DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA

DART-Eval is a suite of benchmarks to assess the utility of human regulatory DNA representations learned by self-supervised DNA Language Models. DART-Eval assesses zero-shot, probed, and fine-tuned performance on five sets of biologically relevant tasks of increasing difficulty.

For more details, see the DART-Eval paper (NeurIPS 2024): https://arxiv.org/pdf/2412.05430

Data

All data is available for download at Synapse project syn59522070.

The Synapse file repository is organized by task. Each task directory contains data.h5, an HDF5 file containing processed inputs and outputs for the task. See the Synapse project wiki for more information on the structure of the HDF5 files.

Additionally, for reproducibility, each task directory contains raw data, inputs, and final outputs for each model evaluated in the manuscript. (See the "Tasks" section for more information on each task.)

Tasks

The commands in this section reproduce the results for each task in the paper. The output files mirror the structure of the Synapse project.

Preliminaries

Prior to running analyses, set the $DART_WORK_DIR environment variable. This directory will be used to store intermediate files and results.

Additionally, download the genome reference files from syn60581044 into $DART_WORK_DIR/refs, keeping the file names. These genome references are used across all tasks.

In the following commands, $MODEL represents the evaluated DNALM architecture, one of caduceus, dnabert2, gena_lm, hyenadna, mistral_dna, and nucleotide_transformer. $MODEL_SPECIFIC_NAME represents the specific version of each model, namely one of caduceus-ps_seqlen-131k_d_model-256_n_layer-16, DNABERT-2-117M, gena-lm-bert-large-t2t, hyenadna-large-1m-seqlen-hf, Mistral-DNA-v1-1.6B-hg38, and nucleotide-transformer-v2-500m-multi-species.

Task 1: Prioritizing Known Regulatory Elements

All inputs, intermediate files, and outputs for this task are available for download at syn60581046.

Inputs

This task utilizes the set of ENCODE v3 candidate cis-regulatory elements (cCREs). A BED-format file of cCRE genomic coordinates is available at syn62153306. This file should be downloaded to $DART_WORK_DIR/task_1_ccre/input_data/ENCFF420VPZ.bed.

Dataset Generation

bash python -m dnalm_bench.task_1_paired_control.dataset_generators.encode_ccre --ccre_bed $DART_WORK_DIR/task_1_ccre/input_data/ENCFF420VPZ.bed --output_file $DART_WORK_DIR/task_1_ccre/processed_inputs/ENCFF420VPZ_processed.tsv

This script expands each element to 350 bp, centered on the midpoint of the element. The output file is a TSV with the following columns:

• chrom: chromosome
• input_start: start position of the length-expanded element
• input_end: end position of the length-expanded element
• ccre_start: start position of the original cCRE
• ccre_end: end position of the original cCRE
• ccre_relative_start: start position of the original cCRE relative to the length-expanded element
• ccre_relative_end: end position of the original cCRE relative to the length-expanded element
• reverse_complement: 1 if the element is reverse complemented, 0 otherwise

Zero-shot likelihood analyses

bash python -m dnalm_bench.task_1_paired_control.zero_shot.encode_ccre.$MODEL

Ab initio models

Extract final-layer embeddings

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.extract_embeddings.probing_head_like

Train probing-head-like ab initio model

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.ab_initio.probing_head_like

Evaluate probing-head-like ab initio model

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.eval_ab_initio.probing_head_like

Probing models

Extract final-layer embeddings from each model

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.extract_embeddings.$MODEL

Train probing models

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.train_classifiers.$MODEL

Evaluate probing models

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.eval_probing.$MODEL

Fine-tuned models

Train fine-tuned models

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.finetune.$MODEL

Evaluate fine-tuned models

bash python -m dnalm_bench.task_1_paired_control.supervised.encode_ccre.eval_finetune.$MODEL

Task 2: Transcription Factor Motif Footprinting

All inputs, intermediate files, and outputs for this task are available for download at syn60581043.

Inputs

This task utilizes the set of HOCOMOCO v12 transcription factor sequence motifs. A MEME-format file of motifs is available at syn60756095. This file should be downloaded to $DART_WORK_DIR/task_2_footprinting/input_data/H12CORE_meme_format.meme.

Additionally, this task utilizes a set of sequences and shuffled negatives generated from Task 1.

Dataset Generation

bash python -m dnalm_bench.task_2_5_single.dataset_generators.transcription_factor_binding.h5_to_seqs $DART_WORK_DIR/task_1_ccre/embeddings/probing_head_like.h5 $DART_WORK_DIR/task_2_footprinting/processed_data/raw_seqs_350.txt

bash python -m dnalm_bench.task_2_5_single.dataset_generators.motif_footprinting_dataset --input_seqs $DART_WORK_DIR/task_2_footprinting/processed_data/raw_seqs_350.txt --output_file $DART_WORK_DIR/task_2_footprinting/processed_data/footprint_dataset_350.txt --meme_file $DART_WORK_DIR/task_2_footprinting/input_data/H12CORE_meme_format.meme

Computing Zero-Shot Embeddings

bash python -m dnalm_bench.task_2_5_single.experiments.task_2_transcription_factor_binding.embeddings.$MODEL

bash python -m dnalm_bench.task_2_5_single.experiments.task_2_transcription_factor_binding.footprint_eval_embeddings --input_seqs $DART_WORK_DIR/task_2_footprinting/processed_data/footprint_dataset_350_v1.txt --embeddings $DART_WORK_DIR/task_2_footprinting/outputs/embeddings/$MODEL_SPECIFIC_NAME.h5 --output_file $DART_WORK_DIR/task_2_footprinting/outputs/evals/embeddings/$MODEL_SPECIFIC_NAME.tsv

Computing Zero-Shot Likelihoods

bash python -m dnalm_bench.task_2_5_single.experiments.task_2_transcription_factor_binding.likelihoods.$MODEL

bash python -m dnalm_bench.task_2_5_single.experiments.task_2_transcription_factor_binding.footprint_eval_likelihoods --input_seqs $DART_WORK_DIR/task_2_footprinting/processed_data/footprint_dataset_350_v1.txt --likelihoods $DART_WORK_DIR/task_2_footprinting/outputs/likelihoods/$MODEL_SPECIFIC_NAME.tsv --output_file $DART_WORK_DIR/task_2_footprinting/outputs/evals/likelihoods/$MODEL_SPECIFIC_NAME.tsv

Further Evaluation Notebooks

dnalm_bench/task_2_5_single/experiments/eval_footprinting_likelihood.ipynb - figure production for likelihood-based evaluation

dnalm_bench/task_2_5_single/experiments/eval_footprinting_embedding.ipynb - figure production for embedding-based evaluation

dnalm_bench/task_2_5_single/experiments/footprinting_pairwise.ipynb - cross-model pairwise production plots

dnalm_bench/task_2_5_single/experiments/footprinting_conf_intervals.ipynb - confidence interval calculation

Task 3: Discriminating Cell-Type-Specific Elements

All inputs, intermediate files, and outputs for this task are available for download at syn60581042.

Inputs

This task utilizes ATAC-Seq experimental readouts from five cell lines. Input files are available at syn60581166. This directory should be cloned to $DART_WORK_DIR/task_3_peak_classification/input_data. <!-- For this task, let $CELL_TYPE represent one of the following cell lines: GM12878, H1ESC, HEPG2, IMR90, or K562. -->

Dataset Generation

Using the input peaks from ENCODE, generate a consensus peakset:

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.make_consensus_peakset

Then, generate individual counts matrices for each sample, using input BAM files from ENCODE and the consensus peakset:

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.generate_indl_counts_matrix GM12878 $BAM_FILE

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.generate_indl_counts_matrix H1ESC $BAM_FILE

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.generate_indl_counts_matrix HEPG2 $BAM_FILE

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.generate_indl_counts_matrix IMR90 $BAM_FILE

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.generate_indl_counts_matrix K562 $BAM_FILE

Concatenate the counts matrices and generate DESeq inputs:

bash python -m dnalm_bench.task_2_5_single.dataset_generators.peak_classification.generate_merged_counts_matrix

Finally, run DESeq for each cell type to obtain differentially accessible peaks for each cell type:

bash Rscript dnalm_bench.task_2_5_single.dataset_generators.peak_classification.DESeqAtac.R

The final output consists of the differentially accessible peaks, available at syn61788656.

Zero-shot baseline clustering

Use FIMO to generate motif scores for each peak sequence.

The following notebook contains information on how to produce the zero-shot clustering results, using the motif counts from FIMO: bash dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.baseline.zero_shot_clustering_baseline.ipynb

Zero-shot embedding clustering

This depends on the final-layer embeddings generated for the probed models.

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.cluster.run_clustering_subset $DART_WORK_DIR/task_3_peak_classification/embeddings/$MODEL_SPECIFIC_NAME.h5 $DART_WORK_DIR/task_3_peak_classification/processed_inputs/peaks_by_cell_label_unique_dataloader_format.tsv $DART_WORK_DIR/task_3_peak_classification/processed_inputs/indices_of_new_peaks_in_old_file.tsv $DART_WORK_DIR/task_3_peak_classification/clustering/$MODEL_SPECIFIC_NAME/

Ab initio models

Here, $AB_INITIO_MODEL is one of probing_head_like or chrombpnet_like (ChromBPNet-like).

Extract final-layer embeddings (probing_head_like only)

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.extract_embeddings.$AB_INITIO_MODEL

Train ab initio models

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.baseline.$AB_INITIO_MODEL

Evaluate ab initio models

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.eval_baseline.$AB_INITIO_MODEL

Probing models

Extract final-layer embeddings from each model

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.extract_embeddings.$MODEL

Train probing models

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.train.$MODEL

Evaluate probing models

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.eval_probing.$MODEL

Fine-tuned models

Train fine-tuned models

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.finetune.$MODEL

Evaluate fine-tuned models

bash python -m dnalm_bench.task_2_5_single.experiments.task_3_peak_classification.eval_finetune.$MODEL

Task 4: Predicting Chromatin Activity from Sequence

All inputs, intermediate files, and outputs for this task are available for download at syn60581041.

Inputs

This task utilizes DNAse-Seq experimental readouts from five cell lines. Input files are available at syn60581050. This directory should be cloned to $DART_WORK_DIR/task_4_peak_classification/input_data.

For this task, let $CELL_TYPE represent one of the following cell lines: GM12878, H1ESC, HEPG2, IMR90, or K562.

Probing models

Extract final-layer embeddings from each model. This should be done for each value of $CATEGORY in ['peaks', 'nonpeaks', 'idr'].

bash python -m dnalm_bench.task_2_5_single.experiments.task_4_chromatin_activity.extract_embeddings.$MODEL $CELL_TYPE $CATEGORY

Train probing models

bash python -m dnalm_bench.task_2_5_single.experiments.task_4_chromatin_activity.train.$MODEL

Evaluate probing models

bash python -m dnalm_bench.task_2_5_single.experiments.task_4_chromatin_activity.eval_probing.$MODEL

Fine-tuned models

Train fine-tuned models

bash python -m dnalm_bench.task_2_5_single.experiments.task_4_chromatin_activity.finetune.$MODEL

Evaluate fine-tuned models

bash python -m dnalm_bench.task_2_5_single.experiments.task_4_chromatin_activity.eval_finetune.$MODEL

ChromBPNet models

Evaluate ChromBPNet Models

bash python -m dnalm_bench.task_2_5_single.experiments.task_4_chromatin_activity.eval_ab_initio.chrombpnet_baseline $CELL_TYPE $CHROMBPNET_MODEL_FILENAME

Task 5: Chromatin Activity Variant Effect Prediction

All inputs, intermediate files, and outputs for this task are available for download at syn60581045.

Inputs

This task utilizes genomic QTL variants from two studies: African caQTLs (Degorter et al.) and Yoruban dsQTLs (Degner et al.). Input TSV files of variants and experimental effect sizes are available at syn60756043 and syn60756039. These files should be downloaded to $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/Afr.CaQTLS.tsv and $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/yoruban.dsqtls.benchmarking.tsv respectively.

Zero-shot embedding-based scoring

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.zero_shot_embeddings.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/Afr.CaQTLS.tsv Afr.CaQTLS $DART_WORK_DIR/refs/GRCh38_no_alt_analysis_set_GCA_000001405.15.fasta

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.zero_shot_embeddings.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/yoruban.dsqtls.benchmarking.tsv yoruban.dsqtls.benchmarking $DART_WORK_DIR/refs/male.hg19.fa

Zero-shot likelihood-based scoring

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.zero_shot_likelihoods.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/Afr.CaQTLS.tsv Afr.CaQTLS $DART_WORK_DIR/refs/GRCh38_no_alt_analysis_set_GCA_000001405.15.fasta

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.zero_shot_likelihoods.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/yoruban.dsqtls.benchmarking.tsv yoruban.dsqtls.benchmarking $DART_WORK_DIR/refs/male.hg19.fa

Supervised probing model scoring

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.probed_log_counts.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/Afr.CaQTLS.tsv $DART_WORK_DIR/task_5_variant_effect_prediction/outputs/probed/$MODEL/Afr.CaQTLS $DART_WORK_DIR/refs/GRCh38_no_alt_analysis_set_GCA_000001405.15.fasta

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.probed_log_counts.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/yoruban.dsqtls.benchmarking.tsv $DART_WORK_DIR/task_5_variant_effect_prediction/outputs/probed/$MODEL/yoruban.dsqtls.benchmarking $DART_WORK_DIR/refs/male.hg19.fa

Supervised fine-tuned model scoring

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.finetuned_log_counts.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/Afr.CaQTLS.tsv $DART_WORK_DIR/task_5_variant_effect_prediction/outputs/fine_tuned/$MODEL_FOLDER/Afr.CaQTLS $DART_WORK_DIR/refs/GRCh38_no_alt_analysis_set_GCA_000001405.15.fasta

bash python -m dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.finetuned_log_counts.$MODEL $DART_WORK_DIR/task_5_variant_effect_prediction/input_data/yoruban.dsqtls.benchmarking.tsv $DART_WORK_DIR/task_5_variant_effect_prediction/outputs/fine_tuned/$MODEL_FOLDER/yoruban.dsqtls.benchmarking $DART_WORK_DIR/refs/male.hg19.fa

Evaluation Notebooks

Helper functions called in the evaluation notebooks: dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.variant_tasks.py

Zero Shot Evaluation Notebook: dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.Zero_Shot_Final.ipynb

Probed Evaluation Notebook: dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.Probed_Final_Counts.ipynb

Finetuned Evaluation Notebook: dnalm_bench.task_2_5_single.experiments.task_5_variant_effect_prediction.Finetuned_Final_Counts.ipynb