End-to-end trainable deep tensor completion algorithms with multimodal data fusion for anticancer drug synergy prediction

Simple item page
dc.contributor.authorZammit, Judah
dc.contributor.examiningcommitteeQian, Yiming (Amazon)
dc.contributor.examiningcommitteeTremblay-Savard, Olivier (Computer Science)
dc.contributor.supervisorHu, Pingzhao
dc.contributor.supervisorLeung, Carson
dc.date.accessioned2023-09-07T14:03:34Z
dc.date.available2023-09-07T14:03:34Z
dc.date.issued2023-08-23
dc.date.submitted2023-08-23T22:34:06Zen_US
dc.date.submitted2023-09-06T20:38:00Zen_US
dc.degree.disciplineComputer Scienceen_US
dc.degree.levelMaster of Science (M.Sc.)
dc.description.abstractComputational algorithms for anticancer drug synergy prediction (ACDSP) have the potential to discover novel anticancer drug combinations. Most of the algorithms are based on deep learning (DL) and use chemical features of anticancer drugs and multiomic features of cancer cell lines. Deep tensor completion (DTC)-based ACDSP algorithms achieve similar performance to DL-based ACDSP algorithms without using chemical or multiomic features. However, current DTC-based ACDSP algorithms are trained in two separate steps, potentially limiting their performance. In this M.Sc. thesis, I develop a novel DTC-based ACDSP algorithm, named Deep Selective CANDECOMP/PARAFAC (DSCP), which can be trained in an end-to-end fashion. It was found that DSCP made significantly better performance on a synthetic tensor classification task, while using less resources than the non-end-to-end trainable version of DSCP. Furthermore, I also propose a new ACDSP algorithm that can utilize multimodal data fusion methods (e.g., intermediate fusion (IF)) to integrate tensor-based drug and cell line data with chemical and multiomic features, which is named SynergyIF. Finally, this thesis found evidence that many ACDSP algorithms do not utilize the chemical and biological structure of their input features, instead treating these features as unique identifiers. This was done by examining the effects of randomizing ACDSP algorithms’ input data on their predictive performance. Overall, although the thesis demonstrates some advantages to integrating deep learning and tensor completion algorithms in an end-to-end fashion, more work is warranted to fuse multimodal data under the learning framework.
dc.description.noteOctober 2023
dc.identifier.urihttp://hdl.handle.net/1993/37597
dc.language.isoeng
dc.rightsopen accessen_US
dc.subjectDeep Learning
dc.subjectCombination Therapy
dc.subjectTensor Completion
dc.subjectMissing Data Imputation
dc.subjectAnticancer Drug Discovery
dc.titleEnd-to-end trainable deep tensor completion algorithms with multimodal data fusion for anticancer drug synergy prediction
dc.typemaster thesisen_US
local.subject.manitobano
oaire.awardNumber5153
oaire.awardTitle2022 Research Manitoba-CancerCare Manitoba Master's Studentship Award
oaire.awardURIhttps://researchmanitoba.ca/funding/programs/masters-studentship-award/
project.funder.identifierhttp://dx.doi.org/10.13039/100008794
project.funder.nameResearch Manitoba
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
zammit_judah.pdf
Size:
2.29 MB
Format:
Adobe Portable Document Format
Description:
Download
License bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
license.txt
Size:
770 B
Format:
Item-specific license agreed to upon submission
Description:
Download
Collections
FGS - Electronic Theses and Practica