Professional Experience

With more than 15 years of experience in computational statistics, data science, and artificial intelligence, I have held technical and leadership roles across pharmaceuticals, healthcare, energy, environmental monitoring, telecommunications, and financial inclusion. I am currently a full-time Statistical Data Science Manager at Pfizer, where I apply data science and statistical programming to clinical trials, particularly in vaccine and immunogenicity research. I work with a multidisciplinary team of statistical programmers, data scientists, and statisticians to support clinical data analysis and the preparation of regulatory documentation.

Before joining Pfizer, I spent five years at DTN as a Senior Principal Data Scientist. In this role, I contributed to the research and development of data-driven products for energy and environmental monitoring, with a particular focus on wildfire applications. I also teach computational statistics and probability theory at the Department of Statistics, University of the Philippines Cebu.

My research integrates computational statistics and artificial intelligence with applications in tropical medicine, pharmaceutical research, and chemical engineering. I am a guest researcher with the University of Copenhagen Drug Safety Group, led by Maurizio Sessa. As part of my civic engagement, I also conduct research with Data and Implementation Sciences for Health, a non-governmental organization that applies technology, artificial intelligence, and data science to population health.

I am currently completing an MSc in Statistics at the University of the Philippines Diliman, my second master’s degree. My thesis investigates uncertainty quantification and empirical likelihood estimation for generative artificial intelligence. I also hold an MSc in Tropical Medicine from the University of the East Ramon Magsaysay Memorial Medical Center and a BSc in Chemical Engineering from the University of San Carlos.

Research

I pursue both applied and theoretical research at the intersection of computational statistics, deep learning, and generative artificial intelligence, with applications spanning tropical medicine, chemical engineering, and drug and vaccine research.

My current collaborations include working with Maurizio Sessa and the University of Copenhagen Drug Safety Group on applications of generative AI in pharmacovigilance and pharmacoepidemiology. I also collaborate with Andrea Rossi of the University of Milan on inappropriate prescribing among older adults, and with Dr. Jason V. Alacapa through Data and Implementation Sciences for Health on data-driven approaches to tuberculosis research and analytics.

Previously, at DTN, I worked extensively on wildfire and storm-risk modeling and contributed to the development of data-driven products for the energy and refined-fuels sectors.

Full-text copies of my publications are available through my ResearchGate profile. If you are unable to access a particular paper, please feel free to contact me on LinkedIn; I would be happy to provide a copy when permitted.

Individualized treatment strategies are critical to enhancing patient outcomes in contemporary clinical environments, where treatment response heterogeneity is the rule rather than the exception. ITE estimation via machine learning provides a means of tailoring interventions more precisely. This study seeks to improve the explainability of subgroupings by combining Virtual Twins (VT) with Shapley Additive Explanations (SHAP). The VT framework estimates ITE by simulating counterfactual outcomes through separate predictive models for treated and control populations based on demographic and clinical variables from medical datasets. To address interpretability challenges, SHAP values were computed to quantify the individual-level influence of each covariate on estimated effects. The integration moves beyond the binary constraints of traditional decision rules, such as arbitrary thresholds like PRAPACHE > 27.5, by using SHAP beeswarm plots to visualize the continuous directional impact and magnitude of every data point. This combination not only allowed for the estimation of who benefits most from a specific intervention, but also provided insights into why the model made certain predictions. The source code for this implementation is available at: https://github.com/tobsilog/SHapley_Additive_exPlanations-Enhanced_Virtual_Twins_for_Explainable_Treatment_Effect_Estimation.git

Optimizing drug treatments in older adults is essentialforimproving health outcomes and reducing drug-related issues. This study aimed to develop an updated and comprehensive explicit indicator of potentially inappropriate medications in older adults, and to evaluate its association with all-cause hospitalization. Administrative healthcare data from Local Health Units (LHUs) in Lombardy, Italy (2015–2018) were used to identify individuals aged ≥65 years who were exposed to potentially inappropriate medications (PIMs), identified through an explicit criteria-based list. Risk-set matching was used to select unexposed controls, and logistic regression models adjusted for high-dimensional propensity scores were used to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for all-cause hospitalization within 90 days after the first exposure date. Among 641 230 older patients, between 52.8% and 56.3% were exposed to at least one PIM. After matching (128 063 pairs) hospitalization rates were higher among the exposed (8.3%–10.2%) compared with the not exposed (5.1%– 6.0%) across the LHUs. Pooled estimates showed an increased risk of hospitalization for those exposed (ORs 1.55, 95% CIs 1.48–1.62), with the highest risk observed in adults aged 65–74 years (1.67, 1.59–1.76) and in women (1.62, 1.55–1.70). Over half of olderindividuals were exposed to at least one PIM, with a 55% increase in short-term hospitalization risk. These findings highlight the urgent need for improved prescribing practices tailored to older populations.

Growing individual case safety report (ICSR) volumes have intensified demand for scalable automated causality assessment. Large Language Models (LLMs) show promise, yet performance on clinically demanding tasks remains suboptimal and inference-time hyperparameter optimization has not been investigated. The objective of this study is to develop a Gaussian Process (GP)-compatible optimization objective and investigate whether temperature optimization improves LLM-expert agreement on Naranjo causality assessment of FAERS ICSRs. Expert causality assessments were performed on 723 stratified FAERS cases. OpenAI's GPT-5.2 was evaluated using chain-of-thought (CoT) prompting. Four composite metrics were developed: Weighted Cosine Similarity (WCS), Information-Weighted Agreement Score (IWAS), Entropy-Weighted Agreement and Cosine Similarity Score (EWACS), and Consensus-Weighted Cosine Similarity (CWCS) and Bayesian optimization using a GP surrogate with Probability of Improvement (PoI) acquisition was applied across temperature [0, 2]. GPT-5.2 outperformed prior biomedical LLMs at baseline (T = 0), achieving 74.1% agreement on question 5 and 65.4% on question 10 of Naranjo algorithm. Entropy analysis identified these as the sole informative optimization targets. Temperature showed no systematic population-level effect (b = 0.002, p = 0.959). EWACS-guided Bayesian optimization improved causality classification agreement from 45.0% to 72.0% (+27 pp), with the largest gain in Doubtful cases (+42.9 pp). EWACS was identified as the optimal GP-compatible metric. The absence of a universal temperature optimum indicates LLM performance is driven primarily by ICSR content, yet case-specific temperature selection produced meaningful improvements, supporting temperature optimization for LLM-assisted pharmacovigilance.

Currently, there is a lack of meta-analysis providing precise incidence estimates of adverse drug reactions (ADRs) and associated risk differences (RDs) linked to pharmacological treatments for muscle degeneration in Duchenne Muscular Dystrophy (DMD). A systematic search of PubMed, Embase, and ClinicalTrials.gov (January 1, 2000–June 5, 2026) identified clinical trials evaluating safety of pharmacological treatments for muscle degeneration in DMD. A random-effects meta-analysis estimated ADR incidence, and RDs versus a common reference (givinostat) were derived from the pooled incidences rather than from conventional head-to-head treatment contrasts. Thirteen clinical trials were included. Heterogeneity in ADR incidence across treatments (p < 0.05) was identified. Prednisone and prednisolone showed the highest proportions of serious ADRs. Discrepancies were found compared to Summary of Product Characteristics (SmPC): deflazacort showed higher rates of erythema (+ 1.5%) and pollakiuria (+ 4.8%), while vamorolone showed increased Cushingoid features and appetite (+ 7.0%). Subgroup analyses showed dose- and duration-dependent ADRs for deflazacort and prednisone. A reference-anchored exploratory comparison of pooled incidences showed lower gastrointestinal and hematological ADR risks with corticosteroid monotherapy than with givinostat plus corticosteroids. Deflazacort showed lower risks of diarrhea (RD: − 0.3860), decreased appetite (RD: − 0.3160), and decreased platelet count (RD: − 0.4110). Precise ADR estimates can support decision-making in DMD care, helping clinicians balance safety and efficacy, address family concerns, and promote adherence by contextualizing treatment risks.

Traditional antibiotic development remains slow and costly, with virtually no new classes targeting resistant gramnegative bacteria (GNB), highlighting the need for innovative approaches in generating new gram-negative antibacterial (GNAB) compounds to combat GNB. This study introduces a computational framework utilizing Junction Tree Variational Autoencoders (JT-VAE) and computational screening to design novel GNAB compounds. Generated compounds underwent Tanimoto similarity analysis and Lipinski’s Rule of Five (LRo5) screening to assess structural novelty and oral bioavailability, followed by agglomerative hierarchical clustering with cophenetic and silhouette score evaluation. Training on curated GNAB compounds from ChEMBL, the GVAE model, specifically the Junction Tree Variational Autoencoder (JT-VAE) model, demonstrated superior performance, producing 10,000 molecules with 100% validity. Subsequent filtering retained 2,141 compounds (21.41%) within optimal Tanimoto similarity thresholds (0.30-0.50) that balance novelty with known antibacterial substructures while meeting LRo5 with ≤2 violations. Property distributions aligned with fragment-based design principles, such as the Rule of Three. Clustering analysis revealed nuanced performance differences among benchmark models: JT-VAE achieved a cophenetic correlation coefficient (CCC) of 0.934 and silhouette score of 0.76, while unconstrained GVAE demonstrated superior clustering performance (silhouette score: 0.86, CCC: 0.98) despite occasionally generating disconnected molecular fragments. JT-VAE’s fragment-based approach successfully addresses limitations of atom-by-atom generation, particularly in handling aromatic systems essential for antibacterial activity. By combining hierarchical graph generation with rigorous cheminformatics filtering, this study provides a reproducible blueprint for accelerating GNAB compound discovery, advancing toward timely, data- driven solutions for the escalating antimicrobial resistance crisis.

Prolonged sea surface temperature (SST) departures above seasonal thresholds—are increasing in the Philippine Sea, threatening fisheries, reefs, and coastal livelihoods. The researchers assemble a daily, area-averaged 1995–2024 record that merges satellite SST, large-scale climate indices (ENSO, PDO, IOD, MJO), and ERA5 atmospheric fields, interpolate short gaps, standardize variables, and use 90-day lookbacks with a one-day (lead-1) forecast horizon. Recursive feature elimination identifies five informative drivers: SST anomaly, Indian Ocean Dipole, surface solar radiation, 2-meter air temperature, and 10-meter meridional wind. A univariate N-BEATS baseline (SST only) achieves lead-1 MAE 0.195 °C, RMSE 0.247 °C, R² 0.959, and MHW F1 0.937 on the held-out test block. A two-stream NBEATSX that encodes exogenous histories reduces lead-1 errors to MAE 0.087 °C and RMSE 0.108 °C (R² 0.992) while raising MHW F1 to 0.971. Kernel SHAP attribution shows the retained drivers—winds, air temperature, solar radiation, and SST anomaly—jointly explain roughly half of the model’s predictive variance alongside the autoregressive SST channel. These results demonstrate that integrating carefully selected exogenous variables into an interpretable deep-learning framework substantially improves near-term MHW predictability and highlights the physical drivers modulating Philippine Sea heat extremes.

The Sessa Empirical Estimator is a data-driven method that constructs treatment episodes from observational data by incorporating the K-Means clustering algorithm. However, K-Means is limited by its assumption of spherical clusters, which often does not hold true for complex, real-world pharmacoepidemiological datasets exhibiting non-convex shapes and varying densities. This study extends the Sessa Empirical Estimator by integrating alternative clustering algorithms to accommodate such data characteristics. Various Cluster Validity Indices were also utilized to determine optimal cluster configurations. Statistical analysis, employing the Mann-Whitney U test, demonstrates a highly significant difference between the distributions of median duration estimates from the novel Sessa Empirical Estimator and the traditional approach. The proposed method showed that the novel Sessa Empirical Estimator consistently yields more compact distributions with fewer outliers, indicating superior consistency, precision, and reliability. This improved Sessa Empirical Estimator provides a more robust and accurate methodology for estimating medication exposure, enhancing the utility of observational data in pharmacoepidemiological research.

Climate change affects public health, particularly through increased Aedes-borne dengue illnesses. This study investigated dengue transmission across time and space in Cebu Province, Philippines using wavelet-based time-series analysis and causal inference. Continuous Wavelet Transform (CWT) revealed strong annual cycles of dengue incidence between 2013 and 2023, with disruptions during the COVID-19 pandemic. Post-pandemic recovery of seasonality was fastest in Cebu city, while Mandaue city and Lapu-Lapu city experienced extended disturbances. Using the Peter and Clark Momentary Conditional Independence (PCMCI) algorithm, lagged causal effects from relative humidity and the minimum infection rates (MIR) of DENV-2 and DENV-4 transmitted by Aedes albopictus (Skuse) in Cebu and Mandaue cities were determined. In contrast, Lapu- Lapu city was influenced by temperature, precipitation, and the MIR of DENV-1 and DENV-4. These findings underscore the value of integrating climate and arboviral data in a dengue vector species in localized outbreak forecasting.

Common data models (CDMs) standardize healthcare data to facilitate reproducible and consistent analyses, supporting decision-making in medicine and vaccine safety and pharmacoepidemiology. Despite their global recognition, the implementation and use of CDMs in the Nordic countries remain underexplored, particularly in the context of cross-national collaborations.

Assessed the ability of two off-the-shelf large language models, ChatGPT and Gemini, to support the design of pharmacoepidemiological studies. While ChatGPT and Gemini show promise in certain tasks supporting pharmacoepidemiological study design, their limitations in relevance and coding accuracy highlight the need for critical oversight by domain experts.

Drug-resistant bacteria pose a significant global health threat, driving the need for innovative antibiotic development. The efficacy of these antibiotics is evaluated through biological potency assays that measure their ability to elicit targeted responses. Beta-lactam bacteria produce beta-lactamases, enzymes that hydrolyze the beta-lactam ring, rendering the antibiotics ineffective. To counteract this mechanism, beta-lactamase inhibitors play a pivotal role by preventing the enzymatic degradation of beta-lactam antibiotics. This study focuses on predicting the chemical compositions and molecular motifs that characterize active beta-lactamase inhibitors. Using k-means clustering, active small-molecule beta-lactamase inhibitors were categorized based on their unique molecular structures. A graph-based modeling approach was employed to represent these molecular structures, then leveraging Graph Attention Networks (GAT) to identify and predict substructural features associated with each distinct cluster. Molecular graph representations served as inputs for the GAT model, enabling precise classification of compounds into distinct clusters. The GAT model's performance in multiclass classification was benchmarked against traditional approaches, demonstrating superior accuracy in identifying key substructures that differentiate active beta-lactamase inhibitors. Additionally, the attention mechanism within the GAT model facilitated the identification of specific molecular motifs by focusing on relevant structural features during the learning process. The findings highlight the effectiveness of the graph-based approach in advancing the understanding and prediction of active betalactamase inhibitors, with implications for drug discovery and combating antibiotic resistance.

Marine heatwaves (MHWs), which are characterized by prolonged periods of anomalously high sea surface temperatures (SST), pose significant threats to marine ecosystems, coastal economies, and global biodiversity. The Philippines—situated in the Coral Triangle—relies heavily on its marine resources, making it vulnerable to these impacts. This study explores the use of deep learning models to predict SST and identify the occurrence of MHWs in the Philippine Sea using data sourced from NOAA Daily OISST. The study evaluated four deep learning architectures: (1) Long Short-Term Memory (LSTM), (2) Convolutional Neural Networks (CNN), (3) a hybrid model combining the strengths of both LSTM and CNN, and (4) Neural Basis Expansion Analysis for Interpretable Time Series (N-BEATS). To optimize performance, all models were fine-tuned using Bayesian optimization (BO). The hybrid model demonstrated a superior performance in SST forecasting, achieving RMSEs of 0.1034 °C at 1-day and 0.0652 °C at 14-day lead times. For MHW detection, N-BEATS achieved an RMSE of 0.0523 °C, with a recall of 98.25%, ensuring reliable identification of rare events. N-BEATS also obtained an F1-score of 97.67%, reflecting a strong balance between precision and recall, minimizing both false positives and false negatives. These results highlight the potential of neural networks in providing accurate SST forecasting and MHW detection to mitigate climate change impacts on the Philippines’ marine ecosystems and coastal communities.

Crop diseases reduce yield, affecting a nation's agricultural sector. This is more pronounced in nations such as the Philippines, where agriculture is the nation's foundation. To decrease the cost and time needed for disease detection, this study utilized Transfer Learning on three Convolutional Neural Networks (DenseNet201, EfficientNetV2M, and InceptionResNetV2) to identify three corn diseases: Common Rust, Blight, and Gray Leaf Spot. Data Augmentation was used to diversify the image dataset, resulting in 4,000 images per category. This research presents a comparison between the different architectures. It also illustrates the effect of the Soft Voting Ensemble Method.

Beyond economic concerns, inflation has also influenced the media narratives of other topics. This study aims to explore how inflation news impacts various aspects of Philippine society by investigating the relationships of inflation-related news articles and other news topics using Granger Causality.

Utilizes a novel complex multi-layered network approach for aggregating drug networks of varying information into a single unified network. Under the hypothesis that drugs that are closely related will appear as neighboring nodes, using a community detection approach allows novel discovery of new applications of old and existing drugs. This paper describes a method to aggregate multi-layered network. Drug data is collated from DrugBank and PubChem; an adjacency matrix was created, from which the Normalized Graph Laplacian is generated. Prior to network aggregation, eigenvectors, and Uniform Manifold Approximation and Projection (UMAP) methods were utilized to reduce the network dimension. Additionally, an inference was implemented to test for complete spatial randomness (CSR) to check for node scattering.