Lin Wan: Learning collective multicellular dynamics with an interacting mean field neural SDE model

Value of the Information & Strength of the Argument

The talk presents a valuable contribution to the field of computational biology by addressing the challenge of modeling cell-cell interactions in high-dimensional gene expression space. The argumentation is solid, with a clear motivation for the need to incorporate interactions and the use of the principle of least action as an inductive bias. The methods are well-founded in mathematical theory, including optimal transport and mean field games. The results are presented with comparisons to existing methods, demonstrating improvements in prediction accuracy and stability. The biological interpretability of the inferred interactions is highlighted, adding to the value of the work.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor through the use of established mathematical frameworks and validation on real biological datasets. The sources cited are primarily the speaker’s own work and related studies, but specific references are not explicitly mentioned in the talk. The title accurately reflects the content, focusing on the interacting mean field neural SDE model. The talk is a seminar presentation, so the level of detail is appropriate for an expert audience, but some technical aspects are simplified for time constraints.

The title accurately reflects the content, which focuses on learning collective multicellular dynamics using an interacting mean field neural SDE model.

8/10

The presentation is a technical seminar by a researcher from a reputable institution, presenting original research with mathematical rigor and validation on biological datasets. The methods are clearly explained, and results are shown with comparisons to existing approaches. However, the talk is a seminar, not a peer-reviewed publication, and some details are omitted due to time constraints.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction to the problem of learning dynamics from snapshot data
• 03:00Background on single-cell RNA sequencing and challenges
• 07:00Comparison with traditional point regression and introduction to distribution fitting
• 10:00Motivation for using the principle of least action and modeling interactions
• 13:00Overview of three methods: discrete, continuous, and interacting mean field
• 16:00Detailed explanation of the discrete interacting mean field model on a graph
• 20:00Extension to continuous space with physics-informed neural SDE
• 24:00Introduction of the scIMF model with transformer-based attention for interactions
• 28:00Results on zebrafish and mouse embryogenesis datasets
• 32:00Discussion on stability and interpretability of the models

• scIMF: A single-cell deep-generative Interacting Mean Field model — The speaker's own work, presented in the talk, but no specific URL was provided.

• Optimal transport for single-cell biology — Related work using optimal transport to model cellular dynamics, supporting the use of distribution matching.
• Neural SDEs for generative modeling — Related work on neural SDEs, providing a foundation for the model architecture.

• Potential limitations of mean field approaches in high-dimensional systems — Mean field approximations may not capture all higher-order interactions, but the talk addresses this by using attention mechanisms.