Food harbors 135,000 ‘dark matter’ molecules that science doesn't yet understand
Network scientists have discovered that besides roughly 150 known nutritional components, food contains 135,000 unclassified molecules—they provide no energy but, like thousands of drugs, perform micro-level regulation inside cells.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Everything derives meaning from connection, not from existing alone
Barabási argues science has long focused on dissecting parts while forgetting the whole—genes, neurons, and people have no meaning in isolation; function emerges only through mutual connection. Consciousness is what neurons collectively do; life is what genes and molecules collectively do; society is what people collectively do. Network science begins by mapping connections first, then asking why they exist this way, rather than studying isolated components.
— BarabásiNetworks grow through winner-take-all mechanisms, not random chance
Traditional random graph theory assumed connections between nodes were purely determined by chance, predicting that node degrees would follow a bell curve. But actual data from the World Wide Web showed that 98% of web pages had only 1 to 5 links, while Google at the time had 800 million—a difference of several orders of magnitude. Barabási and collaborators discovered the cause: ‘preferential attachment’—when new nodes enter a network, they're more likely to connect to nodes already highly connected. This natural process creates hub nodes as networks grow, and the larger the network, the more extreme the gap between the largest hubs and smallest nodes becomes.
— BarabásiRandom graph theory began as pure mathematics, not social reality
Strogatz explains that Erdős and Rényi originally posed the random network model to solve a pure mathematical question—using the metaphor of buttons and string: scatter buttons on a table, then randomly connect them with string. When the ratio of string to buttons reaches a critical point, picking up one button suddenly brings up a whole cluster—a discontinuous phase transition. They were pursuing the elegance of this phase transition proof, not simulating how real people make friends in society. This is the fundamental gap between the random model and actual networks.
— StrogatzEighty-five percent of drugs never bind to the disease gene
Network medicine connects the ‘parts list’ from genomics with protein interactions to create a map. A disease's genes aren't randomly distributed in this network; they cluster in a specific ‘disease module’—like viewing a cell as a New York City map, psychiatric diseases clustered near Times Square, metabolic disease near Wall Street. Barabási's team found that 85% of drugs don't directly bind to the disease protein or gene itself, but instead act downstream. The disease gene usually affects too many critical processes—changing it is too risky. What matters is finding the ‘drug target’ at the right position downstream within the disease module.
— BarabásiOne blood test saves ninety thousand dollars and six months
Take rheumatoid arthritis: the disease module consists of 100 to 200 proteins. The mainstream drug Humira can only disrupt one side of the module. If the patient's disease manifests on the other side, the drug won't work for them—explaining the medical establishment's open secret that ‘most drugs only work for 20% to 40% of patients’. Cypher, the company Barabási helped found, uses a 16-gene biomarker to predict from a single blood sample whether Humira will work for a given patient. With Humira costing $90,000 per course and conventional drug trials taking six months to determine efficacy, this test eliminates the guesswork.
— BarabásiGenes explain at most one-fifth of disease—environment explains the rest
Among America's ten leading causes of death, genetics explains only 10% to 20% of disease risk; the rest comes from environment—exercise, sleep, stress. The biggest single environmental factor is food. Following this logic, the team began mapping every molecule in food into a network. They discovered that conventional nutrition science recognizes only roughly 150 known components, but food actually carries 135,000 previously unclassified molecules. These molecules provide no calories and are ‘invisible’ to metabolism, earning the name ‘dark matter of nutrition’. Take polyphenols: plants can't flee, so they synthesize massive quantities of chemical compounds for defense—the onion genome is twice the size of the human genome. When polyphenols enter the body, they aren't treated as nutrients but, like thousands of drugs, bind to proteins and DNA to fine-tune cellular activity.
— BarabásiBarabási sees neuroscience and programmable materials as network science's next frontiers
Barabási identifies the next major breakthrough for network science: the brain. Neurons are connected permanently by synapses—a ‘physical network’ unlike the optional handshakes of social networks. This dense physical connectivity itself determines brain structure and function. Another direction builds on his early background in materials science: 3D printing now operates at molecular scale. The question becomes: if we can print any network structure, no longer constrained to regular lattices or pure randomness, what networks should we print to create entirely new materials?
— BarabásiIn their own words · checked verbatim
And there was a guy came along and he put a big box on the table and said, you know, I just funded a company and why didn't you just take my t-shirt? And the t-shirt said, Google on it. It was Larry Page handing out as a graduate student, his first t-shirt.
Barabási12:25
This is what we call preferential attachment. And this growth and preferential attachment, we were able to show is responsible for the emergence of the hubs. That is, if a network is growing and the choices are not random, hubs will naturally emerge.
Barabási16:38
But in reality, 85% of the drugs do not bind to anything related to the disease. They don't bind to what we call disease proteins or disease genes.
Barabási29:14
The dirty secret of medicine is that virtually most drugs are only effective to 20, 30, 40% of the patients.
Barabási33:25
this is a $90,000 treatment. And it takes about six months or longer to know whether the drug has an effect for you, whether it works for you or not.
Barabási35:30
Now our databases have about 135,000 molecules that are not nutritional components that are carried by our food.
Barabási39:33
we called it the dark matter of nutrition because it's not existing within the can of nutrition.
Barabási43:43
Figures
| Drugs not binding to disease genes | 85% | 29:14 |
| Time to confirm drug efficacy | approximately 6 months or longer | 35:30 |
| Rheumatoid arthritis disease module proteins | 100 to 200 | 34:29 |
| Genes in the Cypher biomarker for Humira prediction | 16 | 34:29 |
| Known nutritional components in traditional nutrition science | approximately 150 | 39:33 |
| Percentage of disease causation explained by genetics | 10% to 20% | 37:30 |
| Google's link count at the time | 800 million | 17:39 |
| Percentage of web pages with 1 to 5 links | 98% | 17:39 |
Glossary
- preferential attachment
- The mechanism by which new nodes tend to connect to already-highly-connected nodes within a network.
- scale-free network
- A network where node degrees show no typical scale; a few hub nodes possess vastly more connections than others.
- disease module
- A local sub-network formed when disease-related genes cluster together in a biological network.
- foodome
- A research project mapping all molecules in food, not only known nutritional components.
- dark matter of nutrition
- Food molecules numbering in the hundreds of thousands that provide no calories but may regulate health.
How to listen
Researchers and product professionals interested in complex systems, biomedical data, precision medicine, or the intersection of food and health—who want to understand what practical tools network science provides.
The opening discussion of six degrees of separation and the Kevin Bacon game can be skipped without affecting the core content that follows.