genatator-moderngena-base-multispecies-region-model

GENATATOR-ModernGENA (Multispecies Region Model)

Overview

GENATATOR-ModernGENA (Region Model) is a DNA language model fine-tuned for intragenic region detection directly from genomic DNA sequences.

The model performs token-level multilabel classification to identify strand-aware genomic regions, complementing the edge model used for transcript boundary detection.

This model focuses on region-level signal detection, specifically:

• identifying intragenic regions (gene bodies)
• distinguishing them from intergenic background
• modeling strand-specific transcriptional activity

---

Model

Model name on Hugging Face:

genatator-moderngena-base-multispecies-region-model

Architecture properties:

• backbone: ModernBERT (ModernGENA)
• layers: 22
• hidden size: 768
• parameters: ~135M
• tokenization: BPE
• output head: linear projection to 6 classes
• output resolution: token-level

The model predicts six classes.

The correct order of output classes is:

["TSS+", "TSS-", "PolyA+", "PolyA-", "Intragenic+", "Intragenic-"]

Where:

• TSS+ — transcription start site on the forward strand
• TSS- — transcription start site on the reverse strand
• PolyA+ — transcript termination site on the forward strand
• PolyA- — transcript termination site on the reverse strand
• Intragenic+ — intragenic region (forward strand)
• Intragenic- — intragenic region (reverse strand)

---

Training Data

This model was fine-tuned on full genomic sequences, including intergenic regions.

Training data includes annotations for:

• mRNA transcripts
• lncRNA transcripts

Dataset characteristics:

• multispecies dataset (39 mammalian species)
• strand-aware annotations
• genome-wide supervision
• human chromosomes 8, 20, and 21 held out
• for non-human genomes, chromosomes longer than 100 kbp were included

The model is trained to distinguish both transcribed regions and background genomic sequence across species.

---

Method Overview

This model is the region model in the ModernGENA transcript discovery pipeline.

It predicts region-level probability tracks for:

• intragenic regions
• strand-specific transcriptional coverage

The full pipeline consists of:

1. Edge model predicts transcript boundaries (TSS / PolyA)

2. Region model (this model) predicts intragenic regions

3. Post-processing

   • signal denoising
   • peak detection (from edge model)
   • interval construction
   • filtering using region predictions

The region model is used to:

• validate predicted transcript intervals
• remove false positives
• improve reconstruction of full gene structures

---

Key Properties

• strand-aware predictions
• detects intragenic regions
• supports both mRNA and lncRNA
• multispecies generalization
• trained on full genomes
• ab initio inference from DNA only

---

Important Notes

• This model predicts regions, not precise transcript boundaries
• It is intended to be used together with the edge model
• It does not predict exon–intron structure
• It does not produce final transcript annotations by itself

The model outputs region probability tracks, which are used during post-processing to filter and validate transcript predictions.

---

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification

repo_id = "AIRI-Institute/genatator-moderngena-base-multispecies-region-model"

tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    trust_remote_code=True,
)

model = AutoModelForTokenClassification.from_pretrained(
    repo_id,
    trust_remote_code=True,
)

model.eval()

---

Example Inference

import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification

repo_id = "AIRI-Institute/genatator-moderngena-base-multispecies-region-model"

tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForTokenClassification.from_pretrained(repo_id, trust_remote_code=True)

sequence = "ACGTACGTACGTACGTACGTACGTACGTACGT"

enc = tokenizer(sequence, return_tensors="pt")

with torch.no_grad():
    outputs = model(**enc)

logits = outputs.logits

print("Input shape:", enc["input_ids"].shape)
print("Logits shape:", logits.shape)

Example output:

Input shape: torch.Size([1, sequence_length])
Logits shape: torch.Size([1, sequence_length, 6])

Each token receives 6 logits corresponding to the six region and boundary classes.

---

Recommended Inference Workflow

To obtain transcript annotations:

1. run edge model to detect boundaries
2. run region model to detect intragenic regions
3. convert logits to probabilities
4. apply smoothing
5. detect peaks (TSS / PolyA)
6. construct candidate transcript intervals
7. filter intervals using intragenic predictions

---

Limitations

• predicts regions, not full transcript structures
• requires combination with edge model
• does not model exon–intron structure
• output depends on downstream post-processing

---

Summary

GENATATOR-ModernGENA region model is a ModernBERT-based DNA language model for strand-aware intragenic region detection across multiple species.

It complements the edge model and enables:

• filtering of transcript predictions
• improved gene structure reconstruction
• genome-wide transcript discovery
• multi-isoform detection through downstream processing