Novel Preoperative Risk Stratification Using Digital Phenotyping Applying a Scalable Machine-Learning Approach.

BACKGROUND: Classification of perioperative risk is important for patient care, resource allocation, and guiding shared decision-making. Using discriminative features from the electronic health record (EHR), machine-learning algorithms can create digital phenotypes among heterogenous populations, representing distinct patient subpopulations grouped by shared characteristics, from which we can personalize care, anticipate clinical care trajectories, and explore therapies. We hypothesized that digital phenotypes in preoperative settings are associated with postoperative adverse events including in-hospital and 30-day mortality, 30-day surgical redo, intensive care unit (ICU) admission, and hospital length of stay (LOS). METHODS: We identified all laminectomies, colectomies, and thoracic surgeries performed over a 9-year period from a large hospital system. Seventy-seven readily extractable preoperative features were first selected from clinical consensus, including demographics, medical history, and lab results. Three surgery-specific datasets were built and split into derivation and validation cohorts using chronological occurrence. Consensus k -means clustering was performed independently on each derivation cohort, from which phenotypes' characteristics were explored. Cluster assignments were used to train a random forest model to assign patient phenotypes in validation cohorts. We reconducted descriptive analyses on validation cohorts to confirm the similarity of patient characteristics with derivation cohorts, and quantified the association of each phenotype with postoperative adverse events by using the area under receiver operating characteristic curve (AUROC). We compared our approach to American Society of Anesthesiologists (ASA) alone and investigated a combination of our phenotypes with the ASA score. RESULTS: A total of 7251 patients met inclusion criteria, of which 2770 were held out in a validation dataset based on chronological occurrence. Using segmentation metrics and clinical consensus, 3 distinct phenotypes were created for each surgery. The main features used for segmentation included urgency of the procedure, preoperative LOS, age, and comorbidities. The most relevant characteristics varied for each of the 3 surgeries. Low-risk phenotype alpha was the most common (2039 of 2770, 74%), while high-risk phenotype gamma was the rarest (302 of 2770, 11%). Adverse outcomes progressively increased from phenotypes alpha to gamma, including 30-day mortality (0.3%, 2.1%, and 6.0%, respectively), in-hospital mortality (0.2%, 2.3%, and 7.3%), and prolonged hospital LOS (3.4%, 22.1%, and 25.8%). When combined with the ASA score, digital phenotypes achieved higher AUROC than the ASA score alone (hospital mortality: 0.91 vs 0.84; prolonged hospitalization: 0.80 vs 0.71). CONCLUSIONS: For 3 frequently performed surgeries, we identified 3 digital phenotypes. The typical profiles of each phenotype were described and could be used to anticipate adverse postoperative events.

Assessing eligibility for lung cancer screening using parsimonious ensemble machine learning models: A development and validation study.

BACKGROUND: Risk-based screening for lung cancer is currently being considered in several countries; however, the optimal approach to determine eligibility remains unclear. Ensemble machine learning could support the development of highly parsimonious prediction models that maintain the performance of more complex models while maximising simplicity and generalisability, supporting the widespread adoption of personalised screening. In this work, we aimed to develop and validate ensemble machine learning models to determine eligibility for risk-based lung cancer screening. METHODS AND FINDINGS: For model development, we used data from 216,714 ever-smokers recruited between 2006 and 2010 to the UK Biobank prospective cohort and 26,616 high-risk ever-smokers recruited between 2002 and 2004 to the control arm of the US National Lung Screening (NLST) randomised controlled trial. The NLST trial randomised high-risk smokers from 33 US centres with at least a 30 pack-year smoking history and fewer than 15 quit-years to annual CT or chest radiography screening for lung cancer. We externally validated our models among 49,593 participants in the chest radiography arm and all 80,659 ever-smoking participants in the US Prostate, Lung, Colorectal and Ovarian (PLCO) Screening Trial. The PLCO trial, recruiting from 1993 to 2001, analysed the impact of chest radiography or no chest radiography for lung cancer screening. We primarily validated in the PLCO chest radiography arm such that we could benchmark against comparator models developed within the PLCO control arm. Models were developed to predict the risk of 2 outcomes within 5 years from baseline: diagnosis of lung cancer and death from lung cancer. We assessed model discrimination (area under the receiver operating curve, AUC), calibration (calibration curves and expected/observed ratio), overall performance (Brier scores), and net benefit with decision curve analysis. Models predicting lung cancer death (UCL-D) and incidence (UCL-I) using 3 variables-age, smoking duration, and pack-years-achieved or exceeded parity in discrimination, overall performance, and net benefit with comparators currently in use, despite requiring only one-quarter of the predictors. In external validation in the PLCO trial, UCL-D had an AUC of 0.803 (95% CI: 0.783, 0.824) and was well calibrated with an expected/observed (E/O) ratio of 1.05 (95% CI: 0.95, 1.19). UCL-I had an AUC of 0.787 (95% CI: 0.771, 0.802), an E/O ratio of 1.0 (95% CI: 0.92, 1.07). The sensitivity of UCL-D was 85.5% and UCL-I was 83.9%, at 5-year risk thresholds of 0.68% and 1.17%, respectively, 7.9% and 6.2% higher than the USPSTF-2021 criteria at the same specificity. The main limitation of this study is that the models have not been validated outside of UK and US cohorts. CONCLUSIONS: We present parsimonious ensemble machine learning models to predict the risk of lung cancer in ever-smokers, demonstrating a novel approach that could simplify the implementation of risk-based lung cancer screening in multiple settings.

AutoPrognosis 2.0: Democratizing diagnostic and prognostic modeling in healthcare with automated machine learning.

Diagnostic and prognostic models are increasingly important in medicine and inform many clinical decisions. Recently, machine learning approaches have shown improvement over conventional modeling techniques by better capturing complex interactions between patient covariates in a data-driven manner. However, the use of machine learning introduces technical and practical challenges that have thus far restricted widespread adoption of such techniques in clinical settings. To address these challenges and empower healthcare professionals, we present an open-source machine learning framework, AutoPrognosis 2.0, to facilitate the development of diagnostic and prognostic models. AutoPrognosis leverages state-of-the-art advances in automated machine learning to develop optimized machine learning pipelines, incorporates model explainability tools, and enables deployment of clinical demonstrators, without requiring significant technical expertise. To demonstrate AutoPrognosis 2.0, we provide an illustrative application where we construct a prognostic risk score for diabetes using the UK Biobank, a prospective study of 502,467 individuals. The models produced by our automated framework achieve greater discrimination for diabetes than expert clinical risk scores. We have implemented our risk score as a web-based decision support tool, which can be publicly accessed by patients and clinicians. By open-sourcing our framework as a tool for the community, we aim to provide clinicians and other medical practitioners with an accessible resource to develop new risk scores, personalized diagnostics, and prognostics using machine learning techniques. Software: https://github.com/vanderschaarlab/AutoPrognosis.

Incorporating Target-Specific Pharmacophoric Information into Deep Generative Models for Fragment Elaboration.

Despite recent interest in deep generative models for scaffold elaboration, their applicability to fragment-to-lead campaigns has so far been limited. This is primarily due to their inability to account for local protein structure or a user's design hypothesis. We propose a novel method for fragment elaboration, STRIFE, that overcomes these issues. STRIFE takes as input fragment hotspot maps (FHMs) extracted from a protein target and processes them to provide meaningful and interpretable structural information to its generative model, which in turn is able to rapidly generate elaborations with complementary pharmacophores to the protein. In a large-scale evaluation, STRIFE outperforms existing, structure-unaware, fragment elaboration methods in proposing highly ligand-efficient elaborations. In addition to automatically extracting pharmacophoric information from a protein target's FHM, STRIFE optionally allows the user to specify their own design hypotheses.

Deep generative design with 3D pharmacophoric constraints.

Generative models have increasingly been proposed as a solution to the molecular design problem. However, it has proved challenging to control the design process or incorporate prior knowledge, limiting their practical use in drug discovery. In particular, generative methods have made limited use of three-dimensional (3D) structural information even though this is critical to binding. This work describes a method to incorporate such information and demonstrates the benefit of doing so. We combine an existing graph-based deep generative model, DeLinker, with a convolutional neural network to utilise physically-meaningful 3D representations of molecules and target pharmacophores. We apply our model, DEVELOP, to both linker and R-group design, demonstrating its suitability for both hit-to-lead and lead optimisation. The 3D pharmacophoric information results in improved generation and allows greater control of the design process. In multiple large-scale evaluations, we show that including 3D pharmacophoric constraints results in substantial improvements in the quality of generated molecules. On a challenging test set derived from PDBbind, our model improves the proportion of generated molecules with high 3D similarity to the original molecule by over 300%. In addition, DEVELOP recovers 10× more of the original molecules compared to the baseline DeLinker method. Our approach is general-purpose, readily modifiable to alternate 3D representations, and can be incorporated into other generative frameworks. Code is available at https://github.com/oxpig/DEVELOP.

Generating property-matched decoy molecules using deep learning.

MOTIVATION: An essential step in the development of virtual screening methods is the use of established sets of actives and decoys for benchmarking and training. However, the decoy molecules in commonly used sets are biased meaning that methods often exploit these biases to separate actives and decoys, and do not necessarily learn to perform molecular recognition. This fundamental issue prevents generalization and hinders virtual screening method development. RESULTS: We have developed a deep learning method (DeepCoy) that generates decoys to a user's preferred specification in order to remove such biases or construct sets with a defined bias. We validated DeepCoy using two established benchmarks, DUD-E and DEKOIS 2.0. For all 102 DUD-E targets and 80 of the 81 DEKOIS 2.0 targets, our generated decoy molecules more closely matched the active molecules' physicochemical properties while introducing no discernible additional risk of false negatives. The DeepCoy decoys improved the Deviation from Optimal Embedding (DOE) score by an average of 81% and 66%, respectively, decreasing from 0.166 to 0.032 for DUD-E and from 0.109 to 0.038 for DEKOIS 2.0. Further, the generated decoys are harder to distinguish than the original decoy molecules via docking with Autodock Vina, with virtual screening performance falling from an AUC ROC of 0.70 to 0.63. AVAILABILITY AND IMPLEMENTATION: The code is available at https://github.com/oxpig/DeepCoy. Generated molecules can be downloaded from http://opig.stats.ox.ac.uk/resources. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Deep Generative Models for 3D Linker Design.

Rational compound design remains a challenging problem for both computational methods and medicinal chemists. Computational generative methods have begun to show promising results for the design problem. However, they have not yet used the power of three-dimensional (3D) structural information. We have developed a novel graph-based deep generative model that combines state-of-the-art machine learning techniques with structural knowledge. Our method ("DeLinker") takes two fragments or partial structures and designs a molecule incorporating both. The generation process is protein-context-dependent, utilizing the relative distance and orientation between the partial structures. This 3D information is vital to successful compound design, and we demonstrate its impact on the generation process and the limitations of omitting such information. In a large-scale evaluation, DeLinker designed 60% more molecules with high 3D similarity to the original molecule than a database baseline. When considering the more relevant problem of longer linkers with at least five atoms, the outperformance increased to 200%. We demonstrate the effectiveness and applicability of this approach on a diverse range of design problems: fragment linking, scaffold hopping, and proteolysis targeting chimera (PROTAC) design. As far as we are aware, this is the first molecular generative model to incorporate 3D structural information directly in the design process. The code is available at https://github.com/oxpig/DeLinker.

Protein Family-Specific Models Using Deep Neural Networks and Transfer Learning Improve Virtual Screening and Highlight the Need for More Data.

Machine learning has shown enormous potential for computer-aided drug discovery. Here we show how modern convolutional neural networks (CNNs) can be applied to structure-based virtual screening. We have coupled our densely connected CNN (DenseNet) with a transfer learning approach which we use to produce an ensemble of protein family-specific models. We conduct an in-depth empirical study and provide the first guidelines on the minimum requirements for adopting a protein family-specific model. Our method also highlights the need for additional data, even in data-rich protein families. Our approach outperforms recent benchmarks on the DUD-E data set and an independent test set constructed from the ChEMBL database. Using a clustered cross-validation on DUD-E, we achieve an average AUC ROC of 0.92 and a 0.5% ROC enrichment factor of 79. This represents an improvement in early enrichment of over 75% compared to a recent machine learning benchmark. Our results demonstrate that the continued improvements in machine learning architecture for computer vision apply to structure-based virtual screening.