Multi-fidelity Bayesian optimization (MFBO) is a powerful approach that utilizes low-fidelity, cost-effective sources to expedite the exploration and exploitation of a high-fidelity objective function. Existing MFBO methods with theoretical foundations either lack justification for performance improvements over single-fidelity optimization or rely on strong assumptions about the relationships between fidelity sources to construct surrogate models and direct queries to low-fidelity sources. To mitigate the dependency on cross -fidelity assumptions while maintaining the advantages of low-fidelity queries, we introduce a random sampling and partition-based MFBO framework with deep kernel learning. This framework is robust to cross-fidelity model misspecification and explicitly illustrates the benefits of low-fidelity queries. Our results demonstrate that the proposed algorithm effectively manages complex cross-fidelity relationships and efficiently optimizes the target fidelity function.

Selecting an optimal antigen is a crucial step in vaccine development, significantly influencing both the vaccine’s effectiveness and the breadth of protection it provides. High antigen sequence variability, as seen in pathogens like rhinovirus, HIV, influenza virus, complicates the design of a single cross-protective antigen. Consequently, vaccination with a single antigen molecule often confers protection against only a single variant. In this study, machine learning methods were applied to the design of factor H binding protein (fHbp), an antigen from the bacterial pathogen Neisseria meningitidis. The vast number of potential antigen mutants presents a significant challenge for improving fHbp antigenicity. Moreover, limited data on antigen-antibody binding in public databases constrains the training of machine learning models. To address these challenges, we used computational models to predict fHbp properties and machine learning was applied to select both the most promising and informative mutants using a Gaussian process (GP) model. These mutants were experimentally evaluated to both confirm promising leads and refine the machine learning model for future iterations. In our current model, mutants were designed that enabled the transfer of fHbp v1.1 specific conformational epitopes onto fHbp v3.28, while maintaining binding to overlapping cross-reactive epitopes. The top mutant identified underwent biophysical and x-ray crystallographic characterization to confirm that the overall structure of fHbp was maintained throughout this epitope engineering experiment. The integrated strategy presented here could form the basis of a next-generation, iterative antigen design platform, potentially accelerating the development of new broadly protective vaccines.

Deep Symbolic Optimization (DSO) is a novel computational framework that enables symbolic optimization for scientific discovery, particularly in applications involving the search for intricate symbolic structures. One notable example is equation discovery, which aims to automatically derive mathematical models expressed in symbolic form. In DSO, the discovery process is formulated as a sequential decision-making task. A generative neural network learns a probabilistic model over a vast space of candidate symbolic expressions, while reinforcement learning strategies guide the search toward the most promising regions. This approach integrates gradient-based optimization with evolutionary and local search techniques, and it incorporates in-situ constraints, domain-specific priors, and advanced policy optimization methods. The result is a robust framework capable of efficiently exploring extensive search spaces to identify interpretable and physically meaningful models. Extensive evaluations on benchmark problems have demonstrated that DSO achieves state-of-the-art performance in both accuracy and interpretability. In this chapter, we provide a comprehensive overview of the DSO framework and illustrate its transformative potential for automating symbolic optimization in scientific discovery.

We consider Bayesian algorithm execution (BAX), a framework for efficiently selecting evaluation points of an expensive function to infer a property of interest encoded as the output of a base algorithm. Since the base algorithm typically requires more evaluations than are feasible, it cannot be directly applied. Instead, BAX methods sequentially select evaluation points using a probabilistic numerical approach. Current BAX methods use expected information gain to guide this selection. However, this approach is computationally intensive. Observing that, in many tasks, the property of interest corresponds to a target set of points defined by the function, we introduce PS-BAX, a simple, effective, and scalable BAX method based on posterior sampling. PS-BAX is applicable to a wide range of problems, including many optimization variants and level set estimation. Experiments across diverse tasks demonstrate that PS-BAX performs competitively with existing baselines while being significantly faster, simpler to implement, and easily parallelizable, setting a strong baseline for future research. Additionally, we establish conditions under which PS-BAX is asymptotically convergent, offering new insights into posterior sampling as an algorithm design paradigm.

ABSTRACTThe COVID-19 pandemic underscored the promise of monoclonal antibody-based prophylactic and therapeutic drugs1–3, but also revealed how quickly viral escape can curtail effective options4, 5. With the emergence of the SARS-CoV-2 Omicron variant in late 2021, many clinically used antibody drug products lost potency, including EvusheldTMand its constituent, cilgavimab4, 6. Cilgavimab, like its progenitor COV2-2130, is a class 3 antibody that is compatible with other antibodies in combination4and is challenging to replace with existing approaches. Rapidly modifying such high-value antibodies with a known clinical profile to restore efficacy against emerging variants is a compelling mitigation strategy. We sought to redesign COV2-2130 to rescue in vivo efficacy against Omicron BA.1 and BA.1.1 strains while maintaining efficacy against the contemporaneously dominant Delta variant. Here we show that our computationally redesigned antibody, 2130-1-0114-112, achieves this objective, simultaneously increases neutralization potency against Delta and many variants of concern that subsequently emerged, and provides protectionin vivoagainst the strains tested, WA1/2020, BA.1.1, and BA.5. Deep mutational scanning of tens of thousands pseudovirus variants reveals 2130-1-0114-112 improves broad potency without incurring additional escape liabilities. Our results suggest that computational approaches can optimize an antibody to target multiple escape variants, while simultaneously enriching potency. Because our approach is computationally driven, not requiring experimental iterations or pre-existing binding data, it could enable rapid response strategies to address escape variants or pre-emptively mitigate escape vulnerabilities.

We study Bayesian optimization (BO) in high-dimensional and non-stationary scenarios. Existing algorithms for such scenarios typically require extensive hyperparameter tuning, which limits their practical effectiveness. We propose a framework, called BALLET, which adaptively filters for a high-confidence region of interest (ROI) as a superlevel-set of a nonparametric probabilistic model such as a Gaussian process (GP). Our approach is easy to tune, and is able to focus on local region of the optimization space that can be tackled by existing BO methods. The key idea is to use two probabilistic models: a coarse GP to identify the ROI, and a localized GP for optimization within the ROI. We show theoretically that BALLET can efficiently shrink the search space, and can exhibit a tighter regret bound than standard BO without ROI filtering. We demonstrate empirically the effectiveness of BALLET on both synthetic and real-world optimization tasks.

Symbolic optimization methods have been used to solve varied challenging and relevant problems such as symbolic regression and neural architecture search. However, the current state of the art typically learns each problem from scratch and is unable to leverage pre-existing knowledge and datasets that are available for many applications. Inspired by the similarity between sequence representations learned in natural language processing and the formulation of symbolic optimization as a discrete sequence optimization problem, we propose language model-accelerated deep symbolic optimization (LA-DSO), a method that leverages language models to learn symbolic optimization solutions more efficiently. We demonstrate LA-DSO in two tasks: symbolic regression, which allows us to perform extensive experimentation due to its low computation requirements, and computational antibody optimization, which shows that our proposal accelerates learning in challenging real-world problems.

Alchemical free energy perturbation (FEP) is a rigorous and powerful technique to calculate the free energy difference between distinct chemical systems. Here we report our implementation of automated large-scale FEP calculations, using the Amber software package, to facilitate antibody design and evaluation. In combination with Hamiltonian replica exchange, our FEP simulations aim to predict the effect of mutations on both the binding affinity and the structural stability. Importantly, we incorporate multiple strategies to faithfully estimate the statistical uncertainties in the FEP results. As a case study, we apply our protocols to systematically evaluate variants of the m396 antibody for their conformational stability and their binding affinity to the spike proteins of SARS-CoV-1 and SARS-CoV-2. By properly adjusting relevant parameters, the particle collapse problems in the FEP simulations are avoided. Furthermore, large statistical errors in a small fraction of the FEP calculations are effectively reduced by extending the sampling, such that acceptable statistical uncertainties are achieved for the vast majority of the cases with a modest total computational cost. Finally, our predicted conformational stability for the m396 variants is qualitatively consistent with the experimentally measured melting temperatures. Our work thus demonstrates the applicability of FEP in computational antibody design.

Understanding the degree of humanness of antibody sequences is critical to the therapeutic antibody development process to reduce the risk of failure modes like immunogenicity or poor manufacturability. We introduce AbBERT, a transformer-based language model trained on up to 20 million unpaired heavy/light chain sequences from the Observed Antibody Space database. We first validate AbBERT using a novel “multi-mask” scoring procedure to demonstrate high accuracy in predicting complementary determining regions—including the challenging hypervariable H3 region. We then demonstrate several uses of AbBERT at various points along the antibody design process. AbBERT enhances in silico antibody optimization via deep reinforcement learning by utilizing its learned embeddings as additional observations during optimization. Within a larger computational antibody design platform, AbBERT has been successfully applied as an additional design objective, where it displays strong correlations with computational tools predicting antibody structural stability. Finally, mutant antibody sequences that have been scored as unfavorable by AbBERT have shown corresponding low yields when expressed in cells. These use cases demonstrate the power of language modeling within computational antibody design.

Rapid, high throughput measurements of biomolecular interactions are essential across medicine and bioscience, from genetic disease studies to drug screening and vaccine development. For example, de novo antibody development against SARS-CoV-2 during a pandemic requires iterative computational design and experimental testing of hundreds of thousands of candidate antibodies against mutating virus proteins, all within days to weeks. Here we present a platform aimed at production and in situ binding measurements of thousands of different proteins. Plasmids encoding candidate antibodies and the protein antigen of interest, along with a solution of cell-free expression lysate, are enclosed inside microcapsules produced by microfluidic droplet formation. Potential interactions between fluorescently tagged proteins expressed inside the microcapsules are probed via fluorescence correlation spectroscopy (FCS). The use of microcapsules as individual reaction chambers allows for the potential to scale up to millions of measurements. As a first step, we demonstrate FCS measurements (diffusion kinetics and concentrations) taken inside microcapsules while the fluorescent proteins GFP and mCherry are being expressed. The amplitude of correlations decreased over the course of several hours, corresponding to increases in concentrations from ∼10 nM to ∼1 uM, while the sub-millisecond diffusion times remained constant, as expected. We aim to next present FCS measurements of interacting proteins produced inside the microcapsules, including scFV antibodies against the SARS-CoV-2 receptor binding domain.

SummaryRapid assessment of whether a pandemic pathogen may have increased transmissibility or be capable of evading existing vaccines and therapeutics is critical to mounting an effective public health response. Over the period of seven days, we utilized rapid computational prediction methods to evaluate potential public health implications of the emerging SARS-CoV-2 Omicron variant. Specifically, we modeled the structure of the Omicron variant, examined its interface with human angiotensin converting enzyme 2 (ACE-2) and evaluated the change in binding affinity between Omicron, ACE-2 and publicly known neutralizing antibodies. We also compared the Omicron variant to known Variants of Concern (VoC). Seven of the 15 Omicron mutations occurring in the spike protein receptor binding domain (RBD) occur at the ACE-2 cell receptor interface, and therefore may play a critical role in enhancing binding to ACE-2. Our estimates of Omicron RBD-ACE-2 binding affinities indicate that at least two of RBD mutations, Q493R and N501Y, contribute to enhanced ACE-2 binding, nearly doubling delta-delta-G (ddG) free energies calculated for other VoC’s. Binding affinity estimates also were calculated for 55 known neutralizing SARS-CoV-2 antibodies. Analysis of the results showed that Omicron substantially degrades binding for more than half of these neutralizing SARS-CoV-2 antibodies, and for roughly 10 times as many of the antibodies than the currently dominant Delta variant. This early study lends support to use of rapid computational risk assessments to inform public health decision-making while awaiting detailed experimental characterization and confirmation.

Summary Rapidly responding to novel pathogens, such as SARS-CoV-2, represents an extremely challenging and complex endeavor. Numerous promising therapeutic and vaccine research efforts to mitigate the catastrophic effects of COVID-19 pandemic are underway, yet an efficacious countermeasure is still not available. To support these global research efforts, we have used a novel computational pipeline combining machine learning, bioinformatics, and supercomputing to predict antibody structures capable of targeting the SARS-CoV-2 receptor binding domain (RBD). In 22 days, using just the SARS-CoV-2 sequence and previously published neutralizing antibody structures for SARS-CoV-1, we generated 20 initial antibody sequences predicted to target the SARS-CoV-2 RBD. As a first step in this process, we predicted (and publicly released) structures of the SARS-CoV-2 spike protein using homology-based structural modeling. The predicted structures proved to be accurate within the targeted RBD region when compared to experimentally derived structures published weeks later. Next we used our in silico design platform to iteratively propose mutations to SARS-CoV-1 neutralizing antibodies (known not to bind SARS-Cov-2) to enable and optimize binding within the RBD of SARS-CoV-2. Starting from a calculated baseline free energy of −48.1 kcal/mol (± 8.3), our 20 selected first round antibody structures are predicted to have improved interaction with the SARS-CoV-2 RBD with free energies as low as −82.0 kcal/mole. The baseline SARS-CoV-1 antibody in complex with the SARS-CoV-1 RBD has a calculated interaction energy of −52.2 kcal/mole and neutralizes the virus by preventing it from binding and entering the human ACE2 receptor. These results suggest that our predicted antibody mutants may bind the SARS-CoV-2 RBD and potentially neutralize the virus. Additionally, our selected antibody mutants score well according to multiple antibody developability metrics. These antibody designs are being expressed and experimentally tested for binding to COVID-19 viral proteins, which will provide invaluable feedback to further improve the machine learning–driven designs. This technical report is a high-level description of that effort; the Supplementary Materials includes the homology-based structural models we developed and 178,856 in silico free energy calculations for 89,263 mutant antibodies derived from known SARS-CoV-1 neutralizing antibodies.

Objectives Unplanned readmissions to the intensive care unit (ICU) are highly undesirable, increasing variance in care, making resource planning difficult and potentially increasing length of stay and mortality in some settings. Identifying patients who are likely to suffer unplanned ICU readmission could reduce the frequency of this adverse event. Setting A single academic, tertiary care hospital in the UK. Participants A set of 3326 ICU episodes collected between October 2014 and August 2016. All records were of patients who visited an ICU at some point during their stay. We excluded patients who were ≤16 years of age; visited ICUs other than the general and neurosciences ICU; were missing crucial electronic patient record measurements; or had indeterminate ICU discharge outcomes or very early or extremely late discharge times. After exclusion, 2018 outcome-labelled episodes remained. Primary and secondary outcome measures Area under the receiver operating characteristic curve (AUROC) for prediction of unplanned ICU readmission or in-hospital death within 48 hours of first ICU discharge. Results In 10-fold cross-validation, an ensemble predictor was trained on data from both the target hospital and the Medical Information Mart for Intensive Care (MIMIC-III) database and tested on the target hospital’s data. This predictor discriminated between patients with the unplanned ICU readmission or death outcome and those without this outcome, attaining mean AUROC of 0.7095 (SE 0.0260), superior to the purpose-built Stability and Workload Index for Transfer (SWIFT) score (AUROC=0.6082, SE 0.0249; p=0.014, pairwise t-test). Conclusions Despite the inherent difficulties, we demonstrate that a novel machine learning algorithm based on transfer learning could achieve good discrimination, over and above that of the treating clinicians or the value added by the SWIFT score. Accurate prediction of unplanned readmission could be used to target resources more efficiently.

Early detection of pediatric severe sepsis is necessary in order to administer effective treatment. In this study, we assessed the efficacy of a machine-learning-based prediction algorithm applied to electronic healthcare record (EHR) data for the prediction of severe sepsis onset. The resulting prediction performance was compared with the Pediatric Logistic Organ Dysfunction score (PELOD-2) and pediatric Systemic Inflammatory Response Syndrome score (SIRS) using cross-validation and pairwise t -tests. EHR data were collected from a retrospective set of de-identified pediatric inpatient and emergency encounters drawn from the University of California San Francisco (UCSF) Medical Center, with encounter dates between June 2011 and March 2016. Patients (n = 11,127) were 2-17 years of age and 103 [0.93%] were labeled severely septic. In four-fold cross-validation evaluations, the machine learning algorithm achieved an AUROC of 0.912 for discrimination between severely septic and control pediatric patients at onset and AUROC of 0.727 four hours before onset. Under the same measure, the prediction algorithm also significantly outperformed PELOD-2 ( p < 0.05) and SIRS ( p < 0.05) in the prediction of severe sepsis four hours before onset. This machine learning algorithm has the potential to deliver high-performance severe sepsis detection and prediction for pediatric inpatients.

Algorithm-based clinical decision support (CDS) systems associate patient-derived health data with outcomes of interest, such as in-hospital mortality. However, the quality of such associations often depends on the availability of site-specific training data. Without sufficient quantities of data, the underlying statistical apparatus cannot differentiate useful patterns from noise and, as a result, may underperform. This initial training data burden limits the widespread, out-of-the-box, use of machine learning-based risk scoring systems. In this study, we implement a statistical transfer learning technique, which uses a large "source" data set to drastically reduce the amount of data needed to perform well on a "target" site for which training data are scarce. We test this transfer technique with AutoTriage, a mortality prediction algorithm, on patient charts from the Beth Israel Deaconess Medical Center (the source) and a population of 48 249 adult inpatients from University of California San Francisco Medical Center (the target institution). We find that the amount of training data required to surpass 0.80 area under the receiver operating characteristic (AUROC) on the target set decreases from more than 4000 patients to fewer than 220. This performance is superior to the Modified Early Warning Score (AUROC: 0.76) and corresponds to a decrease in clinical data collection time from approximately 6 months to less than 10 days. Our results highlight the usefulness of transfer learning in the specialization of CDS systems to new hospital sites, without requiring expensive and time-consuming data collection efforts.

Background Sepsis is one of the leading causes of mortality in hospitalized patients. Despite this fact, a reliable means of predicting sepsis onset remains elusive. Early and accurate sepsis onset predictions could allow more aggressive and targeted therapy while maintaining antimicrobial stewardship. Existing detection methods suffer from low performance and often require time-consuming laboratory test results. Objective To study and validate a sepsis prediction method, InSight, for the new Sepsis-3 definitions in retrospective data, make predictions using a minimal set of variables from within the electronic health record data, compare the performance of this approach with existing scoring systems, and investigate the effects of data sparsity on InSight performance. Methods We apply InSight, a machine learning classification system that uses multivariable combinations of easily obtained patient data (vitals, peripheral capillary oxygen saturation, Glasgow Coma Score, and age), to predict sepsis using the retrospective Multiparameter Intelligent Monitoring in Intensive Care (MIMIC)-III dataset, restricted to intensive care unit (ICU) patients aged 15 years or more. Following the Sepsis-3 definitions of the sepsis syndrome, we compare the classification performance of InSight versus quick sequential organ failure assessment (qSOFA), modified early warning score (MEWS), systemic inflammatory response syndrome (SIRS), simplified acute physiology score (SAPS) II, and sequential organ failure assessment (SOFA) to determine whether or not patients will become septic at a fixed period of time before onset. We also test the robustness of the InSight system to random deletion of individual input observations. Results In a test dataset with 11.3% sepsis prevalence, InSight produced superior classification performance compared with the alternative scores as measured by area under the receiver operating characteristic curves (AUROC) and area under precision-recall curves (APR). In detection of sepsis onset, InSight attains AUROC = 0.880 (SD 0.006) at onset time and APR = 0.595 (SD 0.016), both of which are superior to the performance attained by SIRS (AUROC: 0.609; APR: 0.160), qSOFA (AUROC: 0.772; APR: 0.277), and MEWS (AUROC: 0.803; APR: 0.327) computed concurrently, as well as SAPS II (AUROC: 0.700; APR: 0.225) and SOFA (AUROC: 0.725; APR: 0.284) computed at admission (P<.001 for all comparisons). Similar results are observed for 1-4 hours preceding sepsis onset. In experiments where approximately 60% of input data are deleted at random, InSight attains an AUROC of 0.781 (SD 0.013) and APR of 0.401 (SD 0.015) at sepsis onset time. Even with 60% of data missing, InSight remains superior to the corresponding SIRS scores (AUROC and APR, P<.001), qSOFA scores (P=.0095; P<.001) and superior to SOFA and SAPS II computed at admission (AUROC and APR, P<.001), where all of these comparison scores (except InSight) are computed without data deletion. Conclusions Despite using little more than vitals, InSight is an effective tool for predicting sepsis onset and performs well even with randomly missing data.

BACKGROUND:Health information technologies can assist clinicians in the Intensive Care Unit (ICU) by providing additional analysis of patient stability. However, because patient diagnoses can be confounded by chronic alcohol use, the predictive value of existing systems is suboptimal. Through the use of Electronic Health Records (EHR), we have developed computer software called AutoTriage to generate accurate predictions through multi-dimensional analysis of clinical variables. We analyze the performance of AutoTriage on the Alcohol Use Disorder (AUD) subpopulation in this study, and build on results we reported for AutoTriage performance on the general population in previous work.METHODS:AUD-related ICD-9 codes were used to obtain a patient population from MIMIC III ICU dataset for a retrospective study. Patient mortality risk score is generated through analysis of eight EHR-based clinical variables. The score is determined by combining weighted subscores, each of which are obtained from singlets, doublets or triplets of one or more of the eight continuous-valued clinical variable inputs. A temporally updating risk score is computed with a continuously revised 12-hour mortality prediction.RESULTS:Among AUD patients, in a non-overlapping test set, AutoTriage outperforms existing systems with an Area Under Receiver Operating Characteristic (AUROC) value of 0.934 for 12-h mortality prediction. At a sensitivity of 90%, AutoTriage achieves a specificity of 80%, positive predictive value of 40%, negative predictive value of 89%, and an Odds Ratio of 36.CONCLUSIONS:For mortality prediction, AutoTriage demonstrates improvements in both the accuracy and the Odds Ratio over current systems among the AUD patient population.

Background: Clinical decision support systems are used to help predict patient stability and mortality in the Intensive Care Unit (ICU). Accurate patient information can assist clinicians with patient management and in allocating finite resources. However, systems currently in common use have limited predictive value in the clinical setting. The increasing availability of Electronic Health Records (EHR) provides an opportunity to use medical information for more accurate patient stability and mortality prediction in the ICU.Objective: Develop and evaluate an algorithm which more accurately predicts patient mortality in the ICU, using the correlations between widely available clinical variables from the EHR.Methods: We have developed an algorithm, AutoTriage, which uses eight common clinical variables from the EHR to assign patient mortality risk scores. Each clinical variable produces a subscore, and combinations of two or three discretized clinical variables also produce subscores. A combination of weighted subscores produces the overall score. We validated the performance of this algorithm in a retrospective study on the MIMIC III medical ICU dataset.Results: AutoTriage 12 h mortality prediction yields an Area Under Receiver Operating Characteristic value of 0.88 (95% confidence interval 0.86 to 0.88). At a sensitivity of 80%, AutoTriage maintains a specificity of 81% with a diagnostic odds ratio of 16.26.Conclusions: Through the multidimensional analysis of the correlations between eight common clinical variables, AutoTriage provides an improvement in the specificity and sensitivity of patient mortality prediction over existing prediction methods. (C) 2016 The Authors. Published by Elsevier Ltd on behalf of IJS Publishing Group Ltd.