scTT: Translating Transcriptional Human-Mouse Single-Cell Signatures using Transformer-based Representational Learning

Example

import torch
from module import ScTT
from dataset import SingleDataset
from torch.utils.data import DataLoader

sctt = ScTT(
    n_genes = 10000,
    n_val = 11,
    n_class = 1000,
    n_species=2
    pooling='mean',
    embed_dim=768,
    n_heads=8,
    n_layers=8,
    lr=1e-4,
)

dataset = SingleDataset(adata,gene2id)
loader = DataLoader(dataset)

sctt.fit(model, loader)

Parameters

• n_genes: int.
  Amount of genes.
• n_val: int.
  Max expression value after preprocessing.
• n_class: int.
  Number of classes to classify.
• n_species: int.
  Number of species.
• embed_dim: int.
  Dimension of the embeddings.
• n_heads: int, default 8.
  Number of heads in Multi-head Attention layer.
• n_layers: int, default 8.
  Number of Transformer blocks.
• lr: float between [0, 1], default 0.0001.
  Learning rate.
• pooling: string, either max pooling, min pooling or mean pooling

References

Please consider citing the following reference:

https://icml-compbio.github.io/icml-website-2020/2020/papers/WCBICML2020_paper_29.pdf https://www.biorxiv.org/content/10.1101/2020.02.05.935239v2