Use deep learning for precision cancer therapy

Nearly 50 new cancer therapies are approved each year. This is good news. “However, it is becoming increasingly difficult for patients and their therapists to maintain tracking and select treatments that have very individual tumor characteristics for the affected individual,” says Altuna Akalin, PhD, head of the Bioinformatics and OMICS Data Science Technology Platform for Berlin Biology, Medical Biologies at MAX Delbrück-Bimsb.

Researchers have been working for some time to develop tools that use artificial intelligence to make more accurate diagnosis and determine the best treatment for individual patients.

Akalin's team has now developed a toolkit called FlexyNesis. This does not rely solely on classical machine learning, but uses deep learning to simultaneously evaluate very different types of data.

“This way, physicians can improve their patients' diagnosis, prognosis and treatment strategies,” says Akalin. Flexynsis is Natural Communication.

“We are running multiple translation projects with physicians who want to identify biomarkers from multiomic data that are consistent with disease outcomes,” says Dr. Borauyal, the publication's first and co-author.

“Many deep learning-based methods have been published for this purpose, but most people have found that they are inflexible, tied to specific modeling tasks or difficult to install and reuse.

This tool finds the root of disease

Deep learning is a subfield of machine learning beyond simple neural networks with one or two computational layers, instead using deep networks that operate on hundreds or thousands of layers. “Cancer and other complex diseases arise from the interaction of a variety of biological factors, such as DNA, RNA, and protein levels,” explains Akalin.

Characteristic changes at these levels are often recorded, such as the amount of HER2 protein produced in breast and stomach cancer, but are not usually analyzed in conjunction with all other treatment-related factors.

This has often been difficult to use so far or can only be useful in answering specific questions,” says Akalin. “In contrast, flexynsis can answer a variety of medical questions at the same time: for example, which type of cancer is involved, drugs that are particularly effective in this case, and how these affect a patient's chances of survival.”

This tool can also help identify suitable biomarkers for diagnosis and prognosis. In other words, to identify primary tumors, metastases of unknown origin are discovered if they are found. “This will facilitate the development of comprehensive and personalized treatment strategies for all types of cancer patients,” says Akalin.

Data integration in the clinic – even if you have no experience with AI

Last year, Akalin introduced another AI-based tool called OnConaut. This will help identify the appropriate cancer therapy as well. “Onconaut relies on known biomarkers, clinical trial results, and current guidelines, so it works on a completely different principle,” explains Akalin. “The tool never gets obsolete, but it actually helps complement your flexibility.”

One hurdle that new tools still have to overcome, at least in Germany, is the fact that multi-omics data is not yet collected on a daily basis in hospitals. “In the US, on the other hand, this data is frequently discussed in hospital tumor boards, where various specialist physicians collaborately plan to treat patients,” says Akalin.

And his team showed that data could be used to accurately predict whether a particular treatment would be effective. “In Germany, so far, detailed multi-omics data has been used only in flagship programs such as the Master Program for Rare Cancer,” he adds. But that may change soon.

Akalin is currently primarily aimed at physicians and clinical researchers, and emphasizes that users of his continuous updates do not need to have a special background in the manipulation of deep learning.

“We hope that hospitals and research groups will lower barriers to implement multimodal data integration and perform simultaneous analysis of OMICS data, written reports and images, even without AI experts,” he says. FlexyNesis is easily accessible online, along with instructions for using the tool.