Heart: +1.05y aging. Brain: -0.35y rejuvenation. Same gene. Same model.

Cross-Tissue Gene Perturbation on Apple Silicon

An MLX-native port of MaxToki (Gladstone + NVIDIA, April 2026), a 1B-parameter cell aging model trained on 175M human cells. We ran 20 gene perturbations across 3 tissues and found tissue-specific reversals in aging direction.

TL;DR

  • NVIDIA + Gladstone released MaxToki, an AI model that predicts how cells age. It's CUDA-only.
  • We ported it to Apple Silicon (MLX) so it runs on a MacBook at 15K tok/s.
  • We tested 20 gene perturbations across heart, brain, and skin using real single-cell data.
  • Key finding: the same gene (RASGEF1B) accelerates aging in heart tissue but reverses it in brain. All 7 paper-validated genes reproduced correctly.
3
Tissues
20
Genes Tested
~0.32
R² (LODO)
15K
tok/s on M2 Max
7/8
Paper Genes Correct
01 //

The Tissue-Specificity Finding

The MaxToki paper validated gene perturbation effects only in cardiac tissue. We extended the analysis to brain microglia and skin fibroblasts using real single-cell data from CELLxGENE (125K cardiomyocytes, 3.8K microglia, ~6K fibroblasts), running 60,000 forward passes across 500 cells per tissue on an M2 Max.

The result: RASGEF1B, the paper's strongest validated pro-aging driver in cardiomyocytes (+1.05y per perturbation - the authors confirmed cardiac function decline in mice within 6 weeks of overexpression), shows the opposite effect in brain microglia (-0.35y). Meanwhile APOE - the #1 Alzheimer genetic risk factor - is the strongest pro-aging signal in microglia (+0.26y) but has minimal effect in cardiac tissue.

7 of 8 paper-validated genes show the correct aging/rejuvenation direction in cardiac tissue. 11 additional literature-supported aging genes (SIRT1, FOXO3, LMNA, TP53, CDKN2A, etc.) also show biologically consistent effects. The cross-tissue divergence is the novel finding.

CELLxGENE cellsRank-value tokenizerMaxToki 217M (MLX)Embedding → Aging axisΔ years
02 //

System

MLX Port

MaxToki ships CUDA-only via NVIDIA BioNeMo. We extracted the HuggingFace safetensors (both 217M and 1B variants), converted to MLX format via mlx_lm.convert, and verified numerical parity against the reference implementation. Max absolute diff: 3.19e-5. Head_dim=154 (non-standard) handled via MLX's SDPA fallback.

Tokenizer

Gene vocabulary (20,271 tokens) extracted from the BioNeMo distcp checkpoint's io.json. Gene medians from Geneformer gc104M (100% overlap verified). Each cell tokenized as [BOS, gene_rank1, gene_rank2, ..., gene_rankK, EOS] with median-normalized expression ranking.

Aging Calibration

Ridge regression (α=1000) from 1,232-dim hidden states to chronological age. Fitted on 2,000 cardiomyocytes from the Litvinukova 2020 atlas (14 donors, ages 40-79). Leave-one-donor-out R²=0.32, MAE=5.67 years. Coefficient vector defines the aging axis for perturbation scoring.

Perturbation

Overexpression: move gene token to rank 1. Knockout: delete gene token. Run backbone forward pass (skip LM head), extract hidden state at EOS. Score: w · (perturbed_emb - baseline_emb), where w is the Ridge coefficient vector. Output is in calibrated year-equivalents.

03 //

Cross-Tissue Perturbation Map

20 genes tested across 3 tissues. Click a gene to see the predicted aging shift. Positive values = the perturbation accelerates cellular aging. Negative = rejuvenation. All results are overexpression direction.

RASGEF1Bvalidated in vivo
+0.00

estimated years of aging in ventricular cardiomyocyte

Validated in vivo alongside P4HA1. Pro-aging driver in cardiac tissue.

n=50040% expressIQR [+0.17, +1.75]
GeneEffect
validated in vivo (MaxToki paper)literature-supportedno dot = exploratory
04 //

Artifacts

maxtoki-mlx - Source code, MLX weights, tokenizer, perturbation pipeline. pip install maxtoki-mlx.

MaxToki Paper - Gomez Ortega et al., bioRxiv 2026. The original model by Gladstone + NVIDIA.

HuggingFace Weights - Apache 2.0 safetensors (217M and 1B variants).