Mr Sach Mukherjee

Report on the IEEE Bioinformatics Conference 2004
Sach Mukherjee
Department of Engineering Science
University of Oxford.
The IEEE Computer Society Bioinformatics Conference (or CSB) is one of a handful of
highly selective meetings in the area of bioinformatics and computational biology. Held every
year at Stanford University, CSB attracts around six hundred attendees, but is single-track.
Bioinformatics is an inherently multidisciplinary field, attracting, among others, engineers,
biologists and statisticians, and CSB attendees tend to reflect this diversity. This year’s
conference took place in mid-August and comprised around thirty talks, a hundred posters
and a half-dozen tutorials.
My own research focuses on the development and application of statistical and machine
learning methods to the analysis of gene expression data. In particular, my work addresses the
problem of finding genes whose levels of activity differ between conditions such as healthy
and diseased, a task called ‘gene selection’. Having shown theoretically that many widelyused
methods for gene selection are simply not very robust, and that an unsupervised learning
approach could be used to significantly ameliorate the problem, I was eager to communicate
my findings to the bioinformatics community. I was interested to hear the views of wet-lab
biologists and statisticians working in broadly the same area. Beyond my immediate research
area, I was also keen to learn as much as possible about where bioinformatics in general is
heading. My goals in attending CSB were therefore threefold: to advertise my own work,
interact with scientists working on similar problems, and also to network more broadly with
researchers from other areas of bioinformatics.
Having previously attended a large international meeting in machine learning (NIPS 2002, in
Vancouver) I was aware that planning would be crucial to fully realising my goals. I had two
separate pieces of work accepted to the conference, a paper and a poster. I was fortunate to
have had my paper selected for podium presentation, which provided me with the opportunity
of giving a 30 minute presentation to the entire conference. My talk turned out to be
scheduled relatively early in the conference, so with regard to promoting my work, I focused
my efforts on simply preparing an effective talk. The likely diversity of the audience led me
to avoid excessive jargon and concentrate on presenting a coherent ‘story’. I also compiled a
short list of people working my sub-area, with whom I specifically wanted to establish
contact, and decided to make every effort to have at least one meaningful conversation with
each of them.

The conference was organized into a total of 14 sessions covering every area of
bioinformatics, from protein structure to gene ontologies. The organizers kept the sessions
short with frequent coffee-breaks in between. This arrangement proved extremely popular as
the breaks provided excellent opportunities to interact one-on-one with speakers, and discuss
ideas in more detail than is possible when asking a question at the end of a talk. Coffee-breaks
also turned out to be an excellent opportunity for ‘targeted’ networking, and over the course
of the conference I was able to have good conversations with most of the researchers on my
list. The setting of the conference could not have been better, with all meals taken in
Stanford's sunny Dohrmann Grove, and posters presented in open-air sessions either side of
the main auditorium. Meal-times proved quite productive too, as I found myself sharing tables
with researchers from a wide variety of backgrounds. This had the effect of opening my eyes
to areas of research that I would otherwise rarely encounter.
My own talk went extremely well, and generated a good deal of interest. I spent several
coffee breaks afterwards discussing my work with interested members of the audience, many
of whom asked extremely good questions. The most useful comments concerned the
biological validation of my largely theoretical results. A key problem in expression analysis is
the scarcity of ‘gold-standard’ datasets, where the truth is known a priori, but several
attendees were able to point me to some very good data upon which to further test my
methods. Indeed, ideas that emerged at CSB have since formed the basis for a new paper
which is currently under review. By and large, my impression was that my work did indeed
raise awareness of some important problems in expression analysis, and that my progress on
addressing those problems was on the right track. I was especially pleased to have a number
of good conversations with biologists and statisticians, and to be invited by several groups to
give seminars to them in the near future. The talk had the further effect of generating interest
in my poster, which I found very well attended once I had given the talk. On my return, I was
also very happy to find that my paper was one of ten or so to have been selected for inclusion
in a special issue of the Journal of Bioinformatics and Computational Biology.
The other talks were generally of a high technical standard, although a few failed to hold my
interest on account of poor presentation. Personally, I found the poster sessions especially
effective, as I was able to understand the work at my own pace and then interact one-on-one
with the authors. Two of the best posters were one by Berger et al. of UCSB on the analysis
of two different kinds of microarray data using a linear algebraic method called ‘Generalized
Singular Value Decomposition’. Another excellent poster was by Xie et al. who showed that
the dynamic range of gene expression measurements depends on the functional category of

the genes. This is an extremely interesting, albeit preliminary, finding, with important
implications for the analysis of expression data.
The best of the keynote speakers were Gene Myers and Michael Eisen, both from the
University of California at Berkeley. Prof Myers' leadership in bioinformatics pre-dates the
term itself, and he provided a superb overview of where the field is heading. His theme was
the importance of seeing the ‘big picture’ in terms of biology. He rightly noted that it is
crucial to go beyond simply ‘cranking the handle’ and observed that there is a growing
tendency in the field to apply ever more complex methods to data without giving much
thought to the underlying scientific questions. Dr. Eisen talked about Drosophila
development, and displayed an infectious enthusiasm for his subject. He emphasised the
importance of bringing together diverse data types to shed light on developmental processes.
With so much work in bioinformatics rather narrowly focused on array analysis (including,
admittedly, my own), his perspective as a biologist was timely and valuable.
Thus, my overall impression was that bioinformatics is heading into a period where key
themes will be the fusion of heterogeneous data types and the integration of biological
knowledge into computational algorithms. To date, a lot of work has focused on one of
expression, protein or sequence information, but increasingly these data will have to be dealt
with together. Bayesian methods have proved adept at handing data fusion in areas like
robotics and sensors, and my feeling is that Bayesian inference will prove if anything even
more popular in bioinformatics, where computational efficiency is generally less of a concern
than in many ‘real-time’ engineering systems. One of the bottlenecks in current
bioinformatics research seems to be the fact that in many ways experimentalists and theorists
do not yet share a common language. One solution to this problem is simply to have the
various communities interact more. Indeed, teams are becoming ever-more multidisciplinary,
with researchers from different fields interacting on a daily, rather than weekly or monthly,
basis. Mixed wet-lab and theory institutes are now appearing at a number of universities,
including the Centre for Gene Function here at Oxford. It was clear at the conference that
only a handful of computer scientists and statisticians really have a good understanding of
biology, and equally that few biologists are really comfortable with the language of
mathematics. Looking at more mature fields like biophysics and biochemistry, it seems that
what is needed are individuals who have both wet-lab training and an understanding of
mathematics and computation. Talking to some of the Stanford undergraduates at the
conference, it seemed that this is precisely the direction in which their training is now headed.

Overall, CSB 2004 was a great experience, enjoyable as well as productive. The lessons I
took away were to learn more biology, take even more care to make my talks and writing
accessible to biologists as well as engineers, and finally to stay in bioinformatics! While
clearly in its early stages as a discipline, the sheer breadth of questions to be answered and the
need for powerful computational algorithms with which to address those questions make it an
exciting field for information engineers to work in.