I was a bit disappointed with the readability of this paper to an outsider of computational biology. Even though the paper employs learning/probabilistic methods, it was difficult to even understand these because of the terminology used throughout the paper. Even seemingly simple algorithmic components such as the PSSM, are made complicated by terminology. I realize that I may not be the target audience, but I think the authors limit their audience by using such specific biological terminology throughout the paper without definition or relating these terms to their high level counterparts. I had a very difficult time figuring out how the different components (such as promoter sequence, transcription factor, TF-binding sites, expressions, etc.) were related.

This paper is in RECOMB, a major bioinformatics conference, though some of the authors are players in ML. So the authors omitted a gentle introduction to gene regulation process.

Shortly speaking, there are ~30000 genes in human, but most of them are "sleeping": their expression level is very low. Gene regulation is a process that enhances or represses the expression level of genes. Since gene (DNA) needs to be transcripted to mRNA to make protein to carry out its function (or to say being turned on), one of main gene regulation mechnism is via transcription. In order to transcript a gene, one or more TFs (transcription factors, a kind of protein) need to bind to particular DNA sequences (called promoters) near the transcription start site. The study of binding between TFs and promoter sequences is a hot topic in biology. People develop both computational algorithms and biological experiments to understand the binding. PSSM is a simple way to model promoters. The localization experiment is a way to measure this binding. And microarray is a way to measure expression level of genes. The goal of this paper is to give a unified model for PSSM, localization, and microarray expression.

I think the paper presented a serious attempt to model via Bayesian Networks. However, I feel the paper delves into details too early and for too long. I guess the author want to make this work reproduceable. It would be more understandable if the author provides a gentle introduction to Bayesina Networks and some rationale to develop such kind of network structure. It would required a more careful organization of the limited space.

Dave, the paper was orginally presented for Recomb which is a computational biology conference. However, it is by no means an easy read even for someone familiar with the biology. I actually think trying to be biologically friendly hurt them, as they made things very cumbersome. But the paper does present an interesting idea, even if takes a few reads to understand it.

I agree with the comments of Dave. I think that authors could have given a stronger background information. Even after reading the full paper that terminology is still not clear to me. I hope that this will change after the presentation today.

Apart from the unclear, basic molecular biology concepts i liked the structure of the paper. The authors clearly separated and explained the components of their computational model and the learning algorithm used for each of the components.