Why a PhD Still Matters in the Age of AI
Every few months, someone declares that the PhD is finished. This year a former AI leader told Business Insider that years spent on a doctorate may be wasted, because AI will have moved on before you graduate. In late September, MIT's Phillip Isola published an essay for his own students making the opposite case: human expertise will matter for a long time, and in the end a PhD can be pursued for the love of the game.
We side with Isola. Here is why, from where we work: single-molecule sensing.
A PhD trains you to find the state of the art
AI can summarise a literature in seconds. What it cannot yet do for you is stand at the edge of a field, tell what is known from what is merely repeated, and decide which experiment is worth a year of your life. That is what a PhD trains. You learn to identify the state of the art, to distrust a beautiful dataset until you know where it came from, and to ask the question nobody has asked yet.
A field built by stubborn people
Single-molecule science has Nobel Prizes, breakthroughs and very human stories. In 1989, W. E. Moerner and his postdoc Lothar Kador at IBM became the first to detect the light absorption of a single molecule: pentacene in a crystal, cooled close to absolute zero. It had long been thought impossible. In 1990, Michel Orrit showed that fluorescence made single-molecule detection far more practical, and fluorescence became the method of choice. When the 2014 Nobel Prize in Chemistry recognised super-resolved fluorescence microscopy, it went to Moerner, Eric Betzig and Stefan Hell. A prize has room for only three names. A field is always built by many more.
Listening to light in a glass bead
Our own corner of the field starts with an old idea. In a whispering gallery, a whisper travels along the curved wall to the far side. Light can do the same inside a glass microsphere only tens of micrometres wide, circling its inner surface. When a molecule lands on the surface, the resonance shifts, and a sensitive instrument can read that shift without any fluorescent label:
2002: label-free protein detection from the shift of a microsphere resonance (bovine serum albumin, and streptavidin binding to biotin).
2008: single influenza A virus particles seen as discrete steps in the resonance, and a Nature Methods review with Arnold on detection down to single molecules.
2014: single-molecule DNA hybridisation, down to 8-mer oligonucleotides, by adding gold nanorods to the microsphere.
2016: single zinc and mercury ions interacting with gold nanorods in water.
Since then: single enzymes at work, from DNA polymerase conformational changes to enzyme heat capacity, and in 2024 optical forces from the sensor used to apply a free-energy penalty and to start regulating a single enzyme.
Chips, tricorders and robot enzymologists
Why chase single molecule sensors? Because they let us watch how biomolecules work, their dynamics and mechanisms, instead of an average over billions of copies. Shrunk onto chips, such sensors could change healthcare, medicine and environmental monitoring.
Are you old enough for Star Trek? The Qualcomm Tricorder XPRIZE, launched in 2012, was won in 2017 by Final Frontier Medical Devices with DxtER, a kit that diagnoses 13 conditions and tracks five vital signs. It is closer to a medical kit than a whirring, beeping tricorder, and a long way from sensing single molecules, but the dream is shared. Add an electronic sense of smell. Then put such sensors inside robot enzymologist laboratories: self-driving labs, like one at the Great Lakes Bioenergy Research Center that pairs machine-learning agents with robotics to engineer enzymes, but with each experiment read out at the single-molecule level.
Where AI comes in
Here is the twist that makes this a great time to be a single-molecule scientist. The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design, and to Demis Hassabis and John Jumper for AlphaFold, which predicts protein structures. AlphaFold learned from the vast archive of known sequences and structures that generations of experimentalists built. That archive is exactly what we can use to decode our own signals. A single-molecule trace is a fingerprint of a protein moving, binding, opening and closing. Machine learning already helps with this kind of decoding: a convolutional neural network for nanopore sensing, QuipuNet, classified events more accurately than earlier methods and increased the number of analysable events fivefold.
But someone has to know what the trace means. Someone has to tell a real conformational change from an artefact, design the control experiment, and notice when a model is confidently wrong. Isola suggests that this role, expertise as a kind of safety infrastructure for society, may be one of the most enduring jobs of the PhD.
So, is a PhD still relevant?
Yes. AI is a superb collaborator, but a PhD teaches you to find the frontier, to read data critically, and to work on problems whose answers are not in the training data yet. In single-molecule sensing those answers are being written now.
Which field would you spend a PhD in?
Comments