Artificial intelligence to unlock real-world evidence in clinical oncology: A primer on recent advances.

PURPOSE: Real world evidence is crucial to understanding the diffusion of new oncologic therapies, monitoring cancer outcomes, and detecting unexpected toxicities. In practice, real world evidence is challenging to collect rapidly and comprehensively, often requiring expensive and time-consuming manual case-finding and annotation of clinical text. In this Review, we summarise recent developments in the use of artificial intelligence to collect and analyze real world evidence in oncology. METHODS: We performed a narrative review of the major current trends and recent literature in artificial intelligence applications in oncology. RESULTS: Artificial intelligence (AI) approaches are increasingly used to efficiently phenotype patients and tumors at large scale. These tools also may provide novel biological insights and improve risk prediction through multimodal integration of radiographic, pathological, and genomic datasets. Custom language processing pipelines and large language models hold great promise for clinical prediction and phenotyping. CONCLUSIONS: Despite rapid advances, continued progress in computation, generalizability, interpretability, and reliability as well as prospective validation are needed to integrate AI approaches into routine clinical care and real-time monitoring of novel therapies.

Using electronic health record metadata to predict housing instability amongst veterans.

Housing instability is considered a significant life stressor and preemptive screening should be applied to identify those at risk for homelessness as early as possible so that they can be targeted for specialized care. We developed models to classify patient outcomes for an established VA Homelessness Screening Clinical Reminder (HSCR), which identifies housing instability, in the two months prior to its administration. Logistic Regression and Random Forest models were fit to classify responses using the last 18 months of document activity. We measure concentration of risk across stratifications of predicted probability and observe an enriched likelihood of finding confirmed false negative responses from veterans with diagnosed housing instability. Positive responses were 34 times more likely to be detected within the top 1 % of patients predicted at risk than from those randomly selected. There is a 1 in 4 chance of detecting false negatives within the top 1 % of predicted risk. Machine learning methods can classify between episodes of housing instability using a data-driven approach that does not rely on variables curated from domain experts. This method has the potential to improve clinicians' ability to identify veterans who are experiencing housing instability but are not captured by HSCR.

High dimensional predictions of suicide risk in 4.2 million US Veterans using ensemble transfer learning.

We present an ensemble transfer learning method to predict suicide from Veterans Affairs (VA) electronic medical records (EMR). A diverse set of base models was trained to predict a binary outcome constructed from reported suicide, suicide attempt, and overdose diagnoses with varying choices of study design and prediction methodology. Each model used twenty cross-sectional and 190 longitudinal variables observed in eight time intervals covering 7.5 years prior to the time of prediction. Ensembles of seven base models were created and fine-tuned with ten variables expected to change with study design and outcome definition in order to predict suicide and combined outcome in a prospective cohort. The ensemble models achieved c-statistics of 0.73 on 2-year suicide risk and 0.83 on the combined outcome when predicting on a prospective cohort of [Formula: see text] 4.2 M veterans. The ensembles rely on nonlinear base models trained using a matched retrospective nested case-control (Rcc) study cohort and show good calibration across a diversity of subgroups, including risk strata, age, sex, race, and level of healthcare utilization. In addition, a linear Rcc base model provided a rich set of biological predictors, including indicators of suicide, substance use disorder, mental health diagnoses and treatments, hypoxia and vascular damage, and demographics.

All-Atom Simulations Uncover Structural and Dynamical Properties of STING Proteins in the Membrane System.

Recent studies have shown that the stimulator of interferon gene (STING) protein plays a central role in the immune system by facilitating the production of type I interferons in cells. The STING signaling pathway is also a prominent activator of cancer-killing T cells that initiate a powerful adaptive immune response. Since biomolecular signaling pathways are complicated and not easily identified through traditional experiments, molecular dynamics (MD) has often been used to study structural and dynamical responses of biological pathways. Here, we carried out MD simulations for full-length chicken and human STING (chSTING and hSTING) proteins. Specifically, we investigated ligand-bound closed (holo) and ligand-unbound open (apo) forms of STING in the membrane system by comparing their conformational and dynamical differences. Our research provides clues for understanding the mechanism of the STING signaling pathway by uncovering detailed insights for the examined systems: the residues from each chain in the binding pocket are strongly correlated to one another in the open STING structure compared with those in the closed STING structure. Ligand-bound closed STING displays ∼174° rotation of the ligand-binding domain (LBD) relative to the open STING structure. The dynamical analysis of residue Cys148 located in the linker region of hSTING does not support the earlier hypothesis that Cys148 can form disulfide bonds between adjacent STING dimers. We also demonstrate that using the full-length proteins is critical, since the MD simulations of the LBD portion alone cannot properly describe the global conformational properties of STING.

An examination of the association between altitude and suicide deaths, suicide attempts, and suicidal ideation among veterans at both the patient and geospatial level.

Suicide is a major public health problem affecting US Veterans and the US in general. Many variables (e.g., demographic, clinical, biological, geographic) have been associated with risk for suicide and suicidal behavior, including altitude; however, the exact nature of the relationship between altitude and suicide remains unclear in part due to the fact that previous studies have used either geospatial data or individual-level data, but not both. Prior research has also failed to consider the full range of suicidal thoughts and behaviors, ranging from suicidal ideation to suicide deaths. Accordingly, the objective of the present research was to use both geospatial data (county and zip codes) and individual-level data to comprehensively assess the association between altitude and suicide mortality, suicide attempts, and suicidal ideation among US Veterans between 2000 and 2018. Taken together, our results demonstrate that there is a strong correlation between altitude and suicide rates at all the levels investigated and using different statistical analyses and even after controlling for significant covariates such as percent of age >50yr, percent male, percent white, percent non-Hispanic, median household income, and population density. We show that there is a positive correlation between altitude and suicide attempts especially when controlling by the covariates and a weak correlation between altitude and suicide ideation and the combination of suicide, suicide attempts and suicide ideation.

A case for developing domain-specific vocabularies for extracting suicide factors from healthcare notes.

The onset and persistence of life events (LE) such as housing instability, job instability, and reduced social connection have been shown to increase risk of suicide. Predictive models for suicide risk have low sensitivity to many of these factors due to under-reporting in structured electronic health records (EHR) data. In this study, we show how natural language processing (NLP) can help identify LE in clinical notes at higher rates than reported medical codes. We compare domain-specific lexicons formulated from Unified Medical Language System (UMLS) selection, content analysis by subject matter experts (SME) and the Gravity Project, to data-driven expansion through contextual word embedding using Word2Vec. Our analysis covers EHR from the Veterans Affairs (VA) Corporate Data Warehouse (CDW) and measures the prevalence of LE across time for patients with known underlying cause of death in the National Death Index (NDI). We found that NLP methods had higher sensitivity of detecting LE relative to structured EHR (S-EHR) variables. We observed that, on average, suicide cases had higher rates of LE over time when compared to patients who died of non-suicide related causes with no previous history of diagnosed mental illness. When used to discriminate these outcomes, the inclusion of NLP derived variables increased the concentration of LE along the top 0.1%, 0.5% and 1% of predicted risk. LE were less informative when discriminating suicide death from non-suicide related death for patients with diagnosed mental illness.

Assessment of chemical-crosslink-assisted protein structure modeling in CASP13.

With the advance of experimental procedures obtaining chemical crosslinking information is becoming a fast and routine practice. Information on crosslinks can greatly enhance the accuracy of protein structure modeling. Here, we review the current state of the art in modeling protein structures with the assistance of experimentally determined chemical crosslinks within the framework of the 13th meeting of Critical Assessment of Structure Prediction approaches. This largest-to-date blind assessment reveals benefits of using data assistance in difficult to model protein structure prediction cases. However, in a broader context, it also suggests that with the unprecedented advance in accuracy to predict contacts in recent years, experimental crosslinks will be useful only if their specificity and accuracy further improved and they are better integrated into computational workflows.

Blind prediction of homo- and hetero-protein complexes: The CASP13-CAPRI experiment.

We present the results for CAPRI Round 46, the third joint CASP-CAPRI protein assembly prediction challenge. The Round comprised a total of 20 targets including 14 homo-oligomers and 6 heterocomplexes. Eight of the homo-oligomer targets and one heterodimer comprised proteins that could be readily modeled using templates from the Protein Data Bank, often available for the full assembly. The remaining 11 targets comprised 5 homodimers, 3 heterodimers, and two higher-order assemblies. These were more difficult to model, as their prediction mainly involved "ab-initio" docking of subunit models derived from distantly related templates. A total of ~30 CAPRI groups, including 9 automatic servers, submitted on average ~2000 models per target. About 17 groups participated in the CAPRI scoring rounds, offered for most targets, submitting ~170 models per target. The prediction performance, measured by the fraction of models of acceptable quality or higher submitted across all predictors groups, was very good to excellent for the nine easy targets. Poorer performance was achieved by predictors for the 11 difficult targets, with medium and high quality models submitted for only 3 of these targets. A similar performance "gap" was displayed by scorer groups, highlighting yet again the unmet challenge of modeling the conformational changes of the protein components that occur upon binding or that must be accounted for in template-based modeling. Our analysis also indicates that residues in binding interfaces were less well predicted in this set of targets than in previous Rounds, providing useful insights for directions of future improvements.

Evaluation of the scale-consistent UNRES force field in template-free prediction of protein structures in the CASP13 experiment.

The recent NEWCT-9P version of the coarse-grained UNRES force field for proteins, with scale-consistent formulas for the local and correlation terms, has been tested in the CASP13 experiment of the blind-prediction of protein structure, in the ab initio, contact-assisted, and data-assisted modes. Significant improvement of the performance has been observed with respect to the CASP11 and CASP12 experiments (by over 10 GDT_TS units for the ab initio mode predictions and by over 15 GDT_TS units for the contact-assisted prediction, respectively), which is a result of introducing scale-consistent terms and improved handling of contact-distance restraints. As in previous CASP exercises, UNRES ranked higher in the free modeling category than in the general category that included template based modeling targets. Use of distance restraints from the predicted contacts, albeit many of them were wrong, resulted in the increase of GDT_TS by over 8 units on average and introducing sparse restraints from small-angle X-ray/neutron scattering and chemical cross-link-mass-spectrometry experiments, and ambiguous restraints from nuclear magnetic resonance experiments has also improved the predictions by 8.6, 9.7, and 10.7 GDT_TS units on average, respectively.

Purely Structural Protein Scoring Functions Using Support Vector Machine and Ensemble Learning.

The function of a protein is determined by its structure, which creates a need for efficient methods of protein structure determination to advance scientific and medical research. Because current experimental structure determination methods carry a high price tag, computational predictions are highly desirable. Given a protein sequence, computational methods produce numerous 3D structures known as decoys. Selection of the best quality decoys is both challenging and essential as the end users can handle only a few ones. Therefore, scoring functions are central to decoy selection. They combine measurable features into a single number indicator of decoy quality. Unfortunately, current scoring functions do not consistently select the best decoys. Machine learning techniques offer great potential to improve decoy scoring. This paper presents two machine-learning based scoring functions to predict the quality of proteins structures, i.e., the similarity between the predicted structure and the experimental one without knowing the latter. We use different metrics to compare these scoring functions against three state-of-the-art scores. This is a first attempt at comparing different scoring functions using the same non-redundant dataset for training and testing and the same features. The results show that adding informative features may be more significant than the method used.

In situ data analytics and indexing of protein trajectories.

The transition toward exascale computing will be accompanied by a performance dichotomy. Computational peak performance will rapidly increase; I/O performance will either grow slowly or be completely stagnant. Essentially, the rate at which data are generated will grow much faster than the rate at which data can be read from and written to the disk. MD simulations will soon face the I/O problem of efficiently writing to and reading from disk on the next generation of supercomputers. This article targets MD simulations at the exascale and proposes a novel technique for in situ data analysis and indexing of MD trajectories. Our technique maps individual trajectories' substructures (i.e., α-helices, ß-strands) to metadata frame by frame. The metadata captures the conformational properties of the substructures. The ensemble of metadata can be used for automatic, strategic analysis within a trajectory or across trajectories, without manually identify those portions of trajectories in which critical changes take place. We demonstrate our technique's effectiveness by applying it to 26.3k helices and 31.2k strands from 9917 PDB proteins and by providing three empirical case studies. © 2017 Wiley Periodicals, Inc.

WeFold: a coopetition for protein structure prediction.

The protein structure prediction problem continues to elude scientists. Despite the introduction of many methods, only modest gains were made over the last decade for certain classes of prediction targets. To address this challenge, a social-media based worldwide collaborative effort, named WeFold, was undertaken by 13 labs. During the collaboration, the laboratories were simultaneously competing with each other. Here, we present the first attempt at "coopetition" in scientific research applied to the protein structure prediction and refinement problems. The coopetition was possible by allowing the participating labs to contribute different components of their protein structure prediction pipelines and create new hybrid pipelines that they tested during CASP10. This manuscript describes both successes and areas needing improvement as identified throughout the first WeFold experiment and discusses the efforts that are underway to advance this initiative. A footprint of all contributions and structures are publicly accessible at http://www.wefold.org.

Miravirsen (SPC3649) can inhibit the biogenesis of miR-122.

MicroRNAs (miRNAs) are short noncoding RNAs, which bind to messenger RNAs and regulate protein expression. The biosynthesis of miRNAs includes two precursors, a primary miRNA transcript (pri-miRNA) and a shorter pre-miRNA, both of which carry a common stem-loop bearing the mature miRNA. MiR-122 is a liver-specific miRNA with an important role in the life cycle of hepatitis C virus (HCV). It is the target of miravirsen (SPC3649), an antimiR drug candidate currently in clinical testing for treatment of HCV infections. Miravirsen is composed of locked nucleic acid (LNAs) ribonucleotides interspaced throughout a DNA phosphorothioate sequence complementary to mature miR-122. The LNA modifications endow the drug with high affinity for its target and provide resistance to nuclease degradation. While miravirsen is thought to work mainly by hybridizing to mature miR-122 and blocking its interaction with HCV RNA, its target sequence is also present in pri- and pre-miR-122. Using new in vitro and cellular assays specifically developed to discover ligands that suppress biogenesis of miR-122, we show that miravirsen binds to the stem-loop structure of pri- and pre-miR-122 with nanomolar affinity, and inhibits both Dicer- and Drosha-mediated processing of miR-122 precursors. This inhibition may contribute to the pharmacological activity of the drug in man.

Creating supersecondary structures with BuildBeta.

BuildBeta is a feature of the ProteinShop software designed to thoroughly sample a protein conformational space given the protein's sequence of amino acids and secondary structure predictions. It targets proteins with beta sheets because they are particularly challenging to predict due to the complexity of sampling long-range strand pairings. Here we discuss some of the most difficult targets in the recent 9th Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (CASP9) and show how BuildBeta can leverage some of the most successful methods in the category "template-free modeling" by augmenting their sampling capabilities. We also discuss ongoing efforts to improve the quality of the supersecondary structures it generates.

BuildBeta--a system for automatically constructing beta sheets.

We describe a method that can thoroughly sample a protein conformational space given the protein primary sequence of amino acids and secondary structure predictions. Specifically, we target proteins with beta-sheets because they are particularly challenging for ab initio protein structure prediction because of the complexity of sampling long-range strand pairings. Using some basic packing principles, inverse kinematics (IK), and beta-pairing scores, this method creates all possible beta-sheet arrangements including those that have the correct packing of beta-strands. It uses the IK algorithms of ProteinShop to move alpha-helices and beta-strands as rigid bodies by rotating the dihedral angles in the coil regions. Our results show that our approach produces structures that are within 4-6 A RMSD of the native one regardless of the protein size and beta-sheet topology although this number may increase if the protein has long loops or complex alpha-helical regions.

SEA domain proteolysis determines the functional composition of dystroglycan.

Post-translational modifications of the extracellular matrix receptor dystroglycan (DG) determine its functional state, and defects in these modifications are linked to muscular dystrophies and cancers. A prominent feature of DG biosynthesis is a precursor cleavage that segregates the ligand-binding and transmembrane domains into the noncovalently attached alpha- and beta-subunits. We investigate here the structural determinants and functional significance of this cleavage. We show that cleavage of DG elicits a conspicuous change in its ligand-binding activity. Mutations that obstruct this cleavage result in increased capacity to bind laminin, in part, due to enhanced glycosylation of alpha-DG. Reconstitution of DG cleavage in a cell-free expression system demonstrates that cleavage takes place in the endoplasmic reticulum, providing a suitable regulatory point for later processing events. Sequence and mutational analyses reveal that the cleavage occurs within a full SEA (sea urchin, enterokinase, agrin) module with traits matching those ascribed to autoproteolysis. Thus, cleavage of DG constitutes a control point for the modulation of its ligand-binding properties, with therapeutic implications for muscular dystrophies. We provide a structural model for the cleavage domain that is validated by experimental analysis and discuss this cleavage in the context of mucin protein and SEA domain evolution.

ProteinShop: a tool for interactive protein manipulation and steering.

We describe ProteinShop, a new visualization tool that streamlines and simplifies the process of determining optimal protein folds. ProteinShop may be used at different stages of a protein structure prediction process. First, it can create protein configurations containing secondary structures specified by the user. Second, it can interactively manipulate protein fragments to achieve desired folds by adjusting the dihedral angles of selected coil regions using an Inverse Kinematics method. Last, it serves as a visual framework to monitor and steer a protein structure prediction process that may be running on a remote machine. ProteinShop was used to create initial configurations for a protein structure prediction method developed by a team that competed in CASP5. ProteinShop's use accelerated the process of generating initial configurations, reducing the time required from days to hours. This paper describes the structure of ProteinShop and discusses its main features.

A physical approach to protein structure prediction.

We describe our global optimization method called Stochastic Perturbation with Soft Constraints (SPSC), which uses information from known proteins to predict secondary structure, but not in the tertiary structure predictions or in generating the terms of the physics-based energy function. Our approach is also characterized by the use of an all atom energy function that includes a novel hydrophobic solvation function derived from experiments that shows promising ability for energy discrimination against misfolded structures. We present the results obtained using our SPSC method and energy function for blind prediction in the 4th Critical Assessment of Techniques for Protein Structure Prediction competition, and show that our approach is more effective on targets for which less information from known proteins is available. In fact our SPSC method produced the best prediction for one of the most difficult targets of the competition, a new fold protein of 240 amino acids.