As the 37th Annual J.P. Morgan (JPM) Healthcare Conference was drawing to a close, a research team led by investigators from the NIH and Global Good generated considerable buzz with the disclosure of a computer algorithm that can analyze digital images of a woman's cervix and accurately identify precancerous changes that require medical attention. Researchers said the artificial intelligence (AI) approach, dubbed automated visual evaluation, offered the potential to transform cervical cancer screening, particularly in low-resource settings.

To develop the method, researchers used more than 60,000 cervical images from a National Cancer Institute archive to train a machine-learning algorithm to recognize patterns in complex visual inputs. The photos were collected during a cervical cancer screening study that was carried out in Costa Rica in the 1990s in which more than 9,400 women participated, with follow-up that lasted up to 18 years. Because of the prospective nature of the study, the researchers gained nearly complete information on which cervical changes became precancers and which did not. The photos were digitized and then used to train the deep learning algorithm so that it could distinguish cervical conditions requiring treatment from those that did not.

Remarkably, the algorithm performed better than standard screening tests at predicting all cases diagnosed during the Costa Rica study. Findings from the collaboration were confirmed independently by experts at the National Library of Medicine and published this month in the Journal of the National Cancer Institute.

The study highlighted just one facet of AI that's rapidly emerging across the life sciences and, specifically, in drug discovery and development, according to Christopher McKenna, global head of professional services and consulting at Clarivate Analytics, parent of BioWorld. McKenna co-authored with Steve Arlington, president of the not-for-profit Pistoia Alliance, The Life Sciences Innovation Report, which examines key trends that are expected to emerge across the life sciences this year.

Over the next few years, McKenna said he expects the big moves between biopharma and AI to occur in applications for data aggregation and automation, codification of human rules and text mining "to make these processes more streamlined and less labor-intensive."

As the industry digests those lessons, "we'll build upon these learnings to better understand biology and start to move to clinical trial computer simulations," McKenna told BioWorld Insight. "We'll get to a point where clinical trials done in humans will be a validation of clinical trial computer simulations."

JPM, Biotech Showcase herald courtship of biopharma, AI

Movement in that direction is occurring rapidly. Timing of the NIH findings, though coincidental, was telling in the wake of JPM and the concurrent Biotech Showcase in San Francisco, both of which featured AI panels and news that seemed to herald the official courtship, if not actual marriage, of biopharma and AI. Consider some of the alliances:

AI specialist Exscientia Ltd., of Oxford, U.K., drew Roche Holding AG to the partnering table in a $68.3 million drug discovery deal as it added $26 million in a series B financing. Exscientia will apply its automated AI drug design platform to generate preclinical candidates against a target that Roche validated. (See BioWorld, Jan. 8, 2019.)

Roche, of Basel Switzerland, also partnered with Helsinki, Finland-based Kaiku Health to combine digital interventions with administered drugs to improve patient care. Kaiku, which is digitizing patient monitoring, uses its algorithms, along with predictive analytics and real-world data mining, to track symptoms and alert a patient's care team to provide personalized support. The Roche collaboration, which the companies called the first of its kind, will include pilot support modules for immune checkpoint inhibitor monotherapy and combination therapy that could also ease the adoption of patient-reported outcomes in oncology.

Recursion Pharmaceuticals Inc., of Salt Lake City, said partner Takeda Pharmaceutical Co. Ltd., of Osaka, Japan, exercised its option from a 2017 alliance for drug candidates in two rare diseases and extended their drug discovery collaboration encompassing AI, experimental biology and automation. The moves came after Takeda evaluated more than 60 unique indications with Recursion's AI-powered drug discovery platform and was especially timely, given Takeda's intent to focus on rare disease indications following the close of its $62 billion takeover of Shire plc. Recursion also disclosed during JPM that it expanded its discovery-stage pipeline by licensing from the Ohio State Innovation Fund exclusive global rights to develop and commercialize REC-2282 (previously OSU-HDAC42), a histone deacetylase inhibitor targeting the rare tumor syndrome neurofibromatosis type 2. The compound was independently identified by Recursion's AI-based discovery platform. (See BioWorld, Oct. 4, 2017, and Jan. 11, 2019.)

Privately held Numerate Inc., of San Francisco, and Lundbeck Pharmaceuticals, a unit of Denmark's H. Lundbeck A/S, entered a multitarget research collaboration to identify candidates for the treatment of central nervous system (CNS) disorders, including depression, psychosis, seizure and neurodegenerative disorders. The pact is combining Lundbeck's expertise in CNS R&D with Numerate's AI-driven D4 drug discovery platform, which also has attracted the attention of Takeda, Amgen Inc. and academic partners.

Sophia Genetics SA, of Lausanne, Switzerland, landed a $77 million round led by Generation Investment Management to extend the reach of its Sophia AI platform into additional hospitals, where its genetic sequencing and digital analysis tools are applied to analyze genomic and radiomic data, with the goal of improving the diagnosis and treatment of individuals with cancer and hereditary disorders.

Elevian Inc., of San Francisco, and Insilico Medicine Inc., of Rockville, Md., inked a collaboration to develop oral medications for diseases of aging targeting the GDF11 pathway and associated targets, with Insilico using its AI platform to analyze biological and structural target data from Elevian.

GE Healthcare and Vanderbilt University Medical Center revealed a five-year partnership to develop multiple diagnostic tools that would improve the predictability both of the efficacy of a cancer immunotherapy treatment and its adverse effects on a specific patient, potentially enabling clinicians to target hot immuno-oncology drugs to the right patients while minimizing potentially damaging, ineffective and costly courses of treatment.

Heart-on-a-chip tissue model company Tara Biosystems Inc., of New York, and AI veteran Insilico partnered to discover and develop therapies targeting cardiac disease and diseases associated with cardiac muscle aging. The collaboration will seek to shave years off discovery and preclinical development by combining Insilico's rapid discovery capabilities and Tara's human-relevant tissue models.

Atomwise Inc., of San Francisco, allied with contract research organization Charles River Laboratories International Inc., of Wilmington, Mass., to offer end-to-end drug discovery solution that incorporates deep learning AI technology, with Atomwise supporting hit discovery, hit-to-lead and lead optimization efforts with its screening platform.

Attention from biopharma unlikely to wane soon – nor is funding

The growing thread between biopharma and AI wasn't limited to JPM. Last year, as it sought to identify linkages among specific genes, diseases and drug sensitivity in advancing its multiple sclerosis (MS) pipeline, Sanofi SA enlisted Linguamatics Ltd. for a biomarker project that would enable the Paris-based pharma to explore a broad and comprehensive knowledge base – in particular, the association of HLA alleles and haplotypes with diseases and drug sensitivity. The Linguamatics AI-based natural language processing (NLP) text-mining software, used to process a collection of literature sources, identified all 22 previously published autoimmune diseases and drug sensitivities associated with HLA alleles and haplotypes and uncovered 33 additional unpublished disease and drug sensitivity associations, more than doubling previously known associations. The curated annotations were fed into a searchable knowledge base for broad use by the Sanofi team, which also is using NLP and text analytics for target identification and prioritization, drug repurposing, interpretation of genes/proteins identified by 'omics experiments and full patent text mining for new targets. Beyond R&D, Sanofi has applied text mining along its pipeline in clinical trial site selection and study design, opportunity scouting, pharmacovigilance, competitive intelligence and customer and social media analysis.

In the run-up to JPM, Australian researchers reported at AusBiotech about harnessing AI to diagnose medical conditions more quickly than ever before. Brisbane's Translational Research Institute is working with Siemens Healthineers at Draper Laboratories using magnetic resonance spectroscopy to learn more about the chemical content of tissues and organs and provide a deeper understanding and earlier detection of conditions ranging from breast cancer to post-traumatic stress disorder, while an AI human sequencing application spun out of Australia's QIMR Berghofer Medical Research Institute into startup Genomiqa is developing clinical assays for cancer patients and diagnostics labs based on data analysis from whole genome sequencing conducted in a matter of days. (See BioWorld, Dec. 28, 2018.)

Also in December, Auron Therapeutics Inc., of Boston, and Elucidata Corp., of Cambridge, Mass., formed a four-year scientific collaboration using an AI-based target discovery platform to identify and validate targets for differentiation-based therapy for acute myeloid leukemia and eight other oncology indications. Elucidata will use its data analytics platform, Polly, to analyze transcriptomic, metabolomics and epigenetic data from biological samples, as well as disease and treatment response data.

As other AI-driven initiatives have made clear, attention from biopharma is unlikely to wane soon – nor is funding. (See BioWorld, April 20, 2018, and Nov. 27, 2018.)

Analysis of 15,000 highly cited papers on AI in The Life Sciences Innovation Report showed some of the hottest areas of development included data classification or pattern recognition, NLP or knowledge representation, artificial life and machine learning, artificial neural nets and fuzzy logic.

AI 'still limited to what we train it to do'

But as relationships bloom, it's important to remember that the science of AI continues to evolve rapidly, which could make early partnerships a game of chance. IBM Watson Health ditched its division head last year in the wake of employee layoffs and corporate restructuring, and IBM CEO Ginni Rometty used her platform as the opening keynoter at the Consumer Electronics Show in Las Vegas this month to plead for the platform's relevancy.

And AI isn't just about the data, which is the measuring stick for many conventional biopharma partnerships. Results also depend on variables such as the way data are fed into a system and the computing time needed to process a dataset. As research scientist Janelle Shane showed in an infamous 2017 neural network experiment on the seemingly simple task of generating and naming paint colors, AI didn't exactly improve on the status quo, churning out mostly browns, beiges and grays and coining names such as Clardic Fug, Stanky Bean and Turdly.

"In the age of humanization of AI, where programs like Watson, Alexa, Siri exist, it's easy to fall into the belief that this AI is able to perform at the level of human intelligence," McKenna said. "However, AI is still limited to what we train it to do. There are many situations in health care where humans don't know first principles, like the etiology of a patient's disease given the reported clinical symptoms. AI cannot be used rigorously to do a task that humans themselves haven't mastered."

Still, the algorithms of AI have been around for more than five decades, McKenna pointed out, and AI systems can be taught to improve. He cited the example of AlphaGo, the computer program written to play the board game Go. Initially, the program could analyze moves better than human contestants but couldn't outpredict its flesh and blood competitors in strategy. But when AlphaGo's developers combined its Monte Carlo search algorithm with neural learning – essentially forcing the computer to play itself over and over again until it learned all of the moves – in 2016 the computer defeated a professional Go player. More advanced versions of AlphaGo were subsequently created and, today, "the AlphaGo players just sit back and watch the program play," McKenna said.

Of course, "we're not there yet with biology," which involves exponentially more dimensions than paint colors or computer games, he quickly added. "But we are continuing to move down the path of understanding biology better and building computational biology into a core function."

AI evolving into hybrid models

Although the early days of AI in biopharma applications are focused on pharmacovigilance and histology image analysis, the evolution of hybrid computing – machine learning combined with text mining or deep learning plus Monte Carlo simulation, for instance – increasingly will be used to tackle problems that were previously inscrutable, McKenna suggested.

"With the speed of computing today and the ability to mix and match algorithms, AI is evolving into these hybrid models," he said. "We're seeing patterns of mathematical algorithms that have existed for decades and bringing them together to solve problems that haven't been tackled before."

At Clarivate, AI is used for indication prioritization, target and biomarker identification, target druggability prediction, drug repurposing, patient stratification, pharmacovigilance monitoring and drug timeline and success rate prediction, McKenna pointed out.

Nevertheless, "AI is still limited in pharma applications, addressing only discrete tasks rather than operating at the level of human intelligence," he said, its benefit measured mostly in providing additional insight in interpreting large datasets for a single type of question or decreasing human effort in collecting, assembling and providing initial review for decisions and actions taken by humans.

Over the next decade, "following Moore's Law," supercomputing levels of AI will become available to the pharmaceutical industry, McKenna predicted.

"I would expect in this decade that AI will be able to choose the best drug target, predict its efficacy and safety in different patient groups, run millions of clinical trial simulations, prepare all the dossiers needed and predict with high accuracy each state of development up to submission and approval," he said. "As this advances and our confidence improves to the level of regulatory reform, entire phases of development will be able to be skipped entirely, greatly improving the speed and ROI of drug development."

Editor's note: The Life Sciences Innovation Report may be downloaded at