AI enables high-throughput profiling of cancer metastases

AI allows profiling of cancer cells

Date: 16th December 2019

For most scientists finding a cure for a given cancer represents the ultimate hope and, whilst we continue to make significant inroads into diagnosing and treating cancer, it still remains one of the leading causes of death worldwide.  Last year, 17 million new cancer cases were diagnosed, with 9.6 millions deaths worldwide.

Furthermore, with 90% of cancer patients dying from distal metastases rather than as a direct result of the primary tumour, assessing the impact of treatment on these cells is critical.

However, whilst traditional methods such as Magnetic Resonance Imaging (MRI) or bioluminescence can detect the primary cancer and larger metastases they lack the resolution required to detect distal metastases, which typically exist as smaller groups of isolated cells that remain elusive and difficult to locate.

Now scientists from Helmholtz Zentrum München, Ludwig Maximilian University of Munich and the Technical University of Munich, Germany, have developed a deep learning algorithm that enables the automated detection of metastases at the level of single cancer cells in whole mice.

The team had previously developed vDISCO.  A pressure-driven, nanobody-based whole-body immunolabeling technology method which enhanced the signal of fluorescent proteins. When applied to mouse bodies that had undergone tissue clearing and fixation it transformed them into a transparent state allowing, in this case, the imaging of whole-body neuronal projections.

By applying vDISCO in this recent study, the team enhanced the fluorescent signal of cancer cells by more than 100-fold and were successfully able to detect the smallest metastases down to individual cancer cells in the cleared tissue of mouse bodies using laser-scanning microscopes.

Where the authors hit a road block, however, was in the manual analysis of the high-resolution imaging data that resulted from the experiment.  Hugely time-consuming and limited by the capability of existing algorithms the team went on to design a bespoke deep-learning based algorithm called DeepMACT (Deep learning based Metastasis Analysis in Cleared Tissue).

DeepMACT was able to automate quantification of metastases down to a single cell and, whilst its accuracy matched human expert manual annotation, it did so 300 times faster.

This meant that the algorithm was able to do the manual detection work, which would typically take months, in less than an hour, enabling the researchers to conduct high-throughput metastasis analysis routinely.

Clinical translation

Whilst the comprehensive detection of cancer cells within the entire body sounds very promising, it is difficult to see how this may translate into an effective, diagnostic or clinical application as samples need to be cleared and fixed.

The authors, however, see this as an enabling, preclinical tool for drug discovery in the characterisation of the types and distribution of metastases from different types of cancers.

By understanding how the primary tumour can metastasise, and the unique way each cancer does so, it potentially enables drug targeting to be tailored more effectively.  Furthermore, by building up a picture of each cancer type and how the disease progresses with time it may also provide opportunities for eliminating or halting progression. Indeed, the team have already gained new insights into the unique metastatic profiles of several different tumours.

An additional exciting application for DeepMACT is its use in mapping the distribution of large molecule therapies such as tumour-specific monoclonal antibodies to determine which metastases are targeted successfully.

The study therefore also interrogated the biodistribution of a therapeutic antibody named 6A10, which had been previously shown to reduce tumour growth.  Interestingly, 23% of the metastases in the bodies of affected mice were shown to be missed by the antibody, suggesting that a subset of the cancer cells had evaded treatment in these instances.

With the success rate of clinical trials in oncology at around 5% it is hoped that this technology or similar could be implemented at the pre-clinical stage to reduce drug attrition rates or improve efficacies.  Using it to determine which drugs can effectively target metastasised cells and thus selecting these for further testing and progression would potentially improve the drug development process by supporting the discovery of more powerful drug candidates for clinical trials.


For more information please read the press release from Helmholtz Zentrum München.


Pan, C., O. Schoppe, A. Parra-Damas, R. Cai, M. I. Todorov, G. Gondi, B. von Neubeck, N. Böğürcü-Seidel, S. Seidel, K. Sleiman, C. Veltkamp, B. Förstera, H. Mai, Z. Rong, O. Trompak, A. Ghasemigharagoz, M. A. Reimer, A. M. Cuesta, J. Coronel, I. Jeremias, D. Saur, A. Acker-Palmer, T. Acker, B. K. Garvalov, B. Menze, R. Zeidler and A. Ertürk (2019). “Deep Learning Reveals Cancer Metastasis and Therapeutic Antibody Targeting in the Entire Body.” Cell 179(7): 1661-1676.e1619.