Novel Predictive Model For Explaining How Anti-Fibrotic Drugs Work

April 10, 2024

By Deborah Borfitz

April 10, 2024 | In drug discovery, the focus of machine learning and artificial intelligence tools has been on predicting outcomes without explanation or understanding of the biochemical pathways mediating those effects. But rigorous translation requires a solid science-based foundation to explain how a drug works, not just that it does, according to Jeff Saucerman, PhD., professor of biomedical engineering and cardiovascular medicine at the University of Virginia. 

To that end, Saucerman and his colleagues created a logic-based mechanistic machine learning (LogiMML) approach combining the strengths of machine learning with a previously developed mechanistic network model of the signaling that happens in cardiac fibroblasts. The behavior of the mechanistic model has been validated in hundreds of conditions and aided the design of later basic science experiments, he says. 

Fibroblasts help repair the heart after injury by producing collagen and causing wounds to contract, says Saucerman. But they can also cause harmful scarring, or fibrosis, as part of the repair process. 

In their most recently published study in Proceedings of the National Academy of Sciences (DOI: 10.1073/pnas.2303513121), he and his colleagues used the LogiMML model to predict mechanisms that mediate the differential effects of 13 clinically relevant drugs on fibroblasts. Specifically, they showed how pirfenidone (treatment for idiopathic pulmonary fibrosis) and Src inhibition affect the regulation of the phenotypic features of, respectively, actin filament assembly and actin–myosin stress fiber formation. Additionally, Src inhibition was predicted by the LogiMML model to regulate actin filaments via the phosphoinositide 3-kinase (PI3K) pathway, which they further validated in additional experiments. 

Pirfenidone was both regulating actin expression in the cells as well as how much of it organized into actin filaments, the tethers supporting contraction of the fibroblasts, explains Saucerman. Similarly, experimental Src inhibitor WH4023 was regulating the assembly of the contractile fibers independent of the expression of actin and myosin proteins responsible for creation of the fibers themselves. 

The more typical way of identifying candidate drugs is high-throughput screening against certain molecular targets or cell phenotypes, Saucerman says. But not knowing how the compounds work has been a major limitation for subsequent studies, including those in cells and animals as well as clinical trials in people. “This knowledge is needed to design clinical trials... to be certain that the drugs are working in the way that we intend and also to be aware of any potential side effects.”   

Fusing of Models

Cardiac fibrosis has been the focus of Saucerman’s lab for the past 14 years, and with good reason, he says. Cardiovascular disease, the biggest cause of death in the U.S. and many other countries, typically involves some level of fibrosis that “stiffens the heart and makes it difficult for it to relax.” A similar pattern is seen with many other chronic diseases such as idiopathic pulmonary fibrosis, fibrotic liver diseases, and autoimmune diseases like rheumatoid arthritis and scleroderma. 

For the latest study, automated microscopy was used to investigate 108 different treatment conditions (a combination of different drugs and environmental stimuli) based on characteristics of fibroblasts—measurement of their size, shape, and expression of collagen, alpha-smooth muscle actin (a marker of the contractile state), and actin filaments. Further high-content image analysis was then done to reveal 137 phenotypic features at the single-cell level that change in response to drug treatments, Saucerman explains. 

With this rich dataset in hand, the research team next applied traditional machine learning approaches such as clustering and principal component analysis to find patterns in the data, he continues. They then turned to their LogiMML model to learn how those drugs were working.  

The mechanistic approach on its own could only have made predictions for the three traditional markers of fibroblast activation, says Saucerman, and they wanted to make calculations for all the phenotypic cell features that could provide insights on how fibroblasts were being impacted by the drugs. They could then “infer the new pathways that link from the drugs through the network down into these new phenotypes.” 

While it is known that actin filaments in fibroblasts may participate in contraction, he says, not much is known about how they are regulated independent of protein expression. “When we looked across our screening data using machine learning, we found patterns that showed... [both] actin filament assembly and actin–myosin stress fiber formation were regulated very differently by the drugs than the traditional markers. 

“That was what clued us in that there would be novel mechanisms to study,” continues Saucerman. The two distinct features were quantified via texture analysis, which identify correlated patterns within an image rather than just looking at its intensity. It was then that the researchers saw that pirfenidone and Src inhibitor WH4023 were strongly associated with each of those two phenotypes. 

In the final study stage, the researchers experimentally validated their approach in human cardiac fibroblasts by applying the Src inhibitor drug and a PI3K inhibitor predicted to be downstream of the action of WH4023 to find they could indeed regulate the actin-myosin stress fiber formation. That confirmed that the LogiMML model could infer new pathways in cells. 

Next Steps

Saucerman says he has collaborators with mouse models that could be used for further testing of pirfenidone and Src inhibitor drugs. Mechanistic studies are also planned to tease apart additional pathways that might be predictive for drug regulation.  

He and his team are also interested in repeating the latest study but in lung, liver, and tumor-associated fibroblasts to learn if they exhibit the same patterns of regulation by the two drugs or use different pathways to regulate their activation, he adds.  

The focus to date has been primarily on drugs approved by the Food and Drug Administration (FDA), to better understand how they work, says Saucerman, making the studies “not yet relevant to new drugs.” But in other collaborations with industry, the same type of approach has been employed using investigational drugs as positive controls for commercial drug development efforts. Additional studies could be done with proprietary compounds combined with FDA-approved therapies to potentially elucidate the body’s cellular-level response to the investigational agents. 

While powerful examples of machine learning exist in biology—most famously AlphaFold in its ability to predict protein structure—the basis of those predictions remains unclear due to the black-box nature of deep learning models, he says. The predictions do not necessarily provide an understanding of how a biological system works and, without that, “it is difficult for the field to reliably build on those results... we need to get beyond the next prediction that gets us to the next paper.”