Data, Task, and Model Relationships

Building successful genomic deep learning pipelines requires using the correct data format, selecting the appropriate task type, and pairing it with a suitable model architecture. OmniGenBench provides a unified framework for aligning these components.

---

OmniGenBench Model Family

OmniGenBench includes model variants specialized for different machine learning task types commonly used in genomics.

Task Type	OmniGenBench Model	Data Format Requirements	Application Scenarios
Sequence Classification	OmniModelForSequenceClassification	{sequence, label}	Promoter identification, functional classification
Sequence Regression	OmniModelForSequenceRegression	{sequence, value}	Expression prediction, stability score prediction
Token Classification	OmniModelForTokenClassification	{sequence, labels_per_position}	Binding site detection, RNA modifications
Multi-label Classification	OmniModelForMultiLabelClassification	{sequence, label_vector}	Multi-functional annotation, localization prediction

These models share the same underlying OmniGenome pre-trained encoder, while the prediction head varies depending on the downstream task.

---

Standard Dataset Formats

OmniGenBench datasets follow standardized CSV schemas to ensure compatibility across all models.

Sequence Classification Format

sequence,label,id,split
ATCGATCGATCG,1,seq_001,train
GCGCGCGCGCGC,0,seq_002,valid

Sequence Regression Format

sequence,value,id,split
ATCGATCGATCG,8.5,seq_001,train
GCGCGCGCGCGC,2.3,seq_002,valid

Token Classification Format

sequence,labels,id,split
ATCGATCGATCG,"0,0,1,1,0,0,0,0,0,0,0,0",seq_001,train

sequence: DNA/RNA/amino acid sequence

labels: comma-separated per-position labels

split: train/valid/test

---

Model Selection Decision Tree

You can determine the correct model type by answering two questions:

• Do I want to classify/predict something about an entire sequence? → Classification or regression model

• Do I want to annotate each nucleotide/residue? → Token classification model

• Do I want to generate a new sequence? → Sequence-to-sequence model (coming soon)

---

Practical Application Examples

Here are typical real-world scenarios showing how to match tasks, models, and data formats.

---

Case 1: Translation Efficiency Prediction

Question: “Is this mRNA sequence’s translation efficiency high or low?”

Task Type: Sequence classification (binary)

Model Choice: OmniModelForSequenceClassification

Data Format Example: { "sequence": "AUGCCC...", "label": 1 }

---

Case 2: Gene Expression Level Prediction

Question: “What expression level will this promoter produce?”

Task Type: Sequence regression

Model Choice: OmniModelForSequenceRegression

Data Format Example: { "sequence": "ATGCCC...", "value": 8.5 }

---

Case 3: Transcription Factor Binding Site Prediction

Question: “Which positions in the sequence bind a specific transcription factor?”

Task Type: Token classification (per-nucleotide labeling)

Model Choice: OmniModelForTokenClassification

Data Format Example: {"sequence": "ATGCCC...", "labels": [0, 0, 1, 1, 0, 0, ...]}

Summary

By aligning: - Task type (classification, regression, token labeling)

- Dataset format (sequence-level or position-level)

- Model architecture (OmniModel variants)

OmniGenBench ensures consistency, reproducibility, and ease of experimentation across genomic deep learning workflows.