Meet MaxToki: The AI That Predicts How Your Cells Age — and What to Do About It

Why It Matters

Temporal AI like MaxToki transforms aging research by enabling precise, longitudinal insight into cellular trajectories, accelerating drug target discovery and personalized interventions for age‑related diseases.

Key Takeaways

• Transformer predicts cell age trajectory from transcriptome snapshots.
• Median age prediction error 87 months, half baseline error.
• Works on unseen cell types, diseases, showing strong generalization.
• Identifies pro‑aging drivers validated in mouse models.
• Scales using FlashAttention‑2 on NVIDIA H100 GPUs.

Pulse Analysis

Traditional single‑cell models treat each cell as a static snapshot, limiting their ability to capture the slow, progressive changes that drive aging and chronic disease. MaxToki overcomes this by adopting a transformer‑decoder architecture—similar to large language models—but training it on rank‑value encoded gene expression data from 175 million healthy cells across diverse tissues. The rank‑based representation downplays ubiquitous housekeeping genes and highlights dynamic transcription factors, while the extended context length enables the model to reason across multiple sequential cell states, effectively learning a temporal dimension of biology.

The model’s novel temporal prompting strategy lets it either predict the elapsed months between observed cell states or generate a plausible future transcriptome for a given time interval. In benchmark tests, MaxToki’s median error of 87 months outperforms linear regressors (≈178 months) and naïve baselines, and it maintains a Pearson correlation above 0.80 on completely unseen cell types and donor ages. When applied to disease samples it quantifies age acceleration—~5 years for smokers’ lung epithelium, ~15 years for pulmonary fibrosis fibroblasts, and ~3 years for Alzheimer’s microglia—while distinguishing resilient patients who show no acceleration, highlighting its potential for clinical stratification.

Beyond scientific insight, MaxToki demonstrates engineering scalability, leveraging FlashAttention‑2 and mixed‑precision bf16 on NVIDIA H100 GPUs to achieve five‑fold training throughput gains. Its open‑source release invites the community to fine‑tune the model for specific tissues or pathologies, accelerating the pipeline from in‑silico hypothesis to in‑vivo validation. As longitudinal single‑cell atlases expand, temporal foundation models like MaxToki are poised to become indispensable tools for pinpointing intervention points before age‑related diseases manifest, reshaping drug discovery and precision medicine.