Most antibody drugs work like a bouncer on the door. They grab their target on the outside surface of a cell and never step through it. For a lot of medicine that is fine. For the biggest neurodegenerative diseases it is close to useless, because the proteins that drive Alzheimer's, Parkinson's, Huntington's and motor neurone disease do their damage inside the cell — misfolding and clumping in the crowded interior of a neuron, exactly where an ordinary antibody cannot follow.

A team at the University of Essex has used AI to get inside. In a study in Nature Communications, Dr Caitlin O'Shea and Dr Gareth Wright report converting 672 different antibodies into "intrabodies" — antibody fragments re-engineered to stay stable and keep gripping their target in the one place the disease actually happens.

The problem was never the target. It was the chemistry of getting there.

Intrabodies are not a new idea. The trouble has always been that antibodies evolved to work in the watery space between cells, and the inside of a cell is a different world — reducing, packed with protein, chemically hostile to a molecule built for the outside. Drop a normal antibody fragment in there and it tends to unfold and stick to itself into a useless clump.

The Essex team traced a big part of that failure to a single, unglamorous property: electrical charge. "Antibodies usually have the wrong charge to exist inside cells without sticking together," O'Shea explained. Get the charge wrong and the fragment aggregates before it can do anything.

That reframing is what made the problem tractable for a computer. Rather than screen antibodies one at a time in the lab, the researchers compared millions of antibody sequences against the proteins found inside human cells, worked out what surface charge a fragment needed to survive intracellularly, and then used protein-design software from the group of Nobel laureate David Baker to redesign each fragment to hit that spec — keeping the part that recognises the disease protein, rewriting the surface so it stays folded and soluble on the inside.

The headline number is the point: 672 antibodies pushed through that pipeline and converted into working intrabodies, rather than a handful hand-tuned over years. AI turned a bespoke, one-off craft into something that runs at scale.

What they aimed it at

The targets are the usual suspects of neurodegeneration — the misfolded, aggregation-prone proteins at the centre of each disease. "We've made intracellular antibodies that stick to proteins that cause neurodegenerative diseases such as Alzheimer's, Parkinson's, Huntington's and motor neurone disease," said Wright. The pitch is to hit those proteins at their source, inside the neuron, before they aggregate — instead of mopping up their consequences from the outside.

The work was funded in part by the MND Association, and the team says the redesigned molecules will be made freely available to other researchers, which matters more than it sounds: a shared library of validated intrabodies is the kind of infrastructure a whole field can build on.

The honest caveats

This is laboratory-stage work, not a treatment. Two large gaps sit between here and a patient. First, an intrabody is only useful if you can get it into the right cells in the body — and the plan for that is gene therapy, delivering the genetic instructions so the neuron makes the intrabody itself. Gene delivery to the brain is real but still hard, and unproven for these particular targets. Second, binding a bad protein in a dish is not the same as changing the course of a disease in a person; that link has to be earned in animals and then in trials, over years.

And a note on timing: the underlying paper is a 2026 Nature Communications study that drew a fresh wave of coverage this week, not a discovery announced today. The science is what is new, not the date.

Why it's worth watching anyway

The interesting shift is what AI protein design is now being used for. The famous demos — designing a protein from a blank sheet — are impressive but abstract. This is narrower and, arguably, more useful: a specific physical barrier (charge-driven aggregation inside the cell) that quietly blocked an entire therapeutic strategy for decades, cleared by pointing a design model at exactly that constraint. That is the pattern to watch as these tools mature — less "AI invents a miracle drug," more "AI removes the boring, stubborn engineering problem that was in the way the whole time."