Chemie | Biochemie | Medizin
Kenneth Arockia, 2006 | Olten, SO
The heterogeneity of cancer is one of the greatest challenges in current medical research. This study aims to tackle this issue by employing Pharmacoscopy and Convolutional Neural Networks (CNN) to establish a personalised treatment option for an Acute Myeloid Leukaemia patient. A patient-derived bone marrow aspirate was prepared and exposed to ten different drug conditions in a microplate. This sample was then stained for cancerous as well as healthy cells using principles of immunohistostaining and imaged using a high-throughput microscope.
The images were then quantified using immunofluorescence (IF) -based and CNN-based image analyses in MATLAB. Both methods demonstrated that Venetoclax had the most on-target effect on the sample through the reduction of cancer cells and preservation of healthy cells. These findings highlight the potential of Pharmacoscopy and AI-driven analyses to advance personalised treatment in Acute Myeloid Leukaemia.
Introduction
This study aims to determine the optimal drug condition for the patient by achieving the following objectives:
1. Preparing a fixed and immunostained sample suitable for high-throughput microscopy.
2. Construction, training and implementation of a CNN on microscopy images.
Methods
The frozen sample from the patient was first thawed, cleaned and distributed over 64 wells of a microplate containing ten different drug conditions and a DMSO well for reference. After fixation and blocking, the sample underwent fluorescence-based immunostaining using DAPI for nuclei and antibodies targeting cancer (CD34, CD33, CD13) and healthy (CD3) cell markers. The images from the following microscopy were then analysed using an IF-based approach that solely classifies cells based on their fluorescence and a CNN-based approach that takes other factors into account such as the structure and appearance of the cells. The programmes for the IF-based analysis as well as the construction, training and implementation of the CNN were done in MATLAB. Finally, a drug score was calculated for each drug and each approach that quantified the effectiveness of the drug. These experiments and microscopy were done at the Snijder Laboratory at ETH Zürich.
Results
Both analysis methods identified Venetoclax as the most effective drug, as it selectively reduced the malignant cell population while preserving healthy cells.
Discussion
The power of Pharmacoscopy in differentiating malignant and non-malignant cells through specific fluorescent antibody-based staining allowed the selection of Venetoclax as the recommended drug, despite the relatively low accuracy of the applied CNN due to a small training data set. This approach, combined with reliable analysis tools such as deep learning, increases the effectiveness and minimises unwanted side effects in the treatment of blood cancer patients, therefore contributing to truly personalized therapy.
Conclusions
This study showcases the value of Pharmacoscopy and AI-based approaches such as CNNs in the context of personalized therapy. Further research e.g. in the form of clinical trials is recommended to test these concepts in real world settings and to possibly transform the field of cancer treatment.
Würdigung durch den Experten
Ramon Pfändler
Mit seiner Arbeit gewährt Kenneth Arockia Einblicke in das Potenzial personalisierter Ansätze zur Behandlung akuter myeloischer Leukämie (AML). Mit der an der ETH entwickelten «Pharmacoscopy» Methode findet er durch Immunfluoreszenz-basierte und KI-gestützte Bildanalyse gezielt wirksame Medikamente. Besonders die Umsetzung und die Interpretation der Resultate des KI-Ansatzes zeugen von Reife und Begeisterung des Teilnehmers. Diese zeigen sich auch in der Diskussion, wo er die Resultate in den aktuellen Wissensstand einbettet, kritisch hinterfragt und zukünftige Nutzungsmöglichkeiten aufzeigt.
Prädikat:
sehr gut
Sonderpreis «Exporecerca Jove – Barcelona Science Fair» gestiftet von der SJf-Trägerschaft
Kantonsschule Olten
Lehrer: Peter Gutierrez