Skip to content

Experiment: CDS from different evolutionary timescales #58

@gonzalobenegas

Description

@gonzalobenegas

Description

Similar to #55 but for CDS.

Perhaps train for 17k steps rather than a full epoch.

Links

Training code
Analysis code
Wandb: mammals, vertebrates, animals

Results

Eval: TraitGym Mendelian traits v2 (validation set, only odd chromosomes).

  • For missense variants, shorter timescales work better initially but the longer timescales surpass them in the end. Animals seems to be the optimal one.
  • For synonymous and splicing variants, there seems to be less differences between models.
Image

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions