DeepChrInteract Model Overview

Switch language: DeepChrInteract 模型总览

This page serves as the navigation hub for the model system rather than trying to explain every architecture in a single long document. The project spans classical convolutional baselines, recurrent sequence models, attention-based architectures, state-space style encoders, DNA foundation models, and self-supervised pretraining.

Model family map

• Classical convolutional encoders: CNN Models

• Residual convolutional branch: ResNet Models

• Recurrent encoders: BiLSTM Model, mLSTM Model

• Attention family: Transformer Model, Linear Transformer Model, iTransformer Model

• Linear-time and state-space style models: Mamba Model, RWKV Model

• Hybrid architectures: Hybrid Models

• Foundation-model and self-supervised routes: DNA Foundation Models, MAE Pretraining Model

• Pairwise representation layer: Fusion Strategies

Registry snapshot

• M1: CNN single-branch Simplest local motif baseline. Detail page: CNN Models

• M2: CNN dual-branch Separate enhancer/promoter encoding. Detail page: CNN Models

• M3: k-mer + CNN Token embedding plus convolution. Detail page: CNN Models

• ResNet18 / ResNet34 Deeper residual convolution route. Detail page: ResNet Models

• M4: BiLSTM Bidirectional sequential dependency modeling. Detail page: BiLSTM Model

• M5: mLSTM Matrix-memory recurrent modeling. Detail page: mLSTM Model

• M6: Transformer Global attention after CNN compression. Detail page: Transformer Model

• M7: Linear Transformer Kernelized linear attention. Detail page: Linear Transformer Model

• M8: iTransformer Channel-wise attention reinterpretation. Detail page: iTransformer Model

• M9: Mamba Selective state-space sequence modeling. Detail page: Mamba Model

• M10: RWKV Linear-recurrence time mixing. Detail page: RWKV Model

• M11: CNN + BiLSTM Local detector plus recurrent context. Detail page: Hybrid Models

• M12: CNN + Transformer Local detector plus global attention. Detail page: Hybrid Models

• M13: DNA LLM encoder External pretrained genomic representations. Detail page: DNA Foundation Models

• M14: MAE-pretrained Transformer Self-supervised pretraining on project data. Detail page: MAE Pretraining Model

Design logic

The project is intentionally broad. It is meant to expose a continuum of sequence modeling assumptions:

• local pattern extraction through CNNs;

• deeper residual convolution as the natural extension of the CNN family;

• ordered dependency modeling through recurrent cells;

• global interaction modeling through attention;

• long-context efficiency through linear-time or state-space style models;

• transfer learning through pretrained genomic foundation models;

• task-adapted representation learning through MAE-style pretraining.

This organization makes it easier to compare not only raw performance, but also inductive bias, runtime behavior, memory tradeoffs, and data efficiency.