Report from the ISGO International Conference on Structural Genomics
(ICSG) 2002

October 10-13th 2002
Berlin, Germany

NB: This report is distilled from my handwritten notes, if anyone wants to
see the programme, participant list, abstract book, or wants more detail
about a particular presentation then, please let me know. If anyone is
interested in some or all of my full notes for the whole workshop +
conference (24 sides of A4!) then please let me know - PJB.

Thursday 10th October

Robert Huber opened the conference with a talk about proteosomes (funtions
include removal of denatured proteins and degradation of short-lived
signalling molecules). These have become a target for drug design but are
large and complex. He described attempts to understand the functional
mechanism based on an understanding of structure.

Chris Dobson talked about the problems of understanding the mechanisms
behind protein folding. Molecules are able to fold accurately and quickly
in a packed cellular environment, and correct folding is essential to
biological function. Large complex molecules appear to be assembled from
smaller pre-folded molecules. Misfolding is the cause of a variety of
diseases: a particular example being protein amyloid diseases (e.g.
Parkinson's disease) where misfolded proteins aggregate and form amyloid
fibres. He suggests that proteins can be thought of as "evolved polymers",
and that certain sequences are more prone to misfolding & aggregation than
others. Since often only a few key residues are responsible for
misfolding, it may be possible to design sequences that preserve function
whilst being more resistant to these problems.

In the session on "Protein Production for Structural Genomics", Dave
Stuart talked about the development of the Oxford Protein Production
Facility, 3yr MRC-funded project to establish HTP protein production &
crystallisation, which is focused on biomedically significant targets
(cancer, immune cells proteome, Herpes virus).

Naomi Chayn talked about methods for HTP xtallisation (drop under oil in
microbatch arrays). Suggests that "containerless" methods (e.g. drop
suspended between two layers of oil) might be suitable for xtallising
membrane proteins. Geoffrey Waldo (Los Alamos) talked about engineering
proteins for improved solubility/crystallisability (points out that 40% of
proteins in TB genome are insoluble). He creates crystallisable constructs
via "directed evolution" i.e. generate mutants and screen for improved
solubility etc, then make mutants from those (screening done using GFP as
a reporter). It wasn't clear whether function was significantly impaired
at the same time as improving crystallisability.

Friday 11th October

Kurt Wuthrich's plenary gave an overview of the history of determining 3D
NMR solution structures of biological macromolecules. Pointed out that NMR
can study proteins in solution, e.g. in vivo fluids, and that it has a
valuable role to play as a complementary method to xtallography: e.g. NMR
studies of confirmational states of individual molecules in supramolecular
structures, like membranes.

Following session examined some of the challenges of studying membrane
proteins, particularly using xtallography (apologies, from my notes I am
unsure whether these are general problems or specific to particular
proteins) for example: lipid co-crystallisation (difficult to remove), and
the xtal packing can be discontinuous or anisotropic. Different ways of
expressing proteins have (dis)advantages: in membranes (correctly folded
but often toxic to the host, difficult to preserve fold whilst purifying);
as inclusion bodies (partially folded or unfolded, so difficult to refold,
but can give high yields); in helical arrays (rarely observed but useful
for EM).

Chris Sander's plenary gave an overview on the "Challenge of Structural
Genomics". His vision was to solve the 3D structure of "all" proteins, to
use for classification, understanding molecular evolution, for drug design
and so on. He believes we only need to experimentally determine a subset
of proteins which is sufficient to cover "protein space", then others can
be determined from homology modelling (estimates 4000 required to cover
90% of Pfam). Can then perform evolutionary classification of all proteins
by structure and functions (it's possible to have structural divergence
and functional convergence). Another challenge is studies of "functional
complexes" i.e. protein-protein interactions, and study of all gene
products in 3D (not just proteins, e.g. also functional RNA).

In the following session on "Structural Genomics and Disease" Wim Hol
talked about the SGPP project ("Structural Genomics of Pathogenic
Protozoa"), and the issues associated with using structural genomics
approaches for fighting disease. He argued that there are great problems
with predicting active sites from structural/functional similarity.
Pointed out that new drugs need to be delivered in pairs, to reduce the
evolution of drug resistance. Eric Adam talked about the platform being
developed by Syrrx to go rapidly from gene to structure to drug (in one
case in just 4 months). Suggested that availability of structures leads to
a diversity (chemically speaking) of potential drugs.

Jean-Paul Renaud's plenary focused on "Characterising Orphan Nuclear
Receptors through Structural Studies". "Orphan receptors" are those for
which no ligand is known, basically they want to find the ligands. Problem
as before is low solubility and/or low crystallisability of the target
proteins, often due to the absence of the ligand.

Sun-Huo Kim talked about the structural genomics ideal of going from
structure to function. Over 60 genomes have been sequenced, containing
10^3 to 10^5 genes, and the functions of over 50% of these are either
unknown or uninferable from the sequence; also 8-20% of functional
annotations are estimated to be in error. He concludes that sequence alone
is insufficient for functional inference, and talked about the idea of
"functional genomics": molecular function (chemistry and physics), and
cellular function (i.e. pathways of molecular functions).

Aled Edwards (Toronto) argued that structural proteomics efforts purify
ten times more soluble protein than is needed for structure determination
via xtallography, and that this excess could be used in rapid biochemical
screens testing for function.

The session on "NMR Methods for Structural Genomics" contained talks on
automating the process of structure determination from NMR data. The
current bottleneck is that assignments of resonances and NOES have be done
manually. Miguel LLinas (Pittsburgh) talked about CLOUDS, which performs
assignments automatically and generates proton densities analogous to
xtallographic maps from which protein structures can be determined.
Geometry and overall quality of CLOUDS structures are comparable with
previous manually-determined structures. Peter Guntert has alternative
software CYANA, which automates two of the stages of conventional NMR
structure solution (collection of confirmational restraints & structure
calculation) - see http://www.guerntert.com. Other talks in the same
session concentrated on NMR as a powerful complementary technique to
xtallography.

Saturday 12th October

Janet Thornton's plenary talked about tools to aid "functional annotation"
- determination of biochemical and biological function and various other
properties (e.g. location of active site, which ligands bind to it) from
the protein structure. Developing program called Profunc (with Roman
Laskowski and James Watson) to do this automatically. There are already
many programs available to do auto identification of protein family from
structure comparison (SAP,SSM,DALI) but same superfamily doesn't
necessarily imply same function. Also, <35% sequence identity almost
definitely alters the function (e.g. binding to different substrate). Uses
3D templates (Andrew Wallace et al, 1997) to search new structures for
possible binding sites, and other methods to determine multeric states,
biological interfaces etc. She argues that "data integration" is required
to pull together information from different sources so it can be used for
functional annotation.

In the "Bioinformatics: Protein Space" session, Sung-hou Kim talked about
the idea of a "molecular function-fold family dictionary" (like a periodic
table of protein fold domains). He introduced the concept of "structural
distance" as a measure of pairwise structural dissimilarity; performing
various analyses he concludes that structures are clustered in particular
regions of 3D fold-space. Lisa Holm talked about a new program "MaxFlow"
for performing structure/sequence alignments. It uses a method called
"transitive alignment", which moves between closely related sequences as
steps to reach more distant homologues - this method seems able to
generate accurate alignments across large evolutionary distances.

Allampura Babu (PSF Berlin) talked about a "web book" (essentially
web-based LIMS) for handling structural genomics data. Has a dictionary of
data items based on a PDB initiative. Basic architecture consists of: web
server (Apache) & database server (MySQL), with scripting done in Python.
Used Netscape as web browser and developed on a Linux platform - so all
components are freely available.

In the second bioinformatics session Olivier Lichtarge (Houston)
introduced the idea of "evolution as a computational filter", to try and
identify functional sites in protein structures via comparisons of
sequences. The concept is that every branch point in the evolutionary tree
is like a virtual screen - so look for changes in trace residues which
alter the function, as these point to changes in the binding sites
(clustering of such residues in the structure are more likely to overlap
with real functional sites - I guess you also need the structure in order
to determine the clustering).

Kengo Kinoshita (Yokohama) introduced eF-site (=electrostatic-surface of
Functional SITE, see http://www.pdbj.org/eF-site), a method for
identifying functional sites. Similar to e.g. DOCK in that it uses
graph-theory and searches against a database - except in this case
compares site against other sites (rather than against possible ligands).
Was able to predict function of a hypothetical protein (in spite of
presence of new fold), later confirmed directly by expt.

In the session on "X-ray methods for structural genomics", Victor Lamzin
talked about the most significant developments in the new version of
ARP/wARP: model-building is faster & models are more complete; can deal
with poorer starting phases; building is possible at lower resoln limit,
e.g. can get 50% completeness at 2.9A. It is important to use good quality
data (e.g. excluding poor data in innermost resoln shell resulted in more
complete model - possibility of feeding this information back into data
collection?). Also can deal with large multimeric structures, and basic
automated ligand building. Future developments include: extension to lower
resoln (~3A); ~100% model completeness (poorly defined loops, NCS);
improved ligand building.

Tom Terwilliger (in addition to plugging PHENIX) talked briefly about
SOLVE (essential feature: a reliable method of scoring putative heavy atom
sites which can discriminate between correct and incorrect solns) and
RESOLVE (iterative model building at moderate resoln, using template
matching into electron density plus probabilistic sequence alignment).

Wayne Anderson (Chicago) gave an overview of his method for automating
structure determination, using existing software (CNS, SOLVE/RESOLVE,
SHELXD). Essentially has a central MySQL database (the "queen ant") which
distributes work to "worker ants" (which are Unix daemons idling until
they recieve some input, one ant for each stage e.g. "scaling by CNS").
Easily extensible to distributed computing environment. Will be released
as open source at some stage.

Dominika Borek (Dallas) suggested a novel data collection protocol. In a
"typical" Se-MAD expt 10% of the errors are due to radn damage, 1-2% are
systematic & random errors; can reduce the latter errors by using a
strategy which maximises the number of photons collected, and correct for
effects of this damage using a model of the resultant changes in
intensities due to radiation damage. In future it is possible that
radiation damage could be used as a source of additional phase information
to aid in the structure determination.

Sunday 13th October

Helen Berman talked about efforts at the PDB to facilitate integration of
structural genomics activities, and to enable HTP deposition. Examples
include: PDB Structural Genomics Portal
(http://www.rscb.org/pdb/strucgn.html); Target database
(http://targetdb.pdb.org, somewhere for groups to register targets or see
if someone else is working on it); Data dictionaries
(http://deposit.pdb.org/mmcif, to enable data exchange between local
databases, includes dictionaries for X-ray/NMR/protein production); tools
for HTP deposition (http://deposit.pdb.org/software, standalone version of
ADIT plus other validation tools available as open source, for individuals
to integrate into their own projects).

Mitch Guss (Sydney Australia) reported back on efforts to facilitate
publication of the results from structural genomics. He conceded that as
yet structural genomics has not delivered an avalanche of new structures,
but suggested it may happen in future - in which case the community needs
to examine the way that macromolecular structures are published. Compares
situation now with the problem that small molecule community faced 30yrs
earlier, and where now most structures are published electronically.

Suggests that scientific publication is something which "adds value" (e.g.
by a discussion of biological importance); which acts as a "statement of
record"; is refereed (nb PDB depositions are validated but not refereed);
is long-lived (won't be changed in future). "Electronic publication" is
taken to refer only to the method of distribution i.e. it cannot be
changed over time.

Tom Terwilliger talked generally about the principles of the ISGO: to
develop standards and policies for structural genomics; sponsor
international meetings and workshops; promote co-operation (including
public-private efforts); promote publication of data. It has a number of
task forces (deposition, publication, IPR, target tracking, etc). Key
principles: free exchange of data and materials; deposition of coordinates
& mandatory data in PDB immediately on completion of structure
determination; public release in short time (<6 months); open exchange of
targets via target database.

There are particular concerns regarding patenting, the ISGO supports the
position that patents "need to have a high degree of utility."

Website is http://www.isgo.org, mailing list news@isgo.org.

Stephen Burley (Structural GenomiX aka SGX) gave a plenary lecture in
which he stressed the need for fast turnaround xtallography in
structure-based drug discovery. The drug discovery procedure is iterative,
at each stage information on structure feeds into chemistry etc to suggest
the next set of leads that will be examined in the next cycle. Essentially
it is necessary to be able to co-crystallise and solve structures of bound
ligands on the time scale of one of these cycles (~4 weeks).

He also introduced a method he called "virtual screening by docking":

1. A large database of binding candidates is reduced to a much smaller
"enriched" database by applying 1st generation methods (e.g. DOCK) as a
rapid filter

2. Use a scoring method which accurately ranks ligands in terms of binding
energies. Historically scoring functions have been poor indicators because
confirmational changes can cause large changes in binding energy, so SGX
method is to take a starting confirmation with the ligand, add solvent and
then run molecular dynamics to generate a series of "snapshots" with
different binding energies. This way they can generate an average binding
energy which can be used for scoring.

Peter Briggs 29/10/2002
--------------------------------------------------------------------------