Mapping ATAC into a fixed RNA space: FlowAlign keeps paired cells and tissue structure with one path-consistency score
Synopsis
FlowAlign turns conditional flow matching from an inference-time generator into a training-time alignment signal: it first learns a source-conditioned velocity field along interpolation paths between paired samples, freezes it, and trains the source-side mapper on the velocity residual that a candidate representation induces. On the paired PBMC and BMMC single-cell multiome benchmarks it maps ATAC representations into a fixed RNA PCA space with lower FOSCTTM than the compared integration methods and architecture-matched paired baselines, while NMI and ARI are higher. Synthetic data and an image–text alignment task report the same kind of improvement.
Interpretation
Multimodal alignment now has another training signal: instead of only asking the mapped point to sit near its paired target or the overall distribution to match, it asks the path velocity induced by a candidate representation in the target space to agree with a velocity field learned from true paired samples. Contrastive losses constrained only relative similarity, distribution-matching losses improved only marginal agreement, and pointwise regression fit each source sample to its paired target without explicitly modeling target-space path geometry. On PBMC, FlowAlign reaches FOSCTTM 0.039 and exact cell-match rate 0.141, against 0.048 and 0.110 for the architecture-matched MSE baseline; on BMMC the two are 0.019 and 0.050, against 0.042 and 0.032 for MSE.
The procedure has two stages: learn a velocity field conditioned on the source representation, then freeze it and use the velocity residual of a candidate representation along the path as a score for training a mapper that sees only the source modality. Conditional flow matching is usually used as an inference-time transport or generative model; here it serves as an alignment criterion during training, and at inference the velocity field and path sampling are discarded so a source sample reaches the fixed target space in one forward pass. On PBMC and BMMC, ATAC is embedded with TF–IDF/LSI and RNA with PCA, each at 25 dimensions, cells are split at the paired-cell level into training, validation and held-out test sets, and only ATAC representations are given at test time; on PBMC FlowAlign reaches NMI 0.872 and ARI 0.885, against 0.821 and 0.814 for the architecture-matched MSE baseline.
The same criterion can act as a plug-in regularizer on an existing task model, reshaping a pretrained ATAC latent space toward a fixed RNA latent space. In a PBMC proof of concept, the scVI RNA latent space is held fixed as the target, and a pretrained PeakVI model is fine-tuned with its peak reconstruction loss plus the FlowAlign path-consistency loss. After regularization, FOSCTTM falls from 0.548 for pretrained PeakVI to 0.068, R@1 rises from 0.000 to 0.046, LTA rises from 0.047 to 0.815, and peak reconstruction AUROC moves from 0.782 to 0.765 and AUPR from 0.637 to 0.630.
Perspective
The demonstrated scope is a synthetic two-dimensional trajectory, the paired single-cell multiome datasets PBMC and BMMC, and the CUBICC paired image–text task; in the single-cell experiments ATAC and RNA are reduced to 25-dimensional representations by TF–IDF/LSI and PCA, and only ATAC-side input is provided at test time. The authors propose future work on more flexible start distributions, field diagnostics, partially paired or unpaired settings, and broader applications.
The method requires paired data, a common vector space in which paths can be constructed, and an accurate Stage-1 velocity field; incompatible source representations or a poorly learned field can weaken the alignment signal. The current analysis assumes Euclidean linear paths, and standalone alignment does not guarantee preservation of source-specific information absent from the target space. Held-out field diagnostics show lower residuals on true paired paths than on paths with randomly shuffled endpoints, but these diagnostics do not directly establish the local smoothness assumption used in the proposition.