MAM: Multinomial Attention Masking for Foundation Models on Sparse Single-Cell RNA-seq Data

Amirreza Naziri, Arash Asgari, Aijun An, Eleftherios Sachlos, Laleh Seyyed-Kalantari
Proceedings of the 7th Conference on Health, Inference, and Learning, PMLR 333:295-311, 2026.

Abstract

Single-cell RNA sequencing (scRNA-seq) has transformed biology by enabling the measurement of gene expression across millions of individual cells, revealing cellular heterogeneity that underlies development, disease progression, and treatment response. This has made scRNA-seq a central data modality in modern biology and drug discovery. Recently, transformer-based foundation models (FMs) have shown strong potential for scRNA-seq analysis, but they often rely on random masking during training. Due to the extreme sparsity of scRNA-seq datasets, conventional uniform masking samples genes without considering their biological importance. In this work, we propose Multinomial Attention Masking (MAM), a module that learns which gene positions are most informative to mask at each training step. We define a set of trainable latent vectors that attend over gene embeddings to produce attention maps, from which a multinomial sampler selects the highest-scoring positions for masking. We show MAM improves FMs pretraining performance and consistently outperforms uniform masking on cell-type classification tasks, while adding negligible computational overhead. Our work benefits researchers building FMs for sparse data and those rely on accurate scRNA-seq analysis to study cell types and disease.

Cite this Paper

BibTeX
@InProceedings{pmlr-v333-naziri26a, title = {MAM: Multinomial Attention Masking for Foundation Models on Sparse Single-Cell RNA-seq Data}, author = {Naziri, Amirreza and Asgari, Arash and An, Aijun and Sachlos, Eleftherios and Seyyed-Kalantari, Laleh}, booktitle = {Proceedings of the 7th Conference on Health, Inference, and Learning}, pages = {295--311}, year = {2026}, editor = {Healey, Elizabeth and Fries, Jason and Pollard, Tom and Tang, Shengpu and Zink, Anna and Hartvigsen, Tom and Agrawal, Monica and Finlayson, Sam and Glicksberg, Benjamin and Beaulieu-Jones, Brett and Wang, Kai and Fontalvo, Daseyra and Sarker, Tasmie and Chen, Irene and Alsentzer, Emily}, volume = {333}, series = {Proceedings of Machine Learning Research}, month = {29--30 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v333/main/assets/naziri26a/naziri26a.pdf}, url = {https://proceedings.mlr.press/v333/naziri26a.html}, abstract = {Single-cell RNA sequencing (scRNA-seq) has transformed biology by enabling the measurement of gene expression across millions of individual cells, revealing cellular heterogeneity that underlies development, disease progression, and treatment response. This has made scRNA-seq a central data modality in modern biology and drug discovery. Recently, transformer-based foundation models (FMs) have shown strong potential for scRNA-seq analysis, but they often rely on random masking during training. Due to the extreme sparsity of scRNA-seq datasets, conventional uniform masking samples genes without considering their biological importance. In this work, we propose Multinomial Attention Masking (MAM), a module that learns which gene positions are most informative to mask at each training step. We define a set of trainable latent vectors that attend over gene embeddings to produce attention maps, from which a multinomial sampler selects the highest-scoring positions for masking. We show MAM improves FMs pretraining performance and consistently outperforms uniform masking on cell-type classification tasks, while adding negligible computational overhead. Our work benefits researchers building FMs for sparse data and those rely on accurate scRNA-seq analysis to study cell types and disease. } }
Endnote
%0 Conference Paper %T MAM: Multinomial Attention Masking for Foundation Models on Sparse Single-Cell RNA-seq Data %A Amirreza Naziri %A Arash Asgari %A Aijun An %A Eleftherios Sachlos %A Laleh Seyyed-Kalantari %B Proceedings of the 7th Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2026 %E Elizabeth Healey %E Jason Fries %E Tom Pollard %E Shengpu Tang %E Anna Zink %E Tom Hartvigsen %E Monica Agrawal %E Sam Finlayson %E Benjamin Glicksberg %E Brett Beaulieu-Jones %E Kai Wang %E Daseyra Fontalvo %E Tasmie Sarker %E Irene Chen %E Emily Alsentzer %F pmlr-v333-naziri26a %I PMLR %P 295--311 %U https://proceedings.mlr.press/v333/naziri26a.html %V 333 %X Single-cell RNA sequencing (scRNA-seq) has transformed biology by enabling the measurement of gene expression across millions of individual cells, revealing cellular heterogeneity that underlies development, disease progression, and treatment response. This has made scRNA-seq a central data modality in modern biology and drug discovery. Recently, transformer-based foundation models (FMs) have shown strong potential for scRNA-seq analysis, but they often rely on random masking during training. Due to the extreme sparsity of scRNA-seq datasets, conventional uniform masking samples genes without considering their biological importance. In this work, we propose Multinomial Attention Masking (MAM), a module that learns which gene positions are most informative to mask at each training step. We define a set of trainable latent vectors that attend over gene embeddings to produce attention maps, from which a multinomial sampler selects the highest-scoring positions for masking. We show MAM improves FMs pretraining performance and consistently outperforms uniform masking on cell-type classification tasks, while adding negligible computational overhead. Our work benefits researchers building FMs for sparse data and those rely on accurate scRNA-seq analysis to study cell types and disease.
APA
Naziri, A., Asgari, A., An, A., Sachlos, E. & Seyyed-Kalantari, L.. (2026). MAM: Multinomial Attention Masking for Foundation Models on Sparse Single-Cell RNA-seq Data. Proceedings of the 7th Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 333:295-311 Available from https://proceedings.mlr.press/v333/naziri26a.html.

Related Material