Downstream Tasks of LucaOne
Downstream tasks and fine-tuning code built on the LucaOne biological foundation model's embeddings, covering protein and nucleic-acid sequence- and token-level classification and regression.
This repository provides downstream tasks built on embeddings from LucaOne, a biological foundation model trained on nucleic-acid and protein sequences. It supports fine-tuning for both sequence-level tasks, such as protein family, solubility or promoter prediction, and token-level tasks, such as secondary structure, binding sites or post-translational modifications.
It supports Hugging Face's AutoModel, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification, AutoConfig and AutoTokenizer interfaces, so a single model architecture can be adapted to different biological modalities and custom downstream tasks through the standard transformers API. It also supports BF16 precision for embedding inference.
Model checkpoints are downloaded automatically from the project's own FTP server based on the chosen LucaOne model type, version and training step, and the code is written in Python under the Apache-2.0 license.