Supplementary data are available at Bioinformatics online.

<div>Fingerprints (FPs) are the most common small molecule representation in cheminformatics. There are a wide variety of fingerprints, and the Extended Connectivity Fingerprint (ECFP) is one of the best-suited for general applications. Despite the overall FP abundance, only a few FPs represent the 3D structure of the molecule, and hardly any encode protein-ligand interactions. Here, we present a Protein-Ligand Extended Connectivity (PLEC) fingerprint that implicitly encodes protein-ligand interactions by pairing the ECFP environments from the ligand and the protein. PLEC fingerprints were used to construct different machine learning (ML) models tailored for predicting protein-ligand affinities (pK_i/d). Even the simplest linear model built on the PLEC fingerprint achieved R_p=0.83 on the PDBbind v2016 "core set”, demonstrating its descriptive power. The PLEC fingerprint has been implemented in the Open Drug Discovery Toolkit (https://github.com/oddt/oddt).</div>

<div>Fingerprints (FPs) are the most common small molecule representation in cheminformatics. There are a wide variety of fingerprints, and the Extended Connectivity Fingerprint (ECFP) is one of the best-suited for general applications. Despite the overall FP abundance, only a few FPs represent the 3D structure of the molecule, and hardly any encode protein-ligand interactions. Here, we present a Protein-Ligand Extended Connectivity (PLEC) fingerprint that implicitly encodes protein-ligand interactions by pairing the ECFP environments from the ligand and the protein. PLEC fingerprints were used to construct different machine learning (ML) models tailored for predicting protein-ligand affinities (pK_i/d). Even the simplest linear model built on the PLEC fingerprint achieved R_p=0.83 on the PDBbind v2016 "core set”, demonstrating its descriptive power. The PLEC fingerprint has been implemented in the Open Drug Discovery Toolkit (https://github.com/oddt/oddt).</div>

Background Despite the fact that tumor microenvironment (TME) and gene mutations are the main determinants of progression of the deadliest cancer in the world – lung cancer, their interrelations are not well understood. Digital pathology data provides a unique insight into the spatial composition of the TME. Various spatial metrics and machine learning approaches were proposed for prediction of either patient survival or gene mutations from this data. Still, these approaches are limited in the scope of analyzed features and in their explainability, and as such fail to transfer to clinical practice. Methods Here, we generated 23,199 image patches from 26 hematoxylin-and-eosin (H&E)-stained lung cancer tissue sections and annotated them into 9 different tissue classes. Using this dataset, we trained a deep neural network ARA-CNN. Next, we applied the trained network to segment 467 lung cancer H&E images from The Cancer Genome Atlas (TCGA) database. We used the segmented images to compute human-interpretable features reflecting the heterogeneous composition of the TME, and successfully utilized them to predict patient survival and cancer gene mutations. Results We achieved per-class AUC ranging from 0.72 to 0.99 for classifying tissue types in lung cancer with ARA-CNN. Machine learning models trained on the proposed human-interpretable features achieved a c-index of 0.723 in the task of survival prediction and AUC up to 73.5% for PDGFRB in the task of mutation classification. Conclusions We presented a framework that accurately predicted survival and gene mutations in lung adenocarcinoma patients based on human-interpretable features extracted from H&E slides. Our approach can provide important insights for designing novel cancer treatments, by linking the spatial structure of the TME in lung adenocarcinoma to gene mutations and patient survival. It can also expand our understanding of the effects that the TME has on tumor evolutionary processes. Our approach can be generalized to different cancer types to inform precision medicine strategies.

Background: Effective target discovery strategies are key to developing new precision medicines against cancer. Large-scale efforts to systematically characterize cancer-associated germline variants, genomic changes in tumors and cancer genetic dependencies via CRISPR and RNAi in cell lines provide valuable resources for machine learning (ML) approaches designed to identify disease-modifying targets. However, fundamental challenges remain in understanding how various data types (including primary tumor and cell line data) can be jointly modeled to decipher the molecular network contexts underpinning the biology and efficacy of new target candidates. Approach: Using a data-driven knowledge graph spanning &gt;2 billion relations over &gt;350,000 biomedical entities, we assembled a comprehensive target prioritization resource with 53 feature types including multi-omics (germline and tumor) and genetic dependency data. Feature categories consisted of raw feature types (e.g. germline genetic, somatic mutation, copy number variation, transcriptomic, proteomic, survival associations, CRISPR/RNAi sensitivity, selectivity, efficacy and essentiality), network-transformed features capturing the underlying molecular network context of each possible target, and multidimensional network integration features combining systems-level information from multiple feature types. Multiple ML models were trained on either omics-based features only or on both omics and genetic dependency features; and using information from each cancer type separately, as well as jointly in multi-disease models. Results: Individual features and the resulting ML models were evaluated using known drug targets in 20 cancer types. We found that a) germline genetic and survival association features along with genetic dependency (efficacy and selectivity) features were the most predictive prior to network transformation; b) network-based feature transformation raised the predictive power of individual features (e.g. increasing mean AUPRC from 0.024 to 0.064 for germline genetic and from 0.018 to 0.048 for cell line sensitivity features); c) network integration of multiple feature types increased mean AUPRC by 21% over the best individual omics features; d) ML models integrating both omics and genetic dependency data outperformed ML omics-only approaches (mean AUPRC 0.219 vs. 0.195); and e) multi-disease ML models provided the best overall performance (mean AUPRC 0.232) underlining the value of multi-dimensional data integration and information transfer across cancer types. Resulting network-based machine learning models provide a highly interpretable view into the top-scoring genes for each cancer, prioritizing established targets, targets recently being evaluated (e.g. FGFR2 and FGFR3 in lung squamous cell carcinoma) and novel target candidates. Citation Format: Janusz Dutkowski, Radosław Bielecki, Karol Nienałtowski, Michał Kukiełka, Roy Ronen. Knowledge graph integration of germline, primary cancer and cancer genetic dependency data prioritizes new target candidates [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr LB148.