Lupus nephritis (LN) is one of the most common and severe complications of lupus, affecting the kidneys and often leading to long-term damage or even kidney failure if not treated properly. The current gold standard for guiding treatment and predicting outcomes in LN is a kidney biopsy, a procedure where a small piece of kidney tissue is collected. The tissue is then stained with special dyes called hematoxylin and eosin (H&E) and examined under a microscope by a pathologist. LN is usually categorized as either proliferative (increase in the number of cells) or non-proliferative (no increase in the number of cells) glomerulonephritis, but studies show that long-term outcomes are more closely tied to inflammation in another part of the kidney, called the tubulointerstitial (TI) area. However, no currently available tools can accurately measure inflammation in these distinct kidney compartments. Dr. Clark’s research will address this gap by first using detailed imaging of kidney biopsies stained with complex methods to train artificial intelligence systems to map and understand the immune environments within the kidney. The trained system could then detect these same immune environments on biopsies stained only with regular H&E.

Dr. Clark will use HD images tagged with up to 60 different cellular and molecular labels to train a machine-learning tool (called a deep convolutional neural network (DCNN)) to identify and classify kidney structures and immune cell neighborhoods —clusters of interacting cells in tissue—within them. Dr. Clark aims to teach the DCNN to identify features on existing kidney biopsy samples that might predict how well someone will do over time or how they might respond to treatment. Unlike traditional blood tests, which only reveal what’s happening in the bloodstream, this method allows researchers to study the local immune environment in the kidneys, where immune cells interact directly with the tissue. Understanding these patterns could enable doctors to make better treatment decisions and improve the accuracy of lupus nephritis diagnoses. Ultimately, the goal is to develop a user-friendly tool that pathologists can use to assess kidney biopsies more effectively and guide treatment choices.

What this means for people with lupus

Lupus nephritis is a life-threatening condition that requires precise treatment decisions to prevent kidney damage or failure. By using AI to segment and analyze kidney biopsies, Dr. Clark’s project could help doctors more accurately predict disease progression and customize treatment for each patient, potentially preventing kidney failure and improving outcomes for people with lupus.

Lupus nephritis (LN) is the most common severe manifestation of systemic lupus erythematosus (SLE) and contributes significantly to overall mortality. Up to 50% of SLE patients develop LN, in many cases necessitating treatment with toxic immunosuppressive therapies. Despite such aggressive treatments, many patients do not respond to therapy, and up to 40% of LN patients progress to renal failure within 5 years of diagnosis. Much work in human LN has focused on peripheral autoimmunity and its manifestation within the kidney, glomerulonephritis (GN). However, prognosis is more tightly linked to tubulointerstitial inflammation (TII) and scarring which are associated with complex in situ immune states. In clinical practice, renal biopsy and conventional histology have remained the gold standard for patient stratification and treatment guidance. LN is categorized as having either proliferative or non-proliferative nephritis based on the activity and frequency of glomerular lesions with therapeutic decisions being based upon this classification. Even in the NIH activity index, which does score TII, it is only worth 3 of 24 points. At best, methods to assess TII are rudimentary. There are no subjective criteria that adequately balances glomerular and TI disease. Beyond this, there are certainly no objective tools that can be used to assess LN activity or chronicity within the glomerular or TI compartments. In this grant application, we demonstrate that we can perform high-dimensional (HD) imaging of human LN biopsies (LN cohorts analyzed with panels of 42 or 61 antibodies). Furthermore, we have developed several analytical tools, based on custom-trained deep convolutional neural networks (DCNNs), and other machine learning models, to segment and annotate immune cells, identify the neighborhoods they form, and understand where these cells and neighborhoods reside within complex renal structures. Finally, we have stained every HD imaged biopsy with H&E, registered these images to the corresponding HD image stack, and have demonstrated that we can use the HD dataset to precisely define structural features on H&E-stained sections. We propose to use our HD datasets to train a DCNN to identify features on H&E associated with specific, prognostically meaningful, immune cell states (Milestones 1-3). We will then validate this tool, the H&E DCNN, in prospective LN clinical and trial cohorts (Milestones 4-5). Finally, we will develop a graphical user interface (GUI) which will facilitate use of our H&E DCNN by other researchers and enable development of a commercial product (Milestone 6).