# Richter et al. find aligning cell data types does not fully integrate them

> A Nature Machine Intelligence paper by Richter et al. reports that aligning different biological data types does not fully integrate them in cell foundation models.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Research
- Published: 2026-11-09T00:00:00+00:00
- Updated: 2026-11-09T11:10:26.144414+00:00
- Web page: https://ai-news.ikangai.com/story/richter-et-al-find-aligning-cell-data-types-does-not-fully-integrate-them/
- Seen in: Nature Machine Intelligence
- Change: research (Research)

Richter et al. report in Nature Machine Intelligence that aligning different types of biological data does not, by itself, fully integrate them in multimodal cell foundation models. The paper introduces a metric, the Synergistic Information Score (SIS). It asks whether fusing two data types adds task-relevant information beyond the strongest single data type, or only repeats signal that one modality already carries.

Unimodal datasets are abundant, but paired measurements capturing several modalities from the same cell or tissue are much scarcer. That scarcity favours compositional foundation models, which reuse frozen pretrained encoders for each modality and combine their outputs through a small learned interface. The authors argue that the usual alignment objectives pull paired representations together in a shared embedding space using contrastive or regression losses. They say these objectives collapse to detect linear redundancy. In their framing, alignment-based fusion can reach a "spectral ceiling" set by the eigenspectrum of the pretrained representations. It recovers dominant linear correlations but may not access synergistic information that appears only through joint nonlinear processing.

SIS is the relative gain in a linear probe's information about a target variable Y when both modalities are used, measured against the better of the two modalities alone. The paper estimates mutual information with linear probes, using F1 Macro for classification and R2 for regression. A positive score indicates accessible synergy. A near-zero or negative score indicates redundancy dominance or fusion failure. The authors note the score is conservative under strong modality imbalance. Fusion has to beat the dominant unimodal baseline to register as positive, so a low score means no extra gain over that baseline, not that the weaker modality is uninformative. The score also depends on the probe. It measures what a fusion interface has made linearly accessible, rather than a complete information decomposition.

The authors benchmarked ten fusion interfaces on three spatial transcriptomics datasets: a pulmonary fibrosis lung atlas, a developmental human thymus dataset, and a breast cancer dataset. The encoders were fixed, with UNI-25 for histopathology and Nicheformer for spatial transcriptomics, so differences in downstream performance come from the fusion step. On the breast dataset, niche classification gave only small positive SIS values. CCA reached 0.058 and CoMM 0.045, while simple concatenation stayed close to the strongest unimodal baseline at 0.011. The authors read this as a redundancy-dominated task, where fusion adds modest signal that individual modalities already largely provide. Lung niche classification behaved similarly, according to the paper.

The thymus dataset illustrates a task that needs synergy. Its gene expression comes from Visium spatial transcriptomics, which captures transcripts over coarse spots spanning several cells. It does not line up precisely with the fine structure visible in histology images. The authors describe this resolution mismatch as creating a regime where complementary spatial information must be combined across modalities, and they report that aggregation fails there. The paper's wider claim is that synergy-aware objectives, such as Contrastive MultiModal (CoMM), can exceed the spectral ceiling on cross-modal-dependent tasks. Alignment-based methods, by contrast, typically saturate near unimodal performance.

## Why it matters

Teams building multimodal biology models may need to test whether each added data source contributes new information rather than duplicating existing signal.

## Original source

- [Multimodal cell foundation models require synergistic integration beyond alignment | Nature Machine Intelligence](https://www.nature.com/articles/s42256-026-01328-3?error=cookies_not_supported&code=ae1eddf0-f76c-48ec-a4a9-7cea6d5c5b12)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Richter et al. find aligning cell data types does not fully integrate them", 2026-11-09, https://ai-news.ikangai.com/story/richter-et-al-find-aligning-cell-data-types-does-not-fully-integrate-them/
