Appendix: Summary of Best Papers Selected for the IMIA Yearbook 2023, Bioinformatics
and Translational Informatics
Grazioli F, Siarheyeu R, Alqassem I, Henschel A, Pileggi G, Meiser A.
Microbiome-based disease prediction with multimodal variational information bottlenecks
PLoS Comput Biol 2022 Apr 11;18(4):e1010050. doi: 10.1371/journal.pcbi.1010050
In this paper the authors addressed the untapped potential of multimodal machine learning
in disease prediction by leveraging the diagnostic potential of gut microbial profiling.
Traditionally, microbial species-relative abundances or strain-level markers extracted
through shotgun metagenomic sequencing have been separately assessed in disease prediction
models. Grazioli et al.'s innovative approach involved the development of a Multimodal
Variational Information Bottleneck (MVIB), a deep learning model capable of integrating
multiple heterogeneous data modalities into a single predictive framework. MVIB was
devised to offer both efficient performance and interpretability. The model creates
a joint stochastic encoding of different input data types, thereby integrating a plethora
of disease-related markers. Through evaluating the model on 11 publicly available
disease cohorts, the researchers achieved high classification performance, with areas
under the ROC curve (AUCs) ranging from 0.80 to 0.95 for five cohorts, while maintaining
medium performance for the remainder. The versatility of MVIB was demonstrated through
cross-study generalization experiments, where training and testing were performed
on different cohorts for the same disease. The results were comparable to a benchmark
Random Forest model. Moreover, the scalability of MVIB was underscored by its ability
to incorporate a third input modality, metabolomic data derived from mass spectrometry,
without compromising efficiency or performance.
Kuppe C, Ramirez Flores RO, Li Z, Hayat S, Levinson RT, Liao X, Hannani MT, Tanevski
J, Wünnemann F, Nagai JS, Halder M, Schumacher D, Menzel S, Schäfer G, Hoeft K, Cheng
M, Ziegler S, Zhang X, Peisker F, Kaesler N, Saritas T, Xu Y, Kassner A, Gummert J,
Morshuis M, Amrute J, Veltrop RJA, Boor P, Klingel K, Van Laake LW, Vink A, Hoogenboezem
RM, Bindels EMJ, Schurgers L, Sattler S, Schapiro D, Schneider RK, Lavine K, Milting
H, Costa IG, Saez-Rodriguez J, Kramann R
Spatial multi-omic map of human myocardial infarction
Nature 2022 Aug;608(7924):766-77. doi: 10.1038/s41586-022-05060-x
In this paper the authors leveraged single-cell-omics profiling to generate a temporal
and spatial map of cardiac cell types in myocardial infarction patients and controls.
Remodeling of cardiac tissues after myocardial infarction significantly contributes
to late-stage mortality and is not well-addressed by current therapies. Limiting the
negative impacts of cardiac remodeling on patients will require the development of
new therapeutic approaches enabled by a more precise molecular understanding of the
cell types involved in the process. The authors combined single-cell approaches, including
single nucleus RNA sequencing and single nucleus assay for transposase-accessible
chromatin sequencing, with spatial transcriptomics in 31 samples that spanned multiple
clinical timepoints. From these data, they were able to identify major cell types
in heart tissue and map these to particular histomorphological regions. Integrating
these multi-omics data identified sets of differentially expressed genes in cells
that marked the border between injured and uninjured tissues, and characterized the
profiles of remodeled versus functional myocardium. The resulting multi-modal atlas
of the human heart generates hypotheses that can facilitate new therapeutic advances
and provides an important resource for the research community.
Weitz P, Wang Y, Kartasalo K, Egevad L, Lindberg J, Grönberg H, Eklund M, Rantalainen
M
Transcriptome-wide prediction of prostate cancer gene expression from histopathology
images using co-expression-based convolutional neural networks.
Bioinformatics 2022 Jun 27;38(13):3462-9. doi: 10.1093/bioinformatics/btac343
The authors developed a novel machine learning to predict gene expression directly
from haematoxylin and eosin-stained whole slide images (WSIs) in samples from patients
with prostate cancer. Molecular phenotyping, especially using gene expression data,
is an increasingly important approach to characterize patient samples, compute clinical
scores from biomarkers, and implement precision care. However, assays to generate
the required data to conduct molecular phenotyping and compute clinical scores are
costly to implement on a large scale. Previous work has shown that molecular phenotypes,
including gene expression, can be accurately predicted from histopathology WSIs and
these WSIs are routinely collected and digitized during care. In their study, Weitz
et al. trained a convolutional neural network (CNN) using prostate cancer samples
from TCGA PRAD to predict gene expression levels for clusters of co-expressed genes.
They then applied their predictions to compute a predicted cell cycle progression
(CCP) score, which correlates with cancer aggressiveness, recurrence, and mortality.
The authors found a significant correlation between predicted gene expression and
transcript levels measured by RNA-sequencing in more than 6,600 genes. The significantly
predicted genes were enriched in pathways relevant to prostate cancer, including those
involved in DNA replication, cell cycle, and metabolism. Weitz et al. computed a CCN-based
CCP score using these results. The CNN-based CCP scores were prognostic in their preliminary
analysis, and were correlated with tumor grade to a similar degree as the RNA-seq-based
CCP. Ultimately, this work suggests that deep learning models could provide a scalable
solution to quantify gene expression phenotypes directly from imaging, particularly
in settings where full molecular phenotyping would be otherwise unattainable.
