The number of chemical compounds discovered each day continues to grow at an exponential rate due to constantly refined and
optimized experimental technologies. However, the experimental determination of the chemical, physical and biological
properties of compounds is often very expensive and time­consuming. As the gap between the existing data and the available
information increases, one of the most pressing research issues in drug development is not the production of new compounds,
but the production of properties. Computational methods that aim to apply machine learning methods to molecular structures so
as to predict its biological properties have been used with some success, many times reducing the time and cost required for the
development of new drugs. However most of these methods do not have shown sufficient precision for the prediction of new
structural scaffolds and their precision is low. One of the reasons that have been attributed to the lack of efficiency of
computational methods is the absence of accurate molecular similarity functions. As it is known that similar molecules should
behave similarly, it is obvious that a reliable similarity metric should provide a reliable tool for better pharmacological property
estimation.
Differently from bioinformatics where protein sequence similarity is a known and solved problem with several reliable methods
being extant, (e.g. BLAST). In chemoinformatics, this is not so. Small molecules are paradoxically more difficult to compare and
the structural characteristics of molecules defeat most graph alignment methods. The common solution is the use of fragment
based heuristics, which are imperfect and prone to several types of errors.
Recently within our lab research team we developed, implemented and tested a new molecular similarity method, NAMS, for
Non­contiguous Atom Matching Structural similarity function. This new algorithm was tested in several known databases and
consistently outperformed other similarity measures [TeixFalc2013]. Also a new inference methodology based on kriging over
metric spaces was developed and proved to be superior to other Quantitative­Structure­Activity Relationship (QSAR) methods
[TeixFalc2014]. NAMS can very reliably assess chemical similarity and on global molecular comparison and is able to give very
good estimates on molecular activities of close molecules (something which was not possible before). Also, due to the nature of
the algorithm, NAMS has the potential to be extended so as to not only predict biological properties but also give
pharmacologists and biochemists insight on why several compounds bind to the same targets, despite being structurally
different. This extension will consequently also improve the quality property predictions by highlighting the most relevant parts
of the molecules that are the cause for its activity.
Furthermore, a common critique for QSAR studies is that many chemical inference methods are only evaluated “ex post facto”
with known databases of molecules and respective properties, and have never been put to the test in an actual drug
development program. In the current proposal we aim to use the predictions from the model to retrieve new molecules from
available chemical databasesand test them in a laboratory within two distinct pharmacological problems. The first one is the
prediction of drugs that are able to improve defective CFTR trafficking to the membrane. CFTR is the protein involved in cystic
fibrosis and deriving new candidate drugs is deemed to have significant impact on its treatment. The second problem is the
prediction of Blood­Brain barrier penetration. The blood–brain barrier defends our central nervous system from extraneous
agents, but makes difficult for many drugs to pass it and is thus a fundamental block in many drug development programs.
For this project we have assembled a trans­disciplinary team that has leading expertise in all the required fields for its
completion. The core is a machine learning team with elements from LaSIGE and BioISI, having been the leading developers of
NAMS. This team will work on database development, model building, testing and model implementation. Two other teams from
iMEd.UL (Faculty of Pharmacy) and a Molecular Biology team (also from BioISI) will supervise the biological and pharmacological
issues and all the laboratory work. These teams have world­leading expertise in the blood­brain barrier penetration and on
CFTR trafficking.
Due to the nature of the project and the quality of the team, that has consistently delivered high impact research, this project
has the potential for high visibility and the results of the research to be published in the best scientific journals. It is further
expected that several molecules can be patented for being lead candidates for future medicines.

Literature Review

Molecules are typical examples of unstructured data for which tasks such as searching, sorting, analyzing and extracting
knowledge are challenging. A molecule can have an arbitrary dimension, structure and composition, and moreover, there is not
an univocal and unequivocal way of coding and comparing these molecules. Several computational tools have been developed
over the years in pursuance of solving this issue. Fundamental observations that justify the amount of methods developed to
compare molecules derive from the fact that similarity has a context [bender2004] and the representation of molecular
structures implies information loss. Researchers have explored the concept of similarity between molecules which provides an
important approach to search databases, predict properties of compounds, design structures with a predefined set of properties and conduct structure­based drug design studies [willett2005, eckert2007]. These studies are based on the “neighborhood” premise, which states that similar molecules usually have similar activities and properties [bender2004]. The definition of
similarity for molecules consists of comparing chemical structures, specifically representing the molecules and quantifying the
similarity between them. Various methods to define structural similarity between molecules are available in the literature
[nikolova2003, bender2004, TeixFalc2013]. The most popular approaches to represent the structure of the molecules under
comparison can be divided in three broad categories, approaches based on structural descriptors (two­ and three­dimensional),
molecular fragments and graph matching (descriptor­independent methods). A descriptor positions each abstract molecular
representation in the descriptor space. It is then possible to compare molecules, considering that the distance of the abstract
molecular representations reflects their similarity in this specific descriptor space [bender2004]. Molecular similarity is a nonlinear
problem for which there is not a set of descriptors or a similarity measure that correlates with every context of
comparisons one can perform [todeschini2009, bender2009].

A commonly used approach to predict chemical, physical and/or biological properties of chemical compounds resorts to the
structure of the molecule using data mining methods through quantitative structure­property/activity relationships (QSPR/QSAR)
[katritzky2000, katritzky2002, doucet2011]. The three major difficulties in the development of QSPR/QSAR models are (1)
quantifying the inherently abstract molecular structure, (2) determining which structural features most influence the given
property (representation problem) [liu2004, gonzalez2008, teixeira2013] and (3) establishing and validating the functional
relationships that most accurately describe the relationship between structural descriptors and the property/activity data
(mapping problem) [tropsha2007, puzyn2009, dearden2009, tropsha2010]. Furthermore, it is acknowledged that it is not
possible to develop a model providing reliable predictions for all possible compounds [tetko2009]. Classical QSPR/QSAR
approaches have several shortcomings, namely (1) the predictive power of the model is highly dependent on the selection of
predictor variables and on the presence of correlation between these variables, (2) the prediction capacity of the model is
limited by the molecular diversity and distribution of the molecules in the training set [oprea2001], (3) the models need to be
re­trained every time new compounds are added or removed. Nevertheless [TeixFalc2014]using a kriging based approach over
NAMS, questioned most of these assumptions and showed that it is possible to produce inference models for which no relearning
is necessary, are able to produce estimation errors individually for each estimation and are able to produce reasonable
estimates ever for molecules widely different. Also the initial conclusions from [Martin200] and Nikolova2003] that compounds
that are similar to known active molecules are themselves far less frequently active than one might expect, have been
challenged by NAMS [TeixFalc2013, TeixFalc2014]

A molecule can also be represented, using graph theory, as a labeled graph whose vertices correspond to the atoms and edges
correspond to the covalent bonds. The representation of molecules using graphs has some advantages, namely, graphs are
intuitive when representing a molecule since they are close to our understanding of a molecule and they have a solid
mathematical background with different existing techniques to compare labeled graphs [ehrlich2011]. However, representing
molecules as graphs raises an important issue, identical graphs do not necessarily represent identical structures and viceversa[ehrlich2011].

The goal of finding common subtructrures for property inference has been pursued[kawabata2011,
rahman2009, batista2006]. Nonetheless and despite the NP­completeness nature of the problem (garey 1979) many
approximated heuristics have been proposed to overcome this complexity[ehrlich2011]. The approach followed by
[TeixFalc2013] is different in that it makes no assumptions on any structural components of any molecule and is able to consider
the characteristics that are not directly faced by graph theory (namely chirality, or cis­trans isomerisms). Having a reliable
structural matching algorithm is only part of the solution as it is necessary to make use the detected similarity for predictions.
This has been accomplished by coupling NAMS with kriging, a metric space based method for inference[TeixFalc2014]. Kriging
models have been used previously in chemoinformatics. [fang2004, hawe2010, sun2011]showed that kriging models were able
to outperform other methods in the development of predictive models of pharmacological properties However, in all of these
studies there was always an explicit use of chemical descriptors arbitrarily chosen according to the nature of the problems.

Plan and Methods

The central goal of this proposal is to improve the molecular similarity algorithm NAMS [TeixFalc2013]. by providing topological
enhancements to the main atom­matching algorithm so that the tool is able to both give pharmacologists and molecular
biologists a direct understanding of the relevant components in a family of active compounds; and secondly to use this newly
derived tool as a standalone tool for the screening phases of drug­development programs. NAMS was designed to compare full
molecules, thus in a way tapping the graph isomorphism problem and providing an extremely reliable polynomial solution.
However for a molecule to function as a drug, many times only specific parts of its structure (not necessarily contiguous) are
necessary. Thus it is envisaged that NAMS can be modified by allowing the weighting of specific compositional and structure
elements that are deemed to be essential for understanding a molecule pharmacological properties. In this proposal it is aimed
to extend NAMS allowing this algorithm to differentiate between parts of the molecule allowing for the discovery of the most
relevant parts as well as its differential amplification for molecular property prediction ad drug inference. Differently from other
methods NAMS is not bound by the existence of any type of chemical descriptors and can conceivably be used in any chemical
property prediction problem. The current inference engine [TeixFalc2014] is based on kriging over the global similarity provided
by NAMS, and although providing results on the level of the best state­of­the­art QSAR algorithms using virtual no information
other than the molecular structure, we believe that a topological differentiation mechanism will be key to extend NAMS over
even more different molecules that share pharmacological characteristics. The current version of the inference engine requires
no learning, as it is a kriging algorithm, taking advantage of the molecular metric space; it is further expected that the new
version will be also able to directly assess the relevant topological characteristics of the most active molecules known and be able to use this intrinsic knowledge to retrieve from large molecular databases the compounds more likely to have the desired characteristics.

The research team for the current project is an assembly of computer scientists and biochemists, pharmacologists and
molecular biologists with world leading expertise in the fields of cheminformatics, molecular modeling, blood­brain barrier
permeability and cystic fibrosis.

This work will be divided in 3 major tasks, and each will be critical for the advancement of the project. First, it is necessary to do
a thorough evaluation of the best existing QSAR methodologies and compare the current inference engine based on NAMS and
Kriging [TeixFalc2014]. This comparison is supposed to be as exhaustive as possible, by testing each existing method over a set
of benchmarks created from data collected in publicly available databases. Secondly, NAMS is to be adapted for including
differential topological features in assessing chemical structural similarity and then this adaptation is to be optimized and tested
within a novel framework, centered on Bayesian learning. The purpose is the empirical identification of the molecular topological
characteristics that for each specific benchmark problem. Finally the topological enhanced model is to be put to the test over two
distinct problems, for which there are world leading experts within the team. The first problem is the development of new lead
compounds for enhancing trafficking of F508del­CFTR to the plasma membrane. This problem is in the center of current drug
development for cystic fibrosis. The second problem aims to determine whether a new molecule has the potential for crossing
the Blood­Brain Barrier. This issue is critical in most drug­development programs for drugs that target the central nervous
system.
Methods

The first task of the project will create a set of benchmarks for fitting a variety of QSAR models from the literature. The
techniques typically used range from usual linear models to sophisticated ensemble and hybrid methodologies [tropsha2010]
Typically on most models there are two essential phases. The first one for finding the best possible descriptors for a given
problem [Tropsha2007] and secondly testing and validating the models from a variety of methods that range from classic neural
networks, support vector machines to hybrid elastic nets and probabilistic graphical models [Koeller2009]. The purpose of this
phase is not to find the best possible model, but instead to assess the capabilities and limitations of current QSAR models, and
to have a global view of the sources of prediction errors and the nature of the problems
This proposal will focus on a strict Bayesian approach for including localized structural similarity in probabilistic models. The
Bayesian emphasis will be critical for realistic results. The majority of molecules do not show pharmacological activity and
missing the prior probabilities is one of the key reasons why many in silico screening studies produce unreliable and nonreproducible
results [MartFalc2012]. For stringent comparison of models and the assessment of the new approach, a through
validation process will be followed, with one independent validation set created for each benchmark set that will not be used in
any phase of the training or cross validation procedures, for a stringent unbiased evaluation of all models.

The central part of this proposal is the second task where the purpose is twofold. A) incorporate the topological differentiation
mechanisms into NAMS and B) define a new inference engine that may be use this information to predict the pharmacological
potential of any molecule The purpose is not to solve the specific benchmarks problems specificities, but rather to identify
algorithmically and statistically how can we identify which structural characteristics are fundamental for each binding problem.
For this objective to be accomplished it is required a rigorous statistical handling within a Bayesian framework for adequately
assessing the relevance of each factor. Structural characteristics of a molecule are not necessarily describable by human terms
and are only topological constructs discovered by using a Markov Chain Monte­Carlo approach over NAMS so that the most
relevant parts of the molecule become prominent and can be used through a dynamic weighting layer over the atoms of each
lead molecule. Secondly, the modeling phase will be centered on fitting a Markov network where the similarities to key instances
within the chemical metric space will condition the posterior probability of any unknown instance to have biological activity. Each
network will be derived in a unsupervised way directly from the existing data. Using a data­driven probabilistic graphical model
for each problem will not ensure per se that the Bayesian requirements are met, as it is anticipated that the priors will be
difficult to assess, and their determination will be one of the key challenges in the forthcoming work. This phase will require a
large computational effort that is to be accomplished with the existing hardware and the new servers to be acquired.

The third task will involve testing the new topologically enhanced similarity metric and the respective inference engine over two
current problems in pharmacology. The first one is the retrieval on molecules with the potential to rescue a genetically
malformed protein (F508del­CFTR) to the plasma membrane. This issue is known to be the central factor causing cystic fibrosis.
The second problem involves discovering factors that may induce molecular penetration through the Blood­Brain barrier (BBB).
This is still an ongoing problem and in silico models have had so far limited success [MartFalc2012]. The results from the new
model will be complemented with a virtual screening similarity approach [Lucas2012]. Efforts will focus on the virtual screening
of large chemical libraries of commercially available compounds libraries (e.g., ZINC, NCI). The NCI database contains around
400,000 drugs, from these, only 250 000 are available for download. The ZINC database as now around 35 million compounds,
however, we will not screen the entire database but only “In –Stock Drug­like” subset (~10 million compounds).

After the lead compounds for both problems is defined these will be put to the test in vitro. The rescue of F508del­-CFTR traffic in
cells treated with lead compounds will be assessed by the F508del­CFTR traffic assay we have established for automated
fluorescence microscopy[Almaça2011,Farinha2013]. On the other hand BBB penetration will be assessed over primary cultures
of human brain microvascular endothelial cells derived from microvessels isolated from temporal tissue removed during
operative treatment of epilepsy. Monolayers of human brain microvascular endothelial cells show characteristically high
transendothelial electric resistance and have proven useful in multiple functional studies for in vitro modeling of the human
blood­-brain barrier[Bernas2010]

Tasks