Gene-perspective Tasks

Figure 4 shows the experimental results for three gene-level tasks:

In Figure 4, (a) and (b) refer to the evaluation of Gene Function Prediction, (c) refers to the evaluation of Perturbation Prediction, and (d) and (e) refer to the evaluation of Gene Network Analysis. Our evaluation based on gene-perspective tasks shows that single-cell LLMs can handle tasks related to functions of genes.

Figure 4. Experimental results of single-cell LLMs and benchmarking methods for gene-level tasks. (a): Comparisions among Geneformer, scGPT and vanilla NN in the Gene Function Prediction task. (b): The effect of hyper-parameters including Loss weight, Bins and Learning rate for scGPT and Geneformer in the Gene Function Prediction task. (c): Correlation of GEARs and scGPT under different settings across different datasets. A higher correlation means lower rank and better performance. The numbers corresponding to the settings represent the average value across two datasets. (d): Dataset-level gene embeddings colored by the marker genes of different cell types. (e): Dataset-level gene embeddings colored by the Leiden cluster.