Harmonizing across datasets to improve the transferability of drug combination prediction.

Combination treatment has multiple advantages over traditional monotherapy in clinics, thus becoming a target of interest for many high-throughput screening (HTS) studies, which enables the development of machine learning models predicting the response of new drug combinations. However, most existing models have been tested only within a single study, and these models cannot generalize across different datasets due to significantly variable experimental settings. Here, we thoroughly assessed the transferability issue of single-study-derived models on new datasets. More importantly, we propose a method to overcome the experimental variability by harmonizing dose-response curves of different studies. Our method improves the prediction performance of machine learning models by 184% and 1367% compared to the baseline models in intra-study and inter-study predictions, respectively, and shows consistent improvement in multiple cross-validation settings. Our study addresses the crucial question of the transferability in drug combination predictions, which is fundamental for such models to be extrapolated to new drug combination discovery and clinical applications that are de facto different datasets.

Motor deficit and lack of overt dystonia in Dlx conditional Dyt1 knockout mice.

DYT1 or DYT-TOR1A dystonia is early-onset generalized dystonia caused by a trinucleotide deletion of GAG in the TOR1A or DYT1 gene leads to the loss of a glutamic acid residue in the resulting torsinA protein. A mouse model with overt dystonia is of unique importance to better understand the DYT1 pathophysiology and evaluate preclinical drug efficacy. DYT1 dystonia is likely a network disorder involving multiple brain regions, particularly the basal ganglia. Tor1a conditional knockout in the striatum or cerebral cortex leads to motor deficits, suggesting the importance of corticostriatal connection in the pathogenesis of dystonia. Indeed, corticostriatal long-term depression impairment has been demonstrated in multiple targeted DYT1 mouse models. Pappas and colleagues developed a conditional knockout line (Dlx-CKO) that inactivated Tor1a in the forebrain and surprisingly displayed overt dystonia. We set out to validate whether conditional knockout affecting both cortex and striatum would lead to overt dystonia and whether machine learning-based video behavioral analysis could be used to facilitate high throughput preclinical drug screening. We generated Dlx-CKO mice and found no overt dystonia or motor deficits at 4 months. At 8 months, retesting revealed motor deficits in rotarod, beam walking, grip strength, and hyperactivity in the open field; however, no overt dystonia was visually discernible or through the machine learning-based video analysis. Consistent with other targeted DYT1 mouse models, we observed age-dependent deficits in the beam walking test, which is likely a better motor behavioral test for preclinical drug testing but more labor-intensive when overt dystonia is absent.

Modeling the behavior of multiple subjects using a Cauchy-Schwarz regularized Partitioned Subspace Variational AutoEncoder (CS-PS-VAE).

Effectively modeling and quantifying behavior is essential for our understanding of the brain. Modeling behavior across different subjects in a unified manner remains a significant challenge in the field of behavioral quantification, which necessitates partitioning the behavioral data into features that are common across subjects, and others that are distinct to each subject. We build on a semi-supervised approach to partition the subspace adequately known as a Partitioned Subspace Variational AutoEncoder (PS-VAE), and propose a novel regularization based on the Cauchy-Schwarz divergence to model the distinct features across subjects. Our model, called the Cauchy-Schwarz regularized Partitioned Subspace Variational AutoEncoder (CS-PS-VAE), successfully models continuously varying differences in behavior, and models distinct features of the behavioral videos across subjects in an unsupervised manner. This method is also successful at uncovering the relationships between recorded neural data and the ensuing behavior.

Micro-dissection and integration of long and short reads to create a robust catalog of kidney compartment-specific isoforms.

Studying isoform expression at the microscopic level has always been a challenging task. A classical example is kidney, where glomerular and tubulo-interstitial compartments carry out drastically different physiological functions and thus presumably their isoform expression also differs. We aim at developing an experimental and computational pipeline for identifying isoforms at microscopic structure-level. We microdissected glomerular and tubulo-interstitial compartments from healthy human kidney tissues from two cohorts. The two compartments were separately sequenced with the PacBio RS II platform. These transcripts were then validated using transcripts of the same samples by the traditional Illumina RNA-Seq protocol, distinct Illumina RNA-Seq short reads from European Renal cDNA Bank (ERCB) samples, and annotated GENCODE transcript list, thus identifying novel transcripts. We identified 14,739 and 14,259 annotated transcripts, and 17,268 and 13,118 potentially novel transcripts in the glomerular and tubulo-interstitial compartments, respectively. Of note, relying solely on either short or long reads would have resulted in many erroneous identifications. We identified distinct pathways involved in glomerular and tubulo-interstitial compartments at the isoform level, creating an important experimental and computational resource for the kidney research community.

Representation of molecules for drug response prediction.

The rapid development of machine learning and deep learning algorithms in the recent decade has spurred an outburst of their applications in many research fields. In the chemistry domain, machine learning has been widely used to aid in drug screening, drug toxicity prediction, quantitative structure-activity relationship prediction, anti-cancer synergy score prediction, etc. This review is dedicated to the application of machine learning in drug response prediction. Specifically, we focus on molecular representations, which is a crucial element to the success of drug response prediction and other chemistry-related prediction tasks. We introduce three types of commonly used molecular representation methods, together with their implementation and application examples. This review will serve as a brief introduction of the broad field of molecular representations.

Deep learning identifies erroneous microarray-based, gene-level conclusions in literature.

More than 110 000 publications have used microarrays to decipher phenotype-associated genes, clinical biomarkers and gene functions. Microarrays rely on digital assaying the fluorescence signals of arrays. In this study, we retrospectively constructed raw images for 37 724 published microarray data, and developed deep learning algorithms to automatically detect systematic defects. We report that an alarming amount of 26.73% of the microarray-based studies are affected by serious imaging defects. By literature mining, we found that publications associated with these affected microarrays have reported disproportionately more biological discoveries on the genes in the contaminated areas compared to other genes. 28.82% of the gene-level conclusions reported in these publications were based on measurements falling into the contaminated area, indicating severe, systematic problems caused by such contaminations. We provided the identified published, problematic datasets, affected genes and the imputed arrays as well as software tools for scanning such contamination that will become essential to future studies to scrutinize and critically analyze microarray data.

Timesias: A machine learning pipeline for predicting outcomes from time-series clinical records.

The prediction of outcomes is a critical part of the clinical surveillance for hospitalized patients. Here, we present Timesias, a machine learning pipeline which predicts outcomes from real-time sequential clinical data. The strategy implemented in Timesias is the first-place solution in the crowd-sourcing DII (discover, innovate, impact) National Data Science Challenge involving more than 100,000 patients, achieving 0.85 as evaluated by AUROC (area under receiver operator characteristic curve) in predicting the early onset of sepsis status. Timesias is freely available via PyPI and GitHub. For complete details on the use and execution of this protocol, please refer to Guan et al. (2021).

A survival model generalized to regression learning algorithms.

Survival prediction is an important problem that is encountered widely in industry and medicine. Despite the explosion of artificial intelligence technologies, no uniformed method allows the application of any type of regression learning algorithm to a survival prediction problem. Here, we present a statistical modeling method that is generalized to all types of regression learning algorithm, including deep learning. We present its empirical advantage when it is applied to traditional survival problems. We demonstrate its expanded applications in different types of regression learning algorithm, such as gradient boosted trees, convolutional neural networks and recurrent neural networks. Additionally, we demonstrate its application in clinical informatic data, pathological images and the hardware industry. We expect that this algorithm will be widely applicable for diverse types of survival data, including discrete data types and those suitable for deep learning such as those with time or spatial continuity.

Assessment of the timeliness and robustness for predicting adult sepsis.

Sepsis is a leading cause of death among inpatients at hospitals. However, with early detection, death rate can drop substantially. In this study, we present the top-performing algorithm for Sepsis II prediction in the DII National Data Science Challenge using the Cerner Health Facts data involving more than 100,000 adult patients. This large sample size allowed us to dissect the predictability by age-groups, race, genders, and care settings and up to 192 hr of sepsis onset. This large data collection also allowed us to conclude that the last six biometric records on average are informative to the prediction of sepsis. We identified biomarkers that are common across the treatment time and novel biomarkers that are uniquely presented for early prediction. The algorithms showed meaningful signals days ahead of sepsis onset, supporting the potential of reducing death rate by focusing on high-risk populations identified from heterogeneous data integration.