Predicting costly pharma flops is AI-aided forecaster’s goal

FRANKFURT • Before Biogen Inc’s experimental Alzheimer’s disease drug failed in a trial last month, analysts surveyed by Bloomberg estimated annual sales of about US$3.7 billion (RM15.21 billion) by 2023. A German artificial-intelligence (AI) firm had come to a different conclusion.

Innoplexus AG, a closely held company based in the outskirts of Frankfurt, uses an algorithm to analyse pharma companies’ drug pipelines that it says take more data and context into account than any other tool. Its assessment of Biogen’s aducanumab gave about a 70% to 90% chance that the trial would miss its goal.

“The system appears to do well in predicting outcomes of clinical trials,” said Hendrik Leber, an MD at fund manager Acatis Investment GmbH who’s evaluating the programme, which has yet to be used for making trades.

Taking the guesswork out of new drug studies would be a pharma investor’s dream. Bringing a single product to market is estimated to cost about US$2.6 billion, and when anticipated blockbusters fail late in development, drugmakers and their shareholders both feel the pain. After aducanumab was halted, Biogen lost US$18 billion in value in one day.

Many analysts also had doubts about the drug, even without AI-driven insights. The plunge in the stock, while significant, indicated that investors had priced in 50% odds for the trial’s success, according to Asthika Goonewardene, an analyst with Bloomberg Intelligence. Concerns were high because of earlier disappointments with similar compounds, and analysts pressed Biogen’s management about updates as recently as its January earnings call.

Innoplexus’ learning software can crawl through as many as five billion webpages a day to assess probabilities for the outcome of clinical trials. The system builds a giant network of relationships between data, using a natural-language processing algorithm trained on medical research to determine whether certain approaches show promise.

The information is fed, along with some 350 other data points, into an AI algorithm that tries to predict a particular drug candidate’s likelihood of success. These other data points include individual hospitals’ track records in conducting trials and how well certain drug components are absorbed by humans, according to Innoplexus founder Gunjan Bhardwaj.

The approach allows the system to assess the scientific risk of a trial’s failure — because it’s the wrong drug or targets the wrong physiological process — as well as the possibility that a particular group of researchers won’t be able to conduct the trial successfully because of low patient recruitment or other factors, said Bhardwaj, a graduate of the Indian Institute of Technology in Mumbai who has worked at Boston Consulting Group and the EY management consulting firm.

AI is far from a magic investment bullet. The Eurekahedge AI Hedge Fund Index, which tracks money pools that utilise AI as part of their core strategies, returned just over 13% in three years through the end of 2018, trailing the 30% gain in the S&P 500 Index with reinvested dividends.

Still, Innoplexus’ own tests on 20,000 completed clinical trials found — albeit retrospectively — that it correctly forecast the outcome in about 85% of cases, including an important Bristol-Myers Squibb Co lung cancer trial’s failure and a successful Johnson & Johnson myeloma treatment study.

Like many similar programmes, Innoplexus’ is something of a black box even to its creators, but they determined a few variables that tilted its software toward a negative view of aducanumab’s chances. For example, track records of some hospitals conducting the trial suggested failure, the company said.

Pipelines are also tricky to evaluate internally, said Klaus Ort, a partner at EY in Germany. Ort said he doesn’t know of any pharma company that has a tool like Innoplexus’, and EY plans to use it to analyse target companies and advise clients on strategic options for their own pipelines.

Innoplexus is bolstering its data set by encouraging researchers to share findings that are unpublished or awaiting publication. These may include negative results, which sometimes don’t see the light of day, but can give important clues to a drug’s chance for success. — Bloomberg