A PyTorch-based framework for benchmarking active learning strategies on DNA sequence-to-expression models.

Trained on human RNA-seq data from ENCODE (with 866 datasets across diverse biosamples, including cell lines and adult tissues) and GTEx data (with 2-3 replicates for each tissue, processed by the recount3 project). The training dataset also includes epigenomic datasets from Enformer, such as CAGE, DNase-seq, ATAC-seq, and ChIP-seq tracks.

GAME was designed for the functional genomics community to create seamless communication across pre-trained models and genomics datasets. It is a product of the feedback from many model and dataset experts and our hope is that it allows for long-lasting benchmarking of models.

Agarwal et al. (2025) MPRA evaluator first requests log-scale predictions for ~56k probes across WTC11, K562, and HepG2 cells. It then calculates Pearson correlations between predicted and measured (log2(rna/dna)). It also assesses cell-type specificity by correlating predicted vs. observed expression differences.

DREAM-RNN model (Rafi et al. 2024, Nature Biotechnology) trained on MPRA data from Agarwal et. al 2023 and can be used for any sequence-to-expression predictions (single task).

The Orca Predictor returns log scaled interaction matrices for chromatin conformation predictions in H1-ESC cells. The predictions are a binned matrix where each value correspond to a 4kbp-by-4kbp interaction region, for example a 1Mbp input sequence prediction will return a 250x250 matrix.

The Enformer module encapsulates the human head of the large-scale multi-task Enformer model and can make predictions for sequences of any length. Its training data includes chromatin accessibility, DNA binding and expression.

CpG predictor serves as a baseline for models which are correlated with many genomic readouts. It ignores the type, cell type and species keys in the Evaluator requests since it is simply counting CpG instances.

The ChromBPNet Predictor encapsulates 9 individual chromatin accessibility models (5 folds for each model, some trained with ATAC-seq data and others with DNase) across five cell types.