Researchers imaged blood-cell nuclei from 150 cancer patients and 35 healthy volunteers, following treatment samples across five time points. Their accepted manuscript, published on 13 October in Scientific Reports, asks whether changing resemblance to healthy cells carries information about recovery.[1]
The information in DNA’s packaging
Chromatin is DNA combined with the proteins that package it. As the National Human Genome Research Institute explains, DNA wraps around histones and gathers into structures that can fit inside a cell nucleus. That organisation can loosen for copying DNA or reading genes. A nucleus therefore contains a physical arrangement of genetic material as well as a genetic sequence.[2]
An image and a sequence answer different questions. Reading the letters identifies the encoded information; examining their physical packaging describes another property of the cell. Neither measurement can simply stand in for the other.[2]
A classifier compared individual nuclei with healthy references. Before-treatment and six-month samples yielded 97 usable trajectories: resemblance was low at six months, rose from low to high, or remained high. Three patients lacked follow-up.[1]
Among 94 followed patients, recurrence or death occurred in 7 of 41 in the low group, 1 of 39 in the rising group and 1 of 14 in the high group. This was an observational association across mixed cancers; the images did not guide treatment.[1]
Learning from a group of cells
The distinction between a person and a cell creates a learning problem. A diagnosis can label an entire sample without identifying which individual images carry the relevant signal. Giving every image the sample’s label would pretend to know something that has not been measured.[3]
Multiple instance learning addresses this by training on groups, often called bags. The foundational attention-based method, published by Maximilian Ilse, Jakub Tomczak and Max Welling in 2018, converts each image into numerical features, learns weights for combining them, then predicts a label for the group. The order of the images does not determine the answer.[3]
Learning those weights differs from taking a simple average or letting every image cast an equal vote. It lets the training process give more influence to informative instances. The weights can also show which images contributed most strongly to a prediction, although a contribution to the model’s answer is not itself a biological explanation.[3]
A preliminary head-and-neck experiment used that approach to forecast trajectories from pretreatment groups of 50 nuclei: 17 of 24 held-out patients were classified correctly. Fisher’s exact test gave p=0.067, above the conventional 0.05 significance threshold. This did not establish a reliable treatment-selection tool.[1]
Sources & further reading
- Deep learning-based analysis reveals patient-level cancer therapy trajectories using single-cell PBMC chromatin images Schlüter and colleagues / Scientific Reports · 13 Oct 2026 · Accessed 19 Oct 2026
- Chromatin National Human Genome Research Institute · Accessed 19 Oct 2026
- Attention-based Deep Multiple Instance Learning Ilse, Tomczak and Welling / Proceedings of Machine Learning Research · Accessed 19 Oct 2026