CNN Models
Switch language: CNN 模型
This page covers the convolutional baseline family: M1, M2, and M3.
These models are the most direct continuation of the original DeepChrInteract
design philosophy and remain important because they anchor the rest of the
benchmark in interpretable local-pattern learning.
Why CNNs matter for EPI
DNA sequence prediction often starts with local motif detection. Convolutional layers are naturally suited for:
scanning the sequence for short recurring nucleotide patterns;
building motif combinations through deeper layers;
offering relatively stable optimization and efficient training;
providing strong baseline performance without requiring external pretraining.
At the operator level, a 1D convolution behaves like a trainable motif scanner:
where \(x\) is the multi-channel sequence input, \(k\) is the kernel width, and \(h_t\) is the activation at position \(t\). Each filter can be interpreted as a learned detector for short sequence patterns and their local variants.
M1: CNN single-branch baseline
M1 uses a single convolutional branch and acts as the simplest baseline in
the project.
Input: one-hot encoded sequence tensor
(B, 5, L)Backbone: two-stage
Conv1d -> BN -> ReLU -> MaxPoolstackHead: global average pooling followed by fully connected layers
Role: minimal local-feature baseline
Interpretation:
Useful when the goal is to test whether local sequence composition alone can already produce non-trivial predictive signal.
It is intentionally simple and easy to compare against more expressive models.
Because it avoids pair fusion and avoids recurrent or attention mechanisms, M1 is the cleanest answer to a foundational question: how far can local feature detectors and hierarchical pooling go on their own?
M2: CNN dual-branch baseline
M2 extends the convolutional baseline to a paired-input setting.
Input: enhancer and promoter are encoded separately
Backbone: the same convolutional encoder is applied to both branches
Fusion: branch outputs are combined through the project fusion module
Role: stronger reproduction of the original paired-region setup
Why it matters:
EPI is a pairwise problem, not a single-sequence classification task.
Separate branches let the model preserve region-specific representations before interaction modeling.
Its computation can be summarized as:
This makes the architecture explicitly separate the two encoding problems before asking a fusion layer to model compatibility, asymmetry, or synergy.
M3: k-mer embedding plus CNN
M3 changes the input representation rather than the high-level convolutional
idea.
Input: k-mer token sequence
Embedding: learnable token embedding table
Encoder: four convolutional blocks over embedding channels
Role: bridge between symbolic tokenization and convolutional feature learning
Why use k-mers:
They expose local compositional units larger than a single nucleotide.
They can sometimes make biologically meaningful subsequence patterns easier to capture than raw one-hot channels alone.
Formally, the input representation becomes:
where \(t_i\) is a k-mer token and \(E\) is the embedding table. The CNN then operates over token vectors rather than raw nucleotide channels, which places M3 between classical motif CNNs and language-model style representations.
Computational complexity
Time: roughly linear in sequence length, \(O(L \cdot k \cdot C_{in} \cdot C_{out})\) per convolutional stage, with constants depending on kernel count and depth.
Memory: usually moderate and dominated by intermediate feature maps rather than pairwise interaction matrices.
Best-fit regime: strong default choice for short-to-medium genomic windows and for large benchmark sweeps where stability and throughput matter.
Relationship to the legacy project
The CNN family is the closest direct descendant of the original DeepChrInteract work. For that reason, it plays two roles at once:
historical continuity with the original project;
practical baseline against which all later architectures are evaluated.