Mapping biology from mouse to man using transfer learning

Abstract

Biomedical research often involves conducting experiments on model organisms in the anticipationthat the biology learnt from these experiments will transfer to the human. Yet, it is commonly the case that biology does not transfer effectively, often for unknown reasons. Despite its importance to translational research this transfer process is not currently rigorously quantified. Here, we show that transfer learning -the branch of machine learning that concerns passing information from one domain to another -can be used to efficiently map biology from mouse to man, using the bone marrow (BM) as a representative example of a complex tissue. We first trained an artificial neural network (ANN) to accurately recognize various different cell types in mouse BM using data obtained from single-cell RNA-sequencing (scRNA-Seq) experiments. We found that this ANN, trained exclusively on mouse data, was able to identify individual human cells obtained from comparable scRNA-Seq experiments of human BM with 83% overall accuracy. However, while some human cell types were easily identified, others were not, indicating important differences in biology. To obtain a more accurate map of the human BM we then retrained the mouse ANN using scRNA-Seq data from a limited sample of human BM cells. Typically, less than 10 human cells of a given type were needed to accurately learn its representation in the updated model. In some cases, human cell identities could be inferred directly from the mouse ANN without retraining, via a process of biologically-guided zero-shot learning. These results show how machine learning can be used to reconstruct complex biology from limited data and have broad implications for biomedical research.Mapping Biology from Mouse to Man

Locomotion in virtual environments predicts cardiovascular responsiveness to subsequent stressful challenges

“…In addition, our model allows incremental training by iterative training with new observations, which allows us to encompass a larger representation of different traits. Furthermore, approaches based on transfer learning can also enable the adaptation of our model to other target domains such as clinical settings or even, in a similar fashion to mapping biological relationships from mouse to human 85 , to facilitate translational approaches mapping animal behavioral/physiological interplay to human models or vice-versa.…”

Single cell RNA sequence analysis of human bone marrow samples reveals new targets for isolation of skeletal stem cells using DNA-coated gold nanoparticles

“…Raw sequencing reads were demultiplexed, grouping reads by cell barcode to generate a digital gene expression (DGE) matrix for downstream analysis, using Drop-Seq tools (v1.0). A modified multi-mapper pipeline was executed to correct multiple alignment [14].…”

“…In total, 9,795 cells were profiled, reduced to 7,385 following quality control processing. To further increase the cell pool for identification of skeletal progenitors, this data was combined with sc-RNAseq data for unsorted bone marrow populations from 3 patients [14] We next sought to identify candidate SSC markers, through examination of localised gene expression in pericytes and endothelial cell clusters, representing 233 and 460 cells respectively. Analysis of differential gene expression between pericytes/endothelial cells, termed 'Skeletal progenitors', and other cell types indicated SPARCL1, as a potential SSC marker due to elevated expression ( Figure 1C).…”

Transferlernen in der Biomedizin

“…denen des Menschen verglichen werden (Abb. 2B), wie wir kürzlich zeigen konnten [7]. Des Weiteren konnten die trainierten Modelle auch vergleichend eingesetzt werden.…”

“…2B). Im Vergleich mit den Modellen, die keinen Informationstransfer nutzen, kann man herausfi nden, welcher Zugewinn bei der Sensitivität oder Präzision des Modells durch das vorherige Lernen von Genexpressionsprofilen der Mauszellen erreicht wird [7]. Dieser Zugewinn kann nun als Maß für die Gemeinsamkeiten zwischen dem Modellorganismus (Ursprungsdomäne) und dem Menschen (Zieldomäne) genutzt werden.…”