usadellab/Helixer

What it solves

Helixer performs ab initio structural genome annotation, also known as "gene calling." It identifies which specific base pairs in a genome sequence belong to different parts of a gene, specifically the Untranslated Regions (UTR), Coding Sequences (CDS), and Introns, allowing researchers to generate primary gene models for various species.

How it works

Helixer combines Deep Neural Networks with a Hidden Markov Model (HMM) to predict gene structures. The process typically follows three steps:

1. Data Conversion: DNA sequences in FASTA format are converted into numerical matrices.
2. Inference: A deep learning model (with specific versions trained for fungi, land plants, vertebrates, and invertebrates) predicts base-wise probabilities for whether a base pair is intergenic, UTR, CDS, or an intron.
3. Post-processing: These probabilities are processed into primary gene models and exported as a GFF3 file.

Who it’s for

It is designed for bioinformaticians and genomic researchers who need to annotate genomes for species across different lineages (fungi, plants, vertebrates, and invertebrates).

Highlights

• Lineage-Specific Models: Provides pre-trained models tailored to four major biological lineages.
• High Performance: Optimized for GPU acceleration (Nvidia and Apple Silicon) to handle realistically sized datasets.
• Flexible Deployment: Available as a command-line tool, a Docker/Singularity container, a web tool, and a Galaxy installation.
• Reproducibility: Includes a --deterministic flag to ensure consistent results across different GPU architectures.