Project

Back to overview

Deep learning for high throughput cryo-electron tomography

English title Deep learning for high throughput cryo-electron tomography
Applicant Castano-Diez Daniel
Number 179041
Funding scheme Project funding (Div. I-III)
Research institution C-CINA Biozentrum Universität Basel
Institution of higher education University of Basel - BS
Main discipline Mathematics
Start/End 01.08.2018 - 31.07.2022
Approved amount 536'667.00
Show all

All Disciplines (2)

Discipline
Mathematics
Radiobiology

Keywords (5)

deep learning; high throughput microscopy; cryo electron tomography; gpu computing; segmentation

Lay Summary (German)

Lead
Kryo-Elektronentomographie erlaubt es, mehrere Proteine in ihrem zellulären Kontext zu visualisieren und liefert so entscheidende Hinweise für die Beschreibung ihrer Funktion.Das Aufkommen neuer hochentwickelter Instrumente hat die automatisierte Erfassung tomographischer Datensätze in einem noch nie dagewesenen Umfang ermöglicht. Während diese gewaltige Zunahme an Informationen die Möglichkeit eröffnet, einen noch tieferen Einblick in die Zellstruktur und -funktion zu gewinnen, bringt sie gleichzeitig auch die heutigen Werkzeuge der Datenanalyse an ihre Grenzen.
Lay summary

Inhalt und Ziel des Forschungsprojekts

 

Unser Vorgehen baut auf moderner Methoden der rechnerbasierten Interpretation von Bildern (Computer Vision) auf, womit spezifische Merkmale in den Rohdaten automatisch erkannt und interpretiert werden können.  Langfristiges Ziel ist es, den Einsatz der Künstlichen Intelligenz (insbesondere der so genannten Deep Learning Methode) zu erforschen, um die Segmentierung der Inhalte von Tomogrammsätzen, d.h. die Identifizierung und Lokalisierung der verschiedenen morphologischen Elemente in der Zelllandschaft, wie z.B. Membranen, Filamente oder charakteristische Makrokomplexe, zu automatisieren.

 

 

Wissenschaftlicher und sozialer Kontext des Forschungsprojekts

 

Mit der durch moderner Kryo-Elektronenmikroskopie erreichbaren hohen Auflösung wird es für Forschungsgruppen zunehmend möglich, sich mit Fragen zu Proteinen mit unmittelbarer pharmakologischer Relevanz zu befassen. Die in unserem Projekt zu entwickelnden Tools zielen auf die Charakterisierung solcher Proteine unter realistischen, physiologischen Bedingungen ab und tragen so dazu bei, die Lücke zwischen Grundlagenforschung und medizinischen Anwendungen zu schließen.

Direct link to Lay Summary Last update: 29.05.2018

Responsible applicant and co-applicants

Employees

Project partner

Abstract

Cryo electron tomography (cryoET) is profiling itself as the method of choice for the study of biological compounds in their cellular context. Recent developments on the instrumentation side have made possible the collection of large amounts of high quality data, bringing up the necessity of establishing efficient frameworks for data analysis that can operate with none to minimal human supervision. Ideal efficiency would be marked by the ability to perform an online analysis of the raw data produced by the microscope.This proposal devises the development and posterior integration of distinct numerical tasks that jointly cover the full computational pipeline that links the raw data gained in a session of tomographic data collection (i.e., a tilt series of micrographs of the sample imaged at different orientations at the electron microscope) to the final results in the form of averaged models of the macromolecules of interest, possibly classified in different functional states, or as automated annotations of the tomogram contents. The final aim of establishing a high-throughput analysis sets the goal of achieving a full and unsupervised completion of this workflow in the order of magnitude of the minutes, matching the time needed by current instrumentation to collect a full tilt series.Already at the level of selection of recording sites on the carbon grid that supports the sample during the imaging session, we intend to refine the current, in-house developed image processing system for computerized selection of viable holes (recording sites) through a deep learning system that learns and automatically applies the selection patterns preferred by experienced human users. After selection of the holes, the next automation requirement is the mutual alignment of the micrographs representing different views. The approach suggested in this approach combines the two main procedural families (location of gold bead fiducials and alignment of micrographs to global tomogram reprojections), joining them through the introduction of an initial model computed by a genetic algorithm type on the central tilts, typically the ones with the highest signal content. This initial model allows for a gradual integration of micrographs of higher tilts through a mutual validation of global reprojection and individual gold bead identifications.The final automation stage concerns the interpretation of the tomogram content, both regarding the off-line discovery of hidden structures, and the on-line identification of copies of compounds of interest, which might appear -depending of the case- in a wide range of supporting geometries defined by the general architecture of the tomogram contents.Algorithmic development will be closely followed by integration in our well established, in-house package Dynamo, ensuring both the availability of an extensive codebase for efficient development of scientific software, and an enormous visibility of the output of this project, due to the widespread used of our software suite.
-