A R T I C L E S
Kim et al.
activity,11b mitochondrial membrane potential,11c and intracel-
lular reducing equivalents.11d These four readouts were expanded
to 40 cell-based assays, each performed in duplicate, by exposing
cells simultaneously to library members and to small molecules
of known biological activity (i.e., chemical-genetic modifier
screens6). A matrix of data was created using 40 (of 72 possible)
combinations of the four readouts with 18 different cellular
conditions (see Figure 2b). In duplicate experiments, each
compound was independently assigned a signed (i.e., “+” or
“-”) Z-score corresponding to the number of standard deviations
it fell from the mean of a well-defined mock-treatment distribu-
tion. This distribution was determined to be consistent with
experimental noise observed under cell-based assay conditions.5
The resulting collection of continuous-valued Z-scores represents
the primary dataset used for further analysis. For analyses
dependent upon discrete (i.e., binned) outcome states, Z-score
data were further subjected to a threshold that resulted in each
measurement being scored as a high- or low-signal outlier, or
as a nonoutlier, from the mock-treatment distribution, based on
the probability that the measurement could be explained by assay
noise (Pnoise < 0.0006).5 To reduce further the possibility of
false-positive scoring, we required that any compound called a
positive be scored as a positive in both replicates for a given
assay, resulting in an overall average “hit” rate for the binned
dataset of less than 1%.
oral bioavailability.14 In that study, limitations on oral bioavail-
ability resulted from the interaction of small molecules with
complex biological processes involving numerous metabolic
enzymes, a situation comparable to the use of broad cell-based
assay measurements, as described herein.
In a preliminary attempt to determine which cell-based assays
might best distinguish between the M and B compound classes,
we represented 40-dimensional average chemical-genetic fin-
gerprints5 both as activity “signatures” (Figure 5c) and as a
mathematically equivalent heat map (Figure 5d). In these
visualizations, activities are represented as Z-scores averaged
over all members (and replicates) within the M and B compound
classes. These data represent two different visual approaches
to defining assays that discriminate between the two classes.
Noting that some assays result in greater variation of data than
others, chemical-genetic fingerprints based on a subset of these
cell-based assays might be used in future related studies. It is
also possible that a subset of these assays might be used as an
empirical surrogate for molecular properties related to confor-
mational flexibility (e.g., oral bioavailability).14
To expand the scope of our global analysis of the impact of
chemical diversity elements on biological outcomes, we prepared
average chemical-genetic fingerprints (see Figure 1) for each
of six stereochemical categories (for each of the M and B
classes) resulting from the combination of three carbohydrate
templates and two R-substitution relationships on the 12-
membered rings (or corresponding acyclic chains). To eliminate
bias in our analysis introduced by different appendage repre-
sentations between the categories, only 3,4-bis-pentenoates (and
their products) were considered in this analysis, resulting in 12
groups of 12 compounds each having a uniform representation
of appendages across all 12 groups. For each of these 12
categories, whose membership was based solely on the chemical
diversity elements of interest, average chemical-genetic finger-
prints were computed as before, and these fingerprints were
hierarchically clustered using linear correlation between fin-
gerprints as a similarity metric (Figure 5e).15 If categories with
a common structural element cluster together based on biological
performance, then there exists a correlation between this
structural element and global biological outcome. Conversely,
if structurally unrelated categories cluster based on performance,
there is no such correlation of biological outcome with stereo-
chemical diversity elements. Our observations suggest the
former scenario for this dataset. Notably, membership in the
M or B compound class has the dominant overall effect on
biological outcome (Figure 5e; similarity < 0.43), extending a
less complex result (based on a binary “hit” versus non-“hit”
metric; see Figure 5a) to a 40-dimensional biological measure-
ment space. Furthermore, the R-substitution relationship on the
12-membered rings determines subclustering with the B class
irrespective of carbohydrate template (Figure 5e, top branch of
dendrogram; similarity < 0.85). Each of these two subclusters
is further divided into trans-fused (Glc,Gal) and cis-fused (Man)
[10.4.0] bicyclic systems. In contrast, subclusters from the M
class are chiefly determined by the identity of the carbohydrate
template itself (Figure 5e, bottom branch of dendrogram;
similarity < 0.87).
Data Analysis. One of our primary observations was that
precursor monocycles (M) generally scored as positive in a
larger number of the 40 assays than did their bicyclic (B)
counterparts (Figure 5a). To test the statistical significance of
this result, we fit compound Gamma-Poisson model distribu-
tions to each dataset and applied a likelihood ratio test for
significance.12 The reported P-value (P < 0.0002; the probability
of observation in a random dataset) was obtained by permutation
testing; matched pairs of compounds were randomly assigned
labels B and M (one of each label per pair) before computing
the likelihood ratio test on each permuted dataset. For only 11
of more than 89 000 permuted datasets was the likelihood ratio
score for randomly permuted data higher than that for the
experimental data, indicating that our experimental results are
extremely unlikely to have occurred by chance. Overall, this
observation suggests that the propensity of these compounds
to score frequently as active in multidimensional screening is
reduced upon imposition of conformational constraints resulting
from the formation of the second ring.
In an attempt to rationalize these results in terms of calculable
properties of small molecules, we calculated certain average
molecular descriptor values we intuitively thought relevant to
conformational restriction, including the number of rotatable
bonds, for each of the M and B compound classes (Figure 5b).13
These tabular data show that the number of rotatable bonds, of
the descriptors computed, are best able to distinguish between
the M and B compound classes, a result reminiscent of a
previous report of the importance of this descriptor in predicting
(11) (a) Stockwell, B. R.; Haggarty, S. J.; Schreiber, S. L. Chem. Biol. 1999, 6,
71-83. (b) Parish, C. R. Immunol. Cell. Biol. 1999, 77, 499-508. (c) Wong,
A.; Cortopassi, G. A. Biophys. Res. Commun. 2002, 298, 750-754. (d)
Perrot, S.; Duterte-Catella, H.; Martin, C.; Rat, P.; Warnet, J. M. Toxicol.
Sci. 2003, 72, 122-129.
(12) See Supporting Information (Section VI) for a detailed description of the
Gamma-Poisson model fit and the likelihood ratio test.
(13) Molecular descriptor values were calculated using Pipeline Pilot (Scitegic,
Inc.; San Diego, CA).
(14) Veber, D. F.; Johnson, S. R.; Cheng, H.-Y.; Smith, B. R.; Ward, K. W.;
Kopple, K. D. J. Med. Chem. 2002, 45, 2615-2623.
(15) Hierarchical clustering and data visualization were performed using Spotfire
Decision Site 7.0 (Spotfire, Inc.; Somerville, MA).
9
14744 J. AM. CHEM. SOC. VOL. 126, NO. 45, 2004