RT @Merijellona: It's Fragment Friday! Happy to share our thoughts on #ctDNA #fragmentomics with @epitkane and @LeppaSM in the latest ctDNA….

@PrimaSanjaya @FIMM_UH @ATGprogram @iCAN_Finland An earlier thread on the key results: For the revision, we validated MuAt in extended WGS cohorts together with @GenomicsEngland and @AaltonenLab - huge thanks!.

Join millions who have switched to Grok.

Wonderful to see our work on (deep) learning how to classify tumor types with WGS/WES data published in Genome Medicine! Great effort by @PrimaSanjaya and coauthors to validate the method in >10,000 whole cancer genomes. @FIMM_UH @ATGprogram @iCAN_Finland

RT @FIMM_UH: Are you our new Group Leader?.We will recruit early-stage group leaders in molecular medicine with notable international exper….

Fantastic work from our collaborators @waaralab on this intriguing hematological disease.

Thread on our key results

Happy to see our paper on reconstructing fluorescent labels in imaging flow cytometry with deep learning published. Excellent work by @TimonenVeera and collaborators! @frcbsresearch @mlbiomed @FIMM_UH

RT @mkuijjer: Would you like to develop your own research line, on the interface of network science, computational tool development, and ca….

We have exciting opportunities in @iCAN_Finland for bioinformaticians and clinical data analysts to transform large-scale molecular profiling and health registry data into breakthroughs in precision cancer medicine!

RT @paula_jouhten: The cool yeasts are tacking to progress against wind or analogously natural selection.

RT @paula_jouhten: Excited to share our method for predictive evolution of metabolic traits using mathematical models of metabolism! Imagin….

RT @MirttiTuomas: Cancer research, histomics and multiomics well presented in the first Finnish LUMI supercomputer….

RT @SWedenoja: Katsauksemme valottaa ongelmia, joita #toisiolaki aiheuttaa rekisteritutkimukselle ja -yhteistyölle @THLorg @HelsinkiUniMed….

DeepIFC code and trained models are available at You can browse our data interactively at

Thank you to the Academy of Finland @SuomenAkatemia for supporting this work, and to all co-authors! @EKerkela Ulla Impola @LeenaPenna @JukkaPartanen1 @OKilpivaara @miarvas.

DeepIFC was trained on IFC images acquired from mononuclear cells. To see if DeepIFC was able to distinguish a previously unseen cell type, we analyzed an independent set of red blood cells. These cells were recognized as a separate cell type. #zeroshot #representationlearning

IFC images sometimes contain more than one cell, allowing investigation of potential cell-cell interactions. DeepIFC was able to reconstruct surface marker fluorescence for individual cells in doublet events. We observed 2x more monocytes in doublets than expected.

We classified cells by gating virtually labeled cells. Balancing training data improved prediction of especially NK, NKT, cytotoxic T and B cells, which were otherwise difficult to predict. For example, data balancing improved F1 score for NKT cells from 0.22 to 0.71.

We evaluated DeepIFC trained on 247k IFC images by virtually labeling mononuclear cells. Cell surface markers CD45, CD3 and CD14, and dead/damaged cell marker 7-AAD were reconstructed well, while CD19, CD8 and CD56 proved more difficult.

DeepIFC is based on Inception U-Net architecture - popular model in cross-modality learning - to reconstruct fluorescent images from brightfield and darkfield images. Cell types are predicted by virtually gating reconstructed images.