AdaGeneBudget: Cell-Adaptive Gene-Token Allocation for Efficient Single-Cell Foundation Models

AdaGeneBudget adaptively selects gene-token prefixes per cell in frozen-backbone scGPT and Geneformer, cutting mean gene-token counts and peak GPU memory while boosting embedding throughput up to 4.63x.

• Why it matters: Fixed input budgets and dataset-level HVG panels don’t jointly choose which genes to keep and how many tokens to allocate per cell, limiting efficiency without sacrificing biological fidelity.
• What they did: AdaGeneBudget is training-free: it ranks genes using each gene’s expression combined with inverse detection frequency from references, then retains the shortest prefix capturing a target fraction of each cell’s expression-specificity score mass.
• The result: Across four scGPT/Geneformer backbone–dataset pairs (Kang and PBMC), it reduced token counts and GPU memory and increased throughput up to 4.63x while preserving annotation utility, fine-grained/low-support identities, lineage-marker programs, and stimulation-associated pathway genes.

AI summary of the abstract · may contain errors · not medical advice · read the paper.