We report the ability of two deep learning-based decision systems to stratify non-small cell lung cancer (NSCLC) patients treated with checkpoint inhibitor therapy into two distinct survival groups. Both systems analyze functional and morphological properties of epithelial regions in digital histopathology whole slide images stained with the SP263 PD-L1 antibody. The first system learns to replicate the pathologist assessment of the Tumor Cell (TC) score with a cut-point for positivity at 25% for patient stratification. The second system is free from assumptions related to TC scoring and directly learns patient stratification from the overall survival time and event information. Both systems are built on a novel unpaired domain adaptation deep learning solution for epithelial region segmentation. This approach significantly reduces the need for large pixel-precise manually annotated datasets while superseding serial sectioning or re-staining of slides to obtain ground truth by cytokeratin staining. The capacity of the first system to replicate the TC scoring by pathologists is evaluated on 703 unseen cases, with an addition of 97 cases from an independent cohort. Our results show Lin's concordance values of 0.93 and 0.96 against pathologist scoring, respectively. The ability of the first and second system to stratify anti-PD-L1 treated patients is evaluated on 151 clinical samples. Both systems show similar stratification powers (first system: HR = 0.539, p = 0.004 and second system: HR = 0.525, p = 0.003) compared to TC scoring by pathologists (HR = 0.574, p = 0.01).

MOTIVATION: High-quality protein sequence alignments are essential for a number of downstream applications such as template-based protein structure prediction. In addition to the similarity score between sequence profile columns, many current profile-profile alignment tools use extra terms that compare 1D-structural properties such as secondary structure and solvent accessibility, which are predicted from short profile windows around each sequence position. Such scores add non-redundant information by evaluating the conservation of local patterns of hydrophobicity and other amino acid properties and thus exploiting correlations between profile columns. RESULTS: Here, instead of predicting and comparing known 1D properties, we follow an agnostic approach. We learn in an unsupervised fashion a set of maximally conserved patterns represented by 13-residue sequence profiles, without the need to know the cause of the conservation of these patterns. We use a maximum likelihood approach to train a set of 32 such profiles that can best represent patterns conserved within pairs of remotely homologs, structurally aligned training profiles. We include the new context score into our Hmm-Hmm alignment tool hhsearch and improve especially the quality of difficult alignments significantly. CONCLUSION: The context similarity score improves the quality of homology models and other methods that depend on accurate pairwise alignments.

Homology modeling predicts the 3D structure of a query protein based on the sequence alignment with one or more template proteins of known structure. Its great importance for biological research is owed to its speed, simplicity, reliability and wide applicability, covering more than half of the residues in protein sequence space. Although multiple templates have been shown to generally increase model quality over single templates, the information from multiple templates has so far been combined using empirically motivated, heuristic approaches. We present here a rigorous statistical framework for multi-template homology modeling. First, we find that the query proteins' atomic distance restraints can be accurately described by two-component Gaussian mixtures. This insight allowed us to apply the standard laws of probability theory to combine restraints from multiple templates. Second, we derive theoretically optimal weights to correct for the redundancy among related templates. Third, a heuristic template selection strategy is proposed. We improve the average GDT-ha model quality score by 11% over single template modeling and by 6.5% over a conventional multi-template approach on a set of 1000 query proteins. Robustness with respect to wrong constraints is likewise improved. We have integrated our multi-template modeling approach with the popular MODELLER homology modeling software in our free HHpred server http://toolkit.tuebingen.mpg.de/hhpred and also offer open source software for running MODELLER with the new restraints at https://bitbucket.org/soedinglab/hh-suite.