The SOT 2025 Annual Meeting and ToxExpo took place from March 16–20, 2025, at the Orange County Convention Center in Orlando, Florida. This premier event brought together over 5,000 toxicology professionals from around the world to share the latest advancements in the field.
Victoria K. Alstat1, Nicholas S. Coungeris1, Andrew S. LaCroix1
1 AxoSim Inc., Maple Grove, MN, USA
Background and Purpose
Drug development pipelines utilize many complimentary biochemical, cell-based, in vitro and in vivo assays, each of which provides a limited amount of safety or efficacy data. Assay development and validation take valuable time and drive up preclinical development costs. Furthermore, data from disparate sources must be combined to select the optimal drug candidates to pursue, an operation frequently plagued by inconsistencies between assays (e.g. cell types, media conditions, endpoints). A single complex in vitro model (CIVM) that combines both safety and efficacy testing offers many advantages over the traditional approach of using separate models. We previously developed a human induced pluripotent stem cell (iPSC)-derived neural organoid system to model CDKL5-deficiency disorder (CDD), which provides a system by which to test for drug efficacy in reducing a neuronal hyperexcitability phenotype (Negraes et al., Mol. Psych. 2021). This model recapitulates key functional features of CDD, a rare X-linked neurodevelopmental disorder characterized by early onset seizures, developmental delay, and severe intellectual disability. While this model was used to identify promising drug candidates, the safety profile and relative safety margins of the proposed drugs were unknown. Gold-standard ATP-based viability assays provide safety information at the end of a drug treatment but are destructive in nature and sacrifice valuable biological samples for a single data point. Longitudinal assays based on membrane rupture and release of cellular components provide a non-destructive means to monitor cell health, but can become costly and time-intensive if performed over the course of weeks and still require perturbation of the experimental system (i.e. collecting media). Here, we sought to develop a non-invasive image-based viability assay that would enable monitoring of CDD patient-derived neural organoids and utilize the combined safety and efficacy data to identify potential safe and effective CDD therapeutics.
Methods
To generate cortical organoids, neural progenitor cells (NPCs) are expanded and seeded into ultra-low attachment plates, wherein they will self-organize into 3D organoids, one organoid per well. Over the course of 6-10 weeks, NPCs co-differentiate into neurons and astrocytes exhibiting spontaneous oscillatory functional activity as measured by high-throughput kinetic plate reader (FLIPR). An imaging-based model to predict Cell Titer Glo (CTG) luminescence value was developed using ImageJ to extract features (circularity, size, intensity, etc.) from brightfield images of CDD and healthy parental control (CTL) organoids treated with a broad range of neurotoxic agents in 7-point dose response for 4 days. These imaging features were then used to train a neural network in JMP to predict luminescence value. To identify disease-specific compound rescue, CDD and CTL were dosed every other day over the course of 3 weeks. Drugs that demonstrated disease-specific functional rescue (i.e. reduction in average peak frequency with higher potency in CDD organoids) without cytotoxicity were considered hits, which would then expand into further target validation studies. Specifically, we defined a metric for disease specificity (Disease Specificity Score = CTL FLIPR IC50 / CDD FLIPR IC50) to rank-order hits in terms of efficacy. Higher values indicate stronger disease specificity and higher priority drugs. We defined a safety metric comparing the potency of predicted toxicity to that of functional changes (Toxicity Score = CDD Tox IC50/CDD FLIPR IC50). Again, higher values are indicative of compounds with a stronger safety profile.
Results
Organoids generated from both healthy and CDKL5-KO lines grow to similar diameters (670 ± 56 µm and 645 ± 59µm for CDD and CTL, respectively) and show appropriate cellular composition based on immunocytochemistry staining for MAP2 (neurons) and GFAP (astrocytes). However, CDD organoids show striking differences in average peak frequency (5.85 ± 0.82 and 3.32 ± 0.57 peaks/min for CDD and CTL, respectively). We verified the reproducibility of this hyperexcitability phenotype across n = 3 independent batches of organoids, showing consistency across organoids (CV = 14.0% and 17.1% for CDD and CTL, respectively) and across plates (CV = 3.88% and 3.97% for CDD and CTL, respectively).
To eliminate the need to perform destructive ATP assays following FLIPR and enable longitudinal cytotoxicity measurements throughout the course of the dosing window, an image-based predictive CTG endpoint was developed based on morphological changes in organoids treated with cytotoxic compounds. The imaging-based cytotoxicity model was able to predict CTG value with an R2 value of 0.92. The model was then validated across n = 2 independent experiments.
Based on an initial screen of 5200 compounds (Negraes et al., Mol. Psych. 2021), 288 promising drugs were screened at 1μM in both CDD and CTL organoids to confirm the decrease in peak count from hit compounds was preferential to CDD organoids. Compounds were ranked based on Disease Specificity Score (higher values mean better disease selectivity) and the top 40 were selected for follow up dose-response screening. To ensure these hit compounds were correcting the hyperexcitable CDD phenotype due to neuromodulation and not cytotoxicity, we applied our image-based Toxicity Score to final dose response screening. Specifically, the top 40 hits were screened in 7-point half-log dose response and ordered based on both Disease Specificity Score and Toxicity Score. Based on this analysis, 22 compounds were determined to be hits. These 22 compounds encompassed 15 unique biological targets. Subsequent screening was used to validate each of these targets individually using known reference compounds. For 1 promising target, 7 out of 8 compounds showed strong disease specificity with minimal toxicity indicating strong potential for a repurposing campaign.
Conclusions
We developed an approach to identifyingdisease-modifying therapeutic candidates using a combination of humaniPSC-derived cortical organoids, high-throughput calcium imaging assay (FLIPR)and a brightfield imaging-based predicted cytotoxicity model. Thisproof-of-concept suggests the capacity for this organoid technology toaccelerate drug discovery and generate compelling preclinical human efficacyand safety data. Furthermore, it provides evidence that an imaging-basedtoxicity assessment can offer a cost-effective and accurate alternative toperforming gold standard viability assays. Using this approach, we were able toscreen over 5000 compounds in less than 1 year, identifying 22 top hits andmultiple promising targets for the treatment of CDD.
Andrew S. LaCroix1, Victoria K. Alstat1, Nicholas S. Coungeris1, David Gallegos3, J. Lowry Curley2
1AxoSim Inc., Maple Grove, MN, USA. 2AxoSim Inc., New Orleans, LA, USA. 3 Takeda Pharmaceuticals, Cambridge, MA, USA
Background and Purpose
The drug development process is fraught with failure due to either safety issues or poor efficacy. When considering safety profile, neurotoxicity is the leading cause of clinical failure (Cook et al., Nat. Rev. Drug Disc., 2014). Furthermore, 12% of drugs withdrawn between 1960-1999 were caused by neuroliabilities (Valentine & Hammond, J. Pharmacol. Toxicol. 2009). The use of complex in vitro models (CIVM) derived from human tissue has dramatically expanded in recent years, promising to provide the necessary biological complexity to improve clinical translation and scale to enable adoption early in drug development pipelines. However, few of these CIVMs have been rigorously tested in terms of their predictive abilities or stability of these predictions over time (reviewed by Kang et al., Biofabrication 2024). Towards this end, we have developed a human induced pluripotent stem cell (iPSC)-derived cortical organoid platform, amenable to high-throughput screening (HTS) on a multiparametric functional endpoint. The functional “waveform” activity in these neural organoids can be reproducibly modified with excitatory and inhibitory modulators, confirming the presence of a fully functional integrated glutamatergic / GABAergic circuitry, ultimately providing a unique high-throughput and translatable drug screening alternative to the traditional pre-clinical screening model. In 2022, this organoid platform was used to develop a predictive clinical neurotoxicity model that showed remarkable specificity (>90%) and good sensitivity (>50%), making it an ideal pre-screening method prior to standard 2-species animal testing (Wang et al., ALTEX 2022). Here, we tested the stability and reproducibility of these predictions over time and used these replicate experiments to refine and automate neurotoxicity score predictions.
Methods
3D cortical organoids were derived from healthy donor iPSCs that underwent a neural progenitor cell (NPC) derivation process involving embryoid body formation, neural rosette plating, and physical isolation of NPCs (Marchetto et al., Cell 2010). NPCs were expanded and seeded into ultra-low attachment 384-well plates, wherein they self-organized into 3D spheroids, one spheroid per well. Culture maintenance medium (BrainPhys supplemented with BDNF and GDNF) was exchanged three times per week, supporting co-differentiation of neurons and astrocytes. After 10 weeks of differentiation, once cultures exhibited strong coordinated network activity, acute (0 - 4 hours) neuromodulation screening of 84 known neurotoxic and safe compounds was performed using a calcium flux assay and high-throughput kinetic plate reader (FLIPR). Compound selection and dosing has been described previously (Wang et al., ALTEX 2022). Briefly, each compound was classified into 1 of 4 categories based on its clinical adverse event rate: 1: negative (<0.01%), 2: rare (0.01-0.1%), 3: infrequent (0.1%-1%), 4: frequent (>1%). Dosing concentrations were selected to span 0.1x – 100x the in vivo Cmax in 7-point dose response. Changes in the number, size, shape, and variability of the spontaneous activity waveforms were quantified using custom written code in Python. A margin of exposure (MOE) value was calculated for each waveform feature as the ratio of total plasma Cmax (tpCmax) to the EC/IC50. MOE values were used to train and test a logistic regression model in R to predict safe (category 1) or neurotoxic (category 2, 3, 4) compounds. Model performance (ROC, sensitivity, specificity) were calculated for each dataset individually as well as the combined datasets.
Results
Cortical organoids exhibited consistent spontaneous functional activity at 10 weeks of differentiation (CVPeak_Frequency = 10.2 ± 2.7%, CVPeak_Height = 16.7 ± 1.7%, across seven 384-well plates). Functional controls 4-AP (30µM) and muscimol (10µM) confirmed the ability to elicit excitatory (>50% increase in peak frequency) and inhibitory (total loss of spontaneous bursting) responses in the organoids, respectively. Consistent with prior screening (Wang et al., ALTEX 2022), diverse compound responses (excitatory, inhibitory, biphasic) were observed for both safe and neurotoxic compound sets. Among the waveform features, peak rise time and decay time were the least stable and showed biphasic responses that required expert inspection to calculate an accurate EC/IC50 value. Adjustments in the waveform analysis pipeline to calculate a rise and decay slope (numerically inverse to time) yielded pharmacologically monotonic responses enabling more reliable potency calculation. Although this adjustment simplified curve fitting for rise and decay features, other waveform features still exhibited complex responses. To further stabilize and remove user-bias from waveform analysis, quantitative criteria in terms of log2 fold-change (log2FC) were selected to automate excitatory response (log2FC ≥ 0.5), inhibitory response (log2FC ≤ -0.5) or no response (│log2FC│ < 0.5) assignments to each waveform feature. These combined improvements, when applied to the original dataset, maintained the performance of the original published work (AUC = 0.729, sensitivity = 53.5%, specificity = 93.3%). Using this stabilized and automated analysis pipeline, we evaluated the updated model performance trained on functional screening data collected in 2024. Remarkably, the model performance was nearly identical to the originally published work (AUC = 0.781, sensitivity = 53.5%, specificity = 90.0%).
Conclusions
This study utilized a human iPSC-derived cortical organoid platform with a functional HTS-compatible endpoint to build and confirm the performance of a clinical neuroliability prediction model. The stability of model predictions was improved through advancements in custom peak detection, waveform feature extraction, and automated potency calculations. Consistent predictive performance was observed between two independent experiments, performed at different sites, separated by a span of years, a strong indicator of a robust model. The key performance metric – high specificity (>90%) was maintained, indicating such predictive neurotox screening would not filter out potentially valuable drugs prematurely. Utilization of this type of CIVM in preclinical settings has the potential to improve drug candidate quality and reducing costly clinical failures due to neuroliability.
Corey Rountree1, Eva Schmidt1, Tyler Rodriguez1, Kevin Simpson1, Michael J Moore1,2,3 and Lowry Curley1
1 AxoSim Inc., New Orleans, LA, USA
2 Department of Biomedical Engineering, Tulane University, New Orleans, LA, USA
3 Brain Institute, Tulane University, New Orleans, LA, USA’
Background and Purpose
Chemotherapy-induced peripheral neuropathy (CIPN) is a debilitating side effect of anticancer therapies, often resulting in chronic pain and loss of sensation as peripheral sensory and motor neurons degenerate from chemotherapy exposure. CIPN ranges in severity from mild annoyance through life-threatening and is a leading cause for patients to switch or discontinue treatments. Depending on the compound, CIPN can be detected with preclinical testing using either 2D cell culture or animal models early in the drug development pipeline before human clinical testing. However, many compounds have passed preclinical testing demonstrating minimal toxicity but were found to cause peripheral neuropathy and other side effects in clinical testing, leading to low clinical trial success rates and limited use of approved compounds. Antibody-drug conjugates (ADCs), highly toxic payloads linked to tumor-specific monoclonal antibodies, are a recent example of compounds designed to have excellent specificity through preclinical testing but unfortunately have shown unexpected peripheral neuropathy in patients. A more predictive human pre-clinical model therefore would significantly reduce the cost and risk of preclinical testing and thereby accelerate the drug discovery process. Microphysiological systems (MPS) offer a promising alternative, bridging the gap between 2D cell cultures and human clinical data by providing an in vitro model with in vivo-like characteristics. We have developed an MPS incorporating human iPSC-derived sensory neurons and human primary Schwann cells to assess toxicity in 2D and 3D formats. The 2D platform uses imaging-based readouts to detect toxicity independently for sensory neurons and Schwann cells. The 3D platform creates a 3D Nerve-on-a-Chip, or NerveSim®, by guiding axonal growth down a channel over an embedded electrode array (EEA) in custom-designed 24-well tissue culture plates. The 3D cultures can deliver functional electrophysiology using the EEA to stimulate and record from 10 locations along the nerve, combined with non-invasive brightfield imaging to track morphology over time. Here, we used both platforms to test and compare the in vitro toxicity of 4 ADCs as well as the common ADC payload monomethyl auristatin E (MMAE).
Methods
Both platforms used human iPSC-derived sensory neurons from Anatomic, Inc cocultured with human primary Schwann cells. The 4 ADCs and 1 payload were commercially sourced from Medchem Express and SelleckChem and consisted of distimab vedotin (HER2-MMAE DAR4), trastuzumab MMAE (HER2 MMAE DAR4), trastuzunab emtansine (Kadcyla, HER2-DM1), trastuzumab deruxtecan (HER2-TOP1i), and MMAE.
The 2D platform plated these cells in 96-well plates and grew them for 4 days before introducing compounds at 4 days-in-vitro (DIV) for 2 days of dosing. After dosing, all wells were fixed and stained for immunocytochemistry focused on cell nuclei (DAPI), neurons (BIII tubulin), and Schwann cells (S100). Plates were imaged with an inverted confocal high content imaging system and the resulting data were analyzed to deliver cell counts, neuron fiber length, Schwann cell circularity, and other metrics.
The 3D platform formed coculture spheroids of the 2 cell types and placed these spheroids into cell culture channels of the EEA plates. Samples were grown for 42 days to create robust nerves in each well atop the EEAs before completing 1 week of compound dosing. Before and after dosing, samples were imaged using brightfield microscopy and evoked electrophysiology was measured. Evoked electrophysiology consisted of eliciting compound action potentials from each well by stimulating with an electrical current of 48 µA at 6 different electrode locations while recording from the 10 electrodes in each well simultaneously. Twelve trials of stimulation were averaged for each unique combination of stimulation and recording electrode, delivering 45 unique recordings. These 45 recordings were converted from the time domain to the velocity domain and the envelope was calculated to produce a purely positive signal. The velocity envelopes were then aligned in velocity space and the maximal velocity projection (MVP) was computed to collapse the recordings into a single measurement. The area-under-the-curve (AUC) of the MVP signal was calculated to produce the velocity density index (VDI) metric that is representative of the population level nerve conduction velocity, amplitude, and number of peaks for each well. The brightfield images were quantified using a custom image filter and then automated segmentation to quantify fiber length, fiber number, and the degeneration index or cell debris caused by degeneration. Both electrophysiology and imaging metrics were normalized using the baseline metrics before dosing on a sample-by-sample basis using IC50 curve fitting and non-parametric statistical tests to measure toxicity.
Results
Disitamab vedotin and trastuzumab MMAE, both containing the MMAE payload, exhibited different results comparing the 2D and 3D platforms. The 2D platform measured significantly reduced metrics for disitamab vedotin but no values were below 50% inhibition for either metric, suggesting low toxicity. Trastuzumab MMAE however showed significant reductions in both neurons and Schwann cells with IC50 values of 60.4 and 145 ug/mL, respectively. In contrast, the 3D assay detected both compounds as toxic through functional electrophysiology with IC50 values of 3.6 ug/mL (disitamab vedotin) and 4.6 ug/mL. The 3D neuron fiber length metric detected significant reduction at the higher concentrations but no IC50 curves as all values were above 50%. Trastuzumab emtansine (Kadcyla), with a DM1 payload, showed moderate toxicity in 2D with neurons (10.2 µg/mL) and Schwann cells (15.4 µg/mL), while the 3D platform exhibited electrophysiological effects at 3.9 µg/mL with significant effects to neuron fiber length but no IC50. Trastuzumab deruxtecan, containing the deruxtecan topoisomerase inhibitor payload, demonstrated the highest tolerability among the ADCs tested, with 2D showing neuron and Schwann cell toxicity measurable at 189 µg/mL and 185 µg/mL, respectively, while the 3D electrophysiology had an IC50 of 83 µg/mL and no IC50 for the 3D neuron fiber length.
The isolated MMAE payload was highly toxic, with 2D toxicity measured for neurons and Schwann cells at 0.9 nM and 2.4 nM, respectively. Similarly, the 3D platform confirmed the neurotoxic impact measuring toxicity in both functional electrophysiology and the 3D neuron fiber length at 0.46 nM and 3.9 nM. These results underscore the heightened sensitivity of our 3D MPS in detecting neurotoxicity and support its utility as a predictive pre-clinical model for peripheral neuropathy risk in ADCs and other chemotherapeutics.
Conclusions
This study highlights that 2D and 3D peripheral nerve assays offer different advantages for pre-clinical compound screening. 2D assays can achieve high throughput screening, capable of testing hundreds to thousands of compounds but are limited to morphological metrics that are less sensitive at detecting some forms of peripheral neuropathy and toxicity. The 3D NerveSim® platform is medium throughput (tens to hundreds of compounds) and allows functional characterization of compound effects that are more sensitive compared to 2D or 3D morphological metrics, as shown by the ADC testing in this study. Together, these data showcase the sensitivity and added value of functional electrophysiological testing for pre-clinical toxicity screening for peripheral neuropathy.
