NEW YORK (GenomeWeb) – Researchers in the Institute for Computational Biomedicine at Weill Cornell Medical College have developed a computational tool for mining combined genomic, chemical, structural, and other datasets to predict new drug targets, identify novel cancer compounds, and find new uses for existing drugs.
Neel Madhukar, a graduate student at Weill Cornell and lead developer of the Bayesian Approach to find Novel Drug Interaction Targets, or BANDIT, software, presented the tool at the annual meeting of the American Association for Cancer Research in Philadelphia last week.
Madhukar is a member of the cancer systems biology laboratory of Olivier Elemento, an associate professor at Weill Cornell. Elemento's lab studies cancer from a systems biology perspective and develops and applies computational methods to match tumor mutations to targeted therapies and enable physicians to select more effective drugs for their patients.
BANDIT is the fruit of one of the lab's projects, which aimed to essentially find a better and faster way to identify novel treatments from publicly available information on small molecules, drugs, and clinical and genomics information, Elemento told GenomeWeb. Madhukar's project, he explained, sought to combine these disparate datasets into a single framework where it could be easily mined and explored.
Currently, the large quantities of structural, chemical, genomic, and other datasets available in the public domain are largely underutilized by many existing computational methods for drug discovery and drug target prediction because these methods typically focus on one or two data types, Madhukar told GenomeWeb.
Such methods can provide useful information on drug targets. For example, by looking at transcriptional data it would be possible to find two drugs with similar transcription response that also have similar targets, he said. But combining information gives a more fine-grained and cohesive picture of what drugs do in biological systems including mechanisms of actions as well as the specific pathways, proteins, and genes that they target, he said. Larger datasets also make it possible to make predictions about untested drugs with no known targets.
That's just what BANDIT is designed to do, he said. It uses a probabilistic approach to compare drug pairs and computes the likelihood that these pairs will have similar targets. The current iteration of BANDIT makes use of a comprehensive database of data on drug efficacies, post-treatment transcriptional responses, drug structures, known adverse effects, and bioassay sensitivities gathered from several publicly available databases, Madhukar said.
Specifically, the drug efficacy data comes from the NCI-60 DTP anticancer drug discovery program, which screens up to 3,000 compounds per year for potential anticancer activity, according to its site. Transcription expression data came from the Broad Institute's Connectivity Map project, which contains more than 7,000 expression profiles representing over 1,300 compounds; while both small molecule structure data and bioassay data were gleaned from the PubChem database. Lastly, it includes data on known side effects for drugs from the Side Effect Resource.
Gathering and combining this data into a single database for BANDIT's consumption was probably the most time consuming task, Madhukar said, since drugs can have multiple names associated with them including chemical and industry names as well as unique identifiers.
"There was a lot mapping names back and forth and trying to figure out if two drugs were the same ... that was probably the biggest part of the project," he said. But the result is what the researchers believe is the most comprehensive database of drug effects and actions combined with known target information, he said.
As a validation step, the researchers first used the method to look at drug pairs with known structures, targets, and mechanisms of action. BANDIT successfully predicted the targets in those cases with 91 percent accuracy compared to other methods used to predict drug targets, Madhukar said.
Next, the researchers used BANDIT to look for drugs that could be used for cancer therapy. Specifically, they explored shared targets between non-cancer drugs and known cancer drugs and also looked for novel targets for known anticancer drugs. According to a conference abstract submitted for AACR, BANDIT predicted a shared target for two cancer drugs, resveratrol and genistein. Previous studies show that "resveratrol enhances the apoptotic effect of genistein … and our prediction reveals that a possible mechanism for the additive effect could be the dual inhibition of a single target," the researchers wrote in the abstract. BANDIT also predicted that the drug vismodegib (Genentech's Erivedge) — used to inhibit the Hedgehog signaling pathway in basal-cell carcinoma cases — could potentially inhibit tyrosine-kinase activity, the results showed.
The researchers then applied BANDIT to roughly 50,000 compounds from the National Cancer Institute's Developmental Therapeutics Program database — many of which haven't made it into development or testing because of high costs associated with drug development — and looked for shared targets between known anticancer agents and unstudied drugs. Among other findings, the researchers identified several candidate molecules that could inhibit the formation of microtubules — drugs that target microtubules are important chemotherapy agents — making them possible alternatives to standard taxane chemotherapies, used for the purpose. That’s important, Elemento noted, because although microtubule-targeting anti-cancer drugs are some of the most successful treatments, there are many patients that don't respond to standard taxane treatments and some develop a resistance to the drug. Finding viable alternatives to taxanes would be immensely useful for cancer therapy.
BANDIT predicted about 17 potential microtubule formation inhibitors — including two drugs, mebendazole and romidepsin, according to the AACR abstract. Elemento's lab then worked with the lab of Paraskevi Giannakakou, an associate professor of pharmacology at Weill Cornell, to test a number of these candidates in an experimental setting. As predicted, their bioassays showed that the molecules did have significant effects on microtubule dynamics. "It's really a nice validation of the approach," he said. The group also identified some molecules that could show promise for crossing the blood-brain barrier, making them potential candidates for brain tumor treatments, Elemento said. The researchers are continuing to test these methods in experimental assays.
BANDIT was well received at AACR, Madhukar said, and some researchers from other institutions have reached out to explore the possibility of using it to explore drug molecules of interest to them. They plan to eventually make BANDIT available for general use but for now, any researchers interested in using the tool are encouraged to contact the team, he said. The researchers are also working on a manuscript detailing BANDIT's development, which they hope to submit for publication in about a month, Elemento said. For now, they are continuing to test the method on various cell lines including some that are resistant to current microtubule targeting anti-cancer agents, he said. Although BANDIT was used specifically for cancer therapies in this study, it can also be applied to identify drug targets for other conditions.
Other planned activities include incorporating additional datasets into BANDIT's database because the power of the classifier increases as the number of included data types grows, according to the developers, and also mining for more predictions about small molecules. They are trying, for example, to predict the potential toxicity of small molecules, Elemento said, using the information contained in BANDIT's database along with results from clinical trials, for instance. This could help researchers identify drug molecules that will fail if moved into human trials. They'll also work on improving BANDIT's ability to predict the mechanisms of action for small molecules, Madhukar said