Caduceus

Processes DNA sequences in both forward and reverse directions, capturing surrounding nucleotide information for more accurate modeling, as configurable via bidirectional_strategy settings in the training scripts.

Ensures consistent predictions regardless of DNA strand orientation, with Caduceus-PS variant designed for this without data augmentation, critical for biological accuracy in tasks like variant effect prediction.

Handles sequences up to 131,000 tokens, enabling analysis of extensive genomic regions, as demonstrated in pre-training on the human reference genome with max_length configurations.

Offers Caduceus-PS and Caduceus-Ph variants on Hugging Face for masked language modeling, allowing quick integration into existing pipelines without starting from scratch.

Provides configurable training pipelines for downstream tasks like GenomicBenchmarks and Nucleotide Transformer datasets, with options for conjoined training and testing to leverage equivariance.