LION Bioscience AG, Im Neuenheimer Feld 517, 69120 Heidelberg, Germany
Columbia, Dep. of Biochem. & Molecular Biophys., 630 West 168th Str, New York, N.Y. 10032, USA
EMBL, 69 012 Heidelberg, Germany; rost@embl-heidelberg.de; http://dodo.bioc.columbia.edu/~rost/
contact e-mail:rost@embl-heidelberg.de
In the wake of the genome data flow, we need - more urgently than
ever - accurate tools to predict protein structure. The problem
of predicting protein structure from sequence remains fundamentally
unsolved despite more than three decades of intensive research
effort. However, the wealth of evolutionary information deposited
in current databases enabled a significant improvement for methods
predicting protein structure in 1D: secondary structure, transmembrane
helices, and solvent accessibility. In particular, the combination
of evolutionary information with neural networks proved extremely
successful. The new generation of prediction methods proved to
be accurate and reliable enough to be useful in genome analysis,
and in experimental structure determination. Moreover, the new
generation of theoretical methods is increasingly influencing
experiments in molecular biology.
Key words: protein structure prediction, evolution, neural networks.
Proteins constitute life's machinery. The first bacterial
genome was sequenced in 1995 [1] ; the first eukaryote (yeast)
followed in 1996 [2] . Meanwhile, more than ten other genomes
have been published [3] , and the human genome (200 times larger
than yeast) is expected to be sequenced as one of the first milestones
in the next millennium. Why bother? Since genomes contain the
blueprint for all parts of life's machinery. The machinery itself
consists of proteins that perform all important tasks in organisms
(catalysis of biochemical reactions, transport of nutrients, recognition,
and transmission of signals). Proteins are formed by joining
20 different amino acids (dubbed residues, when joined in proteins)
into a stretched chain. In water, the chain folds into a unique
three-dimensional (3D) structure ( Fig. 1 ; introduction to protein
structure: [4] ).
Fig. 1.
Representation of the 3D structure of Leishmania
mexicana triose phosphate isomerase (TIM, Protein Data Bank [58]
code 1amk). The trace of the protein chain in 3D is plotted schematically
as a ribbon. Strands are indicated by arrows (yellow), helices
by open coiled-tubes (red). Graph made with RASMOL (Roger Sayle,
ras@32425@ggr.co.uk), and Molscript
(Peer Kraulis, http://www.avatar.se/molscript, [60]
The TIM-barrel is named after the barrel
formed by the strands in the centre of the molecule. The enzyme
is found with a similar structure in most of the known life-forms,
and thus represents a billion year's old perspective at the complexity
of the shapes of life.
Sequence determines structure determines function. The world of proteins is governed by shape: interactions between proteins are mediated by the 'key-hole' principle, i.e., two proteins interact when they fit to one another like a key into a hole. Thus, protein structure determines protein function. What determines structure? All information about the native structure of a protein is coded in the amino acid sequence, plus its native solution environment [5] . Can we decipher the code, i.e., can we predict 3D structure from sequence? In principle, we could; in practice, such approaches are frustrated by the difficulty of the task resulting from the high complexity of protein structure formation [6] . For over 40 years, there has been an ardent search for methods predicting protein structure from sequence (reviews: [7, 8, 6] ; books: [9] ). Many methods were found which looked initially very promising - but always the hope has been dashed [10] . The most successful predictions are achieved by experts starting combining machine-based predictions with their intuition and expertise [11] .
How can neural networks predict protein structure? In practice, the most successful structure predictions extract patterns from data bases of known protein structures. Neural networks comprise a particular tool for pattern recognition and classification [12, 13] . To which extent do neural networks contribute to predicting protein structure, in practice? Initially, researchers applied black boxes, and searched improvements through optimising the internal free parameters (training speed, network architecture). Later, researchers have opened the black boxes by extracting, or implementing rules, by carving specific knowledge into the networks, and by using networks to detect errors or outliers in data bases. More recently, the full potential of the tool has been explored by combining neural networks with evolutionary information. Now, applications of neural networks are amongst the most widely used methods in everyday's bioinformatics.
Here, I sketched neural network based methods (PHD series) for
the prediction of 1D aspects (secondary structure, transmembrane
helices, solvent accessibility) of protein structure. The methods
illustrated that (1) neural networks as black-boxes failed to
improve prediction accuracy, (2) neural networks were sufficiently
flexible to carve expertise from biology into the tool, (3) the
quantum leap in prediction accuracy achieved in the 90's unearthed
from implementing evolutionary information into neural networks,
(4) and that the new generation of prediction methods is extremely
useful in assisting, facilitating, and speeding-up experiments
in molecular biology.
Simplifying the structure prediction problem: The rapidly growing sequence-structure gap (number of known protein structures vs. number of known protein sequences) has enticed theoreticians to solve simplified prediction problems [8] . An extreme simplification is the prediction of protein structure in one dimension (1D), as represented by strings of, e.g., secondary structure, and residue solvent accessibility. Theoreticians are lucky not only because the 1D prediction problem is not only the task they can accomplish best, but in that even partially correct predictions of 1D structure are useful, e.g., for predicting protein function, or functional sites.
Basic idea of secondary structure prediction: The usual goal of secondary structure prediction methods is to classify a pattern of adjacent residues as either H (a-helix), E (for extended b-strand), or L (for loop, i.e., no regular structure). The principal idea underlying most secondary structure prediction methods is the fact that segments of consecutive residues have preferences for certain secondary structure states [4, 14] . Thus, the prediction problem becomes a pattern-classification problem tractable by pattern recognition algorithms. The goal is to predict whether the residue at the centre of a segment of typically 13-21 adjacent residues is in a helix a strand, or in none of the two regular structures.
First and second generation prediction methods: The first
generation of 1D prediction methods was based on physico-chemical
principles, expert rules, and statistics of single residues [15, 16, 17, 4] .
The second generation incorporated the influence of residues
adjacent to the residue for which 1D structure was predicted (local
information). These secondary structure prediction methods shared
three major shortcomings: (1) prediction accuracy was limited
to about 60% accuracy (percentage of residues predicted correctly
in either of the three states H, E, L), (2) strands were predicted
at typically < 40% accuracy, (3) predicted secondary structure
segments were, on average, only half as long as observed segments.
Methods were tailored to overcome one of these problems (long-range
information: [18, 19] ; strand accuracy: [20] ; length:
[21] ). However, the basic assumption was that these problems
originated from using only local information (13-21 adjacent residues).
It was assumed that, in general, 65% of the secondary structure
formation is determined by local interactions, and that strands
are dominantly determined by long-range interactions [22] .
No improvement by simple network: A simple tool that
classified sequence stretches into three secondary structure states
was a neural network (more precisely a multi-layered feed-forward
network) [23, 24, 25] . Input was the sequence vector composed
of 13-21 residues; output the secondary structure state of the
central residue ( Fig. 2 ). However, this simple device was not
better than any other good prediction method. In particular,
none of the three problems (prediction accuracy limited to 60%,
strand accuracy around 40%, short segments) of conventional methods
could be solved by such a device [26] . (However, due to inappropriate
choices of the test sets this was not revealed by the first publications
[27] .)
Fig. 2.
Simple neural network for secondary structure
prediction. For simplification the protein sequence given consists
of two amino acid types (S and P). The protein sequence is translated
into patterns by shifting a window of w adjacent residues
(shown w = 5; typical values in practice are w = 13-21) through
the protein. The output of the network is uniquely determined.
Suppose the output would be: 0.2, 0.4, 0.5 for the three output
states (H, E, L). For known examples the desired output is also
known (1, 0, 0 if the central residue is in a helix). Consequently,
the network error is given by the difference between actual network
output and desired output. The only free variables are the connections.
Training or learning means changing the connections such that
the error decreases for the given examples. A training set typically
comprises some 30,000 examples. If training is successful, the
patterns are correctly classified.
Better prediction of strand by balanced training: Prediction accuracy for each of the three secondary structure classes approximately mirrored the observed occurrence of these classes in the training set [28, 14] . In particular, only 21% of the correctly predicted residues belonged to the class E. Looking at the training dynamics of the network revealed that the network learned H, and L ten times faster than E. Consequently, the idea was to improve the prediction for strand residues by simply increasing the frequency in presenting strand residues during training. Thus, instead of presenting in 1000 iteration time steps 220 examples for E, 310 for H and 470 for L (according to database distribution, dubbed unbalanced training), now at each time step one example for each class was used for training (balanced training). (1) All three classes were predicted almost equally well [28] . (2) Overall accuracy decreased, as the loop residues that were predicted more accurately by the unbalanced network comprised almost 50% of all residues. However, a balanced network proved that the inferior prediction of strand did NOT result primarily from long-range interactions, but from a technical problem.
Better prediction of segment length by 2nd level network:
The average length of a helix is about 10 residues. However,
helices predicted by the network were, on average, four residues
long. The reason was that the network failed to learn correlating
the secondary structure state of adjacent residues. The fact
that, e.g., helices span over, at least, three residues was obscured
by the particular training dynamics necessary to avoid unwanted
database bias: examples presented in time steps t1
and t2 were chosen at random from the training
set (and, thus, were usually not adjacent in sequence). This
problem was corrected by introducing a 2nd level (structure-to-structure)
network [28, 14] . The input of this 2nd level network was
the output of the 1st level (sequence-to-structure) network; the
output was the secondary structure state of the central residue
( Fig. 3 ). The 2nd level network had almost no effect in terms
of overall accuracy. However, the average predicted helix extended
over more than seven residues, i.e., predictions appeared considerably
more protein like than for the 1st level network [28, 14] .
Fig. 3.
Second level neural network [14] . (1) The
window of w adjacent residues is shifted through the
protein (here w = 5). For each window secondary structure is
predicted for the central residue (shown three windows with central
residues S, P, S). (2) The prediction of this first level network
is fed into a second level network. This is again realised by
shifting a window of w adjacent predictions through the
protein (for the second level w = 3). The final prediction of
secondary structure is valid for the central residue of the second
window (here a P).
Better overall accuracy by averaging over many networks: Networks classify patterns separating them by lines. A particular training run results in a particular classification associated with a particular error. Part of this error usually is random noise. Furthermore, unbalanced, and balanced training occasionally yielded quite different predictions, in detail. Which to choose? The answer was to average over both networks, and to attempt reducing the random noise by generating even more differently trained networks (over 2 * 2 = 4 networks: 1st level: balanced, unbalanced, 2nd level: balanced, unbalanced). This 3rd level average over different networks improved prediction accuracy by 1-2 percentage points, and elegantly combined differently focused specialists [28, 29, 30] .
Several problems solved, but accuracy still rather low:
Incorporating facts about protein structures into the specific
choice of the training dynamics and the combination of many independent
neural networks solved two of the problems of conventional prediction
methods (inaccurate prediction of strand, short segments). However,
the overall prediction accuracy was still limited to about 65%
[29, 26] . Long-range information was not incorporated into
the method. Increasing the window size (number of adjacent residues
in protein fragment fed into the network) failed, as the signal-to-noise
ratio increased considerably for longer windows. This problem
was also reflected by that the networks did hardly use higher-order
correlations in the input information: networks with and without
hidden layers performed almost equally well [25, 26] .
Publishing optimistic results? The fact that the early applications of neural networks did not use higher order information was obfuscated by the inappropriate choices of the tests sets. Interestingly, the problem did not arise from 'computer geeks' intruding into the unknown space of biology. Rather, in the early 90's protein structure prediction experts just started to become aware of the importance of the issue of appropriately testing. Part of the problem is 'social' (better results, better journals, more grants). For example, the history of secondary structure prediction has partly been a hunt for highest accuracy scores, with over-optimistic claims by predictors seeding the scepticism of potential users. However, the CASP experiments (protein structures are predicted before the structures are known, then predictors meet every second year and assess how well they did [31, 11] ) have illustrated: exaggerated claims are more damaging than genuine errors. Even a prediction method of limited accuracy can be useful if the user knows what to expect. For the editors of scientific journals this implies that no protein structure prediction method should be published that has not been sufficiently cross-validated. This raises a difficult question: how to evaluate prediction methods?
Full jack-knife test: The prediction of protein structure is an excellent example to illustrate the traps of 'positive thinking'. Given a data set of N proteins of known structure. The ideal testing is by jack-knifing: take N-1 proteins for training, one for testing, and repeat N times. This recipe is simple, but does it suffice? Not, without ascertaining additional constraints. (1) No information of the performance on the one protein used for testing each time ought to steer the training. That is, developers do not want to adjust free parameters (network architecture, training speed, start conditions, training stop)
Complete cross-validation: The full jack-knife (N repeats of the experiment with splitting the N proteins of known structure into N-1 / 1 ) is often prohibited by limited computer resources. A simple alternative to full jack-knifing is the complete cross-validation: the N proteins are split into N-N/F training and N/F test proteins, this is repeated F times. For say F = 10, we refer to this as a complete ten-fold cross-validation. Again, the same constraints apply to each of the separations: (1) no information from F test proteins used for training, and (2) no overlap between training and testing set. In practice, we first have to play around with the neural networks to find a combination of network parameters for which we believe the method would be optimal. How can we thus avoid to look at what we pretend to be unknown (the structure for the testing proteins)? The solution brings forward a third data set, the validation set. Now the N proteins are split into (N-N/F-V ) training proteins, V validation proteins, and N/F test proteins. The V validation proteins are used to find the optimal parameters for the neural networks (and the conditions between training and testing set now apply to all three sets).
Number of cross-validation experiments not important: Are higher values of F (number of cross-validation experiments) better than lower ones? We often find an affirmative answer to this question in the literature. However, the exact number of F is not important provided the test set is representative, comprehensive and the cross-validation results are NOT miss-used to again change parameters. Developers usually have an interest to choose F as high as possible as that permits maximal use of the available information. However, the choice of F is of no meaning for the user provided the cross-validation is done appropriately.
Comprehensive data sets: All available unique proteins
should be used for testing prediction methods (currently about
1000). The reason for taking as many proteins as possible is
simply that proteins vary considerably in structural complexity;
certain features are easy to predict, others harder. Thus, we
can easily select a large subset from the database of known structures
for which any prediction method looks much more accurate than
any other. This problem also raises the issue of comparing apples
and oranges: no matter which data sets are used for a particular
evaluation, a standard set for which results are published by
others should also be included. Finally, the field of protein
structure prediction offers an additional benefit: every month
we witness the publication of about 50 novel protein structures.
Thus, after having finalised the manuscript describing prediction
methods, developers can simply apply the tool to all the new structures
that were unknown by the time they started to invent their method.
Variation in sequence space: The exchange of a few residues can already destabilise a protein [32] . This implies that the majority of the 20N possible sequences of length N form different structures. But, has evolution created such an immense variety? Random errors in the DNA sequence lead to a different translation of protein sequences. These 'errors' are the basis for evolution. Mutations resulting in a structural change are not likely to be accepted, since the protein can no longer function appropriately. Furthermore, the universe of stable structures is not continuous: minor changes on the level of the 3D structure may destabilise the structure (due to high complexity). Thus, residue exchanges conserving structure are statistically unlikely. However, the evolutionary pressure to conserve function has led to a record of this unlikely event: structure is more conserved than sequence [33, 34, 35] . Indeed, all naturally evolved protein pairs that have 35 of 100 pairwise identical residues have similar structures [36, 37] . But, the attractors of protein structures are larger, even: the majority of protein pairs of similar structures has levels of below 15% pairwise sequence identity [38, 39] .
Long-range information in multiple sequence alignments: The residue substitution patterns observed between proteins of a particular family, i.e., changes that conserved structure, are highly specific for the structure of that family. Furthermore, the substitutions realised by evolution, implicitly also carry information about long-range interactions: suppose residues i and i + 100 are close in 3D, then the types of amino acids that can be exchanged (without changing structure) at position i are constrained by that their physico-chemical characteristics have to fit the amino acid types at position i + 100. Indeed, correlated mutations permit to predict inter-residue contacts [40] .
Feeding profiles of residue exchanges into the networks:
The simplest way to use evolutionary information was as following
[14] . (1) A sequence of unknown structure (U) was aligned
against the database of known sequences (i.e. no information of
structure required!). (2) Proteins that had significant sequence
identity to U to assure structural similarity [36, 37] were
extracted and re-aligned by the multiple alignment algorithm MaxHom
[41] . (3) For each position the profile of residue exchanges
in the final multiple alignment was compiled, and was used as
input to the 1st level sequence-to-structure network.
Significant improvement in overall accuracy: Using evolutionary information in the simple way described improved prediction accuracy already from about 65% to over 70%. Further incorporation of specific information compiled from the multiple alignments [14] yielded a further improvement to levels of about 72% accuracy (system dubbed PHDsec). This number represented an average over a distribution (some proteins were predicted more accurately than others), with an approximate Gaussian form, and a standard deviation of about 10% [14] . The neural network system described, here, was the first to surpass the magic line of 70% accuracy [26] , and proved four years after its implementation still to be the most accurate method at the Asilomar prediction contest in 1996 [42] .
Predicting prediction accuracy: I failed to distinguish
proteins predicted well from those predicted poorly based on their
sequence characteristics. However, the strength of the prediction
(measured as the normalised difference between the output unit
with the highest and the one with the next highest value) provided
an extremely useful index for the reliability of the prediction
for each residue [14] , and for the likelihood that the prediction
for the entire protein was below, or above the average of 72%
[14, 42] . This allows in practice to focus on regions predicted
with higher reliability.
Important class problematic for determining 3D: Even in the optimistic scenario that in the near future most protein structures will be either experimentally determined [39] , one class of proteins will still represent a challenge for experimental determination of 3D structure: transmembrane proteins. The major obstacle with these proteins is that they do not crystallise, and are hardly tractable by NMR spectroscopy. Consequently, structure prediction methods are even more needed for this protein class than for globular water-soluble proteins. Fortunately, the prediction task is simplified by strong environmental constraints on transmembrane proteins: the lipid bilayer of the membrane reduces the degrees of freedom to such an extent that 3D structure formation becomes almost a 2D problem. Two major classes of membrane proteins are known: proteins that insert helices into the lipid bilayer, and proteins that form pores by a barrel of b-strands.
Failure of PHDsec to predict transmembrane helices: The neural network system designed to predict secondary structure for globular proteins failed in predicting transmembrane helices. Hence, the networks were trained again on proteins with transmembrane helices. Largely, the resulting prediction system (PHDhtm) was similar to the one used for predicting secondary structure for globular proteins [14] . One difference was the number of output units. PHDhtm distinguished two states: T (transmembrane helix), and non-T (i.e. a globular region). Again information from multiple alignments improved prediction accuracy significantly [14] . The final prediction system was, at least, as accurate as the best alternative prediction schemes [43, 8, 6] .
Problems of PHDhtm: The system described so fare had a major drawback: the 2nd level structure-to-structure network predicted too long membrane helices. This was corrected by introducing cut-off filters that chopped too long segments into several shorter ones. This procedure was relatively sensitive to parameter choices (when and where to cut). Furthermore, the number of transmembrane segments predicted overall was relatively often wrong.
Finding the optimal path through the network output: The problem of predicting transmembrane helices was ideal to incorporate additional aspects of globular information. This was realised by the following algorithm ( Fig. 4 ) [44, 45] . (1) The neural network (PHDhtm) output was converted to preferences. These preferences
Significant improvements by post-processing network output:
The final system (PHDtopology) achieved as significantly higher
performance accuracy than the simple neural network-based system
(PHDhtm) [44] : for about 89% of all membrane proteins all
segments were predicted correctly, and for 86% of all proteins
all segments, and the topology were correctly predicted (compared
to 82% for PHDhtm). Furthermore, the number of false positives
(globular proteins predicted to contain membrane spanning regions
fall from above 4% to below 2%. (Note: this number is extremely
important to analyse entire genomes [44] .)
Fig. 4.
Post-processing neural network output. A system
of neural networks was trained to predict the location of transmembrane
helices. The network output was treated as an 'energy-landscape'
through which the best path was chosen given the constraint that
transmembrane helices have a minimal (span of lipid bilayer) and
a maximal length (longer energetically unfavourable). Thus, the
normalised network outputs were used as input to a dynamic programming
algorithm that found the model (number and locations of HTM's)
representing the best path through all possible models consisting
of HTM's between 18 and 25 residues by optimising the compatibility
of the model with the neural network outputs. The final refined
model output from the dynamic programming was used to apply the
positive-inside rule: positively charged amino acids (Arginine:
R, and Lysine: K) are more often observed inside [59] . A:
pool of all possible membrane helices; B:
successively building the prediction that is most compatible with
the neural network output, given the assumption that the protein
contains 0, 1, 2, 3 helices; C:
final assignment of the helix orientation (topology) by the charge
difference of the non-membrane regions.
Important step towards predicting 3D? If secondary structure segments could be predicted sufficiently accurately, they may be arranged in space as rigid bodies to yield a model for 3D prediction [46] . One criterion for assessing each arrangement could be to use predictions of residue solvent accessibility. The solvent accessibility of a residue embedded in a protein structure can be described in several ways [47, 48, 49] . The simplest is a two-state description distinguishing between residues that are buried (relative solvent accessibility < 16%) and exposed (relative solvent accessibility ³ 16%). The classical method to predict accessibility is to assign either of the two states, buried or exposed, according to residue hydrophobicity [50, 51, 52] .
Evolutionary information improves prediction accuracy:
Solvent accessibility at each position of the protein structure
is evolutionarily conserved within sequence families [53, 54] .
This fact was used to develop another neural network method for
predicting accessibility from multiple alignment information (PHDacc)
[53, 14] . For this method, I skipped the 2nd level network
since accessibility was hardly correlated between adjacent residues.
The network output comprised ten units. Unit n , for n = 0,
..., 9 coded for a relative accessibility A in the interval
n2 ² A < (n+1)2.
This encoding reflected the observation that in protein structures
residues flip more easily between 70% and 100% relative accessibility
than between 0% and 5%. The final network system predicted about
75% of the residues correctly in either of the two states buried,
or exposed. This was more than five percentage points higher
than for methods not using alignment information.
Structure prediction: work in progress... Native 3D structures of proteins are encoded by a linear sequence of amino acid residues. To predict 3D structure from sequence is a task challenging enough to have occupied a generation of researchers. Have we finally succeeded? The bad news is: no, we still cannot predict structure for any sequence. The good news are: we have come closer, and growing databases facilitate the task.
Predictions in 1D: significant improvement by larger databases: The rich information contained in the growing sequence and structure da Table 1 D predictions. Here I sketched, how evolutionary information input to neural network systems yielded better predictions of secondary structure, solvent accessibility, and transmembrane helices. These predictions of protein structure in 1D are significantly more accurate, and more useful than five years ago.
Conditions to become useful: In the field of structure prediction we have witnessed blooming over-optimism [10] , as well as, more, and less intended cheating. The Asilomar meetings [31] to some extent are succeeding in separating the chaff from the wheat. However, Asilomar does not change the basic formula: when you develop a prediction method you ought to spend more than 70% of the time on appropriate evaluation of the performance [55, 8] . The sustained levels of prediction accuracy published for the PHD methods were, supposedly, one of the major reason for their success. Another important issue is that of making the method available. Molecular biologists do NOT have the time to become experts in running programs. Thus, methods should be easy-to-use, and available via the internet [56, 57] .
Learning from evolution to help studying evolution... Mastering protein structure prediction has several impacts on the advance of biology. Firstly, prediction methods have assisted the experimental determination of proteins structures. Secondly, predictions helped unravelling unknown protein functions, and improving our understanding of the mechanisms of partially known functions. Thirdly, prediction methods allow to separate the chaff from the wheat: we can explore which biological information improves the performance of neural networks and which doesn't. This is a simple, yet effective, means of shedding light into the understanding protein structure formation, and protein function (since we fail to predict structure and function, we thoroughly do NOT understand the underlying principles!).
What next? Most breakthroughs in protein structure prediction
were achieved over the last six years. Thus, although we still
cannot solve the general prediction problem, progress has been
made. In general, however, we could ask the question - is it worth
persevering with structure prediction, given that it is clearly
such a difficult task? The answer is: yes. The methods which
have spun off from structure prediction have already given us
considerable insight into the first four complete genomes. Perseverance
with structure prediction will yield fruit in about five years
time when the human genome will be known.
Thanks - in alphabetical order - to all those who contributed
ideas, and helped with motivating discussions: Michael Braxenthaler
(Hoffman-LaRoche, New Jersey), Søren Brunak (CBS, Copenhagen),
Rita Casadio (Univ., Bologna), Sean O'Donoghue (EMBL, Heidelberg),
Piero Fariselli (Univ. Bologna), Terry Gaasterland (Univ. Chicago),
Gunnar von Heijne (Univ. Stockholm), Tim Hubbard (Sanger, Hinxton),
Raimer Kühnen (Univ. Heidelberg), Chris Sander (Millenium,
Boston), Michael Scharf (Take5, Heidelberg), Reinhard Schneider
(LION, Heidelberg), Manfred Sippl (Univ. Salzburg), Sara Solla
(Western Univ., Chicago), Anna Tramantano (IRBM, Rome), Alfonso
Valencia (CNB, Madrid), Gerrit Vriend (EMBL, Heidelberg). Thanks
to Friedrich von Bohlen (LION, Heidelberg) for financial support.