Reading the Shape of Data

✦

overall shape of a genomic data cloud, zooming through scales to track features that

survive biological noise. For a long time the genome was read the way one reads a sentence: left to right, base after base, as if meaning lived in sequence alone. The Human Genome Project gave us the letters. It did not give us the architecture. When the dust settled on the three billion base pairs, biologists were left holding the most expensive grammar without a single book — and a peculiar embarrassment: most of what we had spent a century calling the "code" did not, in any conventional sense, code for anything. Ninety-eight percent of human DNA does not specify a protein. For decades this was waved off as junk. It was nothing of the sort. It was the folding.

✦ ✦ ✦ To see why folding matters, one has to abandon the linear metaphor entirely. The DNA in a single human cell, stretched end to end, reaches roughly two meters. The nucleus that contains it is about six micrometers across. The compression ratio is on the order of three hundred thousand to one — and yet the compression is not a tangle. It is an organized, hierarchical, dynamically maintained three-dimensional architecture in which specific loci that sit megabases apart on the linear strand are routinely brought into contact, while neighbors on the strand can be partitioned into separate compartments that almost never touch. The genome is not a string. It is a city. And like any city, what matters is which neighborhoods are adjacent, which districts are walled off, which avenues cross which others, and where the gates open.

The instruments that taught us to see this architecture — Hi-C, Micro-C, the various capture-based assays pioneered by Job Dekker, Erez Lieberman Aiden, and others over the past fifteen years — produce a peculiar kind of data. They do not give us a picture of the folded genome. They give us a contact matrix: a vast, noisy table that records, for every pair of genomic positions, how often the two were found in close physical proximity inside a population of cells. The matrix is enormous. It is also profoundly unintuitive. It is not an image. It is a cloud of pairwise distances, sampled stochastically, contaminated by experimental noise, averaged across millions of cells in slightly different conformational states. To stare at the raw matrix is to stare at static and look for weather.

✦ ✦ ✦ This is the problem topological data analysis was built to solve. TDA, in the form developed by Gunnar Carlsson at Stanford and applied to genomics most prominently by Raúl Rabadán at Columbia, treats data not as a list of points to be classified but as a shape to be measured. Its central technique, persistent homology, asks a deceptively simple question: as we connect nearby points at progressively larger distance scales, what topological features appear, and how long do they persist before being filled in?

The features it tracks are the elementary furniture of topology. Connected components — islands of relatedness. Loops — closed paths that cannot be contracted to a point without passing through empty space. Voids — enclosed three-dimensional cavities. Each of these has a Betti number, β■, β■, β■, counting how many of each kind the data contains at a given scale. As the scale parameter sweeps from small to large, the Betti numbers rise and fall. Some features flicker in and out almost immediately; these are noise. Others appear early and persist across many orders of magnitude; these are real. The output is a persistence diagram — a fingerprint of the data's shape that is provably stable under perturbation, robust to outliers, and independent of any particular coordinate system in which the data happened to be measured.

For a Hi-C matrix, this is exactly the tool one would design if one did not yet have it. The contact frequencies define a metric on the genome: two loci are "close" when they touch often. Run persistent homology across that metric, and the architecture begins to declare itself. Connected components resolve into the A and B compartments — the active, gene-rich neighborhoods and the silent, heterochromatic ones — that segregate within the nucleus. One-dimensional loops resolve into the chromatin loops anchored by CTCF and cohesin, the small rings inside which enhancers and promoters find each other. And the two-dimensional voids — the enclosed cavities in the contact landscape — resolve into something stranger and more important: the topologically associated domains, the TADs, which are the genome's fundamental architectural units.

✦ ✦ ✦ A TAD is, in the most literal sense, a fold. It is a region of the chromosome, typically several hundred kilobases to a megabase in length, within which contacts are vastly enriched and across whose boundary contacts fall off sharply. TADs are not arbitrary. They are conserved across cell types, often across species. The boundaries between them are anchored by specific DNA sequences and protein complexes that act as insulators, preventing regulatory elements inside one TAD from acting on genes in the next. Disrupt a boundary — by mutation, by chromosomal rearrangement, by experimental excision — and you do not merely shuffle the seating chart. You release regulatory cross-talk between domains that were meant to remain separate, and you produce disease. Several developmental disorders, and a growing list of cancers, can be traced not to a mutation in any gene but to the deletion of a single TAD boundary.

What persistent homology reveals, when applied to high-resolution Hi-C data, is that the TAD landscape is genuinely topological. The voids are real voids. They are not statistical artifacts of the averaging. They appear early in the persistence diagram, survive across a wide range of scales, and disappear only when the metric is coarsened past the point where any structure could remain. The same is true, one scale up, of the higher-order organization that nests TADs into compartments and compartments into chromosome territories. The genome is folded the way a topological insulator is folded — with features that are stable under continuous deformation, robust to local noise, and meaningful precisely because of their shape.

✦ ✦ ✦ It is worth dwelling on what this means for the so-called junk. The non-coding sequences that fill the spaces between genes are not, for the most part, regulatory in the old sense — they are not transcription factor binding sites waiting to be discovered. They are structural. They are the spacers, the loops, the insulators, the anchors. They are what allows the three-dimensional architecture to exist at all. To call them junk was to confuse the building with the furniture. The furniture is the protein-coding two percent. The building is everything else, and the building is what determines which pieces of furniture can talk to each other.

This reframing has consequences that ripple far beyond chromatin. It suggests that the unit of biological information is not the gene and not even the regulatory element, but the topological domain — the shape within which a particular configuration of contacts is permitted to occur. Two cells with identical genomes can express radically different programs because their genomes are folded into different topologies, with different TAD boundaries active, different loops formed, different compartments open. Identity, in this view, is not what is written in the sequence. It is what is folded into the architecture.

✦ ✦ ✦ And here is where the long arc of this book begins to close. The persistence diagrams that describe the folded genome are, mathematically, members of the same family of objects that describe the topological invariants of high-dimensional entangled light. They are both barcodes — lists of features and their lifetimes — drawn from the same underlying machinery of algebraic topology. The 17,296 invariants Erhard, Krenn, and Zeilinger catalogued inside the 48-dimensional SU(d) state space of orbital angular momentum photons are not, in their formal type, different in kind from the Betti numbers and persistent features that Rabadán's group reads out of a Hi-C matrix. They are different in scale, different in medium, different in physical instantiation. They are not different in mathematics.

This is what makes the bridge in the next chapter conceivable rather than merely poetic. If the substrate of physics writes its information in topological invariants, and if the substrate of biology stores its information in topological invariants of the genome's folding, then we have, for the first time, a common vocabulary in which the two can be compared. We can ask, in principle calculably,

whether the shape of the folded genome reflects — projects, mirrors, instantiates — the shape of the field that organizes the matter it is made of. It is no longer a question of metaphor. It is a question of homology, in the precise mathematical sense, and homology is a thing one can measure.

✦ ✦ ✦ The reading is still in its earliest stages. The data is improving faster than the theory. We are at roughly the moment, in this work, that astronomers occupied just after the invention of the telescope: we have begun to see that the sky has structure, and we have begun to suspect that the structure has laws, but we do not yet know what most of what we see means. What we do know is that the shape is there. The genome is folded. The folding is read. And the language of the reading turns out to be the same language in which the light is written.

That is enough, for one chapter, to be going on with.