model_id cell_embedding_dim static_gene_dim tokenizer_or_input max_length cell_layer_pooling perturbation_layer_pooling static_gene_parameter gene_id_mapping MetacellFormer 1024 1024 CP10K/nonzero-median rank; 20,080 genes 1024 final layer non-padding mean incl. CLS target×batch mean minus batch control, then L2 bert.embeddings.word_embeddings.weight Ensembl→HGNC; special-token offset +4 MetacellFormer-SC 1024 1024 same rank tokenizer; raw SC input 1024 final layer non-padding mean incl. CLS same target×batch delta protocol bert.embeddings.word_embeddings.weight Ensembl→HGNC; special-token offset +4 Geneformer-V2 768 768 native rank tokenization 4096 native penultimate-layer mean (emb_layer=-1) native penultimate layer (emb_layer=-1); CLS/EOS excluded input word embedding native token ID→Ensembl; shared gene roster scPRINT 256 256 native random-expression gene selection fixed by cell ID 2000 mean of four official cell-level class tokens same class-token feature then target×batch delta gene_encoder.embedding.weight native Ensembl IDs→shared gene roster scGPT 512 512 native 51-bin count tokenizer 1200 final CLS transformer output same CLS feature then target×batch delta token embedding weight native gene symbol/vocab→shared roster Geneformer-V1 256 256 native rank tokenization 2048 native penultimate-layer mean (emb_layer=-1) native penultimate layer (emb_layer=-1) input word embedding native token ID→Ensembl; shared gene roster CellFM 1536 1536 native weighted gene sampling without replacement; deterministic cell hash 2048 final-layer CLS same CLS feature then target×batch delta gene_emb CellFM native gene order→shared roster; exact-zero rows excluded for all