Ariana Pharmaceuticals, a Paris-based drug discovery software firm, has raised €1.5 million ($2.3 million) that it plans to use to strengthen its sales and marketing efforts and to add new functionality to its flagship data-mining software platform, called KEM, for knowledge extraction and management.
Vizille Capital Innovation, a subsidiary of the Crédit Mutuel - CIC bank, was the lead investor in the funding round. Ariana is headquartered in the Pasteur Institute’s incubator, called Pasteur Biotop, and previously received seed funding from the Pasteur Institute and Aventis.
Ariana, founded in 2003, is pursuing a hybrid business model based on software licenses for KEM, data analysis services, and drug-discovery services such as in silico screening and library design. The firm counts among its customers Pfizer, GlaxoSmithKline, Novartis, and Lundbeck.
KEM was developed to analyze multi-parametric drug-discovery data. The engine mines the data, automatically extracting relations between many different parameters. It has multiple interfaces, which can be applied to ADME evaluation, library screens, bioassays, and other facets of drug development work, Mohammad Afshar, Ariana’s co-founder and CEO, told BioInform.
The company’s technology focus is on decision support and data mining. “It’s an artificial intelligence technology that allows us to analyze data in an exhaustive way, generate hypotheses in a systematic way, and present them to the user to help them make decisions,” said Afshar. “Our tool allows you to identify areas where you need to dig.”
The tool, he said, is applicable at many points along the drug discovery and development
pipeline. Each of these applications must read different data types and must link with varying types of other software.
As a result, he said, “part of what we want to do with the money we have raised, is, for example, [to] improve all the interfaces of our software” to better interoperate with other tools.
“We want to be able to read more data formats and have a seamless interface with a number of industry standard tools,” he said.
Follow the Lead
Existing computational drug-discovery tools on the market use statistical approaches to score molecules based on certain properties and test results, he said. Those methods become problematic in the case of multi-parametric data, when they try to “add things which are not additive,” such as adding the molecular weight and the number of hydrogen bonds in a molecule.
This can create a roadblock for translating in silico results into chemistry, he said, noting that predicting toxicity or absorption with a statistical tool may be practical, but does not tell a chemist what to change in the molecule.
“We want to be able to read more data formats and have a seamless interface with a number of industry standard tools.”
“We use logic” instead of statistics, Afshar said. “We don’t use trends [or] weighed averages; we use logical clauses.” These “clauses” are arranged in logic networks in order to compare different drug discovery parameters. The approach does not expose these parameters to a mathematical formula that weighs and grades them, but instead finds “the subset that is consistent with the multiple endpoint I want to reach,” he said.
In another departure from other software tools on the market, KEM does not compare input data to a database, but instead takes an unsupervised approach. “It looks at the data the chemist has, for example 200 molecules … it can say within the data … that modifying this bit by that bit should improve your [molecule’s] activity.”
Traditional statistical analysis “is done on a few [data] columns, which have been pre-selected,” he said, noting that KEM can analyze all data columns simultaneously in order to identify, for example, patients at risk during a clinical trial or sub-populations that might be more responsive to the drug.
The platform also has a component nicknamed “innocent bystander,” which examines features of a given molecule to identify chemical groups that could prove valuable if tested under different conditions. For example, this module may indicate that a molecule has three different functional groups that have not been studied on their own.
A Logic Engine
The engine itself is derived from theoretical research done at Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier, an artificial intelligence lab within France’s Centre National de la Recherche Scientifique and the University of Montpellier.
The technology is based on the work of Jean Sallantin, a computer scientist and mathematician and the lab’s principal investigator, who had been working on the mathematical theory of Galois lattices — a method of clustering data in a matrix — for over twenty years.
Afshar explained that this approach to data analysis offers scientists “interesting ways of organizing data, allowing you to identify motifs” within the data.
“Together we saw the possibility of applying this to pharma problems.” Galois lattices enable researchers to look at all subsets of instances in a data set that are distinguished by their individual parameters. In drug discovery applications, a scientist could place results from experiments, such as pharmacokinetic properties, in the lattice to derive their relationships with other parameters.
“What is nice about logic is that the dimensionality of the problem does not really apply. You can have as many descriptors as you want because you are identifying logical relations between all these ways of describing molecules,” Afshar said.
Ariana recruited several scientists from the Sallantin lab to apply the theory to practice in drug design.
However, Afshar said, “it has been a long road to work with the pharmas, [to] make sure we have the data and then make progress on both the theoretical side and the engineering side, such as getting the software to work and having the right interfaces.”
Initial work led to a product called eADMET, which profiles large datasets of absorption, distribution, metabolism, and excretion results to optimize molecules. “It took us a while to get [the tool] to do what it was supposed to do,” he admitted.
Working on a molecule is not a linear process. “Often what happens is that you are getting activity, absorption for example, but when you try to get a good half-life for the product you start losing the activity; you [then] catch the activity and you start getting something that starts to be toxic; you get the toxicity right and you realize [the drug] is not absorbed anymore,” he said.
In the course of the iterations two characteristics may be fine while two others are not. This is an issue of “multi-objective optimization,” Afshar explained, noting that pharma companies do not have the tools to address this challenge in a systematic way.
“The tools may predict a particular endpoint but they don’t have tools that would allow them to optimize all of these things at the same time,” he said.
While developing KEM, the company realized the tool offered broader applications, such as establishing safety profiles for drug leads. “We can detect signals very early,” Afshar said. “It is not a substitute [for] statistical analysis; it comes up with early warnings.”
Ariana’s platform can also be used with other data types, such as genomic data. The firm has begun collaborating in this vein with a “major diagnostics company” that Afshar did not identify.
Building in Biotop
Afshar originally trained as a medical doctor in France, and moved toward drug discovery, first working at the interface of biochemistry and software development and then continuing toward computational chemistry and drug design.
He previously co-founded the Cambridge, UK-based structure-based drug discovery company RiboTargets, which used software to create molecules in silico, integrating chemistry and biology information about the molecules, he said.
In 2003, RiboTargets merged with one of the UK’s oldest biotech firms, British Biotech, which shortly thereafter merged with biotech Vernalis. In the merger with British Biotech, according to company statements, RiboTargets was valued at £26 million ($40 million according to the exchange rate at the time.)
As this transaction took place, Afshar began talking to the Pasteur Institute about a new venture, and the institute agreed to offer seed funding and around 1,000 square feet of lab space in its Pasteur Biotop incubator, which is located on its central Paris campus. Aventis, which at the time had not yet merged with Sanofi, also offered an undisclosed amount of seed funding to Ariana.
“The Pasteur Institute, in additional to capital, also brings in a series of services to these companies [in Pasteur Biotop],” Afshar said. “They provided us with a number of contracts with the Pasteur, so we were providing drug discovery services to the Pasteur, so that was generating cash.”
Afshar declined to offer figures but said, “it allowed us to launch the company.”
Afshar said Ariana next landed contracts with GlaxoSmithKline and Pfizer in the UK and was able to use the cash flow from these projects to further develop its software technology.
“We still don’t have any small companies as our clients,” Afshar said.
At Lundbeck Research in Denmark, Ariana Pharma’s software delivered a helpful result in an analysis of large-scale biological data across multiple assays, according to Jonathan Mason, divisional director in computational chemistry and structural biology and chief scientist in predictive technologies and drug design at Lundbeck.
“Even quite related assays were giving a signal that was not predictable from other assay results, either directly or via simple combinations,” Mason told BioInform via e-mail. This result told his team “there was value in producing and using all the broad biological profiling data.”
Ariana’s technology platform has also been put to work in the development of tiludronic acid, or Tildren, a bisphosphonate drug used to treat a degenerative ailment in horses causing lameness called navicular disease and which is similar to human osteoporosis.
As Thierry Bardon, R&D director at Ceva Animal Health in Libourne, France, explained to BioInform in an email, his company “successfully” used Ariana’s platform in its drug development process, in particular in a phase that is comparable to a Phase III clinical trial in humans, he said. The platform brought a result to light that traditional statistical methods did not detect and it shaped this drug’s development process.
“We were able to identify relations between parameters that were not due to chance … and pinpointed a sub-population of animals that was not visible with standard statistical methods,” he said.
“The most interesting outcome for us was being empowered to orient the trials that followed by modifying inclusion criteria to account for the very important individual differences between horses suffering from lameness [in the trial].”