Fix the label directions before training: single-cell reference mapping lifts query macro-F1 from 0.774 to 0.811

Label relationships in a reference atlas are now fixed before training begins, and the encoder only places expression profiles into that pre-set space. The earlier supervised contrastive method SCLSC built each cell-type representative from a mean expression profile and passed cells and representatives through a trainable encoder, so representative positions in the learned space changed during training. Across nine held-out batches from human immune, lung and pancreas datasets, using 30-dimensional outputs and one reference-trained 15-nearest-neighbour classifier, SOLAR reached a mean query macro-F1 of 0.811 against 0.774 for matched PCA, and its advantage over matched PCA was positive within each of the three datasets.

New cells enter the reference space without any optimization on query data, because the trained encoder performs the mapping directly. Mapping runs in two stages: reference cells supply gene selection and principal-component fitting plus encoder training, then the same preprocessing and the frozen encoder encode query cells. The encoder is a three-layer multilayer perceptron taking 40 reference-fitted principal components and producing an ℓ2-normalized output; anchors come from a reduced orthogonal-triangular factorization of one Gaussian matrix, giving each label a unit vector with zero pairwise inner product and Euclidean distance of root two, and anchors share one contrastive pool with cell embeddings during training.

Anchor presence and repetition change how the representation is organized, not only how accurately labels are recovered. At 128 dimensions, three conditions were compared: one anchor copy per cell, one anchor per represented label, and no anchors at all. Repeated anchors raised held-out label average silhouette width and local batch diversity iLISI relative to both controls, while query macro-F1 moved by about 0.007 against no anchors and under 0.001 against unique anchors, and trustworthiness and immune-cell pseudotime smoothness were lower than with either control.

Evaluation covers nine held-out batches from three real datasets, human immune, lung and pancreas, with two simulations as supplementary checks; every evaluated query label occurs in its reference, so the test is closed-label mapping. Local-structure metrics measure agreement with a 50-dimensional PCA representation and with supplied pseudotime, with query trustworthiness scored on six immune and lung batches and joint pseudotime smoothness on three immune batches.

Repeated anchors raised label separation and local batch diversity yet scored lower than both controls on trustworthiness and pseudotime smoothness, so query-label prediction alone does not capture all changes in the representation. The anchor controls vary only the presence and multiplicity of one fixed anchor bank; they do not compare fixed against learned anchors or orthogonal against alternative fixed geometries, and SOLAR and SCLSC differ in inputs, objectives and checkpoint selection, so their gap cannot be attributed to anchor geometry alone. Unknown cell types and incorrect reference annotations were not evaluated.