Zeroshot Variant Effect Prediction

Zeroshot variant effect prediction refers to the task of predicting the functional impact of genetic variants, especially single nucleotide polymorphisms (SNPs), without requiring additional task-specific fine-tuning of the model. AIDO.ModelGenerator implements the procedure proposed by Nucleotide Transformer We use this to predict the effects of single nucleotide polymorphisms (SNPs) in the AIDO.DNA-300M model. This task uses the pre-trained models directly, and does not require finetuning.

To do zeroshot variant effect prediction,

1. Download human reference genome and raw clinvar data to GENBIO_DATA_DIR/ hg38.ml: https://hgdownload.soe.ucsc.edu/downloads.html

clinvar data: https://hgdownload.soe.ucsc.edu/gbdb/hg38/bbi/clinvar/

1. Run python preprocess_clinvar.py to process data into Clinvar_Processed.tsv file with gene_sequence,variant_sequence and label column.
2. Run mgen test --config config.yaml. If you want to take the norm distance between reference and variant sequence embeddings as prediction, the config should be

model:
  class_path: modelgenerator.tasks.ZeroshotPredictionDistance
  init_args:
    backbone: <you-choose>
data:
  class_path: ClinvarRetrieve
  init_args:
    method: Distance
    window: <window size centered around the SNPs>
    test_split_files:
      - <my_sequences.tsv>
    reference_file: <human_reference_genome.ml.fa> # Example: hg38.ml.fa
trainer:
  callbacks:
  - class_path: modelgenerator.callbacks.PredictionWriter
    dict_kwargs:
      output_dir: predictions
      filetype: tsv
      write_cols: ['score','norm_type','labels','num_layer']

If you take the loglikelihood ratio between reference and variant sequence embeddings at the mutation position as prediction, then the config should be

model:
  class_path: modelgenerator.tasks.ZeroshotPredictionDiff
  init_args:
    backbone: <you-choose>
data:
  class_path: ClinvarRetrieve
  init_args:
    method: Diff
    window: <window size centered around the SNPs>
    test_split_files:
      - <my_sequences.tsv>
    reference_file: <human_reference_genome.ml.fa> # Example: hg38.ml.fa
trainer:
  callbacks:
  - class_path: modelgenerator.callbacks.PredictionWriter
    dict_kwargs:
      output_dir: predictions
      filetype: tsv
      write_cols: ['score','label']

The labels and scores are also saved in test_predictions.tsv under the dir specified by --trainer.callbacks.output_dir.

Here are two examples of how to load HF model for inference For norm distance mode

mgen test --config Clinvar_300M_zeroshot_Distance.yaml

For loglikelihood ratio mode

mgen test --config Clinvar_300M_zeroshot_Diff.yaml