Vipul Periwal, Ph.D.

Research Goal

The ultimate goal of our research is to predict quantitatively the effects of therapeutic interventions in human disease.

Current Research

Forecasting HIV drug resistance

When a patient’s human immunodeficiency virus (HIV) becomes resistant to their antiretroviral therapy (ART), that therapy fails and they must be switched to another regimen, often after the failure has already occurred. Drug resistance is a central obstacle in the long-term clinical management of HIV/AIDS, and we aim to make the decision to switch predictable rather than reactive. Doing so means knowing which mutations are reachable from the virus a patient carries now, under the drugs they are now taking, and not only whether that virus is already resistant. The output should be a set of testable statements rather than a resistance probability: a named drug regimen, named mutations, and named dependencies between positions. We aim to develop this into a patient-specific digital twin for HIV drug resistance, extending from the protease enzyme alone to the full POL region, which encodes the enzymes HIV needs to replicate, under the drug combinations used in treatment. We aim to quantify, for an individual patient, when a change of therapy is least likely to be followed by resistance, and to test that against records of real treatment changes and their outcomes. By forecasting the specific mutations, a patient’s virus is likely to develop under a given drug combination, this work could help clinicians choose or change treatment before resistance takes hold and identify antibody therapies that are harder for the virus to escape. In the long term, this approach could improve durable control of HIV and help anticipate treatment resistance in other rapidly changing infections and cancers.

Auditable biomedical machine learning

Single-cell RNA sequencing (scRNA-seq) measures gene expression in individual cells rather than averaging across a tissue and spatially resolved methods add where in the tissue each measurement came from. These data have made it possible to study how cell types differ and how cells change identity during development. They are also hard to work with: each cell yields readings for tens of thousands of genes, most genes register zero in most cells because the technique captures only a fraction of the molecules present, and cells are sampled at a few time points rather than continuously. Our long-term goal is to obtain mechanistic insight into gene regulation from single-cell and spatially resolved data: which genes regulate which, in a form that can be tested experimentally. A model can predict cell types or reconstruct a developmental trajectory accurately without identifying the genes responsible. Reaching mechanistic insight therefore requires every stage of an analysis to be auditable: a reduced representation should name the genes constituting each component, a trained predictor should name the genes it uses, and a mechanistic model should name which gene acts on which. We are pursuing several specific approaches toward auditable analysis.

Personalized predictive machine learning from sparse data

We are developing mechanistic and data-driven machine-learning approaches for forecasting, detecting, and providing early warning of health-relevant anomalies across biological, physiological, and environmental systems, with an emphasis on personalized medical applications. The work integrates dynamical-systems theory, reservoir computing, statistical modeling, time-series forecasting, machine learning, and probabilistic anomaly detection, with an emphasis on interpretable methods that learn meaningful temporal structure from complex nonlinear time-series while reducing dependence on large labeled-datasets. As an example, we developed CASCADE (Chaotic Attractor Sensitivity for Cardiac Anomaly Detection), a personalized framework for cardiac arrhythmia detection from electrocardiographic signals. CASCADE learns patient-specific normal cardiac dynamics and identifies arrhythmic events as persistent failures of short-term predictability relative to individualized baselines.

Research in Plain Language

Insulin resistance is a major risk factor for several common diseases. These include diabetes, heart disease, high blood pressure, and some forms of cancer. Scientists do not yet understand how cells become insulin resistant. Our research group studies three processes related to insulin resistance.

We study problems in insulin’s ability to control the breakdown of fat. This leads to increased levels of free fatty acids (FFA) in the blood. With elevated FFA levels, insulin does not work as well as it should in tissues or cells. A way to measure insulin’s effects on fat break down and blood FFA levels would help monitor various conditions of insulin resistance. We are developing an index to show how FFA levels respond to insulin. Many disease complications relate to oxidative stress. These include problems associated with diabetes, Parkinson's, Alzheimer's, and hardening of the arteries. In oxidative stress, the body cannot remove toxins and repair damage caused by reactive oxygen species (ROS). But ROS are not always bad. ROS signaling is important in how cells work. Cellular structures that produce energy, mitochondria, have ROS but do not show oxidative stress. Using math, we develop models to describe how this happens. The goal is to understand the damage of oxidative stress in insulin resistance and obesity.

Our research also aims to understand how fat tissue growth relates to insulin resistance and diabetes. Fat tissue grows in two ways—when the number of cells increase and when the size of cells increase. How do genetics and diet affect the number and size of fat cells in obesity? We focus on the role played by particular agents that increase insulin sensitivity. Using math, we model changes in fat cell size over time under several conditions. This work will provide a global view of fat cell size and fat tissue growth in conditions featuring insulin resistance.