Tiberius: End-to-End Deep Learning with an HMM for Gene Prediction

For more information, see the Tiberius preprint.

Tiberius is a deep learning-based ab initio gene structure prediction tool that end-to-end integrates convolutional and long short-term memory layers with a differentiable HMM layer. It can be used to predict gene structures from genomic sequences only, while matching the accuracy of tool that use extrinsic evidence.

Currently, we provide only model weights for mammalian species and Tiberius does not predict alternative splicing variants.

Installation

Tiberius can either be installed from source or can be run with a Singularity container

Installation with Singularity container

Build Singularity container with:

singularity build tiberius.sif docker://larsgabriel23/tiberius:latest

Run Tiberius with the Singularity container (use -nv for GPU support):

 singularity run --nv tiberius.sif tiberius.py [options]

Installation from Source

Git Repositories

Clone the repository:

git clone https://github.com/Gaius-Augustus/Tiberius

Install learnMSA either from GitHub or with pip

pip install learnMSA

Python Libraries

The following Python libraries are required:

• tensorflow==2.10.*
• pyBigWig
• biopython
• bcbio-gff
• requests
• gzip
• bz2

They can be installed with:

pip install pyBigWig bio scikit-learn biopython bcbio-gff requests gzip bz2

Tensorflow should be installed with GPU support. If you are using conda, you can install Tensorflow 2.10 with these instructions.

Tiberius does also work with TensorFlow >2.10, however, it will produce an error if you use a sequence length > 260.000 during inference! You can install the current TensorFlow version with

python3 -m pip install tensorflow[and-cuda]

If you want to use GPUs, verify that TensorFlow is installed correctly with GPU support:

python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

Running Tiberius for Gene Prediction

To run Tiberius with bin/tiberius.py, you need to provide a FASTA file containing the genomic sequences. The sequence can either include repeat softmasking (recommended) or be run without softmasking. See softmasking_workflow for recommandations on how to mask repeats for Tiberius. Currently, we only provide weights for mammalian species, they will be downloaded automatically.

Running Tiberius with softmasked genome

If you want to run Tiberius with softmasking, model weights will be downloaded from https://bioinf.uni-greifswald.de/bioinf/tiberius/models/tiberius_weights.tgz into model_weights.

# Run Tiberius with softmasking
python bin/tiberius.py --genome input.fasta --out output.gtf

You can also manually download the weights and provide the path to the weights with the --model argument.

wget https://bioinf.uni-greifswald.de/bioinf/tiberius/models/tiberius_weights.tgz
tar -xzvf tiberius_weights.tgz
python bin/tiberius.py --genome input.fasta --out output.gtf --model path/to/tiberius_weights

Running Tiberius without softmasked genome

If you want to run Tiberius without softmasking, you can use the --no_softmasking argument. Tiberius will download the weights for the non-softmasking model automatically from https://bioinf.uni-greifswald.de/bioinf/tiberius/models/tiberius_nosm_weights.tgz into model_weights.

# Run Tiberius without softmasking
python bin/tiberius.py --genome input.fasta --out output.gtf --no_softmasking

You can also manually download the weights and provide the path to the weights with the --model argument. Note that in this case you have to provide the weights with --model_lstm to Tiberius.

wget https://bioinf.uni-greifswald.de/bioinf/tiberius/models/tiberius_nosm_weights.tgz
tar -xzvf tiberius_nosm_weights.tgz
python bin/tiberius.py --genome input.fasta --out output.gtf --model_lstm path/to/tiberius_nosm_weights

Running Tiberius with evolutionary information

To run Tiberius in de novo mode, evolutionary information data has to be generated with ClaMSA. See docs/clamsa_data.md for instructions on how to generate the data. Afterwards, you should have a directory with files named $clamsa/{prefix}{seq_name}.npz for each sequence of your FASTA file. You can then run Tiberius with the --clamsa argument. Note that your genome has to be softmasked for this mode and that you have to use different training weights than in ab initio mode. If not provided, they will automatically downloaded from https://bioinf.uni-greifswald.de/bioinf/tiberius/models/tiberius_denovo_weights.tgz into model_weights.

# Run Tiberius with softmasking
python bin/tiberius.py --genome input.fasta --clamsa $clamsa/{prefix} --out output.gtf

You can also manually download the weights and provide the path to the weights with the --model argument.

wget https://bioinf.uni-greifswald.de/bioinf/tiberius/models/tiberius_denovo_weights.tgz
tar -xzvf tiberius_denovo_weights.tgz
python bin/tiberius.py --genome input.fasta --out output.gtf --model_lstm path/to/tiberius_denovo_weights

Running Tiberius on Differnet GPUs

Tiberius can run on any GPU with at least 8GB of memory. However, you will need to adjust the batch size to match the memory capacity of your GPU using the --batch_size argument. Below is a list of recommended batch sizes for different GPUs: Here is a list of GPUs to batch siezes:

• A100 (80GB): batch size of 16
• RTX 3090 (25GB): batch size of 8
• RTX 2070 (8GB): batch size of 2

Tiberius produces a GTF file containing the predicted gene structures. It can also generate FASTA-formatted files of coding sequences and protein sequences when locations are specified using the --codingseq and --protseq options, respectively.

If you want to write custom code for Tiberius gene prediction, see example_prediction.ipynb for an example on how to write a custom prediction script.

Training Tiberius

Currently, we provide only model weights for mammalian species. If you want to train Tiberius on your own data, you need at least a genomic seqeunce file (FASTA) and reference annotations (GTF) for each species. Note that you can only train on genes with one transcript isoform per gene. Please remove alternative splicing variants before training. There two ways to train Tiberius:

1. Training Tiberius with a large dataset that does not fit into memory. See training_large_data.md for documentation on how to prepare a dataset and train Tiberius with it.
2. Training Tiberius with a small dataset that fits into memory. See example_train_full.ipynb for an example on how to load data and train Tiberius on a single genome. This can easily be adapted to train Tiberius on several genomes by first loading the data for all genome and then training the model. See training_large_data.md (Step 1) and softmasking_workflow.md for the preparation of the genome and annotation files.

Tiberius Model and Accuracies

Tiberius' model consists of a model that consist of CNN, biLSTM, and a differentiable HMM layer.

The model was trained end-to-end on 32 mammalian genomes and it reaches state-of-the-art prediction accuracies for mammalian genomes, matching the accuracy of extrinsic evidence based gene prediction tools.

Example Annotations from Tiberius

We provide example annotations for Homo sapiens (genome assembly GCF_000001405.40), Bos taurus (genome assembly GCF_000003205.7) and Delphinapterus leucas (genome assembly GCF_002288925.1) that were generated with Tiberius using the default weights:

wget https://bioinf.uni-greifswald.de/bioinf/tiberius/anno/Homo_sapiens.gtf.gz
wget https://bioinf.uni-greifswald.de/bioinf/tiberius/anno/Bos_taurus.gtf.gz
wget https://bioinf.uni-greifswald.de/bioinf/tiberius/anno/Delphinapterus_leucas.gtf.gz

Human genome annotation track on UCSC Genome Browser.

References

• Gabriel, Lars, et al. "Tiberius: End-to-End Deep Learning with an HMM for Gene Prediction." 2024, Bioinformatics, bioRxiv
• Popular science podcast on this paper generated with NotebookLM
• Processed RefSeq annotations used for training, validation and evaluation as described in the paper.
  wget https://bioinf.uni-greifswald.de/bioinf/tiberius/anno/ref_annot.tar.gz