Latent differentiation inference

Abstract

In various multimedia such as biological information, natural language and images, there are many known groups of data that can be interpreted according to potential differentiation structures (branching structures along pseudotime). A typical example is the differentiation structure of cells in living organisms, where human cells are known to have the ability to differentiate from early embryos into various tissues and organs. As the mechanisms of such differentiation are elucidated, it is expected that in the near future it will be possible to induce pluripotent cells into desired tissues and organs, thereby promoting their application in clinical medicine. We are working on algorithms for inferring latent differentiation structures behind various multimedia data.

Contributions

The method is capable of inferring latent differentiation structures, including their uncertainties, from observed multimedia data alone, without the use of supervisory or auxiliary information.Differentiation structure inference is one of the most important topics in machine learning and data mining, especially in biological information processing, where it has been studied for many years and various methods and algorithms have been proposed. The method is characterised in particular by its ability to infer the following two unknown structures from observed data alone in a data-driven manner, without any human intervention by the user/engineer:

- Unknown differentiation structure topology: Differentiated structure topologies, such as the number of branches and the degree of branching equilibrium, are often used as tuning parameters for algorithms. This method provides data-driven inference of these differentiation structure topologies.

- Unknown observation information generation mechanisms: it is not uncommon for algorithms to generally have to be adjusted depending on the media and data of interest. The method has the ability to infer black box transformations between observed information and potential differentiation structures in a data-driven manner.

Future

The elucidation of the differentiation structure mechanisms of human cells is expected to provide important clues for the development of clinical medicine and the realisation of artificial organs in the near future. Our group will continue to develop multimedia data analysis technology.

Reference

M. Nakano, H. Sakuma, R. Nishikimi, R. Shibue, T. Sato, T. Iwata, K. Kashino: Warped Diffusion for Latent Differentiation Inference, International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.
https://proceedings.mlr.press/v238/nakano24a.html

Contact

Masahiro Nakano
Biomedical Informatics Research Group, Media Information Laboratory, NTT Communication Science Laboratories