OBSERVATIONAL DATA ONLY: Inferring Protein Signaling Pathways with Local Information- Rewarded Generative Flow Networks

Haoming Gong, Chengyuan Deng, Yongfeng Zhang, Alexandre Morozov

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Causal Discovery, Signaling Pathway, Generative Flow Networks

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Recovering causal influences in protein signaling networks on the basis of observational rather than interventional data is a key problem in systems biology.
Two established approaches for inferring protein signaling pathways are scoring based methods and constraint based methods. We observe that both approaches lacks learning ingredient and one could complement the other. Therefore, we consider integrating current advances in generative modeling of bayesian structures with a specialized designed information theory-based reward derived from constraints during network generation to combine the best of two approaches. Particularly the reward design considers both two-point and three-point mutual information scores corrected with complexity penalties. The signs and magnitudes of these quantities allow us to penalize or reward Bayesian networks as they are being built edge by edge by the GFlowNet sampler. This allows the search process to quickly focus on structures that are in agreement with mutual information signatures observed in the experimental data. Using this approach leads to better predictions compared to the standard practice of using Bayesian scores, which were previously explored with both GFlowNet and Monte Carlo sampling. Thus, our contribution amounts to a novel scheme for penalizing Bayesian networks that are inconsistent with the data early on in the sampling process based on constraints and integrates it with scoring based methods through generative modeling.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8115