binding shape using a Poisson or multinomial loss
• Count head: Predicts the total number of binding events in the region using a negative binomial loss This multi-task setup forces the model to learn both the spatial distribution and the absolute magnitude of binding, improving generalization and enabling quantitative comparisons across sequences and experimental conditions.

BPNet's predictions can be decomposed using DeepLIFT or integrated gradients to assign importance scores to every nucleotide in the input sequence. These importance scores are then processed by TF-MoDISco to extract consolidated, non-redundant sequence motifs. This pipeline reveals:

• Core binding motifs recognized by the transcription factor
• Spacing and orientation preferences between cooperative factor binding sites
• Syntax rules governing how multiple motifs combine to drive binding

BPNet processes DNA sequences using one-hot encoding and applies reverse complement data augmentation during training. The model architecture enforces strand symmetry by ensuring that the predicted binding profile for a sequence and its reverse complement are consistent. This biological prior—that transcription factors bind DNA regardless of strand orientation—reduces overfitting and improves the model's ability to generalize to unseen genomic regions.

Through in silico mutagenesis, BPNet systematically introduces virtual nucleotide substitutions and measures the predicted change in binding. This enables:

• Allele-specific binding analysis: Quantifying the impact of heterozygous variants on factor occupancy
• Regulatory variant prioritization: Ranking non-coding variants by their predicted effect on binding
• Causal variant identification: Distinguishing driver mutations from passenger mutations in regulatory regions The model outputs a quantitative effect size for every possible single-nucleotide variant in the input sequence.

A convolutional operation where the filter kernel is applied over an input with defined gaps or holes, exponentially expanding the receptive field without increasing parameter count. In BPNet, dilated convolutions enable the model to capture long-range sequence dependencies—such as cooperative TF binding across hundreds of base pairs—while maintaining single-nucleotide resolution. Unlike pooling-based approaches that sacrifice spatial precision, dilation preserves the exact positional information critical for base-resolution ChIP-nexus profile prediction.

Chromatin Immunoprecipitation with Nucleotide Resolution through Exonuclease, Unique barcode, and Single ligation—an enhanced ChIP protocol that achieves single base-pair resolution of protein-DNA binding. Unlike standard ChIP-seq, which produces broad peaks spanning hundreds of bases, ChIP-nexus uses lambda exonuclease to digest DNA precisely to the protein crosslink boundary. BPNet is specifically designed to predict these high-resolution binding profiles, enabling quantitative modeling of TF binding dynamics that lower-resolution assays cannot support.

A computational perturbation method that systematically introduces virtual nucleotide substitutions into a DNA sequence and measures the resulting change in BPNet's binding prediction. By mutating every position to every alternative base and observing the predicted profile change, researchers can identify causal regulatory variants and dissect motif grammar—how multiple TF binding sites interact cooperatively or competitively. This technique transforms BPNet from a predictive model into a mechanistic discovery tool for regulatory genomics.

Transcription Factor Motif Discovery from Importance Scores—an algorithm that clusters high-importance genomic subsequences identified by BPNet's interpretability methods to extract consolidated, non-redundant sequence motifs. Rather than returning thousands of overlapping high-attribution sequences, TF-MoDISco groups them by similarity and produces a compact set of position weight matrices representing distinct binding specificities. This post-processing step is essential for translating BPNet's per-nucleotide importance scores into human-interpretable motif patterns.

A training paradigm where BPNet simultaneously predicts multiple experimental assays—such as ChIP-nexus profiles for different TFs or across cell types—by sharing hidden representations in the dilated convolutional trunk. This approach leverages common regulatory logic (e.g., chromatin accessibility patterns, co-binding factors) to improve generalization on individual tasks, particularly for TFs with limited training data. The shared trunk learns a universal regulatory grammar while task-specific heads specialize in factor-specific binding preferences.

A cross-validation strategy where entire chromosomes are reserved for testing rather than randomly splitting sequences. This prevents information leakage caused by sequence homology between training and validation splits—paralogous regions or segmental duplications that appear on multiple chromosomes would otherwise inflate performance metrics. For BPNet, chromosomes 1 and 8 are commonly held out, ensuring that reported prediction accuracy reflects true generalization to unseen genomic contexts rather than memorization of homologous sequences.