Experimental Research Models

In this newsletter, I explore how three important biomedical research platforms compare: cell models, organoid models, and mouse models. Each model has unique advantages for studying human disease, drug development, and biological mechanisms. I also analyze their structural complexity, relevance to humans, cost, ethical considerations, and scope of application to help researchers choose the right model for their needs. Understanding these differences is critical to advancing translational research and personalized medicine. #CellModels #OrganoidModels #MouseModels #BiomedicalResearch #DrugDevelopment #PrecisionMedicine #LifeSciences #3DCellCulture #TranslationalResearch #ResearchTools #CSTEAMBiotech

I’m excited to share a new paper from our AI researchers at Recursion and Valence Labs, published today in Nature Biotechnology. At its core, this work is about a fundamental challenge in drug discovery: understanding, early and with confidence, how cells will respond to a perturbation — and ultimately, to therapeutic intervention. Today, we largely rely on the wet lab to answer that question. We run experiments, generate data, refine hypotheses, and repeat. This approach works, but it is slow and resource-intensive, especially when exploring a space with an almost infinite number of possible perturbations and combinations. At Recursion, we are working toward a more predictive, in silico approach – and building a deeper, more systematic understanding of how transcriptomics can serve as a bridge between perturbational biology and human disease. We’re modelling biological responses computationally, generating stronger hypotheses, and using the lab to validate and refine — rather than explore from scratch. In this paper, we introduce TxPert: a model designed to predict how a cell’s transcriptomic state, or which genes are turned on or off, changes in response to genetic perturbations. Importantly, TxPert can generalize beyond the data it has seen. It can predict responses to perturbations that were not directly observed during training including unseen single-gene perturbations, novel combinations of perturbations, and known perturbations in new cell types. In some settings, its performance begins to approach the reproducibility of experimental measurements. What makes this possible is not just the model itself, but the combination of: ✅ large-scale perturbation datasets ✅ structured biological knowledge, including multiple knowledge graphs ✅ and the ability to generate predictions that can be directly tested and validated in the lab and fed back to continue improving the model That connection matters. It’s how we move from interesting models to results we can trust. Over time, this is how we enable more computationally driven hypothesis generation, faster iteration, and ultimately better outcomes for patients. This is our vision for the Virtual Cell. TxPert is an important step in that direction and we’ve progressed considerably since this work was first submitted. Stay tuned — much more to come! Huge congratulations to the team behind this work — Frederik Wenkel, Wilson Tu , Cassandra Masschelein, Hamed Shirzad, Liam Hodgson, Ihab Bendidi, Cian Eastwood, Shawn Whitfield, Craig T. Russell, Yassir El Mesbahi, Marta Fay, Berton Earnshaw, Emmanuel Noutahi, PhD, and Alisandra Denton. Read the paper in Nature Portfolio here: https://lnkd.in/eNtv6UB3 #VirtualCell #AI #DrugDiscovery #Biology #NatureBiotechnology

Very excited to share a new paper that has been a long time in the making. This has been a fun collaboration with my co-authors Ruoxuan Xiong (Emory) and Alex Chin (my co-worker at Lyft and now Motif Analytics). Randomized experiments are the gold standard for measuring causal effects, but in marketplaces we are often testing policies that have many plausible spillovers that make it difficult to learn what we need by assigning treatment across users. Instead we randomize over time. This type of experiment seems simple to design, you are implementing a square wave (a type of oscillator) that determines what policy you are running based on time. When I was at Lyft, we had some heuristics for choosing switchback parameters but we rarely had bandwidth to understand their impact. It turns out to be a rich design space, and by choosing how and when you switch policies, you control the bias and variance of the estimates from your experiment. Intuitively, faster switching yields lower variance by increasing your sample size but increases bias because effects tend to persist over time (carryover effects). Your measurements from each time period are also correlated and have heteroskedastic errors due to seasonality (marketplaces tend to have strong daily and weekly cycles). Our approach is effectively a model-based design process where we use historical data to estimate the inputs to the experimental design process. The data allow us to make informed decisions about switching behavior that will yield the lowest error in our estimates. Carryover effects are the hardest quantity to estimate from historical data because on any individual test they are quite noisy, so pooling is necessary to gain some additional precision. We analyze a corpus of hundreds of switchback tests from Lyft's marketplace, and cluster them into an interpretable distribution over impulse responses. A broader point of this research is that all experimental designs lean on prior knowledge to improve the chances of a successful experiment -- even choosing a sample size for desired power in a standard A/B test. In switchback tests, there is an important bias-variance tradeoff we must manage. Without some means to estimate the covariance of errors and the likely size and shape of carryover effects, it is difficult to design an experiment that is likely to be successful.

The Missing Option in the RCT Debate: Government Experiments Several influential debates have emerged about the role of randomized controlled trials (RCTs) in public policy. Abhijit Banerjee and coauthors (2017) defend the “proof of concept” model: start with tightly controlled RCTs—often implemented with NGOs—learn what works, iterate, and gradually move toward government adoption. Lant Pritchett and Justin Sandefur (2015) highlight a related concern. They show the trade-off between internal and external validity: a rigorous RCT conducted elsewhere may provide less useful guidance than an observational study conducted in the policymaker’s own context. Angus Deaton and Nancy Cartwright (2018) go further, warning against exaggerating the virtues of RCTs. In some cases, theory, prior evidence, or other empirical methods may be more informative than a new randomized trial. All three perspectives raise important points. But they largely treat one constraint as given: the difficulty of running randomized evaluations inside government. This leads to a recurring framing of the choice set: • RCTs implemented by NGOs or researchers outside government • Observational or quasi-experimental studies using government data But there is a third option—what John List calls the “C-option”: considering scalability from the start in the research design. In policy contexts, a necessary condition would be to run pilot RCTs directly in government context. In my experience, the feasibility of RCTs inside government is largely endogenous. It depends on institutional design, incentives, institutional culture and collaboration structures. Initiatives such as the Experimental Policy Initiative in Chile’s Budget Office (https://lnkd.in/g6guyvAU) and MEF Lab (https://lnkd.in/gkd_Uuhe) in Peru show that when governments build institutions that enable experimentation and collaboration between internal and external researchers, rigorous program evaluation becomes much more feasible. Interestingly, there are now more than 700 behavioral insights units around the world running thousands of RCTs on nudges and information campaigns. Yet rigorous experimental evaluation of large government programs remains rare. If governments can randomize letters and reminders, why not programs? Building the right structures, RCTs of real policies can become far more common—and the trade-off between rigor and relevance becomes much less severe than the current debate assumes. Scott Cunningham Rachel Glennerster John Floretta Iqbal Dhaliwal Benjamin Krause, Ph.D. Alvaro M. Carreño Mohit Karnani (I am currently working on a paper on this topic with colleagues and will share it here when it is ready.)

Sharing today the latest from our lab, just published in 𝙉𝙖𝙩𝙪𝙧𝙚 𝘽𝙞𝙤𝙢𝙚𝙙𝙞𝙘𝙖𝙡 𝙀𝙣𝙜𝙞𝙣𝙚𝙚𝙧𝙞𝙣𝙜. As stem cell–based neural models gain traction for disease modeling and drug testing, one of the major bottlenecks has been scaling up production. In work led by Yuki Miura and Genta Narazaki, we present a simple and cost-effective way to prevent neural organoid fusion that allows scalable generation of cortical organoids without compromising quality. In brief, this is done by simply adding the cheap food additive xantham gum! This enabled a single experimenter to screen all FDA-approved drugs for neuropsychiatric disorders across 2,400 organoids, identifying compounds that impair human cortical development. Hoping that will be one more step toward scalable human models for brain development and drug discovery. Link to the article here: https://lnkd.in/gBUqNHpb and a short video made by Yuki to show how to dissolve the xantham gum: https://lnkd.in/gX5wipSe #Organoids #StemCells #DrugScreening #Neurodevelopment #TranslationalResearch

Our new article titled “Real-time physiological environment emulation for the Istanbul heart ventricular assist device via acausal cardiovascular modeling” was just published in Artificial Organs. The cost and complexity associated with animal testing are significantly reduced by using mock circulatory loops prior. Novel mock circulatory loops allow us to test biomedical devices preclinically due to their flexibility, scalability, and cost-effectiveness. The presented work describes the development of a hardware-in-the-loop platform to emulate human physiology for the Istanbul Heart (iHeart-II) LVAD. A closed-loop system is developed whereby the effect of the LVAD on the heart and vice versa can be studied. An acausal model of the cardiovascular system is calibrated to emulate advanced-stage heart failure. A new prototype of the iHeart-II LVAD is connected between two air-actuated chambers emulating the left ventricle and aortic chambers with PID controllers tracking numerically modeled pressures from the in-silico model. A lead–lag compensator is used to maintain fluid level. Controllers are tuned using nonlinear Hammerstein-Weiner models identified using open-loop data. The iHeart-II LVAD is operated at various speeds in its operational range, and the resulting hemodynamics are visualized in real time. Hemodynamic variables, such as LVAD flow rate, aortic, left ventricular, and pulse pressure, demonstrate trends similar to clinical observations. The iHeart-II LVAD achieves hemodynamic normalization at ~3500 rpm for the emulated condition. A novel evaluation methodology is adopted to study the performance of the iHeart LVAD under advanced-stage heart failure emulation. The models and controllers used in the platform are readily replicable to facilitate VAD research, pedagogy, design, and development. I would like to thank my doctoral research assistants, Hammad Ur Rahman, Dr. Khunsha Mahmood and MS assistant Farouk Abdulhamid from Koç University Manufacturing and Automation Research Center, and our medical supervisors Prof. Süha Küçükaksu and Prof. Vedat BAKUY from Cardiovascular Surgery Department in the School of Medicine at Başkent University for their contributions to this research. We would like to thank the Scientific and Technological Research Council of Turkey (TÜBİTAK project 318S143) for funding this research. The article is available at the following link; https://lnkd.in/dH_dZQma #artificialorgan #heartpump #ventricularassistdevice #VAD #LVAD #cardiovascularmodel #biomedical #modeling #control #hemodynamics #heartfailure #IstanbulHeartVAD #iHeartVAD

Scientists studying a complex phenomenon don't start with experiments or even hypotheses. They first build a model. They use this model to run simulations and predict what they think will happen. They then run experiments to test the predictive accuracy of their model. If they get a different result than expected, i.e., the experiment invalidates their model, they update it and try again. This is the essence of the scientific method, which can readily be adapted into an equivalent entrepreneurial method: Model - Prioritize - Test 1. When faced with a new idea, we start with a business model describing how we intend to create, deliver, and capture customer value. 2. We then prioritize the riskiest assumptions in the model and make some predictions, which 3. We then attempt to validate through small and fast experiments. Like scientists, we attempt to learn why our predictions fail. Then use those insights to update our model and try again. Model-Prioritize-Test is how you navigate uncertainty in the new world. The Model-Prioritize-Test flywheel powers #ContinuousInnovation.

New paper! Deep-learning-assisted simulation of a cortical circuit: integrating anatomy, physiology and function https://lnkd.in/gDTSUfT9 It comes with a new major model release from Allen Institute: https://lnkd.in/gGmNjv9X We integrated multiple datasets into a biorealistic model of mouse V1 with ~67k neurons (point models) and 19 cell types: • Cell type abundances • Their intrinsic ephys • EM connectomics (MICrONS and more) • Multipatch synaptic physiology • In vivo Neuropixels data Optimizing such models is HARD (used to do it manually 😱). Our simulator offers powerful training capability. Training 67k-neuron V1 model takes ~10 hours on 1 GPU. Inference simulation RUNS FASTER THAN REAL TIME. Now YOU can run such simulations - no supercomputer needed. Importantly, the model includes an approximation of the retina-thalamus-cortex visual pathway, so it can take arbitrary visual stimuli as inputs. One can study cortical processing in response to gratings, images, movies - anything you want or might be showing to real mice. BPTT training adjusts circuit parameters so that model activity matches cell-type-specific statistics measured in vivo 🐭🧠. Importantly, training loss function includes terms PRESERVING BIOLOGICAl CONSTRAINTS such as experimentally measured synaptic weight distributions. After training, the model reproduces a wide range of experimental response properties. Despite being trained only on short snippets of simple stimuli (gray screen, drifting gratings), the network matches cell-type-specific firing rates and tuning statistics observed in vivo. The trained circuit generalizes to new visual inputs. It produces realistic neural responses to previously unseen contrasts and to natural images. This shows that the trained circuit captures meaningful structure rather than overfitting the training stimuli. The N=10 trained networks reveal structured synaptic organization. In particular, excitatory connections show like-to-like wiring, consistent with connectomic observations. We observe diverse relations across cell types, including anti-like-to-like and more exotic patterns. A cool discovery is that training organizes specifically INHIBITORY neurons into functional cohorts. These small cohorts (~1% of the network) exert outsized causal control: e.g., silencing high-outgoing-weight cohort produces widespread disinhibition and selectivity loss. Lastly, and crucially: Models trained without data constraints on synaptic weight distributions reproduce activity statistics, but their SYNAPTIC ORGANIZATION CHANGES. Matching activity is not everything. Bio realism in training strongly affects circuit mechanisms! Congrats and many thanks to Shinya Ito, who spearheaded this work, and to all the amazing co-authors Darrell Haufler Javier Galván Fraile - PhD Kael Dai Joe Aman Guozhang Chen Claudio Mirasso Wolfgang Maass

Pfizer killed danuglipron in April 2025. A Nature paper just published last week explains why the field needed a different mouse. Most people assume that if a drug works in humans, it must work in mice. For the new oral GLP-1 weight-loss drugs (danuglipron, orforglipron, Eli Lilly's just-approved Foundayo), that assumption is wrong. Standard mice have a serine at position 33 of their GLP-1 receptor. Humans have a tryptophan. Small-molecule GLP-1 drugs bind the human version. Not the mouse. So for years, the preclinical pharmacology of an entire class of weight-loss medications was effectively running blind. Godschall and colleagues at UVA fixed that. CRISPR-Cas9. One amino acid swap. A humanized Glp1r mouse that responds to oral small-molecule GLP-1 agonists the way humans do (Godschall et al., 2026, Nature). A second group at Gubra and Terns published a parallel humanized line in EBioMedicine the same window (Sonne et al., 2026). What they found, once the right mouse existed, matters. A new reward circuit. Central amygdala GLP1R neurons project to the ventral tegmental area and dampen dopamine release in the nucleus accumbens. The result: small-molecule GLP-1 drugs selectively suppress palatable food intake without touching standard chow consumption. This is the mechanistic basis for the clinical signals showing GLP-1RAs reduce alcohol and cannabis use disorder incidence (Wang et al., 2024). It is also why long-term effects on motivated behavior in chronic users deserve serious attention. Three things stand out for anyone working in translational science. One. The CRISPR humanized mouse is not a substitute for animal research. It is animal research, done right. Context of use is the governing framework. The wild-type mouse was the wrong model for this molecular class. The humanized mouse is the right one. Two. Conditioned taste avoidance and open field tests could not distinguish nausea from satiety for danuglipron and orforglipron. Continuous home-cage video monitoring with pose tracking (SLEAP) and unsupervised behavioral segmentation (Keypoint-MoSeq) could. Digital phenotyping is not optional anymore. It is the signal-to-noise upgrade that turns weak readouts into discriminative ones. Three. Pfizer discontinued danuglipron after a drug-induced liver injury in a dose-optimization study. The Nature authors state plainly that preclinical profiling in the right model "might have identified" the side-effect risk earlier. Not certain. But plausible. And avoidable. The animal-research-versus-NAMs debate is not the right debate. The right debate is which model for which question, with which endpoints, structured under FAIR data, ready for regulatory scrutiny. This paper is one of the cleanest demonstrations of that argument I have read this year. #TranslationalScience #GLP1 #DigitalBiomarkers #ContextOfUse #AnimalResearch #Pharmacology #DrugDevelopment