A single-cell dataset is never just a cloud of cells. It contains biological neighborhoods, developmental paths, rare identities, experimental batches, and technical noise. A representation model must decide which differences to preserve and which to ignore.
The comparisons define the representation
Contrastive learning trains an encoder by repeatedly deciding which observations should be near one another and which should be separated. Those decisions depend on the examples placed together in each minibatch. Uniform sampling is therefore an implicit modeling choice, even when it is treated as a default.
In multi-dataset single-cell analysis, a uniformly sampled batch can make technical dataset identity one of the easiest signals to learn. It can also underrepresent difficult local boundaries where closely related cell states must be distinguished.
The composition of a minibatch changes the question the model is asked to answer.
Turning a sensitivity into a design tool
CONCORD combines dataset-aware sampling with hard-negative sampling. Dataset-aware batches reduce opportunities to organize the representation around technical batch identity. Hard negatives concentrate learning on biologically close examples, where additional resolution is most valuable.
This approach does not require an increasingly elaborate encoder. A deliberately minimal network can recover detailed cell-state structure when the training comparisons are chosen well. The result is a latent space that can preserve clusters, trajectories, loops, trees, and cross-species relationships while integrating technical batches.
A broader lesson for scientific machine learning
Scientific datasets contain structure in how observations were collected, grouped, and perturbed. Training design should use that structure intentionally. Sampling, batching, and negative selection are not only engineering details; they encode scientific assumptions about relevant variation.
This perspective suggests a practical question for any model: before increasing architecture complexity, can we improve the comparisons that define the learning problem?