Our scientists publish on human immunology, organoids, and the AI that learns from them. Every paper here comes with an easy-to-understand summary you can match to your background, whether you’re curious, a biologist or an AI researcher, so you can see what we found and why it matters.
Published inNature ProtocolsNatureNature Cell BiologyCancer DiscoveryNature BiotechnologyCell Reportsnpj VaccinesPLOS PathogensCellGenome MedicineFrontiers in MicrobiologyScience SignalingInfection and ImmunityJournal of Theoretical BiologyFibrogenesis & Tissue RepaireLifeJournal of Gastroenterology and HepatologyScienceMolecular and Cellular BiologyThe Journal of ImmunologyPNASFEMS Microbiology Letters
Miriana DardanoParallel BioMiriana DardanoScientist, Liam Wilson, Robert Zweigerdt, Lika Drakhlis
Explain at my level
A result other labs cannot reproduce is not much use. This is the step-by-step recipe for growing the blood-making heart organoids and preparing them for imaging, published so that anyone can run it.
A detailed protocol for generating blood-generating heart-forming organoids from human pluripotent stem cells, covering differentiation, culture and sample preparation for advanced imaging modalities. Published as the methods companion to the 2024 Nature Cell Biology paper.
The protocol is the thing that makes a model system reproducible across sites, which in turn is what makes data from it poolable. Worth noting for anyone thinking about where multi-site biological datasets come from: a published, standardized protocol is the precondition, not an afterthought.
Why it matters
Makes the organoid model reproducible in other labs, which is what turns a result into a usable method.
Dardano M, Wilson L, Zweigerdt R, Drakhlis L. Production of human blood-generating heart-forming organoids and sample preparation for advanced imaging. Nat Protoc. 2025.
Anais Elewaut, Guillem Estivill, Felix Bayerl, Leticia Castillon, Maria Novatchkova, Elisabeth Pottendorfer, Lisa Hoffmann-Haas, Martin Schönlein, Trung Viet Nguyen, Martin Lauss, Francesco Andreatta, Milica Vulin, Izabela Krecioch, Jonas Bayerl, Anna-Marie Pedde, Naomi Fabre, Felix Holstein, Shona CroninParallel BioShona Cronin, PhDScientist, Sarah Rieser, Denarda Dangaj Laniti, David Barras, George Coukos, Camelia Quek, Xinyu Bai, Miquel Muñoz i Ordoño, Thomas Wiesner, Johannes Zuber, Göran Jönsson, Jan P. Böttcher, Sakari Vanharanta, Anna C. Obenauf
Explain at my level
T cells need to be re-activated once they are inside a tumor, or they stop working. This study found the cells that do the re-activating are a type nobody expected, and that cancer cells shut them down by releasing one particular molecule. Block that molecule and the T cells come back to life.
Inflammatory monocytes expressing Cxcl9, Cxcl10 and Il15 take a pivotal role in intratumoral T cell stimulation, acquiring and presenting peptide-MHC class I complexes from tumor cells through cross-dressing rather than cross-presentation. Hyperactivated MAPK signaling in cancer cells blunts type I interferon production while inducing PGE2 secretion, which impairs the inflammatory monocyte state. Enhancing IFN-I and blocking PGE2 restores the process and re-sensitizes tumors to T cell-mediated immunity.
A mechanism recovered by connecting a signaling state in one cell type to a functional state in another, within the same tissue. The intervention follows directly from the mechanism and reverses the phenotype, which is the loop that makes a finding actionable. For modeling, the lesson is that the relevant unit of analysis was a cell-cell interaction, not a cell state, and the key player was the cell type nobody had labelled as important.
Why it matters
Identifies how tumors switch off the immune cells that keep T cells working, and how to switch them back on.
Elewaut A, Estivill G, Bayerl F, Castillon L, Novatchkova M, Pottendorfer E, et al. Cancer cells impair monocyte-mediated T cell stimulation to evade immunity. Nature. 2025;637(8046):716–725.
Miriana DardanoParallel BioMiriana DardanoScientist, Felix Kleemiß, Maike Kosanke, Dorina Lang, Liam Wilson, Annika Franke, Jana Teske, Akshatha Shivaraj, Jeanne de la Roche, Martin Fischer, Lucas Lange, Axel Schambach, Lika Drakhlis, Robert Zweigerdt
Explain at my level
Long before bone marrow takes over, a human embryo makes its first blood cells next to the developing heart. Nobody could watch that happen, because it happens inside an embryo. This team grew it in a dish: an organoid with heart muscle, blood vessel lining and blood-forming tissue all developing together, in the right arrangement.
Modulation of the established heart-forming organoid (HFO) protocol with cytokine and growth factor cocktails yielded blood-generating HFOs: hPSC-derived organoids that undergo endothelial-to-haematopoietic transition and produce haematopoietic cells including erythrocytes, monocytes and macrophages. The model supports morphologically structured co-development of cardiac, endothelial and multipotent haematopoietic tissues equivalent to the intra-embryonic haematopoietic region in vivo.
A multi-tissue human model where the interactions between tissues are part of what is being modeled, not abstracted away. Most in vitro systems isolate one lineage; this one keeps three co-developing in their native spatial relationship, which is the regime where emergent behavior lives. For anyone building human-relevant training data, the relevant property is that it generates stage-specific human developmental biology on demand, from a renewable cell source.
Why it matters
Recreates the earliest stage of human blood formation, which has been inaccessible because it happens inside the embryo.
Dardano M, Kleemiß F, Kosanke M, Lang D, Wilson L, Franke A, Teske J, Shivaraj A, de la Roche J, Fischer M, Lange L, Schambach A, Drakhlis L, Zweigerdt R. Blood-generating heart-forming organoids recapitulate co-development of the human haematopoietic system and the embryonic heart. Nat Cell Biol. 2024;26:1984–1996.
Lukas Leiendecker, Tobias Neumann, Pauline S. Jung, Shona CroninParallel BioShona Cronin, PhDScientist, Thomas L. Steinacker, Alexander Schleiffer, Michael Schutzbier, Karl Mechtler, Thibault Kervarrec, Estelle Laurent, Kamel Bachiri, Etienne Coyaud, Rajmohan Murali, Klaus J. Busam, Babak Itzinger-Monshi, Reinhard Kirnbauer, Lorenzo Cerroni, Eduardo Calonje, Arno Rütten, Frank Stubenrauch, Klaus G. Griewank, Thomas Wiesner, Anna C. Obenauf
Explain at my level
The team searched cancer genome data for traces of 6,000 human viruses across 19 kinds of skin cancer. In one rare, aggressive cancer of the fingers and toes they found the same virus in 96% of samples, a strain previously thought harmless.
Off-target reads from routine NGS data were screened for viral pathogens across 19 skin tumor types. HPV42 was identified in 96% (45/47) of digital papillary adenocarcinomas, with viral integration confirmed by sequence capture and FISH. HPV42, previously classed as non-oncogenic and low-risk, recapitulates the molecular hallmarks of high-risk HPVs and elicits a germ cell-like transcriptional program conserved across HPV-driven cancers; reduced to CDKN2A and SYCP2 it serves as a fingerprint of oncogenic HPV.
The signal was already in data everyone had: the discovery came from reading the reads that standard pipelines discard. Worth noting as a case where the bottleneck was the analysis, not the measurement, and where a machine learning approach then compressed a transcriptional program down to a two-gene signature that generalizes across tumor types.
Why it matters
Identifies a new cancer-causing virus by reading data that standard analysis throws away.
Leiendecker L, Neumann T, Jung PS, Cronin SM, Steinacker TL, Schleiffer A, et al. Human Papillomavirus 42 Drives Digital Papillary Adenocarcinoma and Elicits a Germ Cell-like Program Conserved in HPV-Positive Cancers. Cancer Discov. 2023;13(1):70–84.
Christian Umkehrer, Felix Holstein, Laura Formenti, Julian Jude, Kimon Froussios, Tobias Neumann, Shona CroninParallel BioShona Cronin, PhDScientist, Lisa Haas, Jesse J. Lipp, Thomas R. Burkard, Michaela Fellner, Thomas Wiesner, Johannes Zuber, Anna C. Obenauf
Explain at my level
Normally you can label millions of cells and track which ones survive a drug, but you cannot get those specific cells back alive to study them. This method, called CaTCH, lets you reach into the population and retrieve any clone you want, including ones making up one cell in a hundred thousand.
CaTCH (CRISPRa Tracing of Clones in Heterogeneous cell populations) combines a barcode library with a CRISPRa-inducible reporter, allowing precise clonal tracing of millions of cells alongside retrospective isolation of founding clones alive, before and during selection. Clones representing as little as 0.001% of a population were captured, and the method was used to investigate the emergence of resistance to targeted melanoma therapy in vivo.
This turns lineage tracing from read-only into addressable. You can go back to the ancestors of a resistant clone and run functional experiments on them, which means the question of whether resistance was pre-existing or acquired becomes directly testable rather than inferred from barcode frequency dynamics. Generating counterfactual-capable experimental data is the hard part of causal inference in biology, and this is a tool for it.
Why it matters
Makes it possible to retrieve and study the rare cells that survive a treatment, instead of only counting them.
Kevan Shah, Rishi Bedi, Alex Rogozhnikov, Pavan Ramkumar, Zhixiang Tong, Brian Rash, Morgan Stanton, Jordan Sorokin, Cagsar Apaydin, Anthony Batarse, Julia Bergamaschi, Robert Blattner, Spencer Brown, Anthony Bosshardt, Carlos Castrillo, Brenda Dang, Shiron Drusinsky, Luigi Enriquez, David Grayson, Juliana HilliardParallel BioJuliana HilliardCo-Founder and Chief Science Officer, Pei-Ken Hsu, Chili Johnson, Ryan Jones, Andy Lash, Chia-Yao Lee, Kelly Li, Austin McKay, Elliot Mount, Justin Nicola, Ismael Oumzil, Justin Paek, Deborah Pascoe, Arden Piepho, Sean Poust, Daphne Quang, Matthew Schultz, Jessica Sims, Patrick Taylor, Geffen Treiman, Oliver Wueseke, Noah Young, Alex Pollen, Doug Flanzer, Daniel Chao, Gaia Skibinski, Saul Kato, G. Sean Escola
Explain at my level
Lab-grown mini-brains vary a lot from batch to batch, which makes them hard to trust as disease models. This team built a robot system to grow and photograph them at scale, used machine learning to work out exactly where the variability comes from, and then cut it down enough to detect a disease signature nobody had been able to see reliably before.
An automated Organoid Culture and Assay (ORCA) system supporting longitudinal, unbiased phenotyping of organoids across multiple patient lines. Novel machine learning methods were used to decompose variability by donor, clone and seeding batch. Multi-factorial protocol optimization produced a directed forebrain protocol compatible with 96-well culture that preserved tissue complexity at reduced variance, which revealed reproducible morphological and molecular phenotypes in heterozygous TSC+/− organoids at culture day 35.
The variance decomposition is the contribution here. The team separated donor effects from clone effects from batch effects, published the metric as open code (the RTG score), and only then pushed the protocol until the disease signal rose above the noise. That is the discipline that turns organoids into a training substrate rather than a set of anecdotes, and it is the same problem any population-scale human model has to solve first.
Why it matters
Tackles the problem that makes organoids hard to use at scale: knowing how much of what you measure is the biology and how much is the batch.
Shah K, Bedi R, Rogozhnikov A, Ramkumar P, Tong Z, Rash B, et al. Optimization and scaling of patient-derived brain organoids uncovers deep phenotypes of disease. bioRxiv. 2020. doi:10.1101/2020.08.26.251611.
Christian W. Johnson, Yi-Jang LinParallel BioYi-Jang Lin, PhDSenior Scientist, Derion Reid, Jillian Parker, Spiro Pavlopoulos, Patrick Dischinger, Carrie Graveel, Andrew J. Aguirre, Matthew Steensma, Kevin M. Haigis, Carla Mattos
Explain at my level
One particular Ras mutation, G13D, shows up in a quarter of colorectal cancers but almost never elsewhere. This study solved the structures to find out why, and the answer is in the shape of the protein's active site, which is loosened in one version of Ras but not in its near-identical relatives.
Crystal structures of G13D in all three Ras isoforms show an open active site, with adjusted D13 backbone torsion angles and disconnected switch regions. KRas G13D uniquely destabilizes the nucleotide-binding pocket; in the GDP-bound form, A59 occupies the Mg2+ binding site as in the HRas-SOS complex. Structure and biochemistry are consistent with the intermediate GTP-loading of KRas G13D relative to wild type and G12D observed in genetically engineered mouse models.
Three proteins with near-identical G domains respond differently to the same point mutation because of isoform-specific residues outside the active site. It is a concrete case of long-range allosteric context determining the effect of a local change, and therefore of why sequence-local mutation effect prediction fails in exactly the cases that matter clinically.
Why it matters
Explains why one cancer mutation is common in colorectal cancer and rare everywhere else, at the level of protein structure.
Johnson CW, Lin YJ, Reid D, Parker J, Pavlopoulos S, Dischinger P, Graveel C, Aguirre AJ, Steensma M, Haigis KM, Mattos C. Isoform-Specific Destabilization of the Active Site Reveals a Molecular Mechanism of Intrinsic Activation of KRas G13D. Cell Rep. 2019;28(6):1538–1550.e7.
Emily J. Poulin, Asim K. Bera, Jia Lu, Yi-Jang LinParallel BioYi-Jang Lin, PhDSenior Scientist, Samantha Dale Strasser, Joao A. Paulo, Tannie Q. Huang, Carolina Morales, Wei Yan, Joshua Cook, Jonathan A. Nowak, Douglas K. Brubaker, Brian A. Joughin, Christian W. Johnson, Rebecca A. DeStefanis, Phaedra C. Ghazi, Sudershan Gondi, Thomas E. Wales, Roxana E. Iacob, Lana Bogdanova, Jessica J. Gierut, Yina Li, John R. Engen, Pedro A. Perez-Mancera, Benjamin S. Braun, Steven P. Gygi, Douglas A. Lauffenburger, Kenneth D. Westover, Kevin M. Haigis
Explain at my level
KRAS is the most commonly mutated cancer gene, but different versions of the mutation show up in different organs. This study asked why, and found the answer is not chance: each mutant protein behaves differently, and that behavior determines which tissue it can turn cancerous.
A cross-disciplinary comparison of KRAS G12D and KRAS A146T combining structural biology, mass spectrometry and genetically engineered mouse models. KRAS A146T shows marked extension of switch 1 away from the protein body and nucleotide binding site, activating KRAS through a high rate of intrinsic and GEF-induced nucleotide exchange rather than impaired GAP-stimulated hydrolysis. Mice expressing either allele show distinct tissue-specific effects on homeostasis that mirror the mutational frequencies seen in human cancers.
Allele identity carries information that gene-level labels discard. Two mutations in the same gene activate the protein by different biochemical routes and produce different tissue tropism, which means a model keyed on "KRAS mutant" is collapsing two distinct states. The structural mechanism explains the epidemiological pattern, which is a rare case of a causal account recovered all the way from crystal structure to cancer incidence.
Why it matters
Explains why the same cancer gene causes different cancers depending on exactly how it is mutated.
Patricia A. Darrah, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Pauline Maiello, Hannah P. Gideon, Amy J. Myers, Mark A. Rodgers, Joshua A. Hackney, Thomas Lindenstrom, Thomas Evans, Charles A. Scanga, Victor Prikhodko, Peter Andersen, Philana Ling Lin, Dominick Laddy, Mario Roederer, Robert A. Seder, JoAnne L. Flynn
Explain at my level
BCG is the only approved TB vaccine and it protects unevenly. This study tried to strengthen it with a booster shot, testing several boosters in 64 monkeys. None of them worked better than BCG alone, and one even seemed to weaken the protection BCG gave.
A large non-human primate trial of heterologous prime-boost regimens: BCG priming boosted with protein-plus-adjuvant (M72/AS01E or H56/CAF01) intramuscularly, or with replication-defective Ad5 vectors expressing various Mtb antigens delivered intramuscularly and by aerosol. Aerosol Ad5(TB) elicited 10–30% antigen-specific CD4 and CD8 multifunctional cytokine responses in bronchoalveolar lavage, yet no regimen improved protection over BCG. An empty Ad5 vector control appeared to diminish BCG-induced protection.
A well-powered negative result, and a useful one. Strong immune readouts in the lung decoupled entirely from protection, which means the obvious biomarker was not the label. Any model trained to predict vaccine efficacy from cytokine response would have called this wrong. It is a direct argument for measuring outcomes, not proxies, and for keeping negative results in the training distribution.
Why it matters
Shows that a strong immune response does not mean protection, which is a core problem in picking what to advance into trials.
Mark M. Davis, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO
Explain at my level
An essay on closing the gap between immunology research and the clinic: how new ways of measuring the human immune system, many cell types and molecules in many people at once, could make immune health something doctors routinely read and act on.
An essay on translating systems immunology into clinical practice, making the case for high-dimensional, population-scale profiling of human immune status as a routine diagnostic and therapeutic measure.
An essay arguing that the immune system can be measured at scale, across many cell types, molecules and people, and that this data could become a routine clinical signal. For AI readers, it is an early case for why large human immune datasets matter.
Why it matters
Frames the immune system as a measurable, central factor in health and disease, a view that runs through Parallel Bio’s work.
Anthony M. Cadena, Forrest F. Hopkins, Pauline Maiello, Allison F. Carey, Eileen A. Wong, Constance J. Martin, Hannah P. Gideon, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Peter Andersen, Philana Ling Lin, Sarah M. Fortune, JoAnne L. Flynn
Explain at my level
If you already have a tuberculosis infection, are you protected from catching it again? The team tagged two strains of the bacterium with genetic barcodes so they could tell them apart, then infected monkeys a second time. The existing infection blocked the new one almost completely.
Barcoded Mtb libraries allowed primary and secondary infections to be distinguished within the same animal. Macaques with ongoing primary infection formed significantly fewer granulomas attributable to the secondary challenge strain than naive animals, demonstrating robust concomitant immunity, with PET-CT and Q-tag analysis used to attribute new lesions to each library.
Barcoding turns an otherwise unlabelled outcome into per-lesion provenance, so each granuloma carries the identity of the strain that caused it. It is a good example of designing the experiment so the labels come out of it, rather than inferring them afterward, which is what makes a dataset usable for modeling.
Why it matters
Shows the immune system can block reinfection, which sets a benchmark for what a successful vaccine would have to match.
Tom KruitwagenParallel BioTom KruitwagenDirector, Stem Cell Biology, Pierre Chymkowitch, Annina Denoth-Lippuner, Jorrit Enserink, Yves Barral
Explain at my level
Cells have to tell their own DNA apart from foreign DNA that gets in, such as from a virus. This study found yeast cells use a surprising marker to do it: the centromere, a structure that was thought to be only for pulling chromosomes apart during division. DNA without one fails to fold up and gets discarded into just one daughter cell.
Non-chromosomal DNA circles and chromosomes lacking a centromere fail to condense during mitosis, whereas wild-type chromosomes condense normally. The centromere licenses condensation of the chromosome arms in cis. Acentric plasmids instead attach to nuclear pore complexes and segregate asymmetrically into one daughter cell, suggesting the centromere functions as a mark of self-DNA.
A single discrete feature acts as the label for self versus non-self at the level of the whole molecule, and the downstream consequence (condensation, then symmetric segregation) follows from it. It is a tidy example of a biological system implementing authentication with one mark rather than a distributed signature, and of a mechanism that only becomes visible when you remove the mark.
Why it matters
Identifies a defense mechanism cells use to recognize and discard foreign DNA, in a place nobody was looking.
Lisa E. Wagar, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Mark M. Davis
Explain at my level
Much of what we know about the immune system comes from mice, and mouse immunity differs from ours in ways that matter for drugs and vaccines. This review surveys the tools for studying human immunity directly, from organoids and tissue models to detailed single-cell measurements, and argues for making human systems the default.
A review of advanced model systems and tools for human immunology, covering in vitro and ex vivo human models, including organoid and tissue-engineered systems, alongside high-dimensional profiling technologies, and assessing their use in basic and translational research.
Most immunology data comes from mice, a proxy that transfers poorly to humans: a distribution-shift problem in biological form. This review surveys ways to generate data from human immune systems directly, from organoids to high-dimensional single-cell measurements, the kinds of sources human-relevant models can be trained on.
Why it matters
The scientific case for the approach Parallel Bio was founded on: study human immunity in human systems.
Louis R. Joslyn, Elsje Pienaar, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Sara Suliman, Benjamin M. Kagina, JoAnne L. Flynn, Thomas J. Scriba, Jennifer J. Linderman, Denise E. Kirschner
Explain at my level
The same vaccine gives different results in monkeys and in people. This study built a mathematical model of how T cells respond to vaccination, fitted it to both monkey and human blood data, and used it to work out which differences between the species explain the gap.
A two-compartment mathematical model of T cell priming, proliferation and differentiation across lymph node and blood, calibrated separately to NHP and human datasets for the H56 subunit vaccine. Uncertainty and sensitivity analysis isolated the mechanisms driving divergent vaccination responses, including the effect of BCG prime timing on subsequent H56 immunogenicity.
An explicit attempt to learn the transfer function between a model organism and humans rather than assuming one. The same mechanistic model is fitted to both species, and the fitted parameters become the account of why results do not transfer. That is a more honest framing of cross-species prediction than fitting one species and hoping, and it shows what a human-first dataset would make unnecessary.
Why it matters
Makes the gap between monkey and human vaccine responses explicit and measurable rather than assumed.
Joslyn LR, Pienaar E, DiFazio RM, Suliman S, Kagina BM, Flynn JL, Scriba TJ, Linderman JJ, Kirschner DE. Integrating Non-human Primate, Human, and Mathematical Studies to Determine the Influence of BCG Timing on H56 Vaccine Outcomes. Front Microbiol. 2018;9:1734.
Jesse Lyons, Douglas K. Brubaker, Phaedra C. Ghazi, Katherine R. Baldwin, Amanda Edwards, Myriam Boukhali, Samantha Dale Strasser, Lucia Suarez-Lopez, Yi-Jang LinParallel BioYi-Jang Lin, PhDSenior Scientist, Vijay Yajnik, Joseph L. Kissil, Wilhelm Haas, Douglas A. Lauffenburger, Kevin M. Haigis
Explain at my level
Inflammatory bowel disease has few treatment options. This study measured three different layers of biology (genes, proteins and the chemical switches on them) in the same tissue samples, which let them follow the signal from one layer to the next and spot a pathway that no single layer revealed. Blocking it suppressed the disease in mice.
RNA microarray, total protein MS and phosphoprotein MS were collected from the same samples in a mouse colitis model, allowing information flow from RNA to protein to phosphoprotein to be tracked within individual animals. Trans-omics coexpression network analysis combined with human transcriptomic data implicated p21-activated kinase signaling. Chemical inhibition of Pak1 and Pak2 with FRAX597 suppressed active colitis in vivo.
Matched multi-omic measurement on the same samples, rather than separate cohorts per assay, is what makes the cross-layer inference valid here. The finding only appears when the layers are aligned per-animal and combined with human data, which is a good argument for designing data collection around the joint distribution rather than assembling it from marginals later.
Why it matters
Finds a drug target for inflammatory bowel disease that no single type of measurement would have revealed.
Lyons J, Brubaker DK, Ghazi PC, Baldwin KR, Edwards AL, Boukhali M, Strasser SD, Suarez-Lopez L, Lin YJ, Yajnik V, Kissil JL, Haas W, Lauffenburger DA, Haigis KM. Integrated in vivo multiomics analysis identifies p21-activated kinase signaling as a driver of colitis. Sci Signal. 2018;11(519):eaan3580.
Pauline Maiello, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Anthony M. Cadena, Mark A. Rodgers, Philana Ling Lin, Charles A. Scanga, JoAnne L. Flynn
Explain at my level
Two closely related species of monkey are both used to study tuberculosis. This study compared them head to head and found they get strikingly different versions of the disease: one species controls the infection far better than the other. Which animal you pick changes the answer you get.
A quantitative comparison of rhesus and cynomolgus macaques infected with the same low dose of Mtb, using serial PET-CT imaging, granuloma counts and bacterial burden at necropsy. Rhesus showed higher rates of progressive disease, greater lung bacterial burden and more extrapulmonary involvement, while cynomolgus macaques more often contained infection.
A clean measurement of between-species variance inside the animal-model layer itself. Two species separated by a short evolutionary distance give systematically different outcomes from an identical perturbation, which puts a hard floor on how much of the variance in preclinical results is biology of interest versus choice of host. It is the distribution-shift problem one step before the usual mouse-to-human framing.
Why it matters
Quantifies how much the choice of animal, not the biology being studied, determines the result.
Maiello P, DiFazio RM, Cadena AM, Rodgers MA, Lin PL, Scanga CA, Flynn JL. Rhesus Macaques Are More Susceptible to Progressive Tuberculosis than Cynomolgus Macaques: a Quantitative Comparison. Infect Immun. 2018;86(2):e00505-17.
Hayley C. Warsinske, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Jennifer J. Linderman, JoAnne L. Flynn, Denise E. Kirschner
Explain at my level
Scar tissue around a tuberculosis lesion can help by walling off the infection, or hurt by stiffening the lung. This study built a computer simulation of a lesion forming, cell by cell, to work out which signals push it toward each kind of scarring.
A hybrid multi-scale agent-based model spanning molecular, cellular and tissue scales, built to reproduce the two observed morphologies of fibrotic granulomas: peripherally fibrotic with a collagen cuff, and centrally fibrotic. Uncertainty and sensitivity analysis over large simulation sets identified the cytokine conditions, particularly TGF-β1 and IL-10, that differentiate the two outcomes. Model predictions were anchored to stained NHP granuloma sections.
The computational companion to the TGF-β paper already on the page. A mechanistic simulator stands in for experiments that cannot be run in vivo, and large simulation sweeps are used to find which parameters actually separate the outcomes. The pairing of a simulator with a small set of real tissue measurements is the pattern worth noting: the wet data constrains the model, and the model explores where the wet data cannot go.
Why it matters
Explains how the same infection produces two different kinds of lung scarring, which determines how much lung function a patient keeps.
Philana Ling Lin, Pauline Maiello, Hannah P. Gideon, M. Teresa Coleman, Anthony M. Cadena, Mark A. Rodgers, Robert Gregg, Melanie O’Malley, Jaime Tomko, Daniel Fillmore, L. James Frye, Tara Rutledge, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Christopher Janssen, Edwin Klein, Peter L. Andersen, Sarah M. Fortune, JoAnne L. Flynn
Explain at my level
More than two billion people carry a dormant tuberculosis infection, and only some will ever get sick. This study used scans to find which animals were at high risk before anything happened, and the scans predicted reactivation with about 92% accuracy.
Serial FDG PET-CT imaging of 26 latently infected cynomolgus macaques before and during TNF neutralization, with 25 latently infected controls. Reactivation, defined as at least one new granuloma in lung or extrapulmonary sites, occurred in 50% of treated animals. Increased lung inflammation by PET and pre-existing extrapulmonary involvement predicted reactivation with 92% sensitivity and specificity. High-risk animals had higher cumulative and per-lesion bacterial burden and more IL-2, IL-10 and IL-17 producing T cells in granulomas.
An imaging signature that predicts a future binary event at 92% sensitivity and specificity, validated against bacterial burden and granuloma-level cytokine profiles. The interesting part is that the predictive signal sits in a whole-body imaging modality rather than in a blood marker, and that latency turns out to be a spectrum with measurable position rather than a single state.
Why it matters
Reframes latent tuberculosis as a spectrum with measurable risk rather than a single dormant state.
Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Joshua T. Mattila, Edwin C. Klein, Lauren R. Cirrincione, Mondraya Howard, Eileen A. Wong, JoAnne L. Flynn
Explain at my level
The body walls off tuberculosis bacteria inside lung structures called granulomas. This study found that after antibiotic treatment, granulomas shift toward scar-like tissue, and that the signaling molecule TGF-β tracks closely with that change.
We examined pulmonary tuberculosis granulomas after drug treatment and found that active transforming growth factor-β was associated with phenotypic changes in treated lesions, including features of fibrosis.
The immune system contains tuberculosis bacteria inside lung lesions called granulomas. After antibiotic treatment, these lesions shifted toward scar-like tissue, and that shift tracked with levels of active TGF-β, a signaling protein: a measurable molecular marker following a tissue-level outcome.
Why it matters
Shows how the immune system reshapes tissue after an infection is treated, which bears on lasting lung damage and on how we judge whether treatment worked.
DiFazio RM, Mattila JT, Klein EC, Cirrincione LR, Howard M, Wong EA, Flynn JL. Active transforming growth factor-β is associated with phenotypic changes in granulomas after drug treatment in pulmonary tuberculosis. Fibrogenesis & Tissue Repair 9, 6 (2016).
Jessica J. Gierut, Levi B. Wood, Ken S. Lau, Yi-Jang LinParallel BioYi-Jang Lin, PhDSenior Scientist, Casie Genetti, Ahmed A. Samatar, Douglas A. Lauffenburger, Kevin M. Haigis
Explain at my level
Drugs that hit the same target are usually assumed to do the same thing. This study tested several inhibitors of one signaling pathway in the gut lining and found they had quite different effects on cell death, because each one also perturbs the rest of the network in its own way.
Inhibitors of Raf, MEK and ERK exerted distinct effects on the timing and magnitude of TNF-α-induced apoptosis in the mouse intestine. Different MEK inhibitors also diverged: CH5126766 potentiated apoptosis. Computational modeling and experimental perturbation identified Akt as the primary signaling node enhancing apoptosis in the presence of CH5126766.
Same nominal target, different phenotype, because the off-pathway perturbation profile differs per compound. Treating a drug as its annotated target is a lossy encoding, and this paper shows the loss is large enough to flip the direction of the outcome. Network-level modeling is what recovered the responsible node.
Why it matters
Shows that two drugs aimed at the same target can produce opposite outcomes in tissue.
Gierut JJ, Wood LB, Lau KS, Lin YJ, Genetti C, Samatar AA, Lauffenburger DA, Haigis KM. Network-level effects of kinase inhibitors modulate TNF-α-induced apoptosis in the intestinal epithelium. Sci Signal. 2015;8(407):ra129.
Tom KruitwagenParallel BioTom KruitwagenDirector, Stem Cell Biology, Annina Denoth-Lippuner, Bryan J. Wilkins, Heinz Neumann, Yves Barral
Explain at my level
Before a cell divides, it has to pack its long, tangled DNA into compact chromosomes that can be pulled cleanly into the two new cells. Working in yeast, this study found that the packing happens in two separate ways at once: the DNA fiber squeezes itself tighter locally, while a protein machine called condensin shortens each chromosome along its length. Either one alone is enough to get by; lose both and cells divide badly and often die.
Using a live-cell fluorescence assay for local chromatin compaction in budding yeast, we show that mitotic chromosome condensation comprises two temporally and mechanistically distinct processes. Phosphorylation of histone H3 at serine 10 and deacetylation of histone H4 drive nucleosome–nucleosome interactions that compact chromatin over short range in early anaphase, while condensin independently drives axial contraction of chromosome arms, peaking later in anaphase. Disrupting both pathways causes synergistic defects in segregation and viability, and both depend at least partly on the deacetylase Hst2.
Chromosome folding turns out to be two independent mechanisms acting at different length scales and times: local compaction of the chromatin fiber, triggered by chemical marks on histone proteins, and condensin-driven shortening of whole chromosome arms. Removing either one leaves the other intact and cells cope; removing both causes a synergistic failure. It is a clean example of redundancy that single-perturbation experiments miss, and a reason perturbation datasets need combinations, not just one change at a time.
Why it matters
Shows that cells shape their chromosomes for division with two independent backup systems, which is why the role of each only becomes visible when both fail.
Mitotic chromosome condensation involves two separate processes: short-range compaction of the chromatin fiber and condensin-driven axial contraction of chromosome arms.
Short-range compaction is triggered in early anaphase by phosphorylation of histone H3 at serine 10 and deacetylation of histone H4.
Losing either process alone has little visible effect; losing both causes synergistic defects in chromosome segregation and cell viability.
Cite this
Kruitwagen T, Denoth-Lippuner A, Wilkins BJ, Neumann H, Barral Y. Axial contraction and short-range compaction of chromatin synergistically promote mitotic chromosome condensation. eLife 4, e10396 (2015).
Christine M. Fillmore, Chunxiao Xu, Pooja T. Desai, Joanne M. Berry, Samuel P. Rowbotham, Yi-Jang LinParallel BioYi-Jang Lin, PhDSenior Scientist, Haikuo Zhang, Victor E. Marquez, Peter S. Hammerman, Kwok-Kin Wong, Carla F. Kim
Explain at my level
A standard chemotherapy drug works well in only a minority of lung cancer patients. This study found that adding a second drug makes the chemotherapy work in tumors carrying particular mutations, and makes it work worse in tumors without them. Which group a patient is in can be read off their tumor's genetics.
EZH2 inhibition had differential effects on TopoII inhibitor response in NSCLC in vitro and in vivo. BRG1 loss-of-function mutant tumors responded with increased S phase, anaphase bridging, apoptosis and etoposide sensitivity, whereas EGFR and BRG1 wild-type tumors upregulated BRG1 in response and became more resistant. EGFR gain-of-function mutants were also sensitized, through genetic antagonism between EGFR and BRG1.
A drug combination whose effect reverses sign depending on genotype. Averaged over an unstratified population the effect would look like noise or a null result; the signal exists only conditional on the right covariate. It is a compact illustration of why population-average efficacy estimates mislead, and why genotype-stratified data is the minimum useful resolution.
Why it matters
Identifies which lung cancer patients a drug combination would help and which it would harm, from their tumor genetics.
Two liver cancer drugs worked better together than either did alone. The reason turned out to be immune, not direct: the combination brought more killer T cells into the tumor. When those T cells were removed, the benefit disappeared entirely.
In an orthotopic murine liver cancer model, the combination produced synergistic tumor growth delay and survival benefit, associated with increased IFN-γ-expressing CD8+ lymphocytes among tumor-infiltrating lymphocytes and higher numbers of granzyme- and perforin-expressing CD8+ T cells. Apoptosis and vascular normalization both increased. CD8 depletion abolished the synergy.
The depletion experiment is the part that matters: it converts a correlation between T cell infiltration and response into a causal dependency. Perturbation of the proposed mediator, not just measurement of it, is what distinguishes a mechanism from a biomarker, and it is the step most observational datasets cannot support.
Why it matters
Shows a drug combination working through the immune system rather than directly on the tumor.
Bryan J. Wilkins, Nils A. Rall, Yogesh Ostwal, Tom KruitwagenParallel BioTom KruitwagenDirector, Stem Cell Biology, Kyoko Hiragami-Hamada, Marco Winkler, Yves Barral, Wolfgang Fischle, Heinz Neumann
Explain at my level
Before a cell divides, its DNA packs itself down into compact bundles. This study traced the chain of chemical events that sets the packing off, using light-activated crosslinkers to catch proteins in the act of touching each other inside living cells.
Genetically encoded UV-inducible crosslinkers were used to monitor protein-protein interactions with spatiotemporal resolution in living yeast. Phosphorylation of histone H3 serine 10 recruits the deacetylase Hst2p, which removes an acetyl group from histone H4 lysine 16, freeing the H4 tail to contact neighboring nucleosomes and promoting fiber condensation. This provides a condensin-independent driving force for mitotic hypercondensation.
A method paper as much as a mechanism paper: the crosslinkers turn an interaction that is normally inferred into something measured directly in a live cell, with timing attached. Datasets of physical contacts with temporal resolution are rare, and they are what you need to separate a causal chain from a correlation between marks.
Why it matters
Traces the chemical chain that packs DNA for cell division, and provides a way to watch protein contacts happen in living cells.
Wilkins BJ, Rall NA, Ostwal Y, Kruitwagen T, Hiragami-Hamada K, Winkler M, Barral Y, Fischle W, Neumann H. A Cascade of Histone Modifications Induces Chromatin Condensation in Mitosis. Science. 2014;343(6166):77–80.
Boryana Petrova, Sascha Dehler, Tom KruitwagenParallel BioTom KruitwagenDirector, Stem Cell Biology, Jean-Karim Hériché, Kota Miura, Christian H. Haering
Explain at my level
To study how chromosomes fold, you first need a way to measure folding. This study built one: tag two spots on the same chromosome with fluorescent markers, then track the distance between them in a living cell, frame by frame.
A live-cell microscopy assay in Schizosaccharomyces pombe that automatically tracks three-dimensional distance changes between two fluorescently marked chromosome loci at high temporal and spatial resolution, yielding defined parameters for condensation dynamics in mitosis and meiosis. The assay resolved the separate contributions of condensin, topoisomerase II and Aurora kinase, and identified new mutants in condensin, DNA polymerase ε and F-box DNA helicase I specifically defective in pro/metaphase condensation.
The measurement infrastructure that the later condensation work is built on. An automated imaging pipeline converts a qualitative phenotype into continuous parameters, which is what makes a screen possible: once folding is a number per cell per frame, mutants that affect it can be found rather than guessed at.
Why it matters
Turns chromosome folding from something you look at into something you measure, which is what made the later screens possible.
Angela M Green, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, JoAnne L Flynn
Explain at my level
Many kinds of immune cell can make IFN-γ, a signal that helps control tuberculosis. This study asked whether the signal is enough on its own, or whether it has to come from one particular cell type. It had to come from CD4 T cells; the same molecule from other sources did not substitute.
An adoptive transfer system was used to separate the requirement for CD4 T cells from the requirement for IFN-γ itself. IFN-γ from CD4 T cells proved necessary for host survival and optimal long-term bacterial control, and was also required for a robust CD8 T cell response. T cell-derived IFN-γ inhibited intracellular Mtb replication in macrophages.
A cell-type-resolved result: the same molecule carries a different effect depending on which cell secretes it. Bulk measurements of a cytokine would miss this entirely, which is an argument for single-cell resolution in immune datasets and against treating a cytokine level as a sufficient feature.
Why it matters
Shows that where an immune signal comes from matters as much as how much of it there is.
Green AM, DiFazio R, Flynn JL. IFN-γ from CD4 T cells is essential for host survival and enhances CD8 T cell function during Mycobacterium tuberculosis infection. J Immunol. 2013;190(1):270–277.
Daniele V. F. Tauriello, Ingrid Jordens, Katharina Kirchner, Jerry W. Slootstra, Tom KruitwagenParallel BioTom KruitwagenDirector, Stem Cell Biology, Britta A. M. Bouwman, Maria Noutsou, Stefan G. D. Rüdiger, Klaus Schwamborn, Alexandra Schambony, Madelon M. Maurice
Explain at my level
Wnt is one of the signals that tells cells what to become during development, and faults in it drive cancer. This study worked out how two of the proteins in the chain actually grip each other: not at one site, but at three separate stretches that have to work together.
A Frizzled peptide microarray identified a bipartite role for the Dishevelled DEP domain plus C-terminal region (DEP-C) in binding a three-segmented discontinuous motif in Frizzled, spanning two conserved motifs in the third intracellular loop and the classic C-terminal KTxxxW motif. Cooperative use of all three is required for DEP-C binding and Wnt-induced β-catenin activation in cultured cells and Xenopus embryos.
A binding interface that is discontinuous in sequence but contiguous in structure, and only functional when all three segments are present. Sequence-local models would not recover it, and the cooperativity means single-site mutagenesis would have understated each part. A reminder that interaction datasets built from pairwise site mapping encode an assumption that often does not hold.
Why it matters
Maps a developmental signaling step at the level of detail needed to target it with a drug.
Tauriello DVF, Jordens I, Kirchner K, Slootstra JW, Kruitwagen T, Bouwman BAM, Noutsou M, Rüdiger SGD, Schwamborn K, Schambony A, Maurice MM. Wnt/β-catenin signaling requires interaction of the Dishevelled DEP domain and C terminus with a discontinuous motif in Frizzled. Proc Natl Acad Sci USA. 2012;109(14):E812–E820.
Julia J. Williams, Elizabeth M. Halvorsen, Ellen M. Dwyer, Robert DiFazioParallel BioRobert DiFazio, PhDCo-Founder and CEO, Paul J. Hergenrother
Explain at my level
Many bacteria carry a built-in self-destruct system: a toxin kept in check by a matching antitoxin. This study found these systems are common and switched on in drug-resistant hospital bacteria, pointing to a new way to kill them: release the toxin.
We surveyed clinical isolates of Pseudomonas aeruginosa and methicillin-resistant Staphylococcus aureus for toxin–antitoxin loci and found them prevalent and transcribed, supporting TA systems as candidate targets for antibacterial strategies based on toxin activation.
Many bacteria carry paired genes: one encodes a toxin, the other an antitoxin that neutralizes it. This study found these pairs are common and actively expressed in two drug-resistant hospital pathogens, Pseudomonas aeruginosa and MRSA, suggesting that releasing the toxin could be a new way to kill them.
Why it matters
Identified a potential new weak point in “superbugs” that resist today’s antibiotics.
Williams JJ, Halvorsen EM, Dwyer EM, DiFazio RM, Hergenrother PJ. Toxin–antitoxin (TA) systems are prevalent and transcribed in clinical isolates of Pseudomonas aeruginosa and methicillin-resistant Staphylococcus aureus. FEMS Microbiology Letters 322, 41–50 (2011).
Summary written by Parallel Bio, reviewed by Robert DiFazio.
Work with our scientists
We collaborate with academic labs, pharma partners and AI researchers. Datasets and code from our papers are available where marked; for anything else, ask us.