Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: CIHR Strategic Master's Award - Bioinformatics, Comput. Biology/Health Data Sci.

University of British Columbia

$27,000.00

PerturbKAN: Interpretable and Informed Machine Learning Model for Gene Perturbation using Kolmogorov-Arnold Networks

EF 2025-26

Résumé
Background Rapid advances in single-cell technologies have allowed the creation of virtual cell models to simulate and predict cell behaviour. However, the current deep learning approaches have two critical flaws: they are "black boxes" - complex, non-linear architectures which give no insight into reasoning of their predictions, and they are data-centric, where they do not understand biological context and fail to accurately predict effects of gene perturbations that are not seen in the training data (out-of-distribution errors). This lack of transparency and generalization limits their utility in high-stakes fields like drug discovery. Objective and Methodology I propose using Kolmogorov-Arnold Networks (KANs) to develop an interpretable and biologically informed virtual cell model. KANs learn explicit and non-linear activation functions which we can visualize and ascribe meaning to rather than the opaque weights of traditional neural networks. I will also ground my model using prior biological knowledge graphs to encourage it to make biologically plausible predictions. We will validate this approach with a suite of datasets and metrics to benchmark its ability to generalize to novel perturbation scenarios. Significance Leveraging biologically informed and interpretable machine learning, we will develop the first transparent virtual cell model to enable trustworthy in silico experiments. Our emphasis on interpretability represents a shift towards fully mechanistic virtual cell models. We envision that the interpretable virtual cell could act as a virtual laboratory and generate highly confident predictions to guide targeted wet-lab experiments, leading to future breakthroughs in biomedical research, personalized medicine, drug discovery, and programmable biology.
Programme
CIHR Strategic Master's Award - Bioinformatics, Comput. Biology/Health Data Sci.
Emplacement
British Columbia, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
CIHR Funding Decisions Database

Tel que publié

Id
562721
Name
He-Mo Alex
Region
British Columbia
Country
Canada
Orgtype
Institution Paid; Research Institution
Instname2
Genetics
Theme2
Biomedical
Prcname2
Master's Award: Strategic Master's Award
Projecttitle
PerturbKAN: Interpretable and Informed Machine Learning Model for Gene Perturbation using Kolmogorov-Arnold Networks
Programtype2
Studentship Programs
Orgnameinp2
University of British Columbia
Cihramount2
$27,000
Cihrequipment2
$0
Programname2
CIHR Strategic Master's Award - Bioinformatics, Comput. Biology/Health Data Sci.
Keyworddelim
Computational Biology; Deep Learning; Explainable Artificial Intelligence; Gene Perturbation; Gene Regulatory Networks; Genomics And Transcriptomics; Single-Cell Rna Sequencing; Systems Biology
Abstract
Background Rapid advances in single-cell technologies have allowed the creation of virtual cell models to simulate and predict cell behaviour. However, the current deep learning approaches have two critical flaws: they are "black boxes" - complex, non-linear architectures which give no insight into reasoning of their predictions, and they are data-centric, where they do not understand biological context and fail to accurately predict effects of gene perturbations that are not seen in the training data (out-of-distribution errors). This lack of transparency and generalization limits their utility in high-stakes fields like drug discovery. Objective and Methodology I propose using Kolmogorov-Arnold Networks (KANs) to develop an interpretable and biologically informed virtual cell model. KANs learn explicit and non-linear activation functions which we can visualize and ascribe meaning to rather than the opaque weights of traditional neural networks. I will also ground my model using prior biological knowledge graphs to encourage it to make biologically plausible predictions. We will validate this approach with a suite of datasets and metrics to benchmark its ability to generalize to novel perturbation scenarios. Significance Leveraging biologically informed and interpretable machine learning, we will develop the first transparent virtual cell model to enable trustworthy in silico experiments. Our emphasis on interpretability represents a shift towards fully mechanistic virtual cell models. We envision that the interpretable virtual cell could act as a virtual laboratory and generate highly confident predictions to guide targeted wet-lab experiments, leading to future breakthroughs in biomedical research, personalized medicine, drug discovery, and programmable biology.
Competitiondate
202601
Approvedterm2
1 yr 0 mth
Pinamesdelim
He-Mo, Alex
Approvedterm
1 an 0 mois
Cihrcontribution2
~ $27,000~ $0~