← home

$ cat blog/engineering-cell-part-1

The 17,000-Dimensional Elephant

Part 1 of Engineering a Cell · 2026-03-20

Engineering a Cell

I'll be honest upfront: I don't have a deep biology background. My world is applied math, machine learning, and systems architecture. So when I look at a cancer cell, I don't see a biological mystery—I see a distributed system with a rogue routing protocol.

For the last couple of months, I've been trying to mathematically model the Mesenchymal phenotype, a highly aggressive state where cancer cells gain metastatic properties. My goal wasn't just to predict if a cell would enter this state, but to build a control engine capable of calculating the exact perturbations needed to force it back to a healthy baseline.

Building this pipeline—from scraping noise to fighting compilers, to eventually linearizing a biological network—felt exactly like architecting a backend data pipeline. Here is how I hacked the state-space.

The 17,000-Dimensional Elephant in the Room

When you first download transcriptomic data from the Cancer Dependency Map (DepMap), it doesn't look like biological insight; it looks like a wall of static. I was staring at a matrix mapping hundreds of cancer cell lines against 18,435 individual continuous gene expression scores.

You can't build a dynamic Markov model with 18,000 variables. The state-space explosion would instantly melt my compute environment. I needed to find the "Master Regulators"—the exact genes driving the variance of this specific Mesenchymal state.

Before running any models, I had to clean the data. Highly sparse genes introduce massive mathematical noise, so I wrote a strict filtering pipeline dropping any column with more than 90% missing values, followed by mean imputation. With a dense, continuous expression matrix ready, I moved to dimensionality reduction.

The Unscaled PCA "Aha" Moment

Standard data science practice dictates scaling your data (Z-score normalization) before running Principal Component Analysis (PCA). But biological thresholding is a different beast. If a Mesenchymal gene is massively over-expressed, that absolute magnitude matters. It's a physical reality of the cell, not just a statistical outlier.

By running PCA on the unscaled expression data, I isolated the primary phenotypic axis. The results were stark. The data cleanly bifurcated into two distinct clusters: the positive-loading Mesenchymal Activators (like TGFBI, FN1), and the negative-loading Inhibitors.

I had my core 30-gene Mesenchymal signature. Now, I needed to make it move over time.