First, the good. The description of stochastic context free grammars is the best I've seen. I don't know any other reference that even hint at how to use generative grammars to evaluate likelihoods. Once they caught my interest, though, the authors did not carry through with training and evaluation algorithms I could really use. I suspect that parts of the information are there, but I'll have to go back over their opaque notation again to work out just what they've given and just what's been left out.

This same pattern - an interesting introduction with missing or mysterious development - recurs throughout the book. The discussion on clustering and phylogeny goes the same way: a number of techniques are mentioned but not developed. The authors mention a tree drawing problem, not just building the tree's topology, but ordering the branches for the most informative rendering. Again, a critical topic and one that most authors miss - in the end, these authors miss it, too, by mentioning but not filling in the idea.

Their discussion of neural nets suffers badly from the authors' partial presentation. Evaluation of network output for a given input is relatively straightforward, and they present it in some detail. Training the net is the real problem, though, and is given less than a page.

Baldi and Brunak give more of the fundamentals than most authors. For example, they explain the maximum entropy principle well enough that I'll use it in lots of other areas. They give some coverage to topics of intermediate complexity, such as the forward and backward algorithms for HMM training. Finally, they fizzle out at the higher levels of complexity - the Baum-Welch algorithm could have followed from the forward and backward methods, but is left as a reference to another book.

There is some good here, especially in the fundamentals behind important techniques. The discussions I wanted - the more avanced topics, in forms I can use - are often weak, missing, or impenetrable. Just a bit more work, clearly within the authors' capability, would have made this a landmark reference.

I have a good programming background. I also read some papers on neural network and hidden markov models, This book is a lot worse than anything I have read in explaining the stuff. Very disappointed. Save your money and get something else.

In the mathematics and computer science community, Baldi is an internationally recognized expert in the fields of neural nets and Hidden Markov models and their applications (for instance, he holds a patent for neural net recognition of fingerprints). Concerning HMM's Baldi and co-workers have found statistical models for protein families, sequence signals for nucleosome centers, etc. Moreover, Baldi, together with Chauvin, has developed a gradient descent parameter update method for HMM's which has no zero probability absorptions, and allows on-line updates, useful features not supported by the standard EM method.

From these and other applications, I found the text very useful.

As a computer scientist I can say that the book is sometimes difficult to read if you have no previous knowledge of biology. This is because the authors didn't take the simplificative approach of reducing biological problems to abstract mathematics. Rather, they preserved the full biological flavor of the problems. Although this approach costs you more at the beginning, you can eventually get a more accurate and nontrivial picture of the problems.

My conclusion: it is perhaps unlikely that you can learn about bioinformatics using only this book. However, if you want to learn about bioinformatics, this book is a must-have reference.