542
D. Fourches
ended the days when multi-linear regression models were built on small congeneric
sets of chemicals with no rigorous validation protocol, no applicability domain, and
no experimental confirmation. Based on recent requirements set by the editorial
board of the Journal of Chemical Information and Modeling [21], a significant portion of QSAR papers published over the past twenty years would not be accepted
for publication today. This is one of the many consequences of today’s profusion of
experimental data: because it is now easier and inexpensive to measure experimental activities and properties for many chemicals, there is no need for non-predictive,
non- or moderately reliable models built with few compounds that are not transparent and have not been proven to be efficient for virtual screening.
The vast majority of Cheminformatics approaches and associated computational
tools are not ready yet for analyzing and modeling these new generations of very
large, diverse, and complex data streams. Beyond the technical challenges due to
the extreme size levels of the chemical databases requiring new analysis approaches, algorithms, and architectures (such as GPU computing), there are two major related issues that need to be addressed with priority. When it comes to data accuracy,
cheminformaticians are indeed at the mercy of data providers who may inadvertently publish (partially) erroneous data. Several reports on the presence of errors
and inaccuracies in public depositories [22, 23] were published recently, especially
the study from [24] at Bayer describing that only 20–25 % of published data were in
agreement with their in-house results when attempting to reproduce the exact same
experiments. This lack of experimental reproducibility motivated Nature’s editorial
board to reinforce the journal’s acceptance criteria in terms of statistical validation
for reported results in submitted manuscripts as well as abolishing size constraints
for method descriptions. Our group at UNC recently published a list of guidelines
for chemical data curation prior to cheminformatics analysis [25]: we described
several simple but important steps for cleaning chemical records in a database, including the removal of a fraction of the data that cannot be appropriately handled
by conventional cheminformatics techniques. Chemical and biological curation is
critical for enabling the accurate integration of data across different databases and
assuring the correctness of models and hypotheses built on top of these data.
The profusion of chemical and biological data not only enables large scale
cheminformatics studies but also novel areas of applications. Certain of these applications are just emerging and likely to significantly expand in the next few years.
For example, one notable new approach involves the use of text mining of large
collections of articles to extract information relevant to a specific compound, target,
disease, or any other user’s defined term. Recently, Chemotext [26] has been developed on top of Mesh annotations to derive compound-target-disease assertions. Our
group at UNC showed that it was even feasible to derive predictive QSAR models
from curated literature assertions extracted using text-mining protocols [27]. Another emerging trend is the development of hybrid QSAR models that involves a
combination of chemical and biological descriptors. For instance, toxicogenomics profiles determined for 127 drugs were recently used together with chemical
descriptors to generate hepatotoxicity QSAR models [28]. This strategy recently
culminated in the development of the Chemical-Biological Read-Across (CBRA)
D. Fourches
ended the days when multi-linear regression models were built on small congeneric
sets of chemicals with no rigorous validation protocol, no applicability domain, and
no experimental confirmation. Based on recent requirements set by the editorial
board of the Journal of Chemical Information and Modeling [21], a significant portion of QSAR papers published over the past twenty years would not be accepted
for publication today. This is one of the many consequences of today’s profusion of
experimental data: because it is now easier and inexpensive to measure experimental activities and properties for many chemicals, there is no need for non-predictive,
non- or moderately reliable models built with few compounds that are not transparent and have not been proven to be efficient for virtual screening.
The vast majority of Cheminformatics approaches and associated computational
tools are not ready yet for analyzing and modeling these new generations of very
large, diverse, and complex data streams. Beyond the technical challenges due to
the extreme size levels of the chemical databases requiring new analysis approaches, algorithms, and architectures (such as GPU computing), there are two major related issues that need to be addressed with priority. When it comes to data accuracy,
cheminformaticians are indeed at the mercy of data providers who may inadvertently publish (partially) erroneous data. Several reports on the presence of errors
and inaccuracies in public depositories [22, 23] were published recently, especially
the study from [24] at Bayer describing that only 20–25 % of published data were in
agreement with their in-house results when attempting to reproduce the exact same
experiments. This lack of experimental reproducibility motivated Nature’s editorial
board to reinforce the journal’s acceptance criteria in terms of statistical validation
for reported results in submitted manuscripts as well as abolishing size constraints
for method descriptions. Our group at UNC recently published a list of guidelines
for chemical data curation prior to cheminformatics analysis [25]: we described
several simple but important steps for cleaning chemical records in a database, including the removal of a fraction of the data that cannot be appropriately handled
by conventional cheminformatics techniques. Chemical and biological curation is
critical for enabling the accurate integration of data across different databases and
assuring the correctness of models and hypotheses built on top of these data.
The profusion of chemical and biological data not only enables large scale
cheminformatics studies but also novel areas of applications. Certain of these applications are just emerging and likely to significantly expand in the next few years.
For example, one notable new approach involves the use of text mining of large
collections of articles to extract information relevant to a specific compound, target,
disease, or any other user’s defined term. Recently, Chemotext [26] has been developed on top of Mesh annotations to derive compound-target-disease assertions. Our
group at UNC showed that it was even feasible to derive predictive QSAR models
from curated literature assertions extracted using text-mining protocols [27]. Another emerging trend is the development of hybrid QSAR models that involves a
combination of chemical and biological descriptors. For instance, toxicogenomics profiles determined for 127 drugs were recently used together with chemical
descriptors to generate hepatotoxicity QSAR models [28]. This strategy recently
culminated in the development of the Chemical-Biological Read-Across (CBRA)