Single-cell foundation models benefit from cross-modal training: adding proteomics data beats parameter scaling

In plain English

Instead of only making an RNA model bigger, the researchers added protein measurements during training. Their question was whether a second biological view would teach a more useful representation than billions of extra parameters.

What they found

• The authors report that the 70-million-parameter cross-modal model matched or exceeded the larger RNA-only models on most benchmarks.
• The result favors data modality over parameter count in the tested setting.

The catch

• The work is a preprint and has not been peer reviewed.
• All authors are employees and shareholders of Tesorai.

Methods at a glance

Study type

Computational benchmark preprint

Biological system

Single-cell and proteomic measurements

Organism

Multiple studies

Dataset

48,843 proteomic samples from 440 studies plus the original single-cell benchmark suite

Sample size

48,843 proteomic samples across 440 studies

Model architecture

70-million-parameter cross-modal biological foundation model

Training objective

Learn shared representations from transcriptomic and proteomic data

Code

No public code repository is identified in the abstract.

Data

The abstract reports the study count and sample count but not a complete public data manifest.

Tasks and metrics

Tasks: Original single-cell foundation-model benchmark suite

Metrics: Task-specific benchmark metrics reported in the preprint

Baselines: One-billion-parameter RNA-only model; Three-billion-parameter RNA-only model

Principal findings

• The cross-modal model matched or exceeded the larger RNA-only models on most reported benchmarks.

What remains unresolved

External validation: Not reported

• Preprint evidence only.
• The abstract does not establish performance outside the reported benchmark suite.
• The authors report company employment and ownership interests.