Feeding a classifier 4D movies of mitochondria lifts drug-mechanism accuracy from 56% to 75%

Inside the MitoSpace work, the same cancer cells and the same 25 compounds ran with one variable changed, whether the training data was a moving volume or a flat snapshot ^[3][7]. Mechanism grouping came in at 75% against 56%, a gap of 19 percentage points ^[7][20]. Read as errors, that is 25% of calls wrong on 4D data and 44% wrong on 2D, a relative reduction of about 43% ^[21].

MitoSpace trained blind to which drug each cell got ^[5]. It found its own patterns, learning what makes one cell's mitochondria distinct from another's, and produced an organized map that put similarly responding cells together ^[4][5]. "For a century we have believed that mitochondrial form reflects function; this shows the relationship is strong enough that a model can learn it without ever being shown the answer," said Johannes Schoeneberg, the corresponding author and an associate professor of pharmacology at UC San Diego School of Medicine ^[8][19].

Forty thousand movies across 25 compounds is about 1,600 single-cell movies per compound ^[22]. The energetic-state prediction, made from mitochondrial shape and movement alone, is reported across 26 drug conditions ^[6]. A large-scale screen today collects flat 2D images as standard, not lattice light-sheet volumes. The phys.org account does not say what acquiring 40,000 of them took ^[7].

The second study went the other way. Rather than learn the rules, the team wrote them down: they mapped the positions of mitochondria and the microtubule tracks they travel on from a 4D movie, added motor proteins moving at previously established rates, and ran it under the laws of motion, tuning parameters until the simulated mitochondria behaved like the real cell's ^[13]. "We have built a physics-based virtual cell and can compare it side-by-side to the actual 3D microscopy movie, something that has never been possible before," said Schoeneberg ^[14].

Because the parameters were adjusted until the simulation matched, that match is a fit. The prediction comes after, and the researchers report that the virtual mitochondrial networks responded to drugs closely to the way real cells did ^[15]. Another group can test that on a cell outside the tuning set.

The result with the widest reach used the model as trained. MitoSpace organized unfamiliar compounds and sorted human lung organoid cells by developmental stage ^[11]. phys.org describes both studies as a route to less time-consuming bench work and faster drug discovery for cancer, diabetes, Alzheimer's and pediatric mitochondrial disorders ^[18].