Akram Baharlouei is a machine learning engineer at Altos Labs developing single-cell foundation models to illuminate cellular health, rejuvenation, and responses to treatment. Her path into computational biology runs through wireless communications, AI infrastructure at Meta, and research on how faithfully machine learning captures living cells.
She earned a doctorate in electrical and computer engineering at George Mason University and worked at Qualcomm before joining Meta, where she contributed to MultiRay, infrastructure for sharing large-model representations across applications. At Altos, she has applied that systems background to a harder question: whether biological measurements contain enough reliable information to predict what an individual cell will do.
- Measurement quality before model scale. Single-cell RNA sequencing records momentary gene-expression snapshots, not continuous cellular behavior. Natural variation, intermittent gene activity, and differences between laboratories complicate comparisons; microscopy, spatial measurements, and protein data provide essential complementary information.
- Multimodal single-cell benchmarking. Baharlouei coauthored scGeneScope, a NeurIPS 2025 benchmark combining single-cell transcriptomics and microscopy across chemical treatments. Its evaluations found established fit-to-data methods outperforming several zero-shot RNA foundation models on treatment-identification tasks.
- Biological fidelity over language-model analogies. Models such as scGPT and Geneformer treat genes as tokens, but compressing cellular measurements into latent vectors can discard information needed for cellular perturbation-response prediction. Baharlouei argues that expensive transformers must prove their advantage over simpler linear baselines.
- Distributions instead of average cells. Cells respond differently to identical interventions. In her AI Engineer World’s Fair talk, Baharlouei highlighted flow-matching models, including PRiMeFlow, as promising approaches developed by other researchers for modeling the full distribution of possible cellular responses.
