AI develops cancer drug in 30 days – and predicts survival rates
Artificial intelligence has developed a treatment for an aggressive form of cancer in just 30 days and demonstrated it can predict a patient’s survival rate using doctors’ notes.
The breakthroughs were performed by separate systems, but show how the powerful technology’s uses go far beyond the generation of images and text.
University of Toronto researchers worked with Insilico Medicine to develop potential treatment for hepatocellular carcinoma (HCC) using an AI drug discovery platform called Pharma.
HCC is a form of liver cancer, but the AI discovered a previously unknown treatment pathway and designed a ‘novel hit molecule’ that could bind to that target.
The system, which can also predict survival rate, is the invention of scientists from the University of British Columbia and B.C. Cancer, who found the model is 80 percent accurate.
AI developed the cancer treatment (stock) in just 30 days from target selection and after synthesizing only seven compounds
AI is becoming the new weapon against deadly diseases, as the technology is capable of analyzing vast amounts of data, uncovering patterns and relationships and predicting effects of treatments.
Insilico Medicine founder and CEO Alex Zhavoronkov said in a statement: ‘While the world was fascinated with advances in generative AI in art and language, our generative AI algorithms managed to design potent inhibitors of a target with an AlphaFold-derived structure.’
The team used AlphaFold, an artificial intelligence (AI)-powered protein structure database, to design and synthesize a potential drug to treat hepatocellular carcinoma (HCC), the most common type of primary liver cancer.
The feat was accomplished in just 30 days from target selection and after only synthesizing seven compounds.
In a second round of AI-powered compound generation, researchers discovered a more potent hit molecule – although any potential drug would still need to undergo clinical trials.
Feng Ren, chief scientific officer and co-CEO of Insilico Medicine, said: ‘AlphaFold broke new scientific ground in predicting the structure of all proteins in the human body.
‘At Insilico Medicine, we saw that as an incredible opportunity to take these structures and apply them to our end-to-end AI platform in order to generate novel therapeutics to tackle diseases with high unmet need. This paper is an important first step in that direction.’
Another AI system identified characteristics unique to each patient, predicting six month, 36 months and 60 months survival with greater than 80 percent accuracy
The system used to predict life expectancy used natural language processing (NLP)—a branch of AI that understands complex human language—to analyze oncologist notes following a patient’s initial consultation visit.
The model identified characteristics unique to each patient, predicting six month, 36 months and 60 months survival with greater than 80 percent accuracy.
John-Jose Nunez, a psychiatrist and clinical research fellow with the UBC Mood Disorders Centre and B.C. Cancer, said in a statement: ‘The A.I. essentially reads the consultation document like a human would read it.
‘These documents have many details like the patient’s age, the type of cancer, underlying health conditions, past substance use, and family histories.
‘The AI combines all of this to paint a complete picture of patient outcomes.’
Traditionally, cancer survival rates have been calculated retrospectively and categorized by only a few generic factors, such as cancer site and tissue type.
The model, however, is able to pick up on unique clues within a patient’s initial consultation document to provide a more nuanced assessment.
The AI was trained and tested using data from 47,625 patients across all six B.C. Cancer sites located across British Columbia.
‘Because the model is trained on B.C. data, that makes it a potentially powerful tool for predicting cancer survival in the province,’ said Nunez.
‘[But] the great thing about neural NLP models is that they are highly scalable, portable and don’t require structured data sets. We can quickly train these models using local data to improve performance in a new region.’