OR WAIT null SECS
“In theory, the methodology could apply to any medical conditions, as long as we could have enough training data.”
Insilico Medicine, a Baltimore firm that uses artificial intelligence (AI) to discover new drugs, yesterday announced proof-of-concept results of an improved neural network methodology for uncovering anticancer molecules.
The work focuses on the generative adversarial network (GAN) class of neural network architectures, which are increasingly used in drug and biomarker discovery. Unsupervised neural networks are good at finding inferences from smaller data sets and don’t require items to be classified and labelled, but they are often less accurate and efficient than supervised networks. GANs, which were first introduced in 2014, can improve those attributes, according to the company.
The new Insilico study homes in on the advantages of using adversarial autoencoders (AAE) over variational autoencoders (VAE). They call the AAE model they developed druGAN.
“We developed an advanced AAE model that could significantly enhance the capacity and efficiency of development of the new molecules with specific anticancer properties using the deep generative models,” Insilico Chief Operating Officer Qingsong Zhu told Healthcare Analytics News™.
Zhu said these evolving models could significantly change the pharmaceutical industry, saving a lot of time and money.
“After we generate new molecules, we can focus on these few molecules instead of screening from hundreds and thousands compound as the traditional drug discovery method,” he said.
Insilico, based in Johns Hopkins University’s Emerging Technology Centers, has been on the move. In 2017, it has announced 2 funding rounds totaling just under $14 million, formed a drug-development partnership with GlaxoSmithKline, and launched an ALS biomarker and drug discovery platform called ALS.AI.
“Insilico Medicine has a policy of publishing the proof of concept research, which is one year or older to attract more data scientists to work on the healthcare problems. DruGAN is one of these proofs of concept,” company founder and CEO Alex Zhavoronkov said in a statement.
The new study focused on anticancer molecules, Zhu said, because of the wide public availability of cancer genomic data in other studies. He said the core principles are widely applicable, though.
“In theory, the methodology could apply to any medical conditions, as long as we could have enough training data,” he said.
The study, "druGAN: An Advanced Generative Adversarial Autoencoder Model for de Novo Generation of New Molecules with Desired Molecular Properties in Silico,” was published recently in Molecular Pharmaceutics.