Biology subjects

Goh, W. W. B.

Publications and source records attributed to Goh, W. W. B..

6 recordsLinked to original sources

bioRxiv2024

Establishing the foundations for a data-centric AI approach for virtual drug screening through a systematic assessment of the properties of chemical data

Researchers have adopted model-centric artificial intelligence (AI) approaches in cheminformatics by using newer, more sophisticated AI methods to take advantage of growing chemical libraries. It has been shown that complex deep learning methods outperform conventional machine learning (ML) methods in QSAR and ligand-based virtual screening1-3 but such approaches generally lack explanability. Hence, instead of developing more sophisticated AI methods (i.e., pursuing a model-centric approach), we wanted to explore the potential of a data-centric AI paradigm for virtual screening. A data-centric AI is an intelligent system that would automatically identify the right type of data to collect, clean and curate for later use by a predictive AI and this is required given the large volumes of chemical data that exist in chemical databases - PubChem alone has over 100 million unique compounds. However, a systematic assessment of the attributes and properties of suitable data is needed. We show here that it is not the result of deficiencies in current AI algorithms but rather, poor understanding and erroneous use of chemical data that ultimately leads to poor predictive performance. Using a new benchmark dataset of BRAF ligands that we developed, we show that our best performing predictive model can achieve an unprecedented accuracy of 99% with a conventional ML algorithm (SVM) using a merged molecular representation (Extended + ECFP6 fingerprints), far surpassing past performances of virtual screening platforms using sophisticated deep learning methods. Thus, we demonstrate that it is not necessary to resort to the use of sophisticated deep learning algorithms for virtual screening because conventional ML can perform exceptionally well if given the right data and representation. We also show that the common use of decoys for training leads to high false positive rates and its use for testing will result in an over-optimistic estimation of a models predictive performance. Another common practice in virtual screening is defining compounds that are above a certain pharmacological threshold as inactives. Here, we show that the use of these so-called inactive compounds lowers a models sensitivity/recall. Considering that some target proteins have a limited number of known ligands, we wanted to also observe how the size and composition of the training data impact predictive performance. We found that an imbalance training dataset where inactives outnumber actives led to a decrease in recall but an increase in precision, regardless of the model or molecular representation used; and overall, we observed a decrease in the models accuracy. We highlight in this study some of the considerations that one needs to take into account in future development of data-centric AI for CADD.

cancer biology↗

bioRxiv2023

Optimizing Proteomics Data Differential Expression Analysis via High-Performing Rules and Ensemble Inference

In the process of identifying phenotype-specific or differentially expressed proteins from proteomic data, a standard workflow consists of five key steps: raw data quantification, expression matrix construction, matrix normalization, missing data imputation, and differential expression analysis. However, due to the availability of multiple options at each step, selecting ad hoc combinations of options can result in suboptimal analysis. To address this, we conducted an extensive study involving 10,808 experiments to compare the performance of exhaustive option combinations for each step across 12 gold standard spike-in datasets and three quantification platforms: FragPipe, MaxQuant, and DIA-NN. By employing frequent pattern mining techniques on the data from these experiments, we discovered high-performing rules for selecting optimal workflows. These rules included avoiding normalization, utilizing MinProb for missing value imputation, and employing limma for differential expression analysis. We found that workflow performances were predictable and could be accurately categorized using average F1 scores and Matthews correlation coefficients, both exceeding 0.79 in 10-fold cross-validations. Furthermore, by integrating the top-ranked workflows through ensemble inference, we not only improved the accuracy of differential expression analysis (e.g., achieving a 1-5% gain under five performance metrics for FragPipe), but also enhanced the workflows ability to aggregate proteomic information across various levels, including peptide and protein level intensities and spectral counts, providing a comprehensive perspective on the data. Overall, our study highlights the importance of selecting optimal workflow combinations and demonstrates the benefits of ensemble inference in improving both the accuracy and comprehensiveness of proteomic data analysis.

bioinformatics↗

bioRxiv2023

Uncovering the Consequences of Batch-Effect Associated Missing Values in Omics Data Analysis

Statistical analyses in high-dimensional omics data are often hampered by the presence of batch effects (BEs) and missing values (MVs), but the interaction between these two issues is not well-studied nor understood. MVs may manifest as a BE when their proportions differ across batches. These are termed as Batch-Effect Associated Missing values (BEAMs). We hypothesized that BEAMs in data may introduce bias which can impede the performance of missing value imputation (MVI). To test this, we simulated data with two batches, then introduced over 100 iterations, either 20% and 40% MVs in each batch (BEAMs) or 30% in both (control). K-nearest neighbours (KNN) was then used to perform MVI, in a typical global approach (M1) and a supposed superior batch-sensitized approach (M2). BEs were then corrected using ComBat. The effectiveness of the MVI was evaluated by its imputation accuracy and true and false positive rates. Notably, when BEAMs existed, M2 was generally undesirable as the differing application of MV filtering in M1 and M2 strategies resulted in an overall coverage deficiency. Additionally, both M1 and M2 strategies suffered in the presence of BEAMs, highlighting the need for a novel approach to handle MVI in data with BEAMs. Author summaryData in high-throughput omics data are often combined from different sources (batches), which creates batch effects in the data. Missing values are a common occurrence in these data, and their proportions are assumed to be equal across batches. However, instances exist when these proportions vary between batches, such as one batch having more missing values than another, resulting in batch effect associated missing values. Missing values are often dealt with through missing value imputation, but whether the variation in missing value proportions across batches affects imputation outcomes is unknown. In this paper, we investigate the consequence of performing imputation when this issue persists. We simulated data with equal and unequal missing value proportions, then assessed the performance of k-nearest neighbours imputation by its imputation accuracy and downstream analysis outcomes. This revealed that unequal missing value proportions worsens imputation and establishes the need for smarter imputation strategies to handle this complication.

bioinformatics↗

bioRxiv2021

Single-sample proteome enrichment enables missing protein recovery and phenotype association

Proteomic studies characterize the protein composition of complex biological samples. Despite recent developments in mass spectrometry instrumentation and computational tools, low proteome coverage remains a challenge. To address this, we present Proteome Support Vector Enrichment (PROSE), a fast, scalable, and effective pipeline for scoring protein identifications based on gene co-expression matrices. Using a simple set of observed proteins as input, PROSE gauges the relative importance of proteins in the phenotype. The resultant enrichment scores are interpretable and stable, corresponding well to the source phenotype, thus enabling reproducible recovery of missing proteins. We further demonstrate its utility via reanalysis of the Cancer Cell Line Encyclopedia (CCLE) proteomic data, with prediction of oncogenic dependencies and identification of well-defined regulatory modules. PROSE is available as a user-friendly Python module from https://github.com/bwbio/PROSE.

bioinformatics↗

bioRxiv2020

Fate mapping analysis reveals a novel dermal migratory Langerhans-like cell population

Dendritic cells residing in the skin represent a large family of antigen presenting cells, ranging from long-lived Langerhans cells (LC) in the epidermis to various distinct classical dendritic cell subsets in the dermis. Through genetic fate mapping analysis and single cell RNA sequencing we have identified a novel separate population of LC-independent CD207+CD326+ LClike cells in the dermis that homed at a slow rate to the LNs. These LClike cells were long-lived and radioresistant but, unlike LCs, they were gradually replenished by bone-marrow-derived precursors under steady state. LClike cells together with cDC1s were the main migratory CD207+CD326+ cell fractions present in the LN and not, as currently assumed, LCs, which were barely detectable, if at all. These findings bring new insights into the dynamism of cutaneous dendritic cells and opens novel avenues in the development of treatments to cure inflammatory skin disorders.

immunology↗

bioRxiv2019

ER stress sensor Ire1 deploys a divergent transcriptional program in response to lipid bilayer stress

Membrane integrity at the endoplasmic reticulum (ER) is tightly regulated and is implicated in metabolic diseases when compromised. Using an engineered sensor that exclusively activates the unfolded protein response (UPR) during aberrant ER membrane lipid composition, we identified pathways beyond lipid metabolism that are necessary to maintain ER integrity in yeast and are conserved in C. elegans. To systematically validate yeast mutants disrupting ER membrane homeostasis, we identified a lipid bilayer stress (LBS) sensing switch in the UPR transducer protein Ire1, located at the interface of the amphipathic and transmembrane helices. Furthermore, transcriptome and chromatin immunoprecipitation (ChIP) analyses pinpoint the UPR as a broad-spectrum compensatory pathway in which LBS and proteotoxic stress-induced UPR deploy divergent transcriptional programs. Together, these findings reveal the UPR program as the sum of two independent stress events and could be exploited for future therapeutic intervention.

cell biology↗