Course Summary

This class is a joint effort between the BIGIALS group and the Computer Science department. All classes are arranged to reach the heterogeneous crowd interested in physical/biological sciences and informatics. The seminar is research oriented and will prepare the participating students and faculty attending it for collaborative research in this interdisciplinary area.

Bio-informatics includes the use of computational and mathematical techniques (e.g., algorithms, machine learning, AI, complexity theory, dynamical systems) to decipher complicated information in biological structures, gene sequences, and cellular signals. It also includes the construction of new technologies that have been overlooked before and whose roots are in nature, as is the aim of the IBM Blue gene project. This interdisciplinary field is of critical importance for understanding information from humongous biological sources, as well as for the design and construction of new computational methods, software, and hardware. Main topics include adaptive computational approaches and AI algorithms for structural and functional biology (e.g., protein structure prediction, prediction of proteins function and interactions prediction, fold prediction) , neural computation, "dynamical diseases" (e.g., finding structure in cardiac rhythms), building evolutionary robots, DNA and quantum computers, and more.

The course will survey leading trends in bio-computation. Some topics will be introduced by invited lecturers who are active and well known in the field.

Assignments and Grades

Students are advised to read the relevant papers of outside speakers 1 week prior to the presentations, and be ready for a discussion on it. This is the best way of learning from their great experience.

We will also read some chapters from Bioinformatics - The Machine Learning Approach, by Pierre Baldi and Soren Brunak.

Tentative Schedule

Possible additional lecturers:

• Prof. Michael Jordan, Berkeley
  
• Prof. Roger Brocket, Harvard
  

Talk Abstracts

• 2/8   Dr. Tzachi Pilpel
  Combinatorial analyses of genetics regulatory networks
  Several computational methods based on microarray data are currently used to study genome-wide transcriptional regulation. Few studies, however, address the combinatorial nature of transcription, a well-established phenomenon in eukaryotes. Here we describe a new approach using microarray data to uncover novel functional motif combinations in the promoters of Saccharomyces cerevisiae. In addition to identifying novel motif combinations that affect expression patterns during the cell cycle, sporulation and various stress responses, we observed regulatory cross-talk among several of these processes. We have also generated motif-association maps that provide a global view of transcription networks. The maps are highly connected, suggesting that a small number of transcription factors are responsible for a complex set of expression patterns in diverse conditions. This approach may be useful for modeling transcriptional regulatory networks in more complex eukaryotes.
  
  
• 2/15   Prof. Hod Lipson
  Evolutionary robotics and Computational Design
  ** note: this week's meeting starts at 12:30
  The fields of evolutionary computation and evolutionary robotics study adaptive mechanisms based on natural selection, with the aim of algorithmically applying these ideas to solve hard problems like nonlinear optimization and engineering design, as well as shedding light on the evolution of natural systems. One of the difficult questions, both algorithmically and biologically, is the emergence of complexity: Evolutionary processes based on accumulation of random mutations are less likely to result in improvement as individuals grow more complex, because of the exponential nature of the search space. This talk will present some new directions in evolutionary computation that address this scaling problem. Some recent new approaches based on co-evolution, modularity, hierarchical composition and symbiosis will be overviewed, as well as and their application to evolution of robots. The talk will also outline some of the research methodologies used in this field, and some of the currently open questions.

  IMAGINE A LEGO SET AT YOUR DISPOSAL: Bricks, rods, wheels, motors and sensors are your atomic building blocks, and you must find a way to put them together to achieve a given high-level functionality: A machine that can move itself, say. You know the physics of the individual components' behaviors; you know the repertoire of pieces available, and you know how they are allowed to connect. But how do you determine the combination that gives you the desired functionality? This is the problem of Open Ended Synthesis.

  Although we have two proofs of feasibility: Practicing Engineers do it, and Nature does it, we still know little about how this process can be automated for arbitrarily complex tasks.

  For further readings, see:
  

  • Lipson H., 2001, "Uncontrolled Engineering: A review of Evolutionary Robotics", Artificial Life, to appear, book review. pdf
  • Lipson, H., Pollack J. B., 2000, "Automatic Design and Manufacture of Artificial Lifeforms", Nature 406, pp. 974-978. pdf
  
  
  
  
• 2/22   Prof. Mary Beth Ruskai
  An Introduction to Quantum Computation
  Standard methods for both computation and communication encode information in strings of 0 and 1. Quantum particles can be used to encode information in "qubits" which, in addition to having states corresponding to 0 and 1 can be placed in superpositions which contain both 0 and 1 with certain probabilities. Strings of n such qubits can simultaneously encode all possible 2^n combinations of 0 and 1, making quantum computers inherently massively parallel. However, extracting useful output information is constrained by the principles of quantum measurement. In essence, the computational complexity problem has been shifted to the output measurement. It is quite remarkable that algorithms have been constructed which allow one to exploit the power of quantum computers. Shor's 1994 development of an efficient algorithm for factoring large numbers, one of a class of algorithms based on the quantum Fourier transform, stimulated interest and further developments. This talk will introduce the audience to the fundamental principles of quantum theory which make quantum computation possible and potentially more powerful than classical communication. Rather than describing a particular algorithm in detail, we will compare the classical discrete Fourier transform and quantum Fourier transform in a way that illustrates the power of superpositions, its limitations and the fundamental role of the quantum measurement process.
  
  
  
• 3/8   Prof. Simon Kasif
  Systems Biology: Active Learning of Cellular Architectures
  In this talk I will describe the application of computational learning and computer science algorithms to reverse engineering, predictive modeling and analysis of biological cell function.
  Genes and their post-translational interactions provide the basic scripts for biological cell behavior. The problems we address include comparing of the genomic content of human and mouse genomes, creating qualitative interpretations of gene expression data, and constructing functional gene-based models of biological cells.
  In many cases the architectures we construct resemble simplistic computer circuits and networks. I would argue that cells can teach us a lot about computation, and the need to unravel the mystery of living organisms provides an endless supply of challenging basic research problems in computing.
  Ultimately, we would like to develop the capability to program biological systems to do useful things or monitor and repair undesirable events which requires a computational understanding of the biological components and their interaction on different levels of resolution.
  
  
  
• 3/29   Prof. Bruce Tidor
  Solvation effects on protein folding, binding and design: Exploring the Electrostatic Balance
  Electrostatic interactions are prevalent at protein interfaces and are important for protein-protein and protein-ligand association. However, due to the large desolvation penalty incurred by polar and charged groups upon binding, on balance it is unclear whether electrostatic interactions can be strongly stabilizing in aqueous solution near room temperature. Continuum electrostatic calculations suggest that effects of polar and charged groups are generally destabilizing. A novel procedure has been developed for designing ligand-charge distributions that optimize the balance between unfavorable desolvation and favorable interaction. Application of the method to protein systems demonstrates the importance of optimizing electrostatic interactions. Charge optima reveal that electrostatics can be moderately stabilizing and that one strategy to enhance stability involves engineering intra-ligand interactions that become strengthened upon desolvation and binding.
  
  
  
• 4/5   Prof. Leon Glass
  "Dynamical" vs. "genetic" disease - What do complex rhythms reveal about pathophysiology? Recent discoveries in bioinformatics and genomics are revolutionizing our understanding of the genetic basis of disease. Yet, in many diseases the body displays complex abnormal rhythms that are amenable to mathematical analysis. The abnormal rhythms often arise as a consequence of changes in key parameters of an underlying physiological control system, and as such the dynamics give important clues about underlying mechanisms and therapy. This talk reviews applications of the concept of dynamical disease with emphasis on abnormal cardiac and neurological rhythms.
  
  
  
• 4/12   Prof. John Moult
  Protein Structure Prediction
  An analysis of CASP results, identifying what works, what does not, and what the possible ways forward are.

  The Scope of Structural Genomics
  An analysis of how many structures will need to be solved experimentally in order to be able to model 'all' structures with different degrees of accuracy. Combines sequence family analysis, coverage of structure space, and modeling reliability. An emphasis on how useful a model is for obtaining information about function. Tends to go off into evidence for the frequent appearance of new protein families during evolution.
  
  
  

• 4/19   Prof. Leslie Kay
  Context and chemistry in olfactory computation
  The computational architecture of the olfactory bulb is intriguing, as it combines a relatively ordered set of inputs from the peripheral olfactory nerve with a massive array of inputs from many other brain areas. When animals are anesthetized or passively exposed to odors, the activity of the principal neurons (mitral cells) appears to be driven by a relatively ordered set of relationships dependent on similarities in chemical structure, which is suggestive of chemotopy. Neural recordings from awake animals, trained to associate a behavioral meaning with an odor stimulus, present a different picture of odor "representation." In these experiments we find that the activity of individual mitral cells is driven primarily by the behavioral requirements of the stimulus. Odor representation, when seen, is also driven by meaning. When the behavioral requirements of the stimulus changes, so does a cell's odor selectivity. These changes are most likely driven by input from other brain areas, one candidate of which we have found to be the entorhinal cortex as part of the hippocampal system. Behavioral experiments, which we have designed to test the behavioral relevance of chemotopy, also show that an animal's prior experience with odors influences to a large degree the ability to recall a learned odor from among a set of chemically similar odors. However, there is some influence of the odor's chemical class (e.g., all alcohols smell similar, even though individual alcohols can be easily distinguished). We also find that in special circumstances, the chemical structures of mixture components, combined with olfactory receptor biophysics, can determine the perceptual quality of the mixture. These results elucidate the computational structure of the olfactory system, which involves a dynamic interplay between chemistry, anatomy, meaning and behavior.
  
  
  • Kay, L.M. and Laurent, G. (1999) Odor- and context-dependent modulation of mitral cell firing in behaving rats. Nature Neurscience, 2(11): 1003-1009.
    
  • Kay, L.M. and Freeman, W.J. (1998) Bidirectional processing in the olfactory-limbic axis during olfactory behavior. Behavioral Neuroscience, 112(3): 541-553.
    
  • Kay, L.M., Lancaster, L.R., and Freeman, W.J. (1996) Reafference and attractors in the olfactory system during odor recognition. International Journal of Neural Systems, 7 (4): 489-495.
    
  
  
  
  
• 4/26   Prof. Tony Zador
  ni The cocktail party problem: Animal models of computation in the auditory cortex
  The neurons that provide the substrate for cortical computation appear quite noisy (although just how noisy remains a subject of some controversy), suggesting that the cortex uses a strategy rather different from that of a digital computer. The cortex is nevertheless able to solve hard computational problems, such as the cocktail party problem. (sometimes called the "symphony problem," it an auditory analog of the object recognition problem), that remain beyond the capacity of the fastest computers. This problem is a particular interest because it related to the problem of selective attention, and hence consciousness. We have been combining experimental and theoretical approaches to study the special characteristics of cortical "wetware" that endow it with the capacity to solve such hard problems. I will discuss some recent experimental results in which we have studied how acoustic signals are represented in the rodent primary auditory cortex. While I confess these experiments have not yet resolved how rats solves the cocktail party problem, I hope they do provide a first step.