dgs

name

dgs

description

Natural-language control layer for DeepGeSeq (DGS), a deep learning toolkit for genomic sequence analysis. Use when the user asks to call DGS/DeepGeSeq from conversation, create or validate DGS JSON configs, run DGS CLI modes (`config`, `run`, `train`, `evaluate`, `explain`, `predict`), train/evaluate sequence models from FASTA/BED/BigWig data, interpret models with motif/attribution workflows, predict variant effects from VCF, use DGS Python APIs, inspect supported DGS models/data loaders, or handle Chinese requests such as DGS包调用、自然语言运行DGS、训练模型、评估模型、变异效应预测、解释模型、生成配置文件、检查DGS配置.

DGS

Translate a user's natural-language genomics task into a concrete DeepGeSeq action. Prefer the DGS CLI for complete workflows and the Python API for notebooks, exploratory analysis, custom model objects, profile models, sequence design, or low-level data operations.

Quick Workflow

1. Identify the intent: config, run, train, evaluate, explain, predict, Python API, or inspection.
2. Extract required inputs from the request: reference FASTA, intervals BED, target BED/BigWig tasks, VCF, checkpoint, model type, output directory, device, batch size, epochs, and target index.
3. Inspect the local DGS package when the exact API/model/config surface matters:

python /path/to/DeepGeSeq/DGS/skills/dgs/scripts/dgs_helper.py inspect --repo /path/to/DeepGeSeq

4. Create or edit a DGS config instead of assembling long CLI arguments. For new configs, use:

python /path/to/DeepGeSeq/DGS/skills/dgs/scripts/dgs_helper.py new-config --repo /path/to/DeepGeSeq --example full --output config.json

5. Validate before expensive work:

python /path/to/DeepGeSeq/DGS/skills/dgs/scripts/dgs_helper.py validate-config --repo /path/to/DeepGeSeq --config config.json --check-files

6. Run with global flags before the subcommand:

dgs --gpu -1 --seed 42 --no-benchmark train --config config.json

If DGS is not installed as a console command, run from the repository with:

python -m DGS.Main --gpu -1 train --config config.json

Intent Mapping

• "生成配置", "create config", "template": run dgs config --example minimal|full --output <file> or use dgs_helper.py new-config when paths/modes must be filled in.
• "完整流程", "run pipeline", "train then evaluate/predict": create config with modes, validate, then run dgs run --config <file>.
• "训练", "train model": set modes: ["train"] or use dgs train --config <file>.
• "评估", "evaluate/test metrics": require target data and either prior training in the same run or evaluate.checkpoint_path; run dgs evaluate --config <file>.
• "解释", "motif", "attribution", "DeepLIFT", "integrated gradients": require FASTA, intervals, model config/checkpoint context, and optional explain dependencies; run dgs explain --config <file>.
• "变异效应", "variant effect", "VCF", "predict variants": require FASTA and VCF; set predict.vcf_path; run dgs predict --config <file>.
• "Python API", "notebook", "custom model": import DGS modules directly and build loaders/trainers in code.
• "支持哪些模型/API": inspect DGS with the helper and read references/dgs-api.md.

Config Rules

Use JSON config as the handoff between natural language and DGS. Preserve user-provided paths exactly unless converting to absolute paths is useful for reproducibility.

Required by most modes:

• data.genome_path: reference FASTA.
• model.type: exported DGS model class, commonly CNN, DeepSEA, DanQ, Basset, BPNet, scBasset, ResidualNet, Enformer, ChromBPNet, or Borzoi.
• model.args: constructor kwargs; keep output_size aligned to target-task count for supervised scalar outputs.

Required for train/evaluate:

• data.intervals_path: BED-like intervals.
• data.target_tasks: list of {task_name, file_path, file_type} where file_type is bed or bigwig.
• train.optimizer and train.criterion for training. Use nested params.

Required for predict:

• predict.vcf_path.
• predict.sequence_length.

Useful runtime choices:

• Use --gpu -1 or device: "cpu" for CPU runs and smoke tests.
• Use low train.max_epochs, low data.batch_size, and data.num_workers: 0 for quick validation.
• Use data.train_test_split: "chromosome_split" with test_chroms/val_chroms when the user asks for chromosome holdout.

Python API

Use Python API when the user asks for code, notebooks, custom torch models, in-memory data, profile predictions, sequence design, or diagnostics that the CLI does not expose.

Common imports:

from DGS.Data import build_supervised_dataloaders, build_sequence_dataloader, build_profile_dataloaders
from DGS.DL.Trainer import Trainer
from DGS.DL.Predict import vep_centred_from_files, vep_centred_on_ds
from DGS.DL.Explain import calculate_attributions, motif_enrich
from DGS.DL.Design import gradient_ascent_sequence_design, greedy_ism_sequence_design
from DGS.Model import CNN, DeepSEA, DanQ, Basset, BPNet, ChromBPNet, Borzoi

Read references/dgs-api.md when using direct APIs, profile models, sequence design, foundation adapters, data loaders, or output-file conventions.

Execution Discipline

• Confirm missing biological inputs before running expensive workflows; do not invent FASTA/BED/BigWig/VCF paths.
• Prefer config validation with --check-files before training, explaining, or predicting.
• For large training or motif-discovery jobs, show the exact command and expected outputs before running unless the user already asked to execute.
• After running, report created artifacts such as DGS.log, metrics.csv, variant_predictions.csv, checkpoints, motif output directories, and any validation errors.
• If optional dependencies are missing, explain the missing extra: pip install -e ".[explain]" for Captum/tangermeme Python dependencies; modisco must still be available in PATH for motif reports.