The molecular information hidden inside a fluorescent spot

Imagine two tiny films with the same total light in every frame. In one, the brightest part stays near the centre. In the other, light shifts between neighbouring pixels. Reducing each frame to one brightness number would make the films look identical in the resulting record. The discarded positions could still distinguish them.

That illustrative loss of information helps explain smDeepFLUOR. Its neural network accepts a seven-by-seven-pixel patch over ten frames: 490 intensity values with their positions and times intact. The authors’ model card describes an output comprising probabilities for two experimental classes, such as two states of a DNA molecule.[3]

A binding event leaves more than brightness

In research published on 21 June 2026, Jinseob Lee’s team distinguished protein bound to its RNA target from protein stuck nonspecifically to the imaging surface. Average intensities looked alike; a second fluorescent label helped establish the classes.[2]

The model classified specific binding with 93% accuracy and nonspecific binding with 97% in the within-experiment test. For two separately prepared experiments recorded on different days, the respective averages were about 85% and 90%. Training and testing were separated by particle trajectory before movies were divided into short input sequences.[2]

The output probabilities refer to the alternatives supplied during training. Their meaning comes from the experiment: which samples represent each state, and which observations establish that assignment.[3]

Keep the little movie

The public research workflow begins with microscopy images and particle-tracking coordinates. It crops a small image stack around each tracked spot, prepares training and test sets, standardises the intensities and trains a classifier. The released package includes a notebook, an environment file and an example pretrained model for loading and inference.[3][4]

These are materials for laboratories working with comparable single-molecule recordings. The model card says performance depends on imaging conditions, fluorescent labels and how those labels are attached. An experiment’s two classes have to acquire meaning through the samples and controls that define them.[3]

The paper also monitored DNA gap filling using a dye attached to the template. Controls suggest sensitivity to the dye’s surroundings, although the features driving classification remain under investigation.[2]

That leaves a physics question inside a successful classification: what changed around the dye to make the movie distinguishable? The little film preserves clues that its average brightness erases.[2]

Report an error or suggest a correction