Universal Biological Generalization Across Species and Modalities

This direction is merged into another direction: cross-species model generalization. Its last version stays readable and citable here.

Version 1 of 118 papers18 conceptsfrozen 2026-10-04

In recent years, biological AI has shifted from training niche models for specific tasks toward developing foundation models that capture the "universal grammar" of life. A recurring discovery across diverse research—spanning genomics, proteomics, and cellular imaging—is that models trained on one species, tissue, or molecular modality (like RNA) often generalize remarkably well to others (like DNA or different organisms). This phenomenon suggests that deep learning models are not merely memorizing datasets but are uncovering fundamental, evolutionarily conserved biological principles.

The Conservation of Biological Logic

The most striking evidence for this universal grammar is cross-species generalization. Because evolution preserves functional structures, a model that learns the "rules" of biology in a well-studied organism like humans can often apply those rules to mice, flies, or even different kingdoms of life.

For instance, Emmanuel Moebel and colleagues demonstrated in Deep Learning Improves Macromolecule Identification in 3D Cellular Cryo-Electron Tomograms that 3D convolutional networks trained to identify macromolecules in one species can successfully locate similar structures in entirely different organisms. This transferability implies that the morphological features of life’s building blocks are sufficiently conserved that extensive retraining is unnecessary.

Similarly, in the realm of protein interactions, Wenqi Chen showed that the DCSE model can generalize across species by capturing universal sequence motifs rather than species-specific patterns. Even complex generative tasks, such as creating 3D neuronal morphologies, exhibit this property; Franz Rieger found that MoGen, a flow-matching model, could generate high-fidelity neurons for new species with only minimal fine-tuning.

Bridging the Central Dogma

Beyond moving between species, researchers are finding that models can move between biological modalities. Information in a cell flows from DNA to RNA to Protein—the "central dogma." While these layers look different to a computer (four-letter DNA codes vs. twenty-letter protein codes), they share an underlying logic.

Jack Cole and the team behind OmniGenBench observed that genomic foundation models (GFMs) trained on RNA sequences perform competitively on DNA-based tasks. This cross-modal synergy suggests that the "grammar" of a transcriptomic sequence contains enough information to interpret the regulatory logic of the genome.

This idea is taken further by Heng Yang in the development of OmniGenome, which bridges sequence and physical structure. By training on multi-species transcriptomic data, the model gains a comprehensive understanding of genomic logic that applies even to DNA-based regulatory tasks. As Christopher Blum argues in Multi-modal Transfer Learning between Biological Foundation Models, pre-trained weights from one domain (like proteins) can significantly accelerate learning in another (like DNA), proving that these "isolated islands" of biology are actually connected by shared principles.

Robustness to Platform and Tissue Variation

If a model has truly captured a universal principle, it should be immune to the "noise" of how data is collected. This cross-platform robustness is a key indicator of biological reasoning.

  • Tissue Generalization: Yongkang Long demonstrated that APAIQ, a tool for identifying mRNA processing sites, maintains high accuracy when moving from human cell lines to mouse fibroblasts, suggesting it captures the fundamental machinery of the cell rather than tissue-specific artifacts.
  • Technology Generalization: Yuanyuan Chen 0014 found that gene sets identified via deep learning on microarray data remained highly predictive when applied to RNA sequencing (Mining influential genes based on deep learning).
  • Reasoning Generalization: Yuhang Chen introduced CellDuality, which uses reinforcement learning to allow Large Language Models to reason about transcriptomics. The model’s ability to apply human-derived insights to mouse datasets confirms that the underlying biological logic is transferable across out-of-distribution contexts.

Why It Works: The Metric of Life

The success of these transfers often relies on learning a specialized "distance" or "metric" that aligns different distributions. Yonghui Xu showed that Multi-Instance Metric Transfer Learning (MIMTL) could predict protein functions across different kingdoms of life by reweighting data to match target distributions.

By focusing on these shared metrics—whether they are distance maps in nuclei segmentation (Xiaoming Liu 0004, MDC-Net) or sequence concepts in T-cell repertoires (John-William Sidhom, DeepTCR)—AI is moving toward a unified understanding of biology. The research suggests that the "black box" of biological AI is increasingly filled with the same universal rules that have governed evolution for billions of years.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_universal_biological_generalization_acro_c19a9c15,
  title = {Universal Biological Generalization Across Species and Modalities},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/universal-biological-generalization-across-species-and-modalities/rdr_c19a9c154df04b02a10c912e2f2ef58c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_universal_biological_generalization_acro_c19a9c15_v1,
  title = {Universal Biological Generalization Across Species and Modalities},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/universal-biological-generalization-across-species-and-modalities/rdr_c19a9c154df04b02a10c912e2f2ef58c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}