Abstract: Artificial Intelligence (AI) is being adopted across the world and promises a new revolution in healthcare. Our work delves into global regulatory approaches towards AI use in healthcare, with a focus on how common global themes are emerging. We compare these themes to WHO’s regulatory considerations and principles on ethical use of AI for healthcare applications. Our work seeks to take a global perspective on AI policy by analyzing 25 policy, strategy, guidance-based documents, and laws across 14 legal jurisdictions representative of various regions in the world (EU, UK, Australia, Canada, Japan, Italy, Brazil, Egypt, Rwanda, Saudi Arabia, Singapore, India, China and Hong Kong). We analyze the status, direction and existing gaps in AI healthcare regulations. We analyze regulations on four comparative parameters: sector-agnostic generic AI regulations, healthcare-specific AI regulations, non-binding instruments and binding legal instruments. We examine the timeline of jurisprudence evolution in AI globally. We propose solutions to promote international harmonization of AI regulations and examine the requirements for regulating generative AI in healthcare, using China and Singapore as examples of countries with well-developed policies in this area. The need for comprehensive healthcare-AI laws is needed. However, our work notes examples of success of soft-law approaches which allow for the flexibility and adaptability necessary for healthcare regulations, given the differing opinions on generic AI regulation. We discuss how countries such as Singapore are also incorporating AI regulation into existing documents by addressing them as a subsection in guidance documents. We provide nation-specific analysis: most nations are still in the strategy and policy stages of sector-agnostic generic AI regulation, while nations in the Middle East, Southeast Asia and Africa are also picking up momentum in healthcare-specific legislation. Our findings highlight a shared commitment to aligning with the WHO’s ethical AI principles. Link to paper: [2406.08695] Global AI Governance in Healthcare: A Cross-Jurisdictional Regulatory Analysis (arxiv.org)   read more...   read less

Abstract: Background: Drug-induced liver injury (DILI) is a leading cause of acute liver failure and a common reason for the withdrawal of approved drugs from the market. Drug metabolism can play a significant role in the induction of DILI. While some studies have been conducted to investigate how drug interaction with cytochrome P450 (non-CYP) enzymes can be associated with DILI, there has been no comprehensive investigation on the association between non-CYP enzymes and DILI. Methods: For 42 non-CYP enzymes, we gathered data on how 317 drugs interacted with the enzymes as substrates, inhibitors and/or inducers from the Pharmapendium database. The potential of these drugs to induce DILI was obtained from the LTKB database. We performed multivariate logistic regression on the dataset to investigate associations between the drugs’ potential for DILI concern and enzyme families, individual enzymes, types of drug-enzyme interaction. Results: Generally, inhibitors of non-CYP enzymes were associated with high DILI concern. In particular, drug interaction with UDP-glucuronosyltransferase (UGT) enzymes increased odds of a drug having high DILI concern. UGT inhibitors were found to be significantly associated with an increased likelihood of having high DILI concern. Drugs that are UGT substrates were also found to increase likelihood of possessing severe DILI outcomes, but only when they as dual substrate-inhibitors of UGT enzymes. Drug-UGT interaction, when combined with the previously proposed rule-of-two (RO2) had improved predictive performance. Conclusions: This shows that drug interaction with UGT enzymes can serve as a valuable predictor of severe DILI, expanding the tools that are available to predict DILI.   read more...   read less

Abstract: We describe our work in preparing FAIR datasets and knowledge graphs connecting compound information, biological response data, pathways and key events. We have prepared two knowledge toxicology resource examples for case study work: a) a dataset and knowledge graph based on a large European program on New Approach Methods in Toxicology (EU-ToxRisk ), b) a knowledge graph based on a network model for steatosis (supporting ASPIS case study work). The related open knowledge resources include a harmonised data template for organising data generated on the project according to FAIR principles; omics data management best practices and implementation; approach to collected data including harmonisation, referencing and integrity goals; compound database including approach to compound ids, properties, project and case study annotation, linked biological data, batch upload, curation, search and user interface. We discuss through case study examples how the resources can be used in the support of AI-enabled risk assessment case study work and its effective documentation. Our test cases include: a) Enabling effective retrieval of reliable well-referenced linked background knowledge organised into a knowledge graph; b) Enhanced toxicology learning of Machine Learning- and Generative AI- based models; c) Incorporation of AI knowledge in enchancing risk assessment workflow tasks, including analysis, evidence interpretation and decisions. d) Generation of reports documenting risk assessment workflow processes, including high integrity auditable supporting data and knowledge records. We will share our learning from the above cases with regards to good practices, utility and reliability. We will discuss collaboration opportunities include sharing open science resources supporting NAMS-based risk assessment with the regulatory acceptance purpose, the benefit of FAIR practices and approaches to sustainability.   read more...   read less

Abstract: The SSbD4ChEM (Safe and Sustainable by Design framework for the next generation of Chemicals and Materials) project brings together stakeholders from industry, government, academia, and civil society to develop and promote best practices for safe and sustainable product and process design, through demonstration in industrial case studies (2024-2027). SSbD4CheM aims to meet the EU's strategic objectives for digital, enabling, and emerging technologies, sectors, and value chains by developing a comprehensive Safe and Sustainable by Design (SSbD) framework that uses new science-based approaches to identify and address potential hazards and risks, and innovative technologies to support the design of safer and more sustainable products and processes. SSbD4CheM objectives include: 1) SSbD framework implementation to facilitate development of the next generation of chemicals and materials applicable to renewable composites in automotive value chain, PFAS-free coatings for textiles and cellulose nanofibers as additive in cosmetics to replace plastic microbeads; 2) Development of efficient hazard screening tools for alternative assessment of next generation chemicals/materials/products combining in silico tools and multicriteria analysis; 3) Advanced explorative ex-ante Life Cycle Analysis methodology supported by molecular and data-driven modelling to fill data gaps for (novel) materials and chemicals; 4) Integration of chemical/(nano-)material characterization methods to assess products’ environmental safety, quantify emissions to support occupational and consumer safety limits; 5) Development of alternative in silico and in vitro models, as part of New Approach Methodologies for adequate exposure scenarios; 6) Harmonization and validation of analytical and toxicological methods for proposition to regulatory and standardization bodies to contribute to the development of new standards and facilitate market acceptance of SSbD4CheM project results; and 7) Enhance overall SSbD4CheM impact through stakeholder engagement, training, dissemination, exploitation and industrial innovation. We describe the work carried out during the initial phase of the project to define the SSbD4CheM framework for Safe and Sustainable by Design (SSbD). The goal of the framework is to provide structured knowledge guidance on the organisation of SSbD tasks as carried out in case studies. The framework specifies our strategy, methodology and concepts for implementation against SSbD goals. Our strategy involves a tiered iterative approach to SSbD with all activities described in workflows documenting case study tasks, results and decisions. At this stage of the project we are interested in engaging in valuable outreach discussions with stakeholders to obtain valuable feedback and suggestions including potential collaboration and extension of the framework implementation to additional case studies. Keywords: Safe by Design, Sustainability, SSbD, framework, methodology, New Approach Methodologies, workflow, case studies   read more...   read less

Abstract: Titanium dioxide (TiO2) has been incorporated into many consumer and industrial products ranging from food, cosmetics, and medical devices as a photocatalyst and white pigment. Recently, the use of TiO2 in food products was banned by the European Food Safety Authority (EFSA) leading to the need of the re-evaluation of TiO2 for potential health risks (i.e., immune system toxicity). Currently, there is a lack of data using blood derived peripheral blood mononuclear cells (PBMCs) and plasma to examine the potential immunotoxicity of TiO2 and potential differences between male and female immune responses. An in vitro study was conducted to investigate the immunotoxicity and potential sex-based differences from TiO2 exposure using PBMCs and plasma from healthy donors. Two brands of TiO2, which we have defined as E171-1 and E171-2, were dispersed and characterized for testing on inflammasome activation, chemokine and cytokine expression, leukocyte proliferation, plasma coagulation, and complement activation. The results indicated that there are sex-based differences. For E171-1, the females showed more IL-10 and less IL-6 than the males. For E171-2, the females showed more inflammasome activity, IL-8, and MCP-1 than the males. E171-1 exposure increased leukocyte proliferation, whereas E171-2 caused limited effects. Neither TiO2 affected the plasma coagulation nor complement activation. This study demonstrated the importance of considering sex-based differences in immune responses from the exposure to TiO2 in any biocompatibility assessment.   read more...   read less

Abstract: This study investigates the transgenerational effects of perinatal antiretroviral therapy (ART) exposure on gut microbial composition and metabolic function in aged rat offspring. While ART has significantly reduced mother-to-child HIV transmission, its long-term impact on offspring health remains unclear. We examined how indirect exposure to a tricombo-ART during gestation affects gut microbiota and short chain fatty acid (SCFA) production in 12-month-old rat offspring. Pregnant Sprague Dawley rats (n=5/group) were administered with two different doses of ART via oral gavage from gestation day 6 to lactation time (Postpartum 21). Aged rat offspring’s fecal samples were analyzed using shotgun sequencing (for microbial profiling) and HPLC (for SCFA quantification). Dose-dependent and gender-specific alterations in fecal microbial abundances were observed at various taxonomic levels. Key changes include a decline in Firmicutes in males and an increase in Actinobacteria in both genders, with the gender specific reorganizations in Lactobacillus, Bifidobacterium, and Akkermansia distribution. Despite these compositional changes, no significant differences in SCFA concentrations were observed between treatment groups or genders, possibly due to high individual variability and complex interacting factors such as compensatory mechanism of SCFA production via reorganized microbial population. Furthermore, analysis for the functional profiling of microbial community provided insight into the pathway abundance and molecular activities impacted by the ART. This study highlights the importance of considering individual variability and gender specific responses and research on ART’s transgenerational effect on gut microbiome and metabolic function. It also underscores the need for multi-omics approach to unravel the intricate relationships between the gut microbiome, metabolic functions, and the host health. Also, a longitudinal assessment would be valuable to understand if effects are transitory in nature and where the biggest impacts are.   read more...   read less

Abstract: The Singapore Food Agency (SFA) was formed to bring together all food-related resources and capabilities for holistic management of the food industry “from farm to fork”, with the mission to ensure and secure a supply of safe food. To enhance the efficiency of our work processes, SFA has embarked on a digital transformation journey to implement a comprehensive strategy for digitalisation and transformation throughout the organization. This initiative involves strengthening digital capabilities and utilizing innovative technologies and data science to advance research in food safety. SFA aims to facilitate the rapid adoption of emerging technologies by proactively scanning the horizon and encouraging the piloting of new tools. Additionally, SFA is leveraging digital technologies such as CCTVs and Video Analytics to enhance operations, along with an integrated Ops centre and enhanced Enterprise Data Lake to ensure food safety. This holistic approach is designed to improve the organization's ability to monitor and manage food safety through the integration of digital tools and data management systems. Furthermore, SFA is harnessing advancements in GenAI and natural language processing to revolutionize the interpretation of investigation reports and enhance regulatory science capabilities. By continuously pushing boundaries with innovative tools and technologies, SFA is committed to embracing digital transformation, fostering collaboration, and enhancing scientific capabilities in critical areas of food safety while harnessing the potential of advanced AI and natural language processing technologies.   read more...   read less

Background: Pharmacovigilance plays a crucial role in ensuring the safety of pharmaceutical products. It involves the systematic monitoring of adverse events and the detection of potential safety concerns related to drugs. Manual literature screening for pharmacovigilance related articles is a labor-intensive and time-consuming task, requiring streamlined solutions to cope with the continuous growth of literature. This study explores the application of Large Language Models (LLMs) to automate literature screening, aiming to expedite the process and pinpoint articles relevant to the focused area of pharmacovigilance. The overarching objective is to enhance the efficiency and accuracy of article identification, contributing to the ongoing efforts to ensure drug safety. Methods: In this research, we assess the performance of leading LLMs, including GPT-3.5, GPT-4, and Anthropic's Claude2, in automating the categorization of medical publications as relevant or irrelevant for safety signal reviews. We employ a pre-labeled dataset to evaluate the effectiveness of these models. Our analysis encompasses the assessment of N-shot learning, the elucidation of chain-of-thought reasoning, the determination of sensitivity and specificity, and the exploration of factors impacting accuracy. Results: The findings highlight the promising potential of LLMs in literature screening, showcasing notable strengths in terms of reproducibility and sensitivity, although demonstrating a moderate specificity. Notably, when provided with a few examples of abstracts, labels, and corresponding reasoning explanations, we observed an improvement in categorization performance. Moreover, our exploration identified several potential contributing factors influencing prediction outcomes. These factors encompassed the choice of key words and prompt formats, the balance within the training data, and variations in reasoning explanations. Conclusions: By configuring advanced LLMs for swift screening of extensive literature databases, the potential for transformative impact in drug safety monitoring becomes evident. Furthermore, this innovative approach holds promise for extension to various other pharmacovigilance tasks that entail processing large volumes of text data.   read more...   read less

Abstract: This case study is based on 18 months of work with the Quality Control team of one of the largest Pharmaceutical Manufacturing Firm, headquartered in India, operating in 11 countries. Producing medicines for Global consumption, the regulatory and compliance requirements are stringent and non-negotiable. The challenge faced, was despite process automation, detailed SOPs and extensive training, they experienced repeat errors by analysts. Errors seemed random, no identifiable pattern hence ineffective corrective action. The diagnosis revealed the root cause of these errors was mostly behavioural; like lack of attention, unwilling to unlearn outdated practices, complacency etc. The aim of this study was to utilise principles of persuasive technology, nudge technique, gamification and critical mass effect on the participants behaviour resulting in reduced errors. Methodology: Identify participant’s motivation, interests and concerns regarding their work &/ work environment and a persuasion style that would influence their mindset. Basis the persuasion style, using AI to personalise the nudge for participants to choose the right behaviour. Gamification features are used to keep the participants engaged over the 18 months. These included challenges, team competitions, rewards, feedback, streaks, skill development, fun. Critical Mass effect- 65% of the study population consistently participating during this 18 month period made the behavioural changes sustainable. Results: • The first 3 months of the study saw a 27% drop in errors. At 18 months a 46% reduction of errors. • The actively engaged participants during the entire period were 3 times less likely to make mistakes Conclusion: Behavioural compliance is critical in pharmaceutical manufacturing. While automation, SOPs and repetitive training are essential, mindset alignment of a large non-homogenous employee population can be expedited using the principles of persuasive technology, nudge, gamification and critical mass effect.    read more...   read less

Abstract: Lipid composition is a critical quality attribute (CQA) of liposomal formulations since it affects their stability, drug release, biodistribution, safety, and efficacy. We have developed three test methods for the quantitation of liposomal lipid composition [1,2]. Two of them utilized high-performance liquid chromatography with charged aerosol detector (HPLC-CAD), and high-performance liquid chromatography with evaporative light scattering detector (HPLC-ELSD) to quantify three major lipid components that are present in liposomal doxorubicin and other FDA-approved drug products: cholesterol, 1,2-distearoyl-sn-glycero-3-phosphoethanolamine-N-[methoxy(polyethylene glycol)-2000] (DSPE-PEG 2000) and hydrogenated (Soy) L-α-phosphatidylcholine (HSPC) [3,4]. Both test methods were published through ASTM International in collaboration with stakeholders and subject matter experts from industry, regulatory and research agencies, and academia. An interlaboratory study (ILS) in accordance with the standard practice E691 published by ASTM International was conducted with a goal of determining precision and repeatability, which determines the variations in the measured lipid quantities in liposomal formulation, and allow potential users to assess the usefulness of the test method in proposed applications.[5] In this ILS, a total of 47 laboratories from 8 countries reported three replicate test results for each test material provided (lipid in powder form and a liposomal/nanoparticle formulation). The repeatability limit and the reproducibility limits for HPLC-CAD were determined through statistical examination of 384 data points from 32 participants. Similarly, the precision for HPLC-ELSD was determined by analyzing 180 data points, from 15 laboratories. This study provides a substantial data set that successfully built confidence in and demonstrated the effectiveness of multiple ASTM standard test methods that are applicable to pharmaceutical, injectable drug products. Disclaimer: This presentation reflects the views of the authors and does not necessarily reflect those of the U.S. Food and Drug Administration. Acknowledgment: This study was sponsored by ASTM International and supported by an Interagency Agreement between the U.S. FDA/NCTR and the National Institute of Environmental Health Sciences/National Toxicology Program. References: 1. Guidance for Industry: Liposome Drug Products: Chemistry, Manufacturing, and Controls; Human Pharmacokinetics and Bioavailability; and Labeling Documentation, April 2018; 2. Nanotechnology—Over a Decade of Progress and Innovation; 3. ASTM E3323-21—Standard Test Method for Lipid Quantitation in Liposomal Formulations Using High Performance Liquid Chromatography (HPLC) with an Evaporative Light-Scattering Detector (ELSD); 4. ASTM E3297-21—Standard Test Method for Lipid Quantitation in Liposomal Formulations Using High Performance Liquid Chromatography (HPLC) with a Charged Aerosol Detector (CAD); 5. ASTM E177-20 — Standard Practice for Use of the Terms Precision and Bias in ASTM Test Methods.   read more...   read less

Abstract: Analgesics (i.e., oral and inhalation drugs) toxicity can vary significantly based on factors such as dosage, chemical structure, metabolism, and route of administration. This variability poses major challenges to the pharmaceutical industry and human health. Early and accurate prediction of analgesic drug toxicity using graph convolutional network (GCN) models offers a promising solution to mitigate these risks during drug discovery. In this study, we present the development of various classification and regression-based GCN models aimed at evaluating multiple types of toxicities, including acute toxicity, respiratory toxicity, cytochrome (CYP) 450 toxicity (specifically CYP2D6 and CYP3A4), and hepatotoxicity data obtained from TOXRIC database. To address inherent variations in data size, label type, and distribution across different toxicity types, we employed diverse training strategies. First, we used a GCN regression model to predict acute toxicity, achieving notable performance with Pearson correlation coefficients (r) for intraperitoneal, intravenous, and oral administration routes. Second, we trained multiple GCN binary classification models, each tailored to a specific type of toxicity. These models exhibited high area under-the-curve scores for predicting respiratory, CYP450, and hepatotoxicities, respectively. Finally, we utilized an approved analgesic drug dataset to verify the appropriate threshold value for the prediction scores in model usage. Our findings indicate that these GCN models have the potential to significantly reduce the cost and risk associated with drug development by expediting the selection of chemicals with low toxicity profiles.   read more...   read less

Abstract: Amlexanox is an anti-inflammatory asthma drug (Solfa) approved in 1987 by Takeda, and in 1996 as an aphthous ulcer paste (Aphthasol)) by Block Drug. WOREE syndrome is a neurodevelopmental disease with drug-resistant epilepsy and global developmental delay through a nonsense mutation leading to a premature termination codon (PTC) defect of the WWOX gene. This results in the deficiency of the oxidoreductase enzyme transcription regulator involved in tumor suppression, cell growth and differentiation. An Australian parent whose daughter was diagnosed with WOREE syndrome, posited that a readthrough therapy for the nonsense mutation might treat her symptoms, and explored many agents for treating PTC. He eventually discovered Amlexanox, which was preclinically shown to have readthrough effect and inhibit nonsense mediated decay and convinced his daughter’s neurologist to prescribe the drug in an N=1 experiment, as it was approved for chronic use in adults and children. The immediate effect was a remarkable improvement in leg movement, vocalizing and interacting, which had been absent prior to the drug treatment. Improvement lasted four months until they ran out of drug, as it was no longer marketed. After treatment stopped the patient regressed and returned to baseline. Strong scientific evidence exists showing Amlexanox may be efficacious general readthrough therapy for autosomal recessive rare diseases with loss of function nonsense mutation resulting in PTC, and that the drug may be useful in many metabolic and inflammatory diseases. NewFound is exploring access to Amelxanox for WOREE syndrome patients, plus opportunities and collaborations for basket clinical trials and natural history studies in rare diseases with similar genetic mutations. This effort will be discussed in detail and the utility of Amlexanox in multiple PTC related diseases.     read more...   read less

Abstract: AI technologies include expert rule-based systems, knowledge-based symbolic processing, natural language processing, machine learning (including QSAR and deep neural networks), and most recently large language models and generative AI. Many of these techniques were founded and developed in the 90s. Since that time, they have benefited from faster computers with more memory, vastly more data for analysis (e.g., internet sources), and algorithm evolution. However, there are still concerns regarding: dependence on data quality, bias in data and models, licensed source data, regulatory challenges, and performance (e.g., hallucinations). AI applications in regulatory predictive toxicology have been well established since 2007 and have been supporting drug safety assessments including QSAR models and structural alert expert systems for hazard and risk assessment in ICH M7 and other regulatory guidelines (including timely assessments of N-nitrosamines and Extractables and Leachables). Now, AI technologies have again burst on the scene touting algorithmic solutions promising to help solve many scientific problems. Recent OECD guidelines in the QSAR Assessment Framework (QAF) outline criteria (based on earlier validation principles for models) to help ensure data quality, objective measures of QSAR model performance and robustness, a domain of applicability, documentation, transparency, and an underlying mechanistic interpretation where possible. Not all AI techniques follow these guidelines. Consequently, some of these models would not be suitable to support regulatory assessments. Here we illustrate how Leadscope QSAR models and structural alerts adhere to the QAF and support regulatory safety assessment. This includes: extensively curating data from trusted sources (including conflict resolution from multiple sources), defining a fit-for-purpose model endpoint, reducing data bias through balancing training set data, normalizing chemical structure representation, determining the model domain, assessing performance using external validation, and providing transparency and interpretability of chemical descriptors as well as detailing how individual predictions are calculated, including mechanistic and statistical supporting evidence.   read more...   read less

Abstract: As safety concerns about COVID-19 vaccines continue, a method to more systematically evaluate the toxicity of mRNA vaccines using new technologies is needed. Distribution and residue in the body are one of the five factors that WHO must consider when evaluating the safety and toxicity of mRNA-based vaccines in addition to general toxicity tests. Studies using biodistribution and persistence is an important method to evaluate mRNA vaccines, because they are useful for predicting distribution and assessing toxicity after administration. In this study, luciferase tagged mRNA vaccine carriers were analyzed for their shape using transmission electron microscopy (TEM) and assessed for in vivo distribution using an in vivo imaging system(IVIS). Vaccine carriers were administrated once into the right thigh of SD rats. After administration of them, general symptoms were observed, and no other abnormal sysmptoms were observed in any of the subjects during the test period. The vaccine carriers with a stable diameter of about 100-300nm were observed. Movement for the vaccine carriers to the liver along with the administration site was observed at 4 and 8 hours after administration of them and the strongest luminiscence occured at 4 hours after administration. Luminiscence continued to decrease for up to 72hours, and then no further luminiscene was observed on the 7th, the final shooting day. The vaccine carriers confirmed to be distributed in the injection site(muscle), liver and spleen until 3 days after administration. This study suggests that assessment of biodistribution and persistence is expected to be useful in safety studies of mRNA vaccines. Keywords: mRNA vaccine, toxicity, biodistribution   read more...   read less

Abstract: For the difficulties faced by the startup teams in application of medical devices, we build up a solution based on modern AI technology. The whole package includes a basic plan, including simple search for FDA certification, analysis for application, and guides for classification, and a premium for smart search and comparison in similar products. On one hand, based on the technical description an user provides, the AI coupled with a machine learning model will generate the corresponding keywords to improve the accuracy in the smart search module; on the other hand, the module with NLP can identify the specification or characteristics of a product, and the generative AI can therefore compare them among relative products. Thus, our results can accelerate and accurate the search and classification in medical devices, can help any R&D team follow the FDA regulatory compliance better, and can provide anyone with the correct medical device information. We expect that this information technique with the up-to-date openFDA database can assist the regulatory authorities and medical device specialty to reduce the uncertainty in compliance, and can promote innovation and development in this field.   read more...   read less

Abstract: Text summarization is crucial in scientific research, spanning drug discovery, development, review, post-approval, and more. This task demands domain expertise, language proficiency, semantic prowess, and conceptual skill. The recent advent of large language models (LLMs), such as ChatGPT, offers unprecedented opportunities to automate this process. We compared ChatGPT-generated summaries with those produced by human experts using FDA drug labeling documents. The Labeling Highlights of key sections make them an ideal human benchmark for evaluating ChatGPT's performance. Analyzing over 14,000 pairs of summaries, we observed that ChatGPT-generated summaries closely resembled those by human experts. Importantly, ChatGPT exhibited even greater similarity when summarizing drug safety information. These findings highlight ChatGPT's potential to accelerate work in critical areas, including drug safety.   read more...   read less

Abstract: Adverse drug events (AEs) are a leading cause of mortality in the United States, with over 70,000 cases related to death reported annually since 2020. FDA drug labeling documents are a critical resource for AE research, providing comprehensive and reliable drug safety information. However, manually extracting and classifying AE data from these documents is labor-intensive and time-consuming. Recent advancements in natural language processing (NLP) and large language models (LLMs) offer a promising, modernized solutions. This paper introduces AskFDALabel, an LLM-powered framework, to enhance the efficiency and effectiveness of AE research with drug labeling document. We applied AskFDALabel on three experiments related to AE research: (1) DILI Classification, (2) DICT Classification, and (3) Drug AE Profiling. As a result, we observed promising results over all experiments. For DILI and DICT Classification, the overall consistency between AskFDALabel and previous human annotation were 0.984 and 0.929 by F1-score, respectively, significantly better than other toxicity classification approaches. For Drug AE detection, the overall F1-score was 0.911, outperformed than previous NLP approaches. Moreover, even using the drug name instead of actual text contents, the F1-score still reached 0.870. These high consistency performance results indicated AskFDALabel can automate the process of manual review process offering unprecedented speed and accuracy in AE detection and classification. Additionally, AskFDALabel provided the cited reference as well as the detailed explanations for the answer, which could further assist the reviewers to manual check and verify the LLM-generated answers. In conclusion, this study highlights the transformative potential of LLMs in AE research and pharmacovigilance. It demonstrates that generative AI can significantly enhance the efficiency and accuracy of AE detection and classification, improving drug safety and regulatory science.   read more...   read less

Abstract: The detection of duplicate Individual Case Safety Reports (ICSRs) in the FDA Adverse Event Reporting System (FAERS) database is critical. If not properly removed, duplicate reports could lead to biased analyses. This project focused on the exploration of a Generative AI model, Llama 3.1, for detecting duplicate FAERS reports. The study design involved developing a Large Language Model (LLM) approaches based on locally hosted AI model to address Personally Identifiable Information (PII) concerns, and the evaluation of their performance were conducted on 12 human-verified adjudicated benchmark datasets. In this study, three experiments were performed. Experiment 1 focused on prompt design strategies and optimization. The approach for the Experiment 2 was to evaluate the effect of different Llama models and sizes on the performance. Finally, Experiment 3 assembled the modular LLM-powered deduplication pipeline and assessed its performance on an individual data set.   read more...   read less

Background: Patient vital signs in the ICU are traditionally monitored and recorded on an hourly basis. Variation in vital signs is believed to be key in predicting impending patient death or recovery. Although some patients exhibit significant changes in vital signs as death approaches others do not. Within the ICU, there are a number of factors that can influence patient vital signs including medications and treatments along with the condition or injury from which the patients suffers. Providing advance warning of mortality risk to health care providers allows the opportunity for interventions improving the patient’s chances of survival. The increasing number of ICU patients within the US make this topic a significant public health issue. Methods: The ICU data for this study was obtained from the Medical Information Mart for Intensive Care version IV, generated from the Beth Israel Deaconess Medical Center from 2008 through 2019. We selected subjects who were 20 years and older and in one of the 9 ICUs for at least 24-hours but no more than 3 days. The vital signs data, in addition to use of mechanical ventilation, were restructured into an hourly longitudinal format for each eligible patient over the last 24-hours of their ICU stay. A Bidirectional LSTM AI algorithm processed these vital signs to make a mortality prediction for each patient. Results: Our implementation of a bidirectional LSTM AI algorithm coupled with highly granular ICU patient data was predictive of mortality risk. The model could be tuned for a high sensitivity or high positive predictive value. Conclusions: The results confirm that there is potential in predicting mortality risk within ICU patients utilizing routinely collected vital signs. An AI algorithm embedded with high sensitivity offers the opportunity for early warning to health care providers concerning patient mortality risk. Keywords: Mortality; ICU; Death; Artificial Intelligence; Long Short-Term Memory.   read more...   read less

Abstract: The severity outcome of COVID-19 disease resulting from SARS-CoV-2 infection has shown that many cases are related to underlying health conditions like diabetes and cancer. There remains an urgent need to gain mechanistic insights at the molecular level to understand the differences in severity of the infection and discover early biomarkers that enable prediction of severity among COVID-19 patients. These insights could help ease the burden of care and aid in evaluation and development of treatments. In this study, COVID-19-positive patients were categorized into 3 symptom response categories using a 12-point patient response scoring system: mild (score 0-1), moderate (score 2-4), and severe (score 5-12). Blood samples from 93 COVID-19 patients collected at the time of initial diagnoses were processed to plasma and deidentified for proteomic and miRNA analyses. A total of 2,921 proteins and 2,083 miRNAs were analyzed in the plasma samples. Ingenuity Pathway Analysis (IPA) of significant proteins revealed that the top pathways were neutrophil degranulation, cytokine storm, role of osteoblasts in rheumatoid arthritis signaling, interleukin-10 signaling, and wound healing signaling pathways in severe patients. Interleukin-6 (IL6) was involved in nine of the top ten IPA pathways. MicroRNA associations with protein targets were extracted from IPA using experimentally observed and highly predicted criteria. Proteomics and miRNA analyses of plasma from recently diagnosed COVID-19 patients revealed 1,353 protein and 1,167 miRNA biomarkers of severity response to COVID which are associated with preexisting health status and acute response to the COVID infection. Additional experiments are needed to validate the biomarkers.   read more...   read less

Abstract: The rapid advancements in generative AI present transformative opportunities for pharmaceutical regulatory science. This study explores the potential of deconstructing generative AI to employ vectors within large language models (LLMs) as predictive tools for statistical analysis of regulatory precedence, focusing on Real-World Evidence (RWE) and Real-World Data (RWD) in drug approvals. LLMs are a series of sophisticated vector-based algorithms, which we can leverage to analyze historical regulatory decisions. The public dataset comprises a comprehensive range of therapeutic areas and approval scenarios. These models are designed to identify and predict trends, providing nuanced insights into the regulatory pathways using current guidance. A central aspect of this research is the application of RWE and RWD in regulatory decision-making. By examining their historical usage in approvals, our models predict their future impact, aiding strategic planning for pharmaceutical companies and regulators alike. Preliminary findings indicate that vector-based predictive models enhance the accuracy of regulatory outcome predictions (accuracy of model: 85%, CI >0.95). This underscores the potential of generative AI in refining regulatory science and fostering data-driven decision-making processes. This study demonstrates the utility of generative AI and vector-based models for predictive analytics in pharmaceutical regulation. By extracting actionable insights from deconstructed LLMs, we provide a robust framework for understanding regulatory precedence and optimizing the use of RWE and RWD in drug approvals.   read more...   read less

Abstract: Liposomes are an important class of drug products, and account for a significant fraction of the nanomaterial submissions to FDA.1 The liposomal lipid composition and quantitation are critical quality attributes that influence liposome stability and drug release.2 Our lab has developed three standard test methods through ASTM International to quantify the constituting lipids in liposomal doxorubicin formulations, specifically, cholesterol, 1,2-distearoyl-sn-glycero-3-phosphoethanolamine-N-[methoxy (polyethylene glycol)-2000] (DSPE-PEG 2000) and hydrogenated (soy) L-α-phosphatidylcholine (HSPC) using widely available instruments. One of these published test methods utilizes Ultra-High-Performance Liquid Chromatography with Triple Quadrupole Mass Spectrometry (UHPLC-TQMS).3 To evaluate the repeatability of the method and reproducibility in the measurement of the concentration of the three lipids using this standard test method, we have conducted an interlaboratory study (ILS). A total of 17 laboratories participated in this ILS from 7 countries. The laboratories were provided with liposome sample suspensions prepared using a continuous manufacturing process at DIANT Pharma, Inc., along with lipids and internal standards in powder form. Each participant reported three replicate test results which were processed for calculating repeatability and reproducibility limits. A one-way analysis of variance (within- and between-laboratories) was carried out for estimating precision.4 The laboratories utilized UHPLC-TQMS instrument and analytical columns from different vendors which allowed variability in the results to be captured. Thus, the reproducibility and usefulness of the published test method could be established. Data from the analysis of this ILS will be presented. This study is highly relevant for quality control of liposomal formulations and supports FDA’s mission to protect and promote public health. Disclaimer: This presentation reflects the views of the authors and does not necessarily reflect those of the FDA. Acknowledgement: This study was sponsored by ASTM International and supported by an Interagency Agreement between the FDA/NCTR and the National Institute of Environmental Health Sciences/National Toxicology Program. References: [1] Nanotechnology—Over a Decade of Progress and Innovation, A report by FDA, August 2020. [2] Guidance for Industry: Liposome Drug Products: Chemistry, Manufacturing, and Controls; Human Pharmacokinetics and Bioavailability; and Labeling Documentation, April 2018. [3] Standard Test Method for Lipid Quantitation in Liposomal Formulations Using Ultra-High-Performance Liquid Chromatography (UHPLC) with Triple Quadrupole Mass Spectrometry (TQMS). ASTM-International: West Conshohocken, PA 19428-2959. United States, 2022; Vol. 14.02, p 20. DOI: 10.1520/E3324-22. [4] ASTM E177-20 - Standard Practice for Use of the Terms Precision and Bias in ASTM Test Methods.   read more...   read less

Abstract: In vivo and in vitro methods are available to evaluate the amount and rate of skin absorption and contribute to safety and efficacy assessments of topical products. Excised human skin (EHS) is the ‘gold standard’ for in vitro permeation testing; however, due to high cost and supply shortages, there is a need to find suitable alternative skin barrier models. Previously, we evaluated over 6 hours a subset of skin models for in vitro permeation testing (IVPT). In this study, we expanded our research using automated diffusion equipment with flow-through cells to perform IVPT for 24 hours using human epidermal and full-thickness models, Strat-M membrane, and EHS. The cumulative amount, flux, and mass distribution of 14C-labeled salicylic acid, caffeine, mannitol, and testosterone (6 nmol/cm2) were compared across skin models. Transepithelial water loss (TEWL) and histology were also compared. Overall, the TEWL and IVPT data suggest that the reconstructed skin models have a weaker barrier than EHS, although the magnitude of differences varies by skin model and test compound. Work is ongoing to optimize the isolation of the stratum corneum of excised and reconstructed human skin and characterize their lipid composition to help interpret the barrier function data. The lipidomics data, together with the IVPT, TEWL, and histology data, may help guide further optimization of reconstructed human skin to match the barrier properties of EHS. The systematic comparison of alternative skin models and EHS contributes to the evaluation of the potential of these models to predict the dermal absorption of FDA-regulated compounds.   read more...   read less

Abstract: The study aims to develop artificial intelligence (AI)-based models for predicting the activation state of molecular network pathways in diseases to evaluate therapeutic efficacy. We have developed AI-based models to predict the activation state of epithelial-mesenchymal transition (EMT) in cancer. In the current study, a dataset comprising 50 activated and 50 inactivated pathway images for the coronavirus pathogenesis pathway, along with 50 activated and 50 inactivated pathway images for the coronavirus replication pathway, were utilized to train models using the DataRobot Automated Machine Learning platform. The AI application produced an Elastic-Net Classifier (L2 / Binomial Deviance) for the coronavirus pathogenesis pathway and an Elastic-Net Classifier (mixing alpha=0.5 / Binomial Deviance) for the coronavirus replication pathway. To validate the models, 10 additional activated and 10 additional inactivated pathway images were used. The models that successfully predicted the activation state based on the gene expression would be useful for evaluating the efficacy of therapeutics.   read more...   read less

Background: The field of toxicology is undergoing a paradigm shift with the advent of New Approach Methods (NAMs), driven by the need to overcome the limitations of traditional animal toxicity tests, which have shown poor extrapolation to humans. Drug-induced liver injury (DILI) has been a leading cause of drug withdrawals and one of the most studied endpoints with NAMs. In line with the 3R principles to replace, reduce, and refine animal use in scientific research, NAMs are crucial for predicting and assessing DILI; however, their fitness for purpose has yet to be fully established. Our analysis aimed at integrating existing evidence to demonstrate the fitness for purpose of NAMs in DILI assessment. Methods: We conducted a comprehensive comparative analysis with representative in vitro NAM models from diverse NAMs for DILI prediction reported in the literature. The prediction data from these models were converted into binary DILI classifications (positive vs negative) based on their respective thresholds. We first compared the NAMs with three conventional rat models in terms of their prediction accuracy for human hepatotoxicity. Next, we examined the DILI prediction agreement among the NAMs by assessing drug overlap and annotation concordance. Additionally, using the DILIst and Garside DILI annotation reference, we assessed the impact of the DILI annotation reference on the models’ performance evaluation. Furthermore, we conducted an analysis of therapeutic categories to identify patterns in drug distribution within categories that have a high risk of DILI. Results: We observed that the NAMs showed higher concordance with human hepatoxicity than the animal models. Analysis of the in vitro NAM studies revealed a lack of common drugs and minimal concordance in DILI classification. The performance of the NAMs varied slightly when different references were used for evaluation. Additionally, inconsistent distribution of drugs over therapeutic categories was observed in different NAM studies. Conclusion: The current landscape of in vitro NAMs shows great potential to predict human DILI but lack consensus in the drugs investigated, therapeutic category representation, and choice of DILI annotation reference. This inconsistency hinders fair comparisons of different NAMs and prevents drawing statistically reliable conclusions to determine which NAMs can be utilized individually or in combination for accurate DILI prediction.   read more...   read less

Abstract: Drug-induced renal injury (DIRI) can lead to the development of acute kidney injury, chronic kidney disease, or end-stage renal disease, causing over 1.5 million adverse events annually and affecting approximately 26% of the United States population. Currently, the standard biomarkers for DIRI identification are late-stage biomarkers, serum creatinine and blood urea nitrogen, known to lack the sensitivity or specificity to detect nephrotoxicity prior to significant loss of renal function. For the proper development of alternative nephrotoxic methods, a large drug list with annotated DIRI potential is needed. In this study, we describe the creation of DIRIL (Drug-Induced Renal Injury List) – a list of drugs associated with DIRI and nephrotoxicity – from two literature datasets with DIRI annotation, confirmed using FDA drug labeling. DIRIL is comprised of 317 orally administered drugs, were 171 are Nephrotoxic (DIRI Positive) and 146 are Non-Nephrotoxic (DIRI Negative). The DIRIL dataset covers all 14 anatomical, therapeutic and chemical (ATC) classification categories, as classified by the WHO’s ATC Classification System. Approximately 91% of the drugs are from 7 of the 14 categories, with drugs related to the Cardiovascular and Nervous systems categories being most predominant. Additionally, we found that while methods like the Rule-of-Two (RO2) and Biopharmaceutics Drug Disposition Classification System (BDDCS) are known to be successful in the evaluation and severity classification of Drug Induced Liver Injury (DILI), these methods have proven to be ineffective in the classification of nephrotoxic potential. These results indicate that there are some distinct differences in nephrotoxicity as compared to DILI, such as the reabsorption, secretion, or passive filtering of drugs by the kidney. DIRIL will be a relevant and invaluable resource for discovery of new approach methods (NAMs) to predict the occurrence and possible severity of DIRI earlier in drug development.   read more...   read less

Abstract: In the fast-paced pharmaceutical regulatory policy, optimizing workflows and leveraging new technologies is crucial for reporting vast amounts of data originated from external engagements like trade association engagement and other stakeholders while managed by lean regulatory policy teams. Efficiently providing updates on the changing regulatory environment of medicines is top priority to be abreast of the changes across the International Regulatory Environment. Utilization of regenerative AI and Microsoft Power Tools platforms can be a power combination to bring efficiencies to policy teams to communicate. This case study will present how AstraZeneca’s International regulatory policy team employed Microsoft Power Tools platform along with Co-pilot & Large Language Models (LLM) a) to streamline the reporting of trade association interaction from 20 working groups in two International trade associations and 60 documents to process in average b) Using Co-pilot and LLM to create plain language summaries from different document sources ie pdf, power points, word documents. c) creation of summarized content, key talking points and training the AI to detect trends of key policy topics within the material examined. d) Present the dashboard of the metrics in terms of most discussed regulatory science topics in the different fora in scope of the study In summary, this case study will show case on how a lean regulatory policy team utilized Microsoft Power Tools, Co-pilot, and Large Language Models to streamline reporting, create plain language summaries from various sources and detect trends in policy topics.   read more...   read less

Abstract: The anti-microbial properties of silver nanoparticles (AgNPs) and evolving nanotechnology in different fields have increased their applications in consumer, healthcare, and industrial products such as food containers, cosmetics, feminine hygiene products, toothbrushes, bandages, and dental implants. This has led to increase in potential exposures for consumers and workers. Although, sex-based differences in the innate and adaptive immune responses are recognized and documented, the availability of human models that distinguish sex-related immunotoxicological effects of nanomaterials are limited. This study investigated sex-based differences in immune functions caused by exposure of human peripheral blood mononuclear cells (PBMCs) and plasma from healthy donors to AgNPs. Sex-based differences in inflammasome activation, cytokine expression, leukocyte proliferation, chemotaxis, plasma coagulation, and complement activation were evaluated after exposure to 30 nm AgNPs. Females showed significantly more inflammasome activation than males. Silver ions did not present a sex-based difference in inflammasome activation, which indicates that the differences observed with the AgNPs may stem from differences in AgNPs interaction with PBMCs. AgNPs caused significant down regulation of IL-10 and significant modulation of TNF-α, but it did not affect IL-6. AgNPs suppressed leukocyte proliferation in a concentration-dependent manner with no sex-based differences. AgNPs were not chemoattractant and did not activate the complement system or change plasma coagulation times. The findings of this study show the importance of considering sex in biocompatibility assessments of AgNPs.   read more...   read less

Abstract: The Ames assay is required by the regulatory agencies worldwide to assess the mutagenic potential risk of consumer products. As well as this in vitro assay, in silico approaches have been widely used to predict Ames test results as outlined in the International Council for Harmonization (ICH) guidelines. Building on this in silico approach, here we describe DeepAmes, a high performance and robust model developed with a novel deep learning (DL) approach for potential utility in regulatory science. DeepAmes was developed with a large and consistent Ames dataset (>10,000 compounds) and was compared with other five standard Machine Learning (ML) methods. Using a test set of 1,543 compounds, DeepAmes was the best performer in predicting the outcome of Ames assay. In addition, DeepAmes yielded the best and most stable performance up to when compounds were >30% outside of the applicability domain (AD). Regarding the potential for regulatory application, a revised version of DeepAmes with a much-improved sensitivity of 0.87 from 0.47. In conclusion, DeepAmes provides a DL-powered Ames test predictive model for predicting the results of Ames tests; with its defined AD and clear context of use, DeepAmes has potential for utility in regulatory application.   read more...   read less

Abstract: The In-Q Compliance Knowledge Center addresses the complex and dispersed nature of food compliance requirements through advanced natural language processing. This system, akin to models like ChatGPT, is uniquely enhanced by integrating expert-validated data, ensuring higher accuracy in interpreting and managing regulatory information. The platform consolidates a vast array of food compliance regulations into a unified, accessible interface. Users interact with the system using natural language queries, receiving precise, contextually relevant responses derived from an extensive, validated database. This expert validation distinguishes the system by significantly reducing errors and increasing the reliability of the provided information. Moreover, the In-Q Compliance Knowledge Center incorporates robust encryption protocols, ensuring the security and confidentiality of the data. This addresses a critical need for secure data management in compliance processes. The system represents a significant advancement in the application of computational tools for regulatory adherence in the food industry. By integrating expert-validated content and encryption, it offers a reliable, efficient solution for navigating the complex regulatory landscape, thereby enhancing both the accuracy and security of compliance management.   read more...   read less

Abstract: Systematic review methods minimize bias and provide necessary transparency for the process of identifying, critically assessing, and synthesizing research evidence to answer environmental questions, inform policy, and guide public health decisions. However, they are typically labor-intensive and require significant resources. The development of a streamlined workflow that leverages automation and artificial intelligence (AI) is crucial for reducing the time and resource demands of the traditional systematic review approach. By leveraging AI and automation, the Division of Translational Toxicology’s (DTT) enhanced assessment workflow aims to improve efficiency and accuracy while maintaining the rigorous standards of systematic review methodologies. AI technologies, such as natural language processing (NLP), large language models (LLM), and machine learning (ML), can significantly expedite various stages of the process, including literature screening and data extraction. Extraction is one of the most time-consuming steps (~30% of total review time), particularly for environmental health questions, given the breadth of study designs and endpoints represented. This poster outlines an enhanced workflow incorporating AI tools and human-in-the-loop procedures to support the accuracy necessary for public health decisions with the capacity to handle large datasets and facilitate the collection of complex information across environmental health science literature. By automating repetitive tasks, researchers can focus on more critical aspects of the review, such as interpreting results and synthesizing the evidence to reach conclusions. Streamlined workflows support transparency, standardization and reproducibility by providing clear documentation, interactive figures, and standardized procedures for each stage of the process. In this work, we present a framework that demonstrates the potential of AI and automation to significantly reduce level of effort for systematic reviews and thereby remove barriers to their use and accessibility. This poster underscores the potential of informatic tools, such as AI and automation, in revolutionizing environmental health assessments, ensuring that decision-making is both timely and evidence-based.   read more...   read less

Background: As progress in AI regulations increases, it is important especially in industries like healthcare to balance innovation and societal safety. In view of this commitment, various G20 member states have started strategic programmes geared towards developing ethically informed AI strategies. Therefore, this scoping review identifies regulatory frameworks pertaining to AI in healthcare across G20 nations. Methods: We conducted a scoping review following PRISMA-ScR guidelines to map AI health regulations among G20 countries. Our search strategy included five databases and extensive gray literature sources, such as Ministry of Health websites, Google Scholar articles, and Overton database. The inclusion criteria required articles to explicitly reference AI in healthcare. We included commentaries, editorials, and policy briefs, while excluding conference abstracts. Abstract and full-text screening was conducted using Rayyan.ai software. Results and discussion: Our analysis reveals that eleven G20 countries have enacted legislation, decrees, executive orders, or developed action plans and strategies related to AI in healthcare. Several nations are drafting policies focusing on critical infrastructure, safety, transparency, and balancing technological benefits with risk management. This review emphasizes the need for ethically informed, adaptive regulatory frameworks to address the socio-technical complexities of AI in healthcare. The strategic deployment of AI, supported by robust ethical principles and regulatory science, can significantly benefit healthcare stakeholders, including patients. Conclusion: By mapping the regulatory landscape of AI in healthcare across G20 countries, this review contributes to the global dialogue on AI strategy. It provides valuable insights to inform policy, guide future research, and promote a collaborative approach to responsibly harnessing AI's potential in healthcare.   read more...   read less

Abstract: Animal studies are unavoidable in evaluating chemical and drug safety, yielding complex, multidimensional data essential for characterizing the risk and safety profiles of substances regulated by the Food and Drug Administration (FDA). However, there is a shift towards alternatives to traditional animal testing. The FDA Modernization Act 2.0 emphasizes exploring options that support the 3Rs (Replacement, Reduction, and Refinement) of animal use. Artificial Intelligence (AI) offers innovative approaches for risk assessment in support of the 3Rs. Traditionally, AI applications in toxicology have focused on predicting single endpoints, such as toxic versus non-toxic classifications, through methods like Quantitative Structure-Activity Relationships (QSARs). These methods, however, lack the comprehensive, multidimensional insights provided by animal-based toxicity assessments. This study presents AnimalGAN, a generative AI model developed using Generative Adversarial Networks (GANs), designed to simulate virtual animal experiments and generate synthetic multidimensional profiles similar to those obtained from traditional animal studies. AnimalGAN successfully established the relationship between chemical exposure (the combination of chemical, dose and treatment duration) and clinical pathology outcomes (i.e., clinical chemistry and hematologic measures) based on historical animal study data. Our results demonstrate that AnimalGAN is robust in simulating rat clinical pathology measurements, even for drugs that differ significantly in chemical structure, drug class, and FDA approval year from those used in training. Compared to conventional QSAR methods, AnimalGAN outperformed 12 traditional models in predicting clinical pathology measures and produced results comparable to animal studies in evaluating drug-induced hepatotoxicity. Moreover, AnimalGAN’s ability to conduct unlimited synthetic experiments enables the detection of rare toxicological events, enhancing the translation of animal data to human outcomes and addressing one of the key limitations of traditional animal studies. These findings highlight the potential of generative AI to advance toxicological assessments, supporting the transition from animal-based to AI-driven approaches in drug safety evaluation.   read more...   read less

Abstract: Drug-induced cardiotoxicity (DICT) is a significant challenge in drug development and public health. DICT can arise from various mechanisms and New Approach Methods (NAMs) have been extensively developed for individual mechanisms (e.g., QT-Prolongation, hERG Channel Assay, and In Vitro Cardiomyocyte Assays). While these efforts have contributed significantly to our understanding of cardiotoxicity, DICT remains a persistent issue in all stages of drug development, from preclinical screenings and clinical trials to post-market surveillance. This suggests that isolating individual mechanisms may not comprehensively assess cardiotoxicity. To address this, we developed DICTrank, the largest dataset to evaluate overall cardiotoxicity liability. We assessed the utility of DICTrank for in silico NAMs by developing QSAR models using five machine learning methods (Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Random Forest [RF], and extreme gradient boosting [XGBoost]) with varying algorithmic complexities. These models were trained on 80% of 928 drugs (679 DICT-positive and 249 DICT-negative) and tested on the remaining 20%. While both training and test results didn’t indicate a clear relationship between model performance and algorithmic complexity, both XGBoost and RF tended to perform well with DICTrank. Significant feature analysis of both RF and XGBoost models provided insights into DICT mechanisms, with drug properties such as 'Topological Structure Autocorrelation' and 'Eigenvalues from Burden Matrix' contributing significantly to DICT. Model performance varied by therapeutic category, all the models were quite consistent with the highest discrimination power observed for drugs related to the World Health Organization's 'alimentary tract and metabolism' category, suggesting the need to tailor DICTrank-based predictive models based on therapeutic categories. Our study demonstrates the value of DICTrank combined with machine learning as in silico NAMs for early DICT risk identification. Keywords: DICT, Drug safety, Computational approach, QSAR   read more...   read less

Abstract: The Antiretroviral Tri-combination drug has revolutionized HIV treatment by effectively targeting different stages of viral replication. Despite its efficiency, recent cardiac, and neurological effects in the offspring upon exposure to the tri combo drug regimen during gestational period has raised concerns. The intestinal microbiome plays a crucial role in maintaining overall health, the disruption in gut microbiome is linked to various extraintestinal effects such as immune dysregulation and systemic inflammation. This study explores the long-term effects of the in utero and lactational (dams were dosed with 3 different exposures from GD6-PND21) exposure to control or mid dose, and high dose of the tri combo drug on the gut health of 12-month-old rat offspring. We assessed changes in the mucosa-associated microbiome composition, intestinal permeability, and ileal mucosa-associated immune profile that persisted into adulthood. Results of this study demonstrate significant alterations in the gut microbiome, permeability, and cytokines in a concentration-dependent and sex-dependent manner. Female offspring showed profound microbial dysbiosis and dysregulation in cell junction genes even at mid dose treatments while male offspring showed more notable changes in the pro- and anti-inflammatory cytokine expression. These findings suggest that early life exposure to the tri combo drug can result in persistent gut microbiome imbalances, compromised intestinal barrier, and altered immune changes that could participate in the pathogenesis like metabolic disorders, neurodevelopmental conditions, and other chronic inflammatory diseases. With the increasing evidence of the effects of gut health on other organs these pathological conditions could be modulating through the gut. This study emphasizes the need for comprehensive assessment of the tri combo drug safety during pregnancy, given the immediate benefits and potential long-term risks to the offspring.   read more...   read less