AI Watches Worms Eat

Show notes

Welcome to the next episode of The Worm Podcast 🤖

Today we are talking about feeding.

More specifically, how do you measure a worm eating when the worm refuses to stay still?

The answer might be PumpKin — a new machine-learning pipeline that automatically measures pharyngeal pumping in freely moving C. elegans.

🍽️ Watching the grinder

When C. elegans feeds, its pharynx pumps rapidly and a tiny structure called the grinder moves back and forth.

That movement gives us a direct measure of feeding rate.

The problem?

The grinder is less than 30 µm wide, can move around five times per second, and is attached to a worm that is also moving.

Traditionally, researchers often count pumps manually, which is slow and can vary between people.

Other automated methods may require immobilising the worm, specialised equipment or fluorescent reporters — all of which risk changing the behaviour being measured.

🤖 Enter PumpKin

PumpKin takes a different approach.

It uses machine learning to find the pharynx and grinder in ordinary brightfield videos while the worm moves freely.

The pipeline combines:

• Faster R-CNN object detection
• motion compensation
• optical flow
• noise filtering
• automated pump detection

First it follows the pharyngeal bulb.

Then it zooms in on the grinder.

Finally, it separates genuine grinder movement from movement of the worm or the plate.

🍽️ Does it actually work?

The researchers tested PumpKin across eight experimental conditions involving:

• wild-type N2 worms
• eat-2 feeding mutants
• fed and starved animals
• worms recorded on and off food

PumpKin's estimates closely matched those of experienced human observers.

For instantaneous pumping dynamics, its agreement with the human consensus was actually greater, on average, than the agreement between individual human annotators.

🧠 And it revealed some biology

Wild-type worms on food pumped at around 4.7 pumps per second, while eat-2 mutants were much slower at around 1.4 pumps per second.

Take the food away and something interesting happens.

Across genotype and previous satiety state, worms settled around a common basal pumping rate of approximately 0.5 pumps per second.

PumpKin also revealed that feeding is not simply a constant rate. Worms move through bursts and pauses, and those dynamics change depending on food availability and genotype.

🔬 Why this matters

PumpKin is:

automated × non-invasive × freely moving × label-free

That means researchers can measure feeding without fixing the worm in place or genetically modifying it.

And because PumpKin focuses on the movement of a very small body structure inside a moving animal, the approach could potentially be adapted to other behaviours and other organisms.

🧠 The take-home message

Sometimes the interesting behaviour is not the worm moving across the plate.

It is one tiny structure moving inside the worm.

PumpKin gives us a way to watch that behaviour automatically — and lets AI count the pumps so we don't have to.

📄 Paper discussed

Erin Shappell, Debra Buggs, Jennah Walcott and Hang Lu (2026)

PumpKin: A machine-learning pipeline for automatically tracking localized kinematics in freely moving C. elegans

PLOS Computational Biology, 22(7): e1014489

DOI: https://doi.org/10.1371/journal.pcbi.1014489

If you enjoyed this episode, please like, follow and subscribe to The Worm Podcast ⭐

This podcast is generated with artificial intelligence and curated by Veeren. If you would like your publication or product featured on the show, please get in touch.

🔗 www.veerenchauhan.com
📧 veeren.chauhan@nottingham.ac.uk