Herophilus Discovery Engine in action: Targeting synaptogenesis

Pavan Ramkumar
11 min readMar 1, 2023

At Herophilus, we have industrialized the brain organoid — a revolutionary in vitro model of human biology — to discover new therapies for neurological disorders. As we have written about before, brain organoids provide a massive level up in modeling biological complexity and enable us to study complex brain disease in a human-biologically realistic but controlled environment.

Our mission at Herophilus is to take this revolutionary innovation that is sweeping the academic research world and industrialize it for full-strength drug discovery. To do so, we built a scaled technology platform for organoid experimentation, called the Herophilus Discovery Engine, integrating the recent and ever-evolving advances in complex human patient-derived cell reprogramming and culture methods (Orchard), high throughput robotically automated biological assays (OrCA), and state-of-the-art software/machine learning for effective insight generation from complex datasets (Orchestra). This platform enables us to tackle some of the most tantalizing but difficult challenges of finding cures for neurological disease.

This is a case study on the use of our platform to go after a highly compelling but difficult target of neurotherapeutic intervention: the synapse.

Characterizing synapses in human disease models: crucial but challenging

Synapses — junctions between neurons that enable information transmission in brain circuits — are critical windows into brain development and disease. During development, the brain undergoes an explosion of synaptogenesis, followed by a period of competitive synaptic pruning. Unlike neurons that stop dividing after birth, synapses are dynamic: they continue to be formed as new neuron-to-neuron connections are made, removed through a selective process of pruning, and strengthened or weakened throughout life via mechanisms of synaptic plasticity. As synapses change, they change how neurons communicate and how neuronal circuits process and retain information. Synaptic plasticity is thus intimately linked to fundamental cognitive functions including learning and memory.

When normal synaptogenesis, synaptic pruning, or synaptic plasticity becomes dysregulated, the result can be neurodevelopmental (like autism), neurodegenerative, (like Alzheimer’s¹ ²) and neuropsychiatric (like schizophrenia³ ⁴) disease. There are many candidate chemical mediators of synaptic plasticity (see for instance, the emerging interest in psychoplastogens) — that can functionally alter synaptic strength as well as structurally alter dendritic spines. Modulating synaptic maintenance and function can result in improved therapies for mood disorders, PTSD, and depression. Therefore, targeting synaptic dysfunction in neural circuits is a compelling therapeutic strategy in neurology rapidly growing in interest. However, it has been difficult to screen for synaptic drugs in both living animals and traditional human cell culture, for different reasons. On the one hand, animal models make it challenging to study molecular phenomena associated with disease progression and lack translatability to humans. On the other hand, 2D monocultures lack biological realism. Organoids are the goldilocks solution to this problem- controlled and scalable enough to do high-throughput science, but rich and complex enough to capture the biological processes underlying neurological disease.

Why is it hard to study synapses in neuronal cultures?

The development and modulation of synapses throughout their lifetime is critically dependent on the interaction between multiple neuronal cell types: astrocytes, neurons and microglia. Indeed, a recent study that screened small molecules for synapse density changes showed that hits identified using a neuron-astrocyte co-culture were missed in a culture of primary neurons alone. Thus, traditional 2D cultures can lead to false negatives in drug screens and missed opportunities for patients. Modeling physiologically representative synaptogenesis in a dish requires co-culture systems and complex in-vitro models. Enter brain organoids.

At Herophilus, we have developed 3-dimensional cortical brain organoids that develop physiologically realistic astrocytes, neurons, and synapses — a step-function change in in vitro disease modeling technology relative to “classical” 2-dimensional primary neuronal cultures (Fig. 1).

Figure 1: Complex in vitro models for synaptogenesis. Image of a cleared 3D organoid demonstrating robust co-cultures of astrocytes and neurons. Astrocytes are labeled with Glial Fibrillary Acidic Protein (GFAP; green), neurites labeled with Microtubule associated protein 2 (MAP2; white), and presynaptic puncta labeled with Synapsin I (SYN1; red). Image credit: Alejandro López-Tobón

A clear view into synapses: complexity meets scale

Brain organoids develop physiologically realistic cytoarchitecture: i.e. distinct spatial distributions and packing densities for cell types in the brain. Being able to interrogate the biology of interacting cells in this realistic tissue context is one of the key value propositions of 3D culture. At Herophilus, we have developed highly sensitive assays and companion analysis pipelines to interrogate this complex biology. In fact, we have successfully leveraged this unique aspect of organoids to detect phenotypes of cell packing and neurite density from ventricular zones in neurodevelopmental diseases like Rett syndrome.

Automated cleared assay with massive image throughput

To keep the tissue cytoarchitecture intact in imaging-based assays, there are two options on the table: (1) carefully slice organoids, mount the slices onto slides and analyze immunohistochemistry images of physical sections (IHC), or (2) image cleared organoids in 3D using volumetric microscopy, and analyze the 3D reconstruction of the tissue (which we call 3CO). At Herophilus, we have done a significant amount of organoid slicing over the years, but we realized that manual sectioning presents a major human-labor bottleneck for the scale we required. Cryosectioning is an artisanal skill, notoriously resistant to automation, and destroys the 3-dimensional structure of the anatomy, losing a potentially important phenotypic readout. An analysis of the manual aspects of IHC vs 3CO revealed a 96x improvement in labor efficiency and 3–5x higher image throughput (Fig. 2).

Incorporating clearing and imaging knowhow from the field into our currently deployed automated system was no easy feat, and we went through generations of automation development. Beyond the streamlining the clearing protocol itself, we automated the staining process, devised ways to automate the mechanical transfer of organoid between labware, developed and produced mechanical structures to hold organoids in place during imaging, and implemented an automated object detection protocol to precisely locate the organoid before high-resolution imaging.

The end result: massively parallel organoid clearing and immuno-staining protocols and in-situ images of neurites and synapses, acquired without disrupting the three-dimensional morphology of our organoids.

Figure 2: Sectioning vs Clearing. a) Whole-slide images of 20 individual sections from a single organoid, stained with DAPI (blue), MAP2 (white), Synapsin I (red). b) Projection image of 3D cleared organoids, with 72 virtual optical sections, stained with the same markers as in a). The same unit of manual labor resulted in 20 physical sections from one organoid vs. 60–100 optical sections from 96 organoids. Image credit: Brenda Dang, Brian Rash, Alejandro López-Tobón

Computer vision FTW!

The scale unlocked by the cleared assay presents particular challenges for data engineering and analysis. With 5 immunofluorescence channels and 100 optical sections, a single 96-well plate yields 48,000 images. Thus, a small-scale dose-response study can easily result in hundreds of thousands of images. For each of these images we need to extract many measurements of micro-anatomical features, traditionally an extremely laborious manual process susceptible to human bias. How do we turn images at this scale into insights that drive drug discovery decisions? Enter: high-throughput parallel processing computer vision.

The algorithms that power our computer vision methods leverage classical label-free image processing techniques like adaptive thresholding as well as modern label-based convolutional neural network based segmentation techniques. Regardless of algorithm class, our approach is based on three key force multipliers.

The first force multiplier is a comprehensive suite of web-based and cloud-friendly data visualization, annotation, and collaboration tools in our software platform Orchestra. These tools dramatically accelerate data access, data cleanliness, data visualization, data labeling, rapid iteration of algorithms, human-in-the-loop quality control, and pervasive, ubiquitous validation metrics. Without these processes in place, any analysis pipeline would take longer to develop, and is more likely to fall prey to the timeless principle of “garbage in, garbage out”.

The second force multiplier is a streamlined path from development to deployment, liberating our data scientists to devote their attention to mining data for insight. Our ML ops platform manifests in the form of (1) easy access to metadata and images from our centralized cloud metadata and object storage into Jupyter notebooks, (2) easy scalability of prototyped algorithms, via streamlined cloud deployment, orchestration and monitoring of configurable, containerized, autoscaling computational pipelines, and (3) easy sharing of analysis results via browser-based data dashboards.

The third force multiplier is a suite of assisted labeling systems based on corrective editing of the predictions of label-free methods rather than requiring de novo human annotation (i.e. drawing contours around hundreds or thousands of objects from scratch), that drastically reduce the burden of laborious training data labeling.

We built a collection of data-modality-specific segmentation methods and pipelines leveraging these force multipliers — ranging from feature detection of in-situ brightfield and immunohistochemistry sections, to morphology extraction from cleared 3D images and time series extraction from calcium imaging videos — validated by humans and scaled by machines. Our segmentation methods solve tasks across many spatial scales, ranging from finding organoids in wells, through detecting single nuclei, cell bodies, neurites, synapses and sub-cellular organelles (Fig. 3).

We built a collection of data-modality-specific segmentation methods and pipelines leveraging these force multipliers — ranging from feature detection of in-situ brightfield and immunohistochemistry sections, to morphology extraction from cleared 3D images and time series extraction from calcium imaging videos — validated by humans and scaled by machines. Our segmentation methods solve tasks across many spatial scales, ranging from finding organoids in wells, through detecting single nuclei, cell bodies, neurites, synapses and sub-cellular organelles (Fig. 3).

Figure 3. Segmentation pipelines across modalities and objects. From top to bottom, left to right: a) Tracking the growth of organoids requires accurately segmenting them from daily brightfield images. Segmentation masks overlaid on organoid brightfield images through time. b) Animated z-stack fly through of a cleared organoid stained for IBA1 and corresponding microglia masks from a microglia segmentation pipeline. c) Dendritic spines detected on a section of a 3D neurite stack, overlaid with a skeletonized neurite, and segmentation masks for neurites and spines. d) Immunohistochemistry sections cropped around neuronal zones, stained with nuclear markers (DAPI, Sox2, MeCP2), a neuronal marker (MAP2) and presynaptic marker (Synapsin I). Masks from different segmentation methods are overlaid on raw images, and toggled on and off for illustration. e) ROIs of putative single neurons, segmented from a 10-Hz recording of an organoid labeled with a calcium dye. f) Quantifying neuronal health and neurogenesis requires segmentation of neurites. Animated z-stack fly through of a cleared organoid stained for MAP2 and corresponding neurite masks from a neurite segmentation pipeline. Image credit: Dexter Antonio, Nicolas Casabianca, Lucas Cavalcante

Putting it together: HDE powers a small molecule screen for synaptogenesis

To put the core HDE platform into action, we built software interface components to drive screening experiments: management of our continuously growing, richly annotated target-biased compound library (TBL), web based planning tools to configure a screening experiment, and device control of our automated compound-screening work cell (ScreenCell) for high throughput phenotypic screening. Let’s talk about each of these components.

Physically, the target-biased compound library (TBL) is a collection of several tens of thousands of purpose-curated small molecules in well plates for drug screening applications. Orchestra provides access to metadata about the TBL, searchable by target class, mechanism of action, development status and beyond. Using Orchestra, our scientists can query the TBL and design a screening experiment at the level of abstraction needed for experimental design: they can spend most of their effort thinking about cell cultures, number of replicates, number of positive and negative controls, and dosing regime, leaving plate layouts and liquid transfers to what software does best (Fig. 4).

Figure 4. Planning a screening experiment with Orchestra. A user can select a target drug plate, and paint over each well the desired culture entities (such as cell line clones), the protocol that must be used to culture them (Culture Protocol) and the compound treatment regime. Pre-randomized plate layouts can be used from an ever-growing template library. Image credit: Jordan D’Addeo, Andy Lash

If you walk into any automated lab at a platform biotech company today, you will see an array of instruments: liquid handlers that aspirate and dispense liquids, incubators that store plates in controlled conditions, hotels that stack plates, microscopes that image them, and robot arms that move them from one station to another. Each piece of automation equipment is a modern engineering marvel, but making them work together comes with the long-tail challenge of having their machine-interfaces (sometimes analog serial, sometimes TCP/IP; sometimes documented, sometimes not) be mutually interoperable. ScreenCell is the Herophilus solution to this challenge, where these machines are deeply integrated with the Orchestra Platform. Instructions are determined by Orchestra and sent to the instruments, instrument log files proving actions are in turn consumed by Orchestra to document what exactly happened. This deep integration makes it possible for our discovery teams to start from a disease hypothesis, define a scaled assay, execute a screen, analyze images, and call hits all in a matter of weeks.

Having all the lego blocks ready to go, we next set out to validate the biology: i.e. can we first, physiologically model and second, computationally detect an expected change in synapse density in our organoids? By day 40 in vitro, we have observed that organoids develop mature neurons with a dense network of neurites. Given what we know about mature neurons, we hypothesized that as organoids age beyond this timepoint, mature neurons will maintain their morphology, but synapses will continue to develop, and this hypothesis would predict a change in synapse density between younger and older organoids. To test this hypothesis, we cleared, stained, and imaged organoids at two different age groups (young; day 40 and old; day 90). We applied our 3D neurite and synapse segmentation pipelines, and found no difference in neurite density between age groups, but a significant increase in synapse density between day 40 and day 90 (Fig. 5). This study helped us establish that we could detect a true increase in synapse density using the optimized assay parameters (for clearing, staining and imaging).

Putting these blocks together into a full-strength screen, we discovered novel small molecules that increase synapse density (Fig. 5). We are now validating these hits with secondary organoid and cell culture studies using our platform to yield novel targets and chemical starting points for our disease programs.

Figure 5. Synaptogenesis in brain organoids. a) Neurites have already begun to develop by 40 days in vitro (DIV) (young organoids) but synapses continue to develop beyond 90 DIV (old organoids). b) Zoomed-in crop of neurite masks (white) against presynaptic puncta masks (red) for a representative organoid treated with either 0.1% DMSO or a specified concentration of a putative hit from our primary screen demonstrating dose-dependent increase in synapse density: HRP-0010264. Image credit: Lucas Cavalcante

Acknowledgements

The HDE is built by bioengineers, drug discovery scientists, research associates, production technologists, software engineers, and machine learning scientists. Particular call outs to Sneha Rao, Alejandro López-Tobón, Justin Nicola, Corey Landry, Zhixiang Tong for leading the development our automated 3D cleared assay, to Jordan D’Addeo, Justin Nicola, Zhixiang Tong, and Andy Lash for making our ScreenCell a reality, to Brenda Dang and Alejandro López-Tobón for executing the synapse screening studies, to Anthony Bosshardt, Chili Johnson, and Andy Lash for building the computer vision workbench and ML ops features into Orchestra, to Dexter Antonio, Nicolas Casabianca, Lucas Cavalcante, Michael Lin, Noah Young, Jordan Sorokin, and Alex Rogozhnikov for developing our machine learning pipelines.

Footnotes

  1. The Alzheimer’s disease (AD) research community has diversified in recent years from exclusively targeting Amyloid and Tau pathology. For example, as of 2022, 4 out of 21 disease modifying therapies in Phase 3 AD trials target synaptic plasticity, the second most prominent mechanism class after the Amyloid hypothesis (6 out of 21).
  2. Targeting the neuroimmune axis — the interaction between microglia, neurons and astrocytes — has emerged as therapeutic strategy for the broad disease category of neurodegeneration and even neuropsychiatric disease, where inflammation is increasingly implicated as a cause of disease.
  3. Synapse dysregulation is important for schizophrenia because of the known role of microglia in synaptic pruning during development and disease.
  4. At Herophilus, we have recently demonstrated, in collaboration with Cerevel Therapeutics, that a complement (C4A) overexpression model of Schizophrenia in neuroimmune cortical organoids demonstrates increased synaptic pruning by microglia.

--

--

Pavan Ramkumar

Director of Machine Learning at Herophilus. Scaled Biology | Drug Discovery | Machine Learning | Neuroscience