Deep learning-based scoring of tumour-infiltrating lymphocytes is prognostic in primary melanoma and predictive to PD-1 checkpoint inhibition in melanoma metastases.

BACKGROUND: Recent advances in digital pathology have enabled accurate and standardised enumeration of tumour-infiltrating lymphocytes (TILs). Here, we aim to evaluate TILs as a percentage electronic TIL score (eTILs) and investigate its prognostic and predictive relevance in cutaneous melanoma. METHODS: We included stage I to IV cutaneous melanoma patients and used hematoxylin-eosin-stained slides for TIL analysis. We assessed eTILs as a continuous and categorical variable using the published cut-off of 16.6% and applied Cox regression models to evaluate associations of eTILs with relapse-free, distant metastasis-free, and overall survival. We compared eTILs of the primaries with matched metastasis. Moreover, we assessed the predictive relevance of eTILs in therapy-naïve metastases according to the first-line therapy. FINDINGS: We analysed 321 primary cutaneous melanomas and 191 metastatic samples. In simple Cox regression, tumour thickness (p < 0.0001), presence of ulceration (p = 0.0001) and eTILs ≤16.6% (p = 0.0012) were found to be significant unfavourable prognostic factors for RFS. In multiple Cox regression, eTILs ≤16.6% (p = 0.0161) remained significant and downgraded the current staging. Lower eTILs in the primary tissue was associated with unfavourable relapse-free (p = 0.0014) and distant metastasis-free survival (p = 0.0056). In multiple Cox regression adjusted for tumour thickness and ulceration, eTILs as continuous remained significant (p = 0.019). When comparing TILs in primary tissue and corresponding metastasis of the same patient, eTILs in metastases was lower than in primary melanomas (p < 0.0001). In therapy-naïve metastases, an eTILs >12.2% was associated with longer progression-free survival (p = 0.037) and melanoma-specific survival (p = 0.0038) in patients treated with anti-PD-1-based immunotherapy. In multiple Cox regression, lactate dehydrogenase (p < 0.0001) and eTILs ≤12.2% (p = 0.0130) were significantly associated with unfavourable melanoma-specific survival. INTERPRETATION: Assessment of TILs is prognostic in primary melanoma samples, and the eTILs complements staging. In therapy-naïve metastases, eTILs ≤12.2% is predictive of unfavourable survival outcomes in patients receiving anti-PD-1-based therapy. FUNDING: See a detailed list of funding bodies in the Acknowledgements section at the end of the manuscript.

mlf-core: a framework for deterministic machine learning.

MOTIVATION: Machine learning has shown extensive growth in recent years and is now routinely applied to sensitive areas. To allow appropriate verification of predictive models before deployment, models must be deterministic. Solely fixing all random seeds is not sufficient for deterministic machine learning, as major machine learning libraries default to the usage of nondeterministic algorithms based on atomic operations. RESULTS: Various machine learning libraries released deterministic counterparts to the nondeterministic algorithms. We evaluated the effect of these algorithms on determinism and runtime. Based on these results, we formulated a set of requirements for deterministic machine learning and developed a new software solution, the mlf-core ecosystem, which aids machine learning projects to meet and keep these requirements. We applied mlf-core to develop deterministic models in various biomedical fields including a single-cell autoencoder with TensorFlow, a PyTorch-based U-Net model for liver-tumor segmentation in computed tomography scans, and a liver cancer classifier based on gene expression profiles with XGBoost. AVAILABILITY AND IMPLEMENTATION: The complete data together with the implementations of the mlf-core ecosystem and use case models are available at https://github.com/mlf-core.

A data management infrastructure for the integration of imaging and omics data in life sciences.

BACKGROUND: As technical developments in omics and biomedical imaging increase the throughput of data generation in life sciences, the need for information systems capable of managing heterogeneous digital assets is increasing. In particular, systems supporting the findability, accessibility, interoperability, and reusability (FAIR) principles of scientific data management. RESULTS: We propose a Service Oriented Architecture approach for integrated management and analysis of multi-omics and biomedical imaging data. Our architecture introduces an image management system into a FAIR-supporting, web-based platform for omics data management. Interoperable metadata models and middleware components implement the required data management operations. The resulting architecture allows for FAIR management of omics and imaging data, facilitating metadata queries from software applications. The applicability of the proposed architecture is demonstrated using two technical proofs of concept and a use case, aimed at molecular plant biology and clinical liver cancer research, which integrate various imaging and omics modalities. CONCLUSIONS: We describe a data management architecture for integrated, FAIR-supporting management of omics and biomedical imaging data, and exemplify its applicability for basic biology research and clinical studies. We anticipate that FAIR data management systems for multi-modal data repositories will play a pivotal role in data-driven research, including studies which leverage advanced machine learning methods, as the joint analysis of omics and imaging data, in conjunction with phenotypic metadata, becomes not only desirable but necessary to derive novel insights into biological processes.

The Eukaryotic CO2-Concentrating Organelle Is Liquid-like and Exhibits Dynamic Reorganization.

Approximately 30%-40% of global CO2 fixation occurs inside a non-membrane-bound organelle called the pyrenoid, which is found within the chloroplasts of most eukaryotic algae. The pyrenoid matrix is densely packed with the CO2-fixing enzyme Rubisco and is thought to be a crystalline or amorphous solid. Here, we show that the pyrenoid matrix of the unicellular alga Chlamydomonas reinhardtii is not crystalline but behaves as a liquid that dissolves and condenses during cell division. Furthermore, we show that new pyrenoids are formed both by fission and de novo assembly. Our modeling predicts the existence of a "magic number" effect associated with special, highly stable heterocomplexes that influences phase separation in liquid-like organelles. This view of the pyrenoid matrix as a phase-separated compartment provides a paradigm for understanding its structure, biogenesis, and regulation. More broadly, our findings expand our understanding of the principles that govern the architecture and inheritance of liquid-like organelles.

Native architecture of the Chlamydomonas chloroplast revealed by in situ cryo-electron tomography.

Chloroplast function is orchestrated by the organelle's intricate architecture. By combining cryo-focused ion beam milling of vitreous Chlamydomonas cells with cryo-electron tomography, we acquired three-dimensional structures of the chloroplast in its native state within the cell. Chloroplast envelope inner membrane invaginations were frequently found in close association with thylakoid tips, and the tips of multiple thylakoid stacks converged at dynamic sites on the chloroplast envelope, implicating lipid transport in thylakoid biogenesis. Subtomogram averaging and nearest neighbor analysis revealed that RuBisCO complexes were hexagonally packed within the pyrenoid, with ~15 nm between their centers. Thylakoid stacks and the pyrenoid were connected by cylindrical pyrenoid tubules, physically bridging the sites of light-dependent photosynthesis and light-independent carbon fixation. Multiple parallel minitubules were bundled within each pyrenoid tubule, possibly serving as conduits for the targeted one-dimensional diffusion of small molecules such as ATP and sugars between the chloroplast stroma and the pyrenoid matrix.