Herophilus Aims To ‘Industrialize’ Use Of Organoids For Drug Discovery
By Deborah Borfitz
October 14, 2021 | A San Francisco-based startup is using robotic automation to “industrialize” the organoid and scale up use of the next-generation human in vitro models in drug discovery experiments for brain diseases. The plan is to use these intact systems, inclusive of microglia as well as neurons, to go after genetically complicated diseases and apply machine learning techniques to find connections between the multi-phenotypic data that gets generated, according to Saul Kato, Ph.D., co-founder and CEO of Herophilus.
The company is named after the Alexandrian physician who pioneered the use of what is now called biophenotyping around 300 BC. Herophilus specializes in “deep phenotyping,” what Kato defines as “the multi-modal description of disease expression in a rich biological model” based on the aggregation of data from various phenotyping modalities—gene expression, protein composition, cell morphology, and neural activity—which collectively and interdependently make up the complicated biological picture of a disease.
Microglia, the primary immune cells of the central nervous system, are a newer addition to the organoids being grown by Herophilus, launched in 2017, says Kato. This is a direct reflection of emerging evidence that inflammatory responses in the brain are heavily involved in the early steps of neurodegenerative diseases, notably Alzheimer’s.
Machine learning experts were on board from day one because the company was co-founded in part by “hardcore Silicon Valley types,” Kato says. But software is only part of the Herophilus triad, which is also equal parts biology and automation engineering.
In addition to its own internal drug programs—notably for Rett syndrome, a rare neurological disorder—Herophilus works with pharmaceutical companies developing treatments for neurodegenerative diseases. It now has several partnerships that are specific to Alzheimer’s disease, says Kato.
Organoids offer a higher level of “biological realism” than can be produced in vitro relative to the traditional method of growing cell cultures on a flat, hard surface, he says. They are produced by culturing stem cells in more naturalistic ways so they aggregate and self-assemble into tissue.
Around 2013, scientists noticed that three-dimensional (3D) tissue cultures could recapitulate far more healthy brain activity and brain-tissue-like functionality than was possible with disassociated 2D cultures. “Neurons would survive, and they would form some connections with other neurons, but never resemble anything close to brain tissue, so this was a major step forward,” says Kato.
Kato was building the foundation for Herophilus when he joined the neuroscience faculty at the University of California, San Francisco (UCSF) in 2016 to start the Foundations of Cognition Laboratory. This unrelated basic science work at UCSF involves studying the brain of a worm (C. elegans) to understand how the organism produces cognitive function.
The Herophilus Discovery Engine has three pillars that have been branded Orchard (biobanking component), OrCA (organoid culture and assay), and Orchestra (biodata platform). Step one is collecting blood from patients and “reprogramming” somatic cells into pluripotent stem cells for growing the organoids, he explains. This is done in a standardized way with multiple quality checks because “the ultimate behavior of the culture is highly dependent on how the reprogramming is done.”
At present, the biobank primarily contains samples specific to Rett syndrome, 22.q11.2 deletion syndrome (a severe form of psychosis), and idiopathic schizophrenia. The first two are genetically determined neurodevelopmental diseases, good starting points for the company four years ago, says Kato.
With growing appreciation that neurodegenerative diseases can be studied with organoids that recapitulate the brain-immunology axis, Orchard has enlarged its focus to biosample collection from Alzheimer’s patients. The biobank already houses samples expressing presenilin 1 and presenilin 2 mutations most commonly responsible for familial Alzheimer's disease, he says.
While organoids are typically cultured using spinning bioreactors, Herophilus grows the 3D tissue cultures in their own individualized environment under “careful watch,” continues Kato. “We are big fans of nondestructive methods… to generate time series [data]” throughout the life cycle of the organoids. The process involves liquid handling robots that are integrated with advanced microscopes.
Discovery efforts get coordinated and analyzed on the Orchestra platform, which lives primarily in the cloud, he says. This includes software that controls the robots and tracks data generated by experiments (still mostly done by humans) across many different cultures, donor types, and diseases.
One of the principal outputs of data aggregation gets served up in a dashboard to in-house biologists (the company counts 13 Ph.D.'s on staff) who explore the data to generate “disease-related intuitions and insights” about the information, says Kato. Another data stream goes to a deep learning classifier that distinguishes the characteristics of cultures grown from a disease versus a healthy control.
Herophilus has tasked itself with increasing production of “boutique approaches” to engineering organoids so more experiments can be done—for example, to try drug compounds on cultures derived from patients having the same or different genetic background of the same disease. “We spend a lot of time optimizing and integrating aspects of different academic protocols,” says Kato, noting that the current culturing protocol integrates four different organoid growth practices.
The goal here is largely “reliability and reproducibility” of these models of human biology, Kato says. “One cool thing about organoids is they will resemble one brain area or another depending on how they are cultured,” so they can be customized to the brain region (e.g., dorsal forebrain) of scientific interest.
Inclusion of microglia in the models allows for scientific inquiry into the neural-immune interactions in neurodegenerative diseases. Traditional organoid cultures exclude this component because neurons and microglia are derived from different germ layers (interacting cells in an embryo); respectively, the ectoderm and mesoderm, he explains.
Deep phenotyping seeks to extract as much relevant biological information as possible out of the 3D cultures. This includes gene expression from RNA-seq (RNA profiling based on next-generation sequencing), tissue morphology using a microscope, and data about proteins and metabolites, as well as the electrical activity patterns of neurons as they fire.
It would be impossible to make sense out of all that data without the aid of machine learning, says Kato, referencing the well-known “curse of dimensionality” that data scientists experience when including so many predictor variables in a model. Machine learning models reveal and quantify deep phenotypes and how different data readouts relate to one another.
These deep phenotypes can aid in the drug development process in a multitude of ways, Kato says, including to inform translational research, companion diagnostic development, patient stratification, and clinical trial design. “The drug industry wants to tell a convenient story of systematic… steady progress, a march forward, and, in reality, it is a far more iterative process involving a lot of trial and error.” Every cycle of preclinical learning requires a return to a biologically realistic model.
Although Herophilus has not yet been engaged in clinical trials, the potential utility of its approach in this realm is for patient selection to maximize the likelihood of a good outcome, he says. “We can make a connection between the genotypes of potential patients and microscopic phenotypes that would be expressed in a biological model like ours.”
Study sponsors could sub-select patients based on the phenotypic cluster into which they fall. In Rett syndrome, Herophilus has 42 patient samples and 18 of those have common truncating mutations, Kato cites as an example.