Bridging the gap between embryo euploidy, pregnancy potential and morphology using artificial intelligence for ploidy estimation: a retrospective evaluation.

Sex determination from skeletal remains presents significant challenges, particularly when bones are damaged or incomplete. In such cases, histomorphological analysis of fragmented bone pieces becomes essential. The medial clavicle is recognized as a valuable anatomical marker in forensic and anthropological research due to its pronounced sex-related morphological variations. This study aimed to develop a deep learning-based method for sex determination using histological images of the medial clavicle in Thai population, and to evaluate its performance with both validation and blind test sets utilizing the GoogLeNet convolutional neural network architecture. A total of 100 pairs of clavicles were included, with 70 cases (35 males,35 females) assigned to the training group and 30 (15 males,15 females) to the test group. Histological images underwent pre-processing and were standardized in size before being input into the training model. Validation accuracy was assessed using MATLAB, while descriptive statistics for the test set were calculated with SPSS software. GoogLeNet demonstrated superior performance, achieving a validation accuracy of 96.43% and a blind test accuracy of 90%. These results highlight the potential of a deep learning approach using 2D histological images of the medial clavicle as a straightforward and effective tool for sex determination in forensic anthropology, offering a high degree of accuracy. This method paves the way for rapid, objective, and accessible sex identification, even from fragmented human remains, and demonstrates promise for broader applications in the forensic and anthropological sciences.

Avian aspergillosis, caused by Aspergillus fumigatus (Af), lacks sensitive antemortem diagnostics. Existing microbial cell-free DNA (cfDNA) tests are prone to contamination and require a high pathogen load. We hypothesized that infection-induced tissue damage in chickens creates differentially methylated regions (DMRs) in host cfDNA, enabling machine learning (ML) diagnostics. Serum cfDNA samples (n = 124) were obtained from broiler chickens (n = 76) with Af and non-Af infections (Escherichia coli or Gallibacterium anatis) and controls. Oxford Nanopore sequencing enabled DMR detection and ML training. Performance was evaluated using an independent set (n = 49) and 10-repeat Monte Carlo cross-validation (CV) (n = 490 evaluations per test) as quality control. A High Accuracy test (93 DMRs, neural network) achieved 98.0% accuracy (sensitivity 95%, specificity 100%, AUC 0.974, PR-AUC 0.928) in the independent set, with CV accuracy 92.0% [95% CI: 89.7%-94.4%]. A Fast test (35 DMRs, SVM) achieved 81.6% accuracy and CV accuracy 79.6% [74.9%-84.3%]. An In Situ test (5 DMRs, random forest) designed for field deployment achieved 71.4% accuracy and CV accuracy 62.9% [58.7%-67.0%]. Stratified CV accuracy showed 84.6% [65.1%-95.6%] correct classifications for E. coli and 100% [80.5%-100%] for G. anatis. Markers showed high bootstrap stability and predominantly overlapped EMARs and enhancers. In conclusion, we present MethylSense (https://github.com/markusdrag/MethylSense), an automated open-source software. The High Accuracy test achieved 92.0% [89.7%-94.4%] CV accuracy (CV sensitivity 94.5% [91.4%-97.6%], CV specificity 90.3% [87.8%-92.9%]). While validated in chickens, MethylSense is adaptable to other species and pathogens, offering scalable, contamination-resilient diagnostics for veterinary and conservation applications.IMPORTANCEMethylSense is an automated software for training machine learning diagnostics using differentially methylated regions (DMRs) in cell-free DNA from Oxford Nanopore sequencing. We applied MethylSense to develop three Aspergillus fumigatus tests for chickens, each optimized for different clinical scenarios. The High Accuracy test (93 DMRs, neural network) demonstrated 98.0% accuracy, in a blinded test set (n = 49) with sensitivity 95%, specificity 100%, ROC-AUC 0.974, and PR-AUC 0.928. Stratified 10-repeat Monte Carlo cross-validation (n = 490) showed correct classifications of 84.6% [CI: 65.1%-95.6%] Escherichia coli and 100% [80.5%-100%] Gallibacterium anatis infected specificity samples. A Fast test for rapid <1 h sequencing (35 DMRs, support vector machine) achieved 81.6% accuracy (sensitivity 80%, specificity 82.8%). An In Situ test (5 DMRs, random forest) for field deployment via methylation-specific PCR achieved 71.4% accuracy (sensitivity 45%, specificity 89.7%). Bootstrap analysis demonstrated exceptional marker stability (80.6%-100%) with minimal batch effects, confirming robust host-based diagnostics.

BackgroundClinical studies have shown that aortic arch pulse‐wave velocity (PWVaa), a measure of local aortic stiffness, is a strong independent predictor of subsequent white matter hyperintensity volume and white matter integrity, both associated with cognitive decline, elevated stroke risk, vascular dementia, and neurodegenerative diseases. Total arterial compliance (TAC), a measure of global arterial stiffness, has been recognized as a marker of preclinical vascular disease. This study introduces a smartphone‐based method for the noninvasive measurement of PWVaa and TAC using carotid pressure waveforms acquired via smartphone.MethodsThis method uses intrinsic frequency analysis of smartphone‐acquired (iPhone) carotid pressure waveforms to assess PWVaa and TAC. The method was trained, validated, and blind‐tested on a cohort of 132 participants aged 20 to 90 years, including both healthy individuals and those with cardiovascular disease, all of whom underwent cardiac magnetic resonance imaging, tonometry, and iPhone waveform measurements.ResultsIn the blind test set, our method achieved Pearson correlations of 0.81 and 0.80 for PWVaa and TAC, with biases of −0.20 m/s and −0.06 mL/mm Hg and limits of agreement of −4.09 to 3.68 m/s and −0.52 to 0.40 mL/mm Hg, respectively. In the heart failure population, correlations were 0.81 for both, with a PWVaa a bias of −1.07 m/s and TAC bias of −0.06 mL/mm Hg.ConclusionsOur smartphone‐based method enables accurate assessment of local and global arterial stiffness metrics (PWVaa and TAC). It offers easy‐to‐use monitoring of vascular aging and arterial health, with important implications for identifying patients at higher risk of neurodegenerative and cardiovascular diseases.RegistrationURL: clinicaltrials.org; Unique Identifier: NCT02240979.

This study aimed to develop deep learning (DL) models for CT-free attenuation correction and Monte Carlo-based scatter correction in 99mTc-macroagregated albumin (99mTc-MAA) SPECT imaging, with the goal of enhancing quantitative accuracy for improved treatment planning and pre-therapy dosimetry in 90Y-selctive internal radiation therapy (SIRT). Data from 222 patients who underwent 99mTc-MAA SPECT imaging prior to 90Y-SIRT were included in this study. Uncorrected SPECT images (without attenuation and/or scatter correction) were used as input to a modified 3D shifted-window UNet Transformer (Swin UNETR) architecture. Three separate models were trained to predict attenuation corrected (AC), scatter corrected (SC), and joint attenuation and scatter corrected (ASC) SPECT images. The dataset was split into a training set (~ 80%) and an independent test set (~ 20%). Model training was performed using a five-fold cross-validation framework, with final evaluation conducted on the blind test set. To clinically assess model performance, 3D voxel-wise dosimetry was performed on the test set using the local energy deposition method, assuming 99mTc-MAA as a surrogate for 90Y distribution. Quantitative evaluation included organ- and voxel-level metrics, along with Gamma analysis using three combinations of distance-to-agreement (DTA, mm) and dose-difference (DD, %) criteria. The average (± SD) of the voxel-wise mean error (ME) was ≤ 0.003Gy for all tasks. The Relative Error (RE (%)) for AC, SC, and ASC tasks were 4.64 ± 7.52%, 8.99 ± 26.35%, and 16.45 ± 25.83%, respectively. Voxel-level Gamma evaluations within the whole body using three different criteria sets, including "DTA: 4.79mm, DD: 1%"; "DTA: 10mm, DD: 5%"; and "DTA: 15mm, DD: 10%" yielded pass rates of over 99.60%. The mean absolute error (MAE) for lesions, normal liver and lungs across all tasks were 3.16 ± 3.39, 0.35 ± 0.36, 0.41 ± 0.47Gy for AC, 1.97 ± 2.79, 0.19 ± 0.16, 0.22 ± 0.20Gy, for SC and 5.16 ± 7.10, 0.45 ± 0.51, and 0.34 ± 0.37Gy for ASC, respectively. Multiple models were developed for key SPECT quantification tasks, with potential value in clinical setting lacking reliable CT data or sufficient computational resources for Monte Carlo simulations. The models look promising for potential clinical translation and integration into commercial reconstruction software.

Introduction: This study is part of the broader Stem Line project Mito-Cell-UAB073, specifically focusing on "Stem Cell Lines-Quality Control," and aims to innovate in the field of Quality Control (QC) through a unique, artificial intelligence (AI)-powered model known as Life Cell AI UAB. This model utilizes deep learning algorithms and computer vision, allowing it to make accurate viability assessments of cell and stem cell lines based solely on static images captured through standard optical microscopes. Aim: The aim of this study was to develop and validate an AI-driven, image-based model that reliably predicts cell line viability. Methods: Our methodology involved training the Life Cell AI UAB model on single static images of cell lines using advanced computer vision and deep learning techniques. Performance evaluation was conducted on three independent blind test sets sourced from various biotechnology laboratories, allowing for assessment across diverse environments. Results: The Life Cell AI UAB model achieved a sensitivity of 82.1% in identifying viable cell lines and a specificity of 67.5% for non-viable lines across the test sets. Each blind test set exhibited a weighted accuracy above 63%, with a combined accuracy of 64.3%. Notably, predictions showed a clear distinction between correctly and incorrectly classified cells. The model outperformed traditional QC methods by improving accuracy in binary classification tasks by 21.9% (p = 0.042) and demonstrated a 42.0% enhancement over conventional Standard Operation Procedure (SOP) procedures (p = 0.026). Conclusion: The Life Cell AI UAB model represents a notable advancement in biobanking QC, offering a precise, standardized, and non-invasive method for assessing cell line viability. This model has the potential to streamline QC processes across laboratories, minimizing the need for time-lapse imaging and promoting uniformity in QC practices for both cell and stem cells.

Nitro-polycyclic aromatic hydrocarbons (nitro-PAHs) are potent environmental pollutants with known mutagenic and carcinogenic properties. This study integrates molecular docking and quantitative structure–activity relationship (QSAR) modeling to investigate the DNA-binding potential and mutagenicity of 30 structurally diverse nitro-PAHs. Using the high-resolution DNA dodecamer crystal structure (PDB ID: 1D63), molecular docking was performed via a hierarchical Glide workflow, followed by MM-GBSA binding free energy calculations to refine binding affinity estimates. The top ligands exhibited substantial binding potential, with ΔG bind values ranging from –22.0 to –35.9 kcal/mol, particularly favoring dinitro-substituted derivatives such as 2,7-dinitropyrene. A five-descriptor QSAR model was developed using multiple linear regression (MLR), random forest (RF), Gaussian process regression (GPR), and neural networks (MLP), with MLR showing the best predictive accuracy (RMSE = 0.86) on a blind test set. The most influential descriptors included molecular surface area, polarizability, and Connolly surface parameters. In silico toxicity assessments using ProTox-III revealed high mutagenic potential, aryl hydrocarbon receptor activation, and blood-brain barrier permeability for top-scoring ligands, with variable predictions for carcinogenicity and neurotoxicity. Collectively, these results provide insight into the molecular determinants of DNA intercalation and toxicological risks associated with nitro-PAHs, providing a computational foundation for environmental hazard assessment and structure-based screening.

Nitro-polycyclic aromatic hydrocarbons (nitro-PAHs) are potent environmental pollutants with known mutagenic and carcinogenic properties. This study integrates molecular docking and quantitative structure–activity relationship (QSAR) modeling to investigate the DNA-binding potential and mutagenicity of 30 structurally diverse nitro-PAHs. Using the high-resolution DNA dodecamer crystal structure (PDB ID: 1D63), molecular docking was performed via a hierarchical Glide workflow, followed by MM-GBSA binding free energy calculations to refine binding affinity estimates. The top ligands exhibited substantial binding potential, with ΔG bind values ranging from –22.0 to –35.9 kcal/mol, particularly favoring dinitro-substituted derivatives such as 2,7-dinitropyrene. A five-descriptor QSAR model was developed using multiple linear regression (MLR), random forest (RF), Gaussian process regression (GPR), and neural networks (MLP), with MLR showing the best predictive accuracy (RMSE = 0.86) on a blind test set. The most influential descriptors included molecular surface area, polarizability, and Connolly surface parameters. In silico toxicity assessments using ProTox-III revealed high mutagenic potential, aryl hydrocarbon receptor activation, and blood-brain barrier permeability for top-scoring ligands, with variable predictions for carcinogenicity and neurotoxicity. Collectively, these results provide insight into the molecular determinants of DNA intercalation and toxicological risks associated with nitro-PAHs, providing a computational foundation for environmental hazard assessment and structure-based screening.

Accurate prediction of antibody paratopes is a critical challenge in structure-limited, high-throughput discovery workflows. We present ParaDeep, a lightweight and interpretable deep learning framework for residue-level paratope prediction directly from amino acid sequences. ParaDeep integrates bidirectional long short-term memory networks with one-dimensional convolutional layers to capture both long-range sequence context and local binding motifs. We systematically evaluated 30 model configurations varying in encoding schemes, convolutional kernel sizes, and antibody chain types. In five-fold cross-validation, heavy (H) chain models achieved the highest performance (F1 = 0.856 ± 0.014, MCC = 0.842 ± 0.015), outperforming light (L) chain models (F1 = 0.774 ± 0.023, MCC = 0.772 ± 0.022). On an independent blind test set, ParaDeep attained F1 = 0.723 and MCC = 0.685 for H chains, and F1 = 0.607 and MCC = 0.587 for L chains, representing a 27% MCC improvement over the sequence-based baseline Parapred. Chain-specific modeling revealed that heavy chains provide stronger sequence-based predictive signals, while light chains benefit more from structural context. ParaDeep approaches the performance of state-of-the-art structure-based methods on heavy chains while requiring only sequence input, enabling faster and broader applicability without the computational cost of 3D modeling. Its efficiency and scalability make it well-suited for early-stage antibody discovery, repertoire profiling, and therapeutic design, particularly in the absence of structural data. The implementation is freely available at https://github.com/PiyachatU/ParaDeep, with Python (PyTorch) code and a Google Colab interface for ease of use.

Introduction: Accurate assessment of cardiac output (CO), a standard cardiovascular performance index, is essential for diagnosing and managing a wide range of cardiovascular conditions, including heart failure, shock, and valvular disease (Eur Heart J.PMID: 2092985). However, standard methods such as echocardiography or thermodilution are either operator-dependent, resource-intensive, or invasive. This limits their use in routine screening and outpatient monitoring. A noninvasive, rapid method to estimate CO from a single arterial pressure waveform could transform cardiovascular care by enabling continuous or point-of-care evaluation. Aim: This study aimed to develop and validate a noninvasive method for estimating cardiac output using features extracted from a single carotid pressure waveform captured with a tonometry measurement device. Methods: A cohort of 2448 individuals (age range: 19–90 years) from the Framingham Heart Study was analyzed. All participants had consistent CO measurements across multiple echocardiographic recordings. Carotid pressure waveforms were obtained using an arterial tonometry device and calibrated using cuff-based brachial pressures. Reference aortic flow values were computed by first measuring the left ventricular outflow tract diameter from 2D echocardiography (parasternal long-axis view) to calculate the cross-sectional area. Then, the pulsed Doppler velocity waveform from the apical 5-chamber view is multiplied by this area to generate the aortic flow waveform over time. CO values were computed by averaging the flow waveform over the entire cardiac cycle. Intrinsic frequency (IF) parameters were computed from the carotid waveforms and used as inputs for machine learning models. Eighty percent of the data was used for model training, and the remaining twenty percent was reserved for blind testing. Results: Single-waveform CO estimation showed a Pearson correlation of 0.76, limits of agreement of ±1.09, and a bias of 0.00 compared to reference values in the blinded test set (Fig. 1 and 2). Conclusions: Estimating cardiac output from a single carotid pressure waveform offers a non-invasive and scalable tool for hemodynamic monitoring. This method may improve early detection and management of various cardiovascular conditions where cardiac output is critical, such as heart failure, cardiogenic shock, and myocardial infarction. This method is well-suited for both in-patient and remote patient monitoring.

Introduction: Left ventricular pulsatile workload (LVPW) is a clinically established marker of cardiac afterload and function, and it is strongly associated with cardiovascular morbidity and mortality. Elevated LVPW contributes to adverse ventricular remodeling, impaired cardiac performance, and the development of heart failure (HF) (Eur Heart J. PMID: 29947746). However, clinical adoption of LVPW assessment remains limited due to the requirement for simultaneous pressure and flow measurements. This study introduces a smartphone-based approach for estimating LVPW noninvasively from only carotid pressure waveforms, enabling accessible and scalable cardiovascular monitoring using only smartphone camera-derived signals. Methods: A clinical cohort of 115 participants (41% women, BMI 25.9 ± 5.5, age range 20–92 years, mean 53 ± 18) was studied, including 43 individuals with cardiovascular disease (17 ambulatory HF patients). Reference LVPW values were calculated using ascending aorta flow from phase-contrast MRI combined with carotid pressure waveforms acquired via applanation tonometry. Carotid pressure waveforms were also recorded using a custom iPhone 5S (Apple Inc.) application by placing the camera against the neck, over the carotid artery (Crit Care Med. PMID: 28441235). These waveforms were calibrated using cuff-based brachial pressures, and their intrinsic frequency (IF) parameters were extracted. Using these IF metrics, a physics-based machine learning model was trained on 80% of the dataset to approximate LVPW and evaluated on the remaining 20% in a blinded test. Results: Smartphone-derived LVPW estimates showed a strong correlation with the gold standard reference computed from pressure-flow values. The Pearson correlation coefficient was 0.83 for the blind test set and 0.86 among the HF patient subgroup, as shown in Fig. 1. Conclusions: LVPW can be reliably and non-invasively estimated using only carotid pressure waveforms captured with an unmodified smartphone camera (iPhone in this study). This noninvasive, low-cost approach may enable routine assessment of pulsatile afterload for both clinical and at-home cardiac monitoring. This may facilitate the delivery of precision medicine with timely treatment plan modifications in patients with HF whose outcomes are highly sensitive to increases in pulsatile afterload. This technique may also expand access to early cardiovascular risk stratification in other diseases.

Borehole image logs are essential for characterizing subsurface formations, particularly in identifying fractures that influence reservoir behavior and productivity. Manual interpretation of such logs, however, remains time consuming and susceptible to subjectivity and inconsistency. To address these challenges, the SPWLA Petrophysical Data-Driven Analytics special interest group (PDDA SIG) launched its 4th Annual Machine-Learning Competition, aimed at developing automated methods for accurate, efficient, and reproducible fracture detection. The competition utilized resistivity image logs and conventional quadruple-combo logs from eight wells in the Western Canadian Sedimentary Basin (WCSB), accompanied by expert-labeled fracture annotations for training. A separate blind test data set from two additional wells in the same basin was reserved for final evaluation. Participants were provided with a Jupyter Notebook containing preprocessed data and a baseline framework to facilitate model development. Submissions were evaluated based on F1 score and averaged root mean squared error (RMSE) on the blind test set, reflecting both classification accuracy and predictive reliability. This paper reviews the top five approaches submitted, highlighting key methodologies, feature engineering strategies, and model architectures that led to improved fracture detection performance. Our results demonstrate that advanced machine-learning techniques can substantially enhance the consistency and accuracy of fracture identification from borehole image logs. These findings support the integration of data-driven solutions into petrophysical workflows, offering scalable and objective tools to augment or replace manual interpretation, ultimately improving decision making in exploration and production operations.

The rapid evolution of viruses like SARS-CoV-2 and itsemergingvariants requires advanced diagnostic techniques for effective pandemicmanagement. This study introduces a machine learning (ML)-based surface-enhancedRaman scattering (SERS) methodology for the precise strains, substrains-baseddetection, and differentiation of SARS-CoV-2 in clinical nasopharyngealswab samples. Pristine silver nanorod substrates fabricated usingthe glancing angle deposition method were used for the sensitive detectionof the wildtype, kappa, delta, and omicron variants of SARS-CoV-2.Also, four different substrains of omicron strain (BA.1, BA.2, BA.5,and XBB) were detected and distinguished using the developed platform.A detection limit of around 100 pfu/mL was established for the 4 variantsand 4 covariants of the COVID-19 virus. However, challenges arisein the clinical samples due to the subtle spectral variations betweenclosely related variants of SARS-CoV-2. To address this, ML modelswere integrated with SERS data to discern intricate patterns, enhancingthe differentiation capabilities. In this study, we employed two differentclassifiers, support vector machine (SVM) and bidirectional long short-termmemory network (BiLSTM), for identifying the targeted variants fromnasopharyngeal swabs of 122 positive patients, who were previouslyidentified as the specific strain of SARS-CoV-2 through next-generationsequencing. The SVM classifier achieved an accuracy of 88.79% (95%CI: 83.18–94.39) and the BiLSTM model 85.98% (95% CI: 79.44–92.52)for variant classification on the validation set. Further, the modelswere validated on a blind test set, where an accuracy of 74.77% (95%CI: 67.29–83.18) and 70.09% (95% CI: 62.59–78.50) wasachieved, respectively. Furthermore, the SVM classifier, trained forsubvariant classification of omicrometer variants, obtained an accuracyof 95.83% (95% CI: 87.50–100.00) on the validation set. Thisintegrated ML-SERS approach not only enhances detection efficacy butalso provides on-site disease prediction ability, which will be immenselyhelpful for disease management.

The integrity of subsea oil and gas pipelines is essential for offshore safety and environmental protection. Conventional leak detection approaches, such as manual inspection and indirect sensing, are often costly, time-consuming, and prone to subjectivity, motivating the development of automated methods. In this study, we present a deep learning-based framework for detecting underwater leaks using images acquired in controlled experiments designed to reproduce representative conditions of subsea monitoring. The dataset was generated by simulating both gas and liquid leaks in a water tank environment, under scenarios that mimic challenges observed during Remotely Operated Vehicle (ROV) inspections along the Brazilian coast. It was further complemented with artificially generated synthetic images (Stable Diffusion) and publicly available subsea imagery. Multiple Convolutional Neural Network (CNN) architectures, including VGG16, ResNet50, InceptionV3, DenseNet121, InceptionResNetV2, EfficientNetB0, and a lightweight custom CNN, were trained with transfer learning and evaluated on validation and blind test sets. The best-performing models achieved stable performance during training and validation, with macro F1-scores above 0.80, and demonstrated improved generalization compared to traditional baselines such as VGG16. In blind testing, InceptionV3 achieved the most balanced performance across the three classes when trained with synthetic data and augmentation. The study demonstrates the feasibility of applying CNNs for vision-based leak detection in complex underwater environments. A key contribution is the release of a novel experimentally generated dataset, which supports reproducibility and establishes a benchmark for advancing automated subsea inspection methods.

This study presents an automated, objective method for eyelid movement assessment in de-novo Parkinson’s disease(PD) using a one-dimensional camera setup during monologue. These measurements were related to Dopamine Transporter Single Photon Emission Tomography and clinical scores. State-of-the-art computer-vision technologies and deep-learning neural networks were utilized to measure fourteen eyelid movement markers describing blinking and eyelid kinematics. Video-recordings were collected from a total of 120 de-novo patients with PD and 55 healthy controls. Abnormal blinking was present in 38% of PD, indicated by a reduced blink rate (p < 0.001), an increased inter-blink interval (p < 0.001), and an increased rigidity of the palpebral aperture (p < 0.001). The classification experiment reached an area under the curve of 0.81 on a blinded test set. The blink rate correlated with the loss of nigral dopaminergic neurons (r = 0.35, p < 0.001). These findings suggest eyelid movement markers as strong reflections of striatal dopaminergic activity levels, underscoring the method’s potential as a reliable early PD biomarker.

Transthyretin (TTR) plays a vital role in thyroid hormone transport and homeostasis in both the blood and target tissues. Interactions between exogenous compounds and TTR can disrupt the function of the endocrine system, potentially causing toxicity. In the Tox24 challenge, we leveraged the data set provided by the organizers to develop a deep learning-based consensus model, integrating sPhysNet, KANO, and GGAP-CPI for predicting TTR binding affinity. Each model utilized distinct levels of molecular information, including 2D topology, 3D geometry, and protein-ligand interactions. Our consensus model achieved favorable performance on the blind test set, yielding an RMSE of 20.8 and ranking fifth among all submissions. Following the release of the blind test set, we incorporated the leaderboard test set into our training data, further reducing the RMSE to 20.6 in an offlineretrospective study. These results demonstrate that combining three regression models across different modalities significantly enhances the predictive accuracy. Furthermore, we employ the standard deviation of the consensus model's ensemble outputs as an uncertainty estimate. Our analysis reveals that both the RMSE and interval error of predictions increase with rising uncertainty, indicating that the uncertainty can serve as a useful measure of prediction confidence. We believe that this consensus model can be a valuable resource for identifying potential TTR binders and predicting their binding affinity in silico. The source code for data preparation, model training, and prediction can be accessed at https://github.com/xiaolinpan/tox24_challenge_submission_yingkai_lab.

Glioblastoma Multiforme (GBM) is the most common and aggressive primary adult brain tumor, with a median overall survival of 15 months. The peritumoral brain zone (PBZ), the region surrounding the GBM, may be associated with tumor infiltration and aggressiveness and has been recognized as clinically and prognostically important. This study investigates quantitative imaging analysis of the PBZ and its impact on survival in GBM using deep learning (DL). We conducted a retrospective study using the BraTS 2021 (1,251 subjects) and BraTS 2020 (235 subjects) datasets, both containing tumor segmentations, with the latter also including overall survival data (median: 370 days; 49% died within one year). We developed a unified DL-based image reconstruction model to extract image features. For each patient, we utilized four MRI sequences—T1, T2, T1CE, and FLAIR. For each sequence, we considered three regions of interest (ROIs) that included the tumor, tumor plus 5 mm dilation, and tumor plus 1 cm dilation, with the dilation region commonly used to define the PBZ. This combination resulted in 12 images per patient (4 sequences × 3 ROIs). Tumor reconstruction was trained by SwinUnetR on BRATS2021, using a loss combining Mean Squared Error (MSE) and Structural Similarity Index (SSIM). The image feature vector from tumor reconstruction was fed into a separate, fully connected network to predict survival probabilities over time intervals. We trained and evaluated three survival models with the same architectures and patients but varying ROIs. The models were trained and evaluated on the BraTS 2020 dataset with an 80/20 split for training/validation and blinded testing, using the logistic hazard as the loss function. The survival prediction models were evaluated using the time-dependent concordance index (Ctd) and area under the receiver operating characteristic curve (AUC) at one year.Of 47 patients in the blinded test set, 23 died in the first year of diagnosis (median survival: 387 days). The Ctd for models based on tumor features alone, tumor plus 5 mm dilation, and tumor plus 1 cm dilation were 0.56 [95% CI: 0.43-0.70], 0.63 [95% CI: 0.50-0.74], and 0.59 [95% CI: 0.46-0.71]. The corresponding AUCs were 0.56 [95% CI: 0.39-0.72], 0.63 [95% CI: 0.47-0.79], and 0.58 [95% CI: 0.42-0.75] (no significant differences among the three models). While no significant differences were observed among the three ROIs, the results suggest that incorporating analysis of the PBZ into DL models can enhance prognostic performance for GBM, with the 5 mm surrounding tissue outperforming the 1 cm tissue. Further independent validation of these findings should be performed using additional external GBM patient datasets. As tumor segmentation is crucial for this study, future work should investigate the impact of minor variations in tumor segmentation on model performance. The codes for this study are on github.com/AIM-KannLab/surrounding_prognosis. Citation Format: Maryamalsadat Mahootiha, Hemin Ali Qadir, Ilangko Balasingham, Benjamin H. Kann. Impact of peritumoral zone on glioblastoma prognosis based on deep learning approach [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2022.

Background: Chinese hamster ovary (CHO) cell metabolism is complex, influenced by nutrients like glucose and glutamine and metabolites such as lactate. Real-time monitoring is necessary for optimizing culture conditions and ensuring consistent product quality. Raman spectroscopy has emerged as a robust process analytical technology (PAT) tool due to its non-invasive, in situ capabilities. This study evaluates Raman spectroscopy for monitoring key metabolic parameters and IgG titer in CHO cell cultures. Methods: Raman spectroscopy was applied to five 10 L-scale CHO cell cultures. Partial least squares (PLS) regression models were developed from four batches, including one with induced cell death, to enhance robustness. The models were validated against blind test sets. Results: PLS models exhibited high predictive accuracy (R2 > 0.9). Glucose and IgG titer predictions were reliable (RMSEP = 0.51 g/L and 0.12 g/L, respectively), while glutamine and lactate had higher RMSEP due to lower concentrations. Specific Raman bands contributed to the specificity of glucose, lactate, and IgG models. Predictions for viable (VCD) and total cell density (TCD) were less accurate due to the absence of direct Raman signals. Conclusions: This study confirms Raman spectroscopy's potential for real-time, in situ bioprocess monitoring without manual sampling. Chemometric analysis enhances model robustness, supporting automated control systems. Raman data could enable continuous feedback regulation of critical nutrients like glucose, ensuring consistent critical quality attributes (CQAs) in biopharmaceutical production.

B-type rapidly accelerated fibrosarcoma (BRAF) is a key oncogene that regulates cell signaling and proliferation, rendering it a crucial target for cancer therapeutics. Traditional QSAR methods are hindered by their reliance on a singular model, their inability to grasp complex nonlinearities, and limited generalization, undermining predictive efficacy. To address these challenges, we introduce BRAFPred, a novel framework that leverages stacked ensemble learning to integrate both classical machine learning and advanced deep learning techniques for the precise prediction of BRAF inhibitors. We utilized 12 handcrafted molecular descriptors derived from PaDeL, in conjunction with small molecule sequence features, as foundational inputs. Furthermore, we employed extreme gradient boosting (XGB), support vector regression (SVR), and deep learning architectures based on Chemprop and a pretrained BERT model (FG-BERT) to generate additional predictive features. These multisource features were subsequently integrated within a meta-ensemble random forest regression model, which utilized 26 input variables. Empirical results demonstrate that BRAFPred significantly outperforms benchmark models, achieving a mean absolute error (MAE) of 0.383 and a coefficient of determination (R 2) of 0.855, surpassing Chemprop (MAE = 0.443, R 2 = 0.803), FG-BERT (MAE = 0.460, R 2 = 0.785), and Stack_BRAF (MAE = 0.403, R 2 = 0.839). Extensive evaluation on benchmark data sets affirms BRAFPred's superiority over state-of-the-art methodologies, with robust generalization capabilities demonstrated on blind test sets. Additionally, ablation studies and case analyses underscore the robustness of the model's design. The source code, data sets, and prediction results for BRAFPred are available for further research at https://github.com/EvanZhang1216/BRAFPred.

Retention of molecules on immobilized artificial membrane (IAM) chromatography is a key physicochemical property for predictive models of permeability across biological barriers, with applications in drug design and ecotoxicology. Currently, IAM retention is solely experimentally determined, which limits its utility for screening virtual compound libraries or for predictions of yet not synthesized molecules. The present study focuses on developing predictive models of IAM retention factors (logkw(IAM)) for a structurally diverse set of drug compounds, scrutinizing the role of lipophilicity, experimental and calculated, as well as the contribution of additional molecular parameters, selected from a pool of physicochemical, constitutional, topological and 3D descriptors. After obtaining a data overview by principal component analysis, both multiple linear regression (MLR) and partial least squares (PLS) analyses were used to construct lipophilicity-based models and lipophilicity-independent models. Bulk, polarity and fraction of anionic species were common descriptors in all models. It was demonstrated that calculated lipophilicity values introduced additional uncertainty, depending on the software used. On the other hand, lipophilicity-independent MLR and PLS models, which relied solely on computational descriptors, showed comparable performance with lipophilicity-based models, while offering the advantage to more useful for screening large libraries in early drug discovery. The reliability of lipophilicity-independent MLR and PLS models was assessed by external validation as well as by using a blind test set. Error distribution between lipophilicity-based and lipophilicity-independent models was also investigated and found to be comparable, while it was better than the differences between experimental and calculated lipophilicity values.