🧠 Excited to share my main PhD project! We mapped the regulatory rules governing Glioblastoma plasticity using single-cell multi-omics and deep learning. This work is part of a two-paper series with @bayraktar_lab , @OliverStegle and @MoritzMall groups. Preprint at end🧵👇

Exciting times.

RT @NovakovskyG: Excited to share my first contribution here at Illumina! We developed PromoterAI, a deep neural network that accurately id….

@bayraktar_lab Thread from @bayraktar_lab

Thanks to all our collaborators- specially Laura (co-lead on computational analysis) and Elisa, Tannia and Fani (lead on experimental data).Thanks to funding agencies who made this work possible and the patients for agreeing to donate samples.

15/ We believe this approach can be applied to other cancers to uncover—and exploit—plasticity brakes. Get in touch if interested! #GBM #MultiOmics #CancerResearch #deeplearning #cancerneuroscience #GRNs #singlecell #Cancer.

14/ Want to learn more? Read our regulatory paper and companion spatial multi-omics study .Watch out for a separate 🧵from @bayraktar_lab on conserved spatiotemporal trajectories in GB.

13/ Our framework unifies regulatory and spatial logic of GB heterogeneity. While our companion paper showed subclones are intermixed across conserved tissue niches, our regulatory model explains WHY—they follow the same trajectory because they're constrained by the same rules!.

12/ Beyond chromatin changes, MYT1L transformed GB cells into less aggressive neuronal-like cells with: • Enhanced neurite-like morphology • Reduced tumor microtube connectivity • Decreased proliferation • In vivo: slower growth, less invasion, longer survival!

11/ Our experimental validation was striking• MYT1L overexpression closed >80% of differential chromatin regions • MYT1L knockout reopened access to plastic fates • MYT1L directly binds and represses regulators of other states •85% of scDORI's predicted target TFs confirmed!

10/ 🔥 Can we use these GRNs to manipulate tumor identity and push GB cells into less plastic states? YES! We predicted MYT1L as the key regulatory bottleneck- a master repressor that locks cells into neuronal-like states by directly repressing regulators of plastic states.

9/ Remarkably, our regulatory roadmap explains tumor architecture! States with easy transitions exist in close spatial proximity, while states separated by regulatory barriers are spatially distant. The regulatory rules we've uncovered directly shape how tumors are organized!

8/ This roadmap revealed striking asymmetry in GB plasticity: • OPC/NPC-like and AC-like states can easily activate multiple alternate fates • Neuronal-like states are "locked" by strong repression barriers • Transitions follow preferred directions

7/ 🔑 The big question: Which tumor states can easily transition to others? We developed metrics to quantify both activation potential (what enables transitions) and repression barriers (what prevents them), creating the first regulatory roadmap of GB plasticity.

6/ The power of multi-omics: We can distinguish what a cell IS versus what it COULD BECOME. While only 16% of Topic Regulators are expressed across different tumor states, over 54% are epigenetically accessible—revealing "primed drivers" ready for activation during transitions!

5/ Applied to GB, scDORI uncovered Topics that redefine tumor heterogeneity through regulatory logic. Each Topic links specific TFs, enhancers, and target genes that work together across tumor states. We identified key "Topic Regulators" (TRs)—master TFs per Topic.

4/ scDORI • Scales to millions of cells • Models continuous cell state GRNs • Incorporates repression signatures • Each cell is modeled as a mixture of regulatory Topics • Can be applied to ANY multi-omic dataset (happy to hear your feedback, separate 🧵soon).

3/ To decode GB's regulatory logic, we developed scDORI—an autoencoder that decomposes multi-omic profiles into "regulatory Topics." Each Topic represents specific TF-target gene relationships, modeling cells without requiring predefined cell-types. Code:

2/ Key finding: Malignant GB cells are dramatically more plastic than non-malignant cells, BUT neuronal-like tumor states show surprisingly LOW plasticity. This hints at regulatory constraints we could potentially exploit therapeutically!

1/ Glioblastoma (GB) is driven by cellular plasticity—tumor cells switching between states. With @bayraktar_lab team, we profiled >1M nuclei with sc multi-ome (RNA+ATAC) from 12 GBs, capturing full tumor heterogeneity. Core Q: what regulatory mechanisms guide these transitions?