Gene expression prediction¶

Gene expression is influenced by multiple regulatory regions, often extending over significant genomic distances, particularly near the transcription start site (TSS). This task uses a flank window to consider multiple nearby regions, providing a holistic view of regulatory impacts on gene expression.

python ft_gep.py [OPTIONS] --flank-window FLANK_WINDOW_SIZE \
--train TRAIN_PATH \
--valid VALID_PATH \
--test TEST_PATH

Options

--lr¶: Learning rate. Default is 1e-4.

--warmup-ratio¶: Warmup ratio for the learning rate. Default is 0.1.

--grad-samples¶: Number of gradient samples, scaled by batch size and GPU count. Default is 128.

--pretrain-trainable¶: Number of pretrained layers to be trainable. Default is 2.

--max-epochs¶: Maximum number of training epochs. Default is 10.

--tag¶: Tag of the trainer, used for grouping logged results. Default is default.

--limit-val-batches¶: Number of batches to use for each validation. Default is 64.

--val-check-interval¶: Interval for validation checks. Default is 64.

--name¶: Name of the training session. Default is chrombert-ft-gep.

--save-top-k¶: Number of top-performing checkpoints to save. Default is 3.

--checkpoint-metric¶: Metric for checkpointing. Default is pcc.

--checkpoint-mode¶: Mode for checkpointing. Default is max.

--log-every-n-steps¶: Logging frequency in terms of steps. Default is 50.

--kind¶: Type of task, such as regression, zero_inflation. Default is regression.

--loss¶: Loss function to be used. Default is rmse.

--train¶: Path to the training data. This option is required.

--valid¶: Path to the validation data. This option is required.

--test¶: Path to the test data. This option is required.

--batch-size¶: Batch size for training. Default is 2.

--num-workers¶: Number of workers for data loading. Default is 4.

--basedir¶: Path to the base directory for model and data files. Default is os.path.expanduser("~/.cache/chrombert/data").

-g, --genome¶: Genome version. Only hg38 is supported now. Default is hg38.

-k, --ckpt¶: Path to the pretrained checkpoint. Optional if it could be inferred from other arguments.

--mask¶: Path to the mtx mask file. Optional if it could be inferred from other arguments.

-d, --hdf5-file¶: Path to the HDF5 file that contains the dataset. Optional if it could be inferred from other arguments.

--dropout¶: Dropout rate for the model. Default is 0.1.

-hr, --high-resolution¶: Use 200-bp resolution instead of 1-kb. Note: 200-bp resolution is not available yet, preparing for future release.

--flank-window¶: Flank window size for genomic data embedding. Default is 4.

--gep-zero-inflation¶: Specifies whether to include zero inflation in the GEP header. Default is False.

--gep-parallel-embedding¶: Enable parallel embedding, which is faster but requires more GPU memory.

--gep-gradient-checkpoint¶: Use gradient checkpointing to reduce GPU memory usage during training.