GENATATOR-ModernGENA is a DNA language model fine-tuned for transcript boundary detection directly from genomic DNA sequences.
The model performs token-level multilabel classification to identify strand-aware transcript boundaries, enabling reconstruction of transcript intervals for both:
mRNA genes
lncRNA genes
This model focuses on transcript position discovery, namely:
transcription start sites (TSS)
transcript termination sites (PolyA)
Model
Model name on Hugging Face:
genatator-moderngena-base-human-edge-model
Architecture properties:
backbone: ModernBERT (ModernGENA)
layers: 22
hidden size: 768
parameters: ~135M
tokenization: BPE
output head: linear projection to 4 classes
output resolution: token-level
The model predicts four classes.
The correct order of output classes is:
["TSS+", "TSS-", "PolyA+", "PolyA-"]
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
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:
human genome only
strand-aware annotations
genome-wide supervision
human chromosomes 8, 20, and 21 held out
for non-held-out data, chromosomes longer than 100 kbp were included
The model is therefore trained to distinguish both transcript boundaries and non-transcribed background sequence.
Method Overview
This model is the edge model in the ModernGENA transcript discovery pipeline.
It predicts boundary probability tracks for:
transcript start sites
transcript termination sites
both strands independently
The full pipeline consists of:
Edge model
predicts strand-specific transcript boundary tracks
Region model
predicts strand-specific intragenic coverage
Post-processing
signal denoising
peak calling
interval construction
filtering using region predictions
This design allows recovery of full transcript intervals and multiple transcript boundary isoforms for the same gene.
Key Properties
strand-aware predictions
supports both mRNA and lncRNA
trained on human genome
ab initio inference from DNA only
designed for genome-wide transcript discovery
Important Notes
This model predicts transcript boundaries only
It does not predict exon–intron structure
It does not produce final transcript annotations by itself
Full inference requires post-processing and, typically, the paired region model
The model outputs boundary logits, which are intended to be converted into probability tracks and further processed into transcript intervals.
