Pohjonen*, Batouche*, Kantola, Rannikko, Sandeman, Erickson, Pitkänen#, Mirtti#
Journal of Pathology Informatics 100715
A foundation model pre-trained on 48 million prostate tissue tiles outperforms natural-image pre-training and improves prostate cancer survival models when combined with clinical nomograms.
Langohr*, Kaaja*, Douglas, Nebelung, J. Koski, Ikonen, Katainen, Maljanen, Hakkarainen, Räisänen, Niinimäki, Kakko, Siitonen, Adhikari, Vähä-Koskela, Heckman, Lahtela, Wartiovaara-Kautto#, Pitkänen#, Kilpivaara#
HemaSphere
Single-cell transcriptomics of patient marrow finds ferroptotic erythroid stress arising before TP53 mutation in both ERCC6L2 disease and Shwachman-Diamond syndrome.
Tagmazian#, Schwarz, Lange, Pitkänen#, Vuoksimaa#
bioRxiv
A variational autoencoder over 3,110 amyloid PET scans separates four subgroups along the Alzheimer's continuum without preset positivity thresholds, two of them progressing faster.
Sanjaya, Pitkänen
medRxiv
MuAt2 classifies tumour type and subtype jointly across 14,527 Genomics England whole genomes, with embeddings that stratify glioma prognosis and infer origins of unknown primaries.
Sanjaya, Pitkänen
bioRxiv
The MuAt models packaged for Docker and Bioconda, reproducing published accuracy and reaching 81% in the Genomics England secure environment without retraining.
Y. Koski, Patel, Kakko von Koch, Jouhten, Aaltonen, Palin#, Sahu#, Pitkänen#
bioRxiv
A statistical toolkit for nanopore signal that separates four adduct-inducing compounds by their effects on base quality, ionic current and translocation dynamics.
Tagmazian, Schwarz, Lange, Pitkänen#, Vuoksimaa#
European Journal of Neuroscience
A residual network predicts cerebrospinal fluid amyloid-beta directly from amyloid PET scans (r = 0.92), without preselected reference regions or tracer-specific thresholds.
J. Koski, Langohr, Räisänen, Lahtinen, Hakkarainen, Heckman, Wartiovaara-Kautto#, Pitkänen#, Kilpivaara#
medRxiv
An ACMG-guided pipeline over 511 Finnish hematological malignancy patients finds CHEK2 variants enriched 1.7-fold and candidate alleles in CUX2, RNPC3 and MFSD2A.
Högel-Starck*, Timonen*, Atarsaikhan, Mogollon, Polso, Hassinen, Honkanen, Soini, Ruokoranta, Ahlnas, Juvila, Miettinen, Rodosthenous, Arvas, FinnGen, Heckman, Partanen, Daly, Palotie, Paavolainen#, Pietiäinen#, Pitkänen#
bioRxiv
Fluorescence imaging of 50 million mononuclear cells from 390 blood donors yields 18 morphology clusters and 93 genetic associations across 30 loci.
Ji, Li, Sun, Dong, Taalas, Zhang, Wu, Pitkänen, Marttinen
ACM Computing Surveys
A review that decomposes deep learning models for medical coding into encoder, architecture, decoder and auxiliary-information components, with benchmarks and open challenges.
Sanjaya, Maljanen, Katainen, Waszak, Genomics England Research Consortium, Aaltonen, Stegle, Korbel, Pitkänen
Genome Medicine 15:47
Attention over individual somatic mutations types tumours from whole genomes at 89% accuracy and recovers subtypes that were never given as training labels.
Timonen, Kerkelä, Impola, Penna, Partanen, Kilpivaara, Arvas, Pitkänen
Cytometry Part A
An Inception U-Net generates fluorescent marker images from brightfield imaging flow cytometry, typing blood cells without staining, including types unseen during training.
Meriranta, Pitkänen, Leppä
Seminars in Hematology
A review of cell-free DNA fragmentomics in B-cell lymphoma liquid biopsy: how fragment patterns form, the technical limits, and where machine learning fits.
Pohjonen, Stürenberg, Rannikko#, Mirtti#, Pitkänen#
iScience 25, 2
Regularising prediction scores with spectral decoupling counters overfitting to easy features and adds up to 9.5 percentage points on external medical imaging datasets.
Pohjonen, Stürenberg, Föhr, Pitkänen*, Rannikko*, Mirtti*
arXiv
Distribution-shifted datasets expose generalisation failures that validation against external cohorts misses, and strong augmentation trains models that hold performance under those shifts.
The ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium
Nature 578, 82–93
Integrative analysis of 2,658 whole cancer genomes across 38 tumour types, covering driver mutations, chromothripsis, telomere maintenance and germline effects on somatic mutation.
Katainen, Donner, Cajuso, Kaasinen, Palin, Mäkinen, Aaltonen, Pitkänen
Nature Protocols 13, 2580
Interactive desktop software for variant analysis: a genome-wide scan of 3 million somatic variants from 200 cancer genomes runs in about ten minutes.
* equal contribution · # shared senior authorship