AI models detect brain tumors with high accuracy using epigenetic fingerprints

MRI images show brain tumors in creepy positions. And brain biopsies increase the risk of patients who consult us because of double vision. The situation, such as the discussion of this case, cited as an example in an interdisciplinary team of cancer medicine experts, prompted researchers at Charite-University Berlin along with cooperative partners to look for new diagnostic procedures. The result is an AI model. This model utilizes specific properties in tumor genetic material, such as epigenetic fingerprints obtained from cerebrospinal fluid, for example. As the team shows in the journal Natural Cancer,The new model classifies tumors quickly and very reliably.

Today, there are far more types of tumors known than the organs they develop. Each tumor has its own unique properties, specific tissue characteristics, growth rate, and metabolic specificity. Nevertheless, tumor types with similar molecular properties can be grouped together. Treatment of individual diseases depends critically on the type of tumor. New targeted therapies address specific structures of tumor cells or block signaling pathways to stop the growth of pathological tissue. Chemotherapy can be chosen according to the type of tumor and adjusts their dosage accordingly. In particular, for rare tumor types, it may be possible to pursue innovative therapies as part of their research.

With the increasingly personalized and rapidly developing cancer medicine, accurate diagnosis at a certified oncology center is a way to make treatment successful. ”

Professor Martin E. Clayce, Chief Medical Officer of Charite

Comprehensive molecular, cellular, and functional analysis of tumors based on tissue samples provide the necessary information, but physicians also face cases where it is impossible or extremely dangerous to extract tissue samples from tumors. Moreover, even histological examination alone cannot provide as accurate a diagnosis as newer AI models.

Examining the genome, not within the organization

Methods for characterizing brain tumors based on traditional microscopic diagnosis are based on epigenetic properties, which are modifications to the tumor's genetic material. They are part of every cell's memory and determine which part of the genetic information is being read. “Hundreds of thousands of epigenetic modifications act as switches on and off of individual gene sections. Those patterns form unique and unmistakable fingerprints,” explains Dr. Philip Euchilchen, a scientist at the Berlinsite of the German Cancer Consortium and the recently published Institute of Neuropathology in Karite. “In tumor cells, epigenetic information is altered in a distinctive way. Based on its profile, tumors can be distinguished and classified.” In the case of brain tumors, even a sample of cerebrospinal fluid is sufficient and can be obtained relatively easily – fully dispensing the surgery.

Comparing unknown fingerprints with thousands of known fingerprints of different cancers and assigning them to a particular tumor type requires machine learning methods, namely artificial intelligence, given the very broad and complexity of the data. Furthermore, different DNA sequencing methods have been applied in the past. Furthermore, epigenetic analysis is usually limited to defined patterns and gene segments typical for individual tumor types. “As a result, our aim was to develop a model that accurately classifies tumors, even if they were based solely on a portion of the entire tumor epigenome. Or the profiles were collected through a variety of techniques and different accuracy.”

Reliable and traceable

The newly developed AI model is listed under the name Crossnn, and its architecture is based on a simple neural network. This model was trained on a large number of reference tumors and subsequently tested on more than 5,000 tumors. “Our model allows for a very accurate diagnosis of brain tumors in 99.1% of all cases, and is more accurate than previous workplace AI solutions.” “In addition, we were able to train AI models in the same way that can distinguish more than 170 tumor types from all organs, achieving an accuracy of 97.8%. This means that it can be used for cancers of all organs, in addition to relatively rare brain tumors.” The critical factor for future approval in clinical applications is that the model is fully explained, i.e. it is necessary to understand how decisions will arrive.

The molecular fingerprints that the AI ​​model receives for measurements can be attributed to tissue samples or body fluids. For certain brain tumors, Charité's School of Neuropathology already offers non-invasive diagnosis based on cerebrospinal fluid, known as liquid biopsy. This allows for diagnosis to be made without stressful manipulation even in difficult situations. The patient who consulted Double Vision and us was one of the beneficiaries. “We examined cerebrospinal fluid using nanopore sequences, a novel, extremely fast and efficient form of genetic analysis. The model classification revealed that it is a central nervous system lymphoma, allowing for rapid initiation of appropriate chemotherapy.

Cross in clinical trials

The accuracy of the methodology has even surprised researchers. “The AI ​​model architecture is much simpler than previous approaches, so it remains explainable, but provides more accurate predictions and thus increases diagnostic certainty,” says Sören Lukassen. Therefore, together with the German Cancer Consortium (DKTK), the research team is planning clinical trials with CrossNN at all locations in the eight DKTKs in Germany. Additionally, intraoperative use should be tested. The aim is to transfer accurate and relatively inexpensive tumor decisions based on DNA samples into routine care.