Fig. 2: The Nucleotide Transformer model accurately predicts diverse genomics tasks after fine-tuning.
From: Nucleotide Transformer: building and evaluating robust foundation models for human genomics
a, Performance results across downstream tasks for fine-tuned NT models as well as HyenaDNA, DNABERT and Enformer pre-trained models based on the MCC. We also trained BPNet models from scratch for comparison (original, 121,000 parameters; large, 28 million parameters). Data are presented as mean MCC ± 2 × s.d. from the tenfold cross-validation procedure (n = 10 for each point). b, Normalized mean of MCC performance across downstream tasks (divided by category) for all LMs after fine-tuning. c, The Multispecies 2.5B model performance on DNase I hypersensitive sites (DHSs), histone marks (HMs) and TF site predictions from different human cells and tissues compared to the baseline DeepSEA model. Each dot represents the receiver operating characteristic (ROC) AUC for a different genomic profile. The average AUC per model is labeled. d, The Multispecies 2.5B model performance on predicting splice sites from the human genome, compared to the SpliceAI and other splicing models. e, The Multispecies 2.5B model performance on developmental and housekeeping enhancer activity predictions from Drosophila melanogaster S2 cells, compared to the baseline DeepSTARR model.
