Bioinformatics has a new AI model called Corgi (Context-aware Regulatory Genomics Inference), and it promises a sizeable jump in how well we can predict gene expression and epigenetic marks. Unlike earlier models, Corgi extrapolates beyond its training sets, predicting gene behaviour in cell types it has never encountered. That is the kind of breakthrough that can speed up drug development and our grasp of complex biological processes.
What happened
Corgi, built to predict genomic regulation, gets past a fundamental limitation of existing models: their inability to extrapolate to new cell types and conditions. Corgi combines DNA sequence with trans-regulator expression to forecast chromatin accessibility, histone modifications and gene expression coverage. It was trained on a large sequencing dataset covering both bulk and single-cell data.
Corgi's architecture mimics cellular gene regulation, using the expression of trans-regulators (transcription factors, coactivators, chromatin modifiers and RNA-binding proteins) as a context vector. The model is highly accurate at predicting gene expression in new cell types, as well as DNase-seq, ATAC-seq, DNA methylation and a range of histone marks. In the hardest setting, where the model has to generalise to both new sequences and new cell types, Corgi holds up well, reaching 0.76 for DNase-seq and 0.69 for RNA-seq.
Why it matters
Corgi's ability to predict genomic regulation in unstudied cell types is a serious step forward for bioinformatics and medicine. Researchers can get valuable data on gene behaviour without running expensive, labour-intensive experiments for every new cell type. That accelerates work on genetic disease, oncology and personalised drug development.
For AI developers working in biotech, Corgi is a case study in applying deep learning to hard biological problems. The model uses a hybrid convolutional-transformer architecture plus FiLM (Feature-wise Linear Modulation) to fuse DNA sequence and the context vector efficiently. It shows what inventive architectures can do when you need to combine heterogeneous data and still generalise well.
What it means in practice
- Faster biomedical research: developers can use Corgi to get quick predictions of gene activity across different cellular contexts, cutting both time and experiment costs.
- New drug development: the model helps identify candidate drug targets and predict their effect on gene regulation in specific cell types, which is critical for personalised medicine.
- FiLM and hybrid architectures: studying Corgi's design may inspire AI developers to apply the same approaches (FiLM, hybrid convolutional-transformer networks) in other domains that need heterogeneous data fused together.
- New options for sequencing data analysis: Corgi+ (the extended version of Corgi) can impute epigenomic tracks from RNA-seq data alone, making analysis cheaper and more accessible for labs with limited resources.
- Predicting genetic variant effects: the model can predict the effects of genomic variants in unstudied cell types, which matters for understanding disease mechanisms and building diagnostic tools.