The embryo game uncovers hidden cell behaviours Authors: Maria Abou Chakra, Joshua Hislop, Ipek Egilmez, Rafael Alkalai, Omar Bashth, Kieran Maheden, Akash Gundagathi, Mo R. Ebrahimkhani, Nika Shakiba Published: 2026 ABSTRACT While cells are driven to preserve their own survival, they must cooperate to ensure embryo viability. To explore how cellular behaviours shape human development, we intersect evolutionary game theory and in vitro embryo modelling. We focus on cell competition, a conserved cell-cell killing thought to enable embryonic error correction, though few studies directly link it to embryo success. Our collective risk game shows that killing is costly for a cell, but beneficial when the embryo exceeds its target size and the probability of survival is low. Analogously, successful heX-embryoids form within the target size range of the native human embryonic disc and inhibiting cell death produces larger, defective embryoids. Killing and proliferating are responsive cellular behaviours that maintain optimal embryoid size, consistent with their emergence as conditional strategies in our game. Further, conditional killing is a “threat” to restrain over-proliferating cells. Counter-intuitively, cell competition may underpin cooperation in the embryo. Introduction Cellular behaviours play a critical role in shaping the success of the developing embryo1. It is a great feat to grow from one cell (the zygote) to the trillions of cells that compose our bodies; an absence of coordination and cooperation between cells risks embryo viability. Embryonic cells need to work together to coordinate cell divisions to achieve the correct population size at requisite developmental checkpoints2–6. For example, the formation of the amniotic cavity following implantation of the mouse embryo to the uterine wall depends on embryo size falling within a target range that enables proper morphogenesis and cavity formation7–9. But working together does not necessarily mean equal contributions – one cell may divide more than others. For instance, beginning from the first two cells (blastomeres) that emerge from the zygote, various scenarios can allow the embryo to reach the correct size: both blastomeres could contribute equally, but unequal contributions could also yield viable cell numbers. Indeed, blastomeres are known to make remarkably skewed contributions to the developing tissues of the human body10–12. How cells respond to their neighbours’ uneven or “unfair” contributions could provide new insights into embryo health. The success of the individual cell depends not only on its own behaviour, but that of the collective embryo13. Cells that behave unfairly can defect from the collective by dividing uncontrollably to overtake the population. These cells can drive the embryo to failure, for example, as cancers14–17. To protect against the loss of the whole embryo, cells can become “killers” and eliminate their neighbours18–20. This “killer” behaviour has the potential to dramatically remodel the embryo environment, as in mouse development, where 35% of epiblast cells are purged following embryonic implantation18. In fact, cell death has been shown to peak at this stage, shortly before gastrulation gives rise to the body plan21. The elimination of cells during this and other stages of development catalyses skewed contributions between cells19,20,22–25, each seeking to maximise their own progeny in the embryo. However, it is not clear if this strategy is a symptom of healthy cell coordination for embryo success or a competitive behaviour uncorrelated with embryo viability. The highly conserved nature of killing across tissue systems and species suggests that the emergence of this process during development may be for the greater good of the embryo19,20,23,25–28. Indeed, this process has been found to enable error correction in a variety of processes across model developmental systems, including during Drosophila wing development29, zebrafish pattern formation27, and mouse skin development25. While possibly advantageous to the collective cells composing the embryo, the elimination of cells is costly for the individual cell. For example, cells expend energy to engulf and eliminate neighbours20 as well as to remove debris of dead cells nearby30,31. The time spent on these activities could be considered as missed opportunities to increase their own fitness. There are also open questions about the context in which “killers” emerge: Do all cells kill for the same reason? Does killing help make space for a killer’s own progeny or to reach a target size to ensure group success? The challenging nature of tracking killing in embryonic development has limited our ability to answer these questions, given that killed cells are erased from the embryo32. Technical and ethical considerations have further hindered studies of early human development, leaving the mechanisms by which cells in the human embryo regulate their numbers largely unknown33,34. Exciting advancements in models of human development provide a timely opportunity to revisit these fundamental questions35–42. We use a combination of in vitro and in silico modelling to shed light on whether cell killing is a strategy used to hit target embryo size34, exploring the link between this cellular strategy and embryo outcomes. In vitro strategies for deriving embryo-like structures provide a window into development by recapitulating key aspects of their in vivo counterparts, including embryonic lineages and morphogenesis43. For example, heX-embryoids41 are derived from human induced pluripotent stem cells (hiPSCs)44,45. Specifically, wild-type hiPSCs are co-cultured with hiPSCs carrying an inducible human GATA6 transgene, a key regulator of extra-embryonic endoderm46,47. Following co-culture, wild-type hiPSCs form clusters and undergo self-organisation, giving rise to three-dimensional cellular structures that mimic amniotic cavity formation, akin to early embryos41. Like human embryos12, heX-embryoids have variable sizes, allowing us to explore whether deviations in size impact their ability to correctly form an amniotic cavity, a known size checkpoint in mouse development7–9. Since little is known about the effect of size deviations on underlying cell behaviours at this stage in human development, the heX-embryoid is a compelling model with which to explore this question9. In silico models, such as evolutionary game theory (EGT), provide another window into the cellular behaviours that emerge during embryonic development. EGT is widely used to explore how individuals will respond under social conditions. EGT reframes the cellular interaction problem as a game between competing strategies. The model then compares how a player with one strategy will manage against another. EGT has been applied to a broad range of unintuitive problems across fields, including understanding cooperative behaviours in the context of human behaviour48–52, climate change53–55, and ecology56,57. Public goods games58 in particular have been used to model multicellular systems59–62 in terms of their gene interactions63 and social structure, in which cells contribute to and share a common pool of resources64 or punish others that defect65. For instance, EGT has been applied to model cells as players to understand how they interact during cancer development, when they cooperate or defect for access to common resources, and how to drive the population into a state accessible by current radiation therapies17,66–68. These studies have revealed that cancer cells often exploit their environment to increase their fitness69, but some have also evolved to cooperate in order to evade therapies14. While EGT models have been used to explore basic and clinical questions in the context of multicellular systems such as cancer, no study to date has applied it to explore cell interactions within a developing system that grows in time, such as embryogenesis. This gap is particularly significant for human development, where direct observation of the cell-level behaviours that drive embryo success or failure has not been accessible. EGT allows us to probe how cellular strategies that confer a fitness advantage emerge over evolutionary time, enabling predictions about when and why killing should emerge as a beneficial behaviour at the collective level. Here, we use a collective risk game to connect cell interactions during embryonic development to embryo viability. The collective risk game is a threshold public goods game70 where individuals must work together to avoid collective loss. It is well-suited for capturing dynamic systems such as the developing embryo, where cells interact repeatedly over time, a process naturally captured by the multiple rounds of the game. By intersecting EGT and the heX-embryoid, we create and apply a new framework for predicting the impact of cellular behaviours during inaccessible stages of human embryonic development. We find that embryoids that are able to produce an amniotic cavity with a single lumen (which we label as “viable” structures) fall within a size range that matches human embryos at this developmental stage. This defines a “target size” that embryoids must achieve to maximise viability in the game. Our results further suggest that killing is a responsive behaviour used to enable success when the embryoid exceeds its target size by removing extra cells. Overall, we find that killing is a global strategy during early embryonic development and may be a regulative signal to ensure robustness and viability. Results A collective risk model of embryonic development We model embryonic development as a collective risk dilemma (CRD; Fig. 1a and “Methods”). The CRD allows us to capture cell behaviours as they interact under various conditions to achieve a collective target. The target represents the number of cells needed at developmental milestones for correct embryonic structures to emerge (target size). Developmental checkpoints in which embryos must hit a target size have been reported in vivo in mouse embryos7–9, in vitro in mouse and in human gastrulation models71–73, as well as in zebrafish whole embryo explants6. We model a single checkpoint where the collective embryo pot needs to achieve the target size at the end of the game, after r rounds. If the two players collectively reach the set target size of the embryo pot, all players obtain their individual payoffs (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π), which directly influences the fitness of the cellular strategy played by the player (whether to proliferate, kill or do nothing). We correspond achieving the target size to the probability of embryo survival (p), which determines the outcome of the collective behaviour at the end of the game and represents the probability that the embryo is viable. We model an “exact-or-nothing” game (see “Methods”), where the final embryo pot size must not deviate from the target size, otherwise risking their payoff with a survival probability, p. For example, in a zero survival probability scenario, a player contributing one cell obtains a payoff of 1 when the target is reached and 0 when the target is missed (Fig. 1b). We explore low survival probabilities that are observed in early human development (60% of embryos reportedly fail to implant74,75), represented by risk in the game. Throughout the game, players face a social dilemma where they are tempted to contribute as many cells as possible to increase their individual payoff, but where doing so risks the overall viability of the organism, and with it their payoff.Fig. 1The collective risk game of embryonic development.a We imagine embryonic development as the setting for a collective risk game, where each player is external to the embryo pot, the rounds of the game are steps in developmental time, and the target cell number is a checkpoint in development. At each round of the game, players choose between three actions: contributing progeny (C), killing cells already contributed (K), or doing nothing (N). At the end of this exact-or-nothing game, each player obtains a payoff (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π) proportional to the number of cells it contributed that survived until the end of the game. If they collectively fail to reach the target or exceed it, players receive their payoff proportional to the embryo’s survival probability, p. b Schematic of a two-player, one-round game with moderate difficulty of achieving the target and low survival probability. One example depicts a viable game (left), where both players contribute one cell to reach the target of two cells. The other example is a nonviable game (right), where only one cell is contributed, and the target is missed. c Frequency of strategies that evolve over generations of the game as a function of p. Players can play either K, C, or 2 C strategies. d Payoff matrix and simplex plot with streamlines showing the evolutionary direction of the system towards stable states in a limited strategy space for this two-player, one-round game. Blue and yellow indicate regions of low and high velocity, respectively. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing; = degree of game hardness. The target icon was taken from Flaticon.com, with attribution to Good Ware. The collective risk game of embryonic development. a We imagine embryonic development as the setting for a collective risk game, where each player is external to the embryo pot, the rounds of the game are steps in developmental time, and the target cell number is a checkpoint in development. At each round of the game, players choose between three actions: contributing progeny (C), killing cells already contributed (K), or doing nothing (N). At the end of this exact-or-nothing game, each player obtains a payoff (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π) proportional to the number of cells it contributed that survived until the end of the game. If they collectively fail to reach the target or exceed it, players receive their payoff proportional to the embryo’s survival probability, p. b Schematic of a two-player, one-round game with moderate difficulty of achieving the target and low survival probability. One example depicts a viable game (left), where both players contribute one cell to reach the target of two cells. The other example is a nonviable game (right), where only one cell is contributed, and the target is missed. c Frequency of strategies that evolve over generations of the game as a function of p. Players can play either K, C, or 2 C strategies. d Payoff matrix and simplex plot with streamlines showing the evolutionary direction of the system towards stable states in a limited strategy space for this two-player, one-round game. Blue and yellow indicate regions of low and high velocity, respectively. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing; = degree of game hardness. The target icon was taken from Flaticon.com, with attribution to Good Ware. Using this CRD embryo game, we explore the strategies that lead to viable embryos and identify conditions under which killers emerge76. The strategies define how players behave under various conditions. Beneficial strategies are expected to increase in frequency, while detrimental ones are expected to be eliminated from the system. For a single player, it is most beneficial to divide and produce many progeny, whereas the group must ensure the target size is achieved. The evolutionary dynamics of the model allow change over time, where strategies fluctuate in frequency owing to evolutionary pressures, exploration in strategy space, and payoffs. This setup allows us to study the dynamics of cellular decision-making without assuming that players act rationally. We use this framework to understand the conditions in which killing, as opposed to proliferation or noncompetition, is a beneficial cellular strategy. We simulate pairwise games where two players meet repeatedly over developmental time (rounds of the game, r; Fig. 1a). We assume that each round corresponds to an arbitrary duration that can be calibrated to developmental time by adjusting the maximum number of cells that players can contribute in each round, mirroring the rate of cell division. For example, a round of the game can correspond to a developmental day by setting the maximum cell contribution to two cells, matching a cell cycle rate of one cell division per day. In each round, players individually choose among three actions (Fig. 1a): contributing progeny to the embryo pot (C), which represents cell proliferation; killing cells already contributed (K), which represents competitive cell killing; or doing nothing (N). A killer can only remove cells that have been contributed to the embryo pot in the preceding round(s) of the game. The killer will randomly target one of the cells, including their own. For an individual player, contributing progeny improves fitness, which is proportional to increases in their payoff. Killing, on the other hand, comes at the cost of missing an opportunity to contribute and potentially sacrificing their own progeny. For the group, contributing increases the final embryo size, whereas killing reduces it. Total current embryo pot size is calculated by taking the difference between total contributed and killed cells, and players are not included in the total cell number (Fig. 1a). The CRD allows interactions across multiple rounds, where players not only acquire their payoffs at the end of the game, but their contributions directly affect the embryo pot and thus collective success, which corresponds to a milestone when there is an embryo size checkpoint9. CRDs have a large set of Nash equilibria77, where any strategy state that allows the players to meet the target is an equilibrium, irrespective of the individual’s payoff. For instance, a target embryo size of four cells can be achieved when one player contributes three cells, and the other only contributes one cell. This is considered a stable state despite the unfair contributions between the cells. Inconveniently, as the number of rounds or actions increases, the number of strategies also increases exponentially, expanding the possible stable states of the system. To better understand how players will behave across various conditions, we will gradually study the system using replicator dynamics (see “Methods”). We define the difficulty of the game by the number of Nash equilibria present and the type of strategy coordination required. Games requiring a greater number of strategies to lead to a Nash equilibrium are defined as harder games, owing to the higher difficulty of player coordination (Supplementary Fig. S1a and “Methods”). Contributing is the best strategy in one-round games We start with a one-round game; while it is not the ideal scenario for a growing embryo, it allows us to explore how the probability of survival and target size influence the frequency of four strategies: a non-contributor (plays N); a one-cell contributor (plays C); a two-cell contributor (plays 2 C); and a killer (plays K). The one-round game is ideal to understand the intuitive aspects of the strategies in a basic scenario. Our simulations show that a two-cell contribution is a better strategy for players, because it provides the highest payoff, though their payoff reduces as the probability of survival decreases (Fig. 1c). One-cell contributors increase in frequency as the probability of survival decreases. This basic scenario demonstrates that contributing is the most beneficial, irrespective of survival probability (Fig. 1c and Supplementary Fig. S1b). A two-player, one-round game with a final target size of two cells is represented by the payoff matrix in Fig. 1d, where each entry represents a cell’s payoff after playing a particular strategy (defined by each row) against a competitor’s strategy (defined by each column). Irrespective of the other player’s strategy, a non-contributor and a killer will end up with a payoff of zero. On the other hand, contributors will receive a payoff that is proportional to their contributions. We visualise the dynamics of the game and co-existence of three strategies (K, C and 2 C) or (K, N, C) using a simplex plot, where arrows illustrate paths that strategies follow as the system evolves from various initial states (Fig. 1d). The average payoff of a one-cell contributor is Cp and 2Cp for a two-cell contributor (Fig. 1d), indicating that over-contributing is the best strategy across all low-risk scenarios and when the target is easy to achieve. When the target size is moderately difficult, where over-contributing causes failure and risks a zero payoff for everyone, one-cell contributors (fair contributors) increase in frequency (Fig. 1c and Supplementary Fig. S1c, d). On the other hand, when achieving the target size is hard, such that only one cell can contribute, most games fail (Supplementary Fig. S1c, d). Analysis of the one-round game shows that contributing is evolutionarily stable, and this holds regardless of probability of survival (Fig. 1d and Supplementary Fig. S1b) or target size (Supplementary Fig. S1c, d). While the one-round game helps us gain an intuitive understanding and establish a player’s drive to contribute, it cannot answer the question of what maintains killing in a developing multicellular system. So far, the role of killers in this one-round game is not clear because they cannot react when they play simultaneously. Even if we relax this assumption and consider a sequential game in which players take turns, making killing possible following a contribution in the same round (see alternative game in Supplementary Fig. S2a), we observe that killers only slightly increase in frequency when the probability of survival is low or the target size is hard to achieve (Supplementary Fig. S2b). Furthermore, embryonic development proceeds across time and the one-round game does not capture this; with time, we can explore when killing versus contribution is more beneficial. Will killers emerge with the inclusion of time, and if killers are beneficial, when should such a strategy be used? To answer these questions, we turn to a two-round game. Time allows killing to emerge when targets are hard to achieve The CRD is a muti-round game that can be used to explore the timing of strategies. We set up a two-player, two-round game where a player can choose among three actions in each round (C, K, or N). For example, players can choose to contribute in both rounds (C C) or contribute in the first and kill in the second (C K). An example of a game with these strategies is shown in Fig. 2a. In such a two-round game there are six possible strategies (Fig. 2b): over-contributors that contribute in both rounds (C C); fair-contributors that contribute in one round and do nothing in the other, including early contributors (C N) and late contributors (N C); killers that contribute (C K); killers that do not contribute (N K); as well as non-contributors (N N). Note that players cannot kill in the first round of the game since the embryo pot is empty, as was seen in a one-round game. The payoff matrix for this game (Fig. 2b) suggests that players fare best when they contribute all the time (C C) and the worst when they choose either an (N K) or (N N) strategy. It is also clear that (C K) is only a good strategy if other players choose (C C) or (C N), as shown in Supplementary Fig. S3a, b. As expected from the payoff matrix, simulations show that over-contributors (C C) are most successful when the probability of survival is high (Fig. 2c). On the other hand, as the probability of survival decreases, we see the emergence of fair-contributors and killers. This means that killers have a role to play in such a game. Specifically, killers that proliferate in the first round (C K) are more abundant than non-contributors (N K). All combinations of (C K) / (N C), (C K) / (C N), (N K) / (N C), and (N K) / (C N) are Nash equilibria when the target is hard to achieve. However, evolutionary dynamics show that they are not equally successful with respect to their average payoffs (Fig. 2b, c and Supplementary Fig. S3b, c). A wait-and-see strategy (N C) is common in low-survival (high-risk) scenarios77 and seems to be good for evading killers. However, it is unclear if such a strategy makes sense for a developing system.Fig. 2Killing emerges in two-round games when the target is difficult to achieve.a Schematic of a two-player, two-round game with moderate difficulty of achieving the target and low survival probability (p). A viable game is depicted, where both players contribute in round one, and one player contributes while the other kills in round two to reach the target of two cells. b Payoff (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π) matrix of the game. c Frequency of strategies that evolve over generations of the game as a function of p, comparing a hard and moderate game. d, e Payoff matrix and simplex plot with streamlines showing the evolutionary direction of the system towards stable states in a limited strategy space and varied target size. Target size difficulty is changed between conditions shown, and the strategy of early (d) versus late (e) contribution is compared. Blue and yellow indicate regions of low and high velocity, respectively. Abbreviations indicate: C = contribute, K = kill, N = do nothing. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing; = degree of game hardness. The target icon was taken from Flaticon.com, with attribution to Good Ware. Killing emerges in two-round games when the target is difficult to achieve. a Schematic of a two-player, two-round game with moderate difficulty of achieving the target and low survival probability (p). A viable game is depicted, where both players contribute in round one, and one player contributes while the other kills in round two to reach the target of two cells. b Payoff (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π) matrix of the game. c Frequency of strategies that evolve over generations of the game as a function of p, comparing a hard and moderate game. d, e Payoff matrix and simplex plot with streamlines showing the evolutionary direction of the system towards stable states in a limited strategy space and varied target size. Target size difficulty is changed between conditions shown, and the strategy of early (d) versus late (e) contribution is compared. Blue and yellow indicate regions of low and high velocity, respectively. Abbreviations indicate: C = contribute, K = kill, N = do nothing. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing; = degree of game hardness. The target icon was taken from Flaticon.com, with attribution to Good Ware. To understand how players coordinate their efforts and identify the conditions which promote co-existence in viable embryos, we next run simulations among three-strategy systems under many initial conditions. A game with such few players is ideal since the early days in development involve a limited number of embryonic cells78. We use this limited strategy space to understand the possible outcomes in a small population setting. In the low-survival scenario, killers emerge in the presence of over-contributors when the target is small, making it hard to achieve (Fig. 2d, e and Supplementary Fig. S3a, b). This leads killers to co-exist with either fair- or over-contributors, depending on the target size (Fig. 2d, e). Killers also indirectly influence the timing with which cellular contributions are most beneficial to the individual. Amongst the fair-contributors, the late-contributors (N C) achieve a higher payoff than those that contribute in the first round (C N). Contributing late allows the player to make its contribution while evading the possibility that killers will eliminate their progeny. Both combinations of (C K) / (N C) and (C K) / (C N) are Nash equilibria when the target is one cell. However, exploring the dynamics of a low-survival (high-risk) game where the target is hard (one cell), we see that (C K) / (C N) co-exist in a stable state (Fig. 2d). For the same parameter set, the combination of (C K) / (N C) does not co-exist (Fig. 2e). When killers interact with players that contribute early (C N), this splits their own losses by reducing the chance that their progeny will be removed. On the other hand, interactions with a late contributor (N C) result in a payoff of zero for the killer. As a result, we find that killers co-exist with early contributors and not with late contributors, driving the timing of contributions to the early stages of the game and killing to the later rounds (Fig. 2d, e and Supplementary Fig. S3a, b). In a similar manner, we find that non-contributing killers do not co-exist with over-contributors as they have no chance to earn a payoff (Supplementary Fig. S3c). Overall, these simulations reveal that killing can emerge in systems that rely on collective coordination across time, such as early embryonic development. One-round games do not yield killers, suggesting that killing is a function of developmental time. Killing emerges as an evolutionary stable strategy when the probability of embryo survival is low, and over-contribution is possible. Therefore, target size shapes strategies, where the abundance of killers increases as the target size is decreased relative to the possible contributions of players, making them more difficult to achieve. We also find that the timing of contributions influences how players coordinate. Contributing early ensures that killers can fix any over-contributions in later rounds. Early contributors share the potential losses with late killers. Human embryoids reveal a relationship among cell death, target size, and successful cavity formation Our CRD embryo model predicts a role for killing to support viability in embryo development. While this is aligned with studies of mouse9,25, zebrafish27, and fruit fly29 development, little is known about the role of cell death in early human development. To explore this, we use the heX-embryoid to model the post-implantation embryo41 (Fig. 3a). Post-implantation development in humans is a key period, as embryos must meet size and morphological milestones in advance of the onset of gastrulation, where the body plan will be patterned. We leverage the published timecourse single-cell (sc)RNA-seq dataset of heX-embryoids41 to check for heterogeneous expression of fitness-associated genes, which would support the possibility that cell competition can occur in this embryoid model. As before41, clustering of the scRNA-seq data reveals an epiblast cell population before (day 0-2) and after (day 3-5) lumen formation (Supplementary Fig. S4a), which exhibits strong expression of pluripotency markers (NANOG, POU5F1, and MYC; Supplementary Fig. S4b). We also observe amnion- (expressing ISL1 and TFAP2A, Supplementary Fig. S4c) and posterior-like fates (expressing MIXL1 and TBXT, Supplementary Fig. S4d), which segregate to opposite ends of the clustering. We use a hypergeometric test to assess whether orthologous mouse winner/loser genes79 appear as differentially expressed across our heX-embryoid cell clusters (Supplementary Fig. S4e). We find that the early (day 0-2) epiblast has a high winner-associated gene signature (p < 1 x 10−150). Following lumen formation, divergent winner/loser signatures emerge in the posterior- and amnion-like compartment, with posterior-like cells adopting a winner state (p < 1 x 10−30) and amnion-like cells adopting a loser state (p < 1 x 10−30; Supplementary Fig. S4e). We observe a global enrichment for differentially expressed winner genes in the posterior-like cell cluster, while that of loser genes is seen in the amnion-like cell cluster (Supplementary Fig. S4f and Supplementary Data 1). Co-expression analysis of top differentially expressed winner/loser-associated genes further illustrates that MIXL1-expressing (posterior-like) cells express winner genes (Supplementary Fig. S5a) while ISL1-expressing (amnion-like) cells express higher levels of loser genes (Supplementary Fig. S5b). Consistent with this, the posterior-like cells are also enriched for the expression of MYC, a well-known marker of winner pluripotent state20,80,81. These results support the possibility that cell competition may be conserved in the heX-embryoid, an in vitro model of early human development. Further, they suggest that the extraembryonic amniotic cells may be more likely to be eliminated than embryonic cells, particularly the posterior cells that give rise to the primitive streak and later undergo gastrulation to pattern the embryo.Fig. 3In vitro heX-embryoid model of human post-implantation development links target size and viability.a The heX-embryoid model recapitulates key aspects of the human post-implantation embryo, including amniotic cavity formation. b For simplicity, we consider embryoids that form a single lumen to be successful (viable), while those containing multiple lumens have a defect and are considered unsuccessful (nonviable). c heX-embryoid size is related to viability, where viable heX-embryoids are within a target size range that corresponds to the size of the bona fide human embryonic disc at days 8-9 post-fertilisation (grey) and nonviable heX-embryoids are larger than this range. Box-plot reports median as centre line, upper and lower quartiles and error bars indicate standard error. Exact p-values: single lumen control versus caspase inhibitor, p = 0.017; multi-lumen control versus caspase inhibitor, p = 0.001604; control single versus multi-lumen, p = 6.68 × 10−8; caspase inhibitor single versus multi-lumen, p = 1.56 × 10−10. d, Example confocal images of heX-embryoids (taken from n = 3 biological replicates) below, within and above the target size range. e Frequency of viable and nonviable heX-embryoids below, within and above the target size range in control and caspase-inhibited (with Z-VAD-FMK) conditions. Exact p-value: p = 0.0063. f Total frequency of viable and nonviable heX-embryoids in control and caspase-inhibited conditions. Exact p-values: control above versus within, p = 1.33 × 10−12; control above versus below, p = 2.97 × 10−6; caspase inhibitor above versus within, p = 1.56 ×10−7; caspase inhibitor above versus below, p = 1.66 × 10−4; control versus caspase inhibitor within, p = 5.53 × 10−3. Data shown in (c, e, f) includes: in “control”, n = 215 (single lumen), n = 92 (multi-lumen); and in “caspase inhibitor”, n = 143 (single lumen), n = 92 (multi-lumen). Data from the heX-embryoids is combined across n = 3 biological replicates. Pairwise comparisons were made using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons (c) and chi-squared tests of independence or Fisher’s exact test with Holm-Bonferroni correction for multiple comparisons (e, f). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Source data are provided as a Source Data file. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing. The target icon was taken from Flaticon.com, with attribution to Good Ware. In vitro heX-embryoid model of human post-implantation development links target size and viability. a The heX-embryoid model recapitulates key aspects of the human post-implantation embryo, including amniotic cavity formation. b For simplicity, we consider embryoids that form a single lumen to be successful (viable), while those containing multiple lumens have a defect and are considered unsuccessful (nonviable). c heX-embryoid size is related to viability, where viable heX-embryoids are within a target size range that corresponds to the size of the bona fide human embryonic disc at days 8-9 post-fertilisation (grey) and nonviable heX-embryoids are larger than this range. Box-plot reports median as centre line, upper and lower quartiles and error bars indicate standard error. Exact p-values: single lumen control versus caspase inhibitor, p = 0.017; multi-lumen control versus caspase inhibitor, p = 0.001604; control single versus multi-lumen, p = 6.68 × 10−8; caspase inhibitor single versus multi-lumen, p = 1.56 × 10−10. d, Example confocal images of heX-embryoids (taken from n = 3 biological replicates) below, within and above the target size range. e Frequency of viable and nonviable heX-embryoids below, within and above the target size range in control and caspase-inhibited (with Z-VAD-FMK) conditions. Exact p-value: p = 0.0063. f Total frequency of viable and nonviable heX-embryoids in control and caspase-inhibited conditions. Exact p-values: control above versus within, p = 1.33 × 10−12; control above versus below, p = 2.97 × 10−6; caspase inhibitor above versus within, p = 1.56 ×10−7; caspase inhibitor above versus below, p = 1.66 × 10−4; control versus caspase inhibitor within, p = 5.53 × 10−3. Data shown in (c, e, f) includes: in “control”, n = 215 (single lumen), n = 92 (multi-lumen); and in “caspase inhibitor”, n = 143 (single lumen), n = 92 (multi-lumen). Data from the heX-embryoids is combined across n = 3 biological replicates. Pairwise comparisons were made using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons (c) and chi-squared tests of independence or Fisher’s exact test with Holm-Bonferroni correction for multiple comparisons (e, f). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Source data are provided as a Source Data file. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing. The target icon was taken from Flaticon.com, with attribution to Good Ware. Establishing that competition is possible enables us to explore the purpose of cell death. We next ask whether cell death is involved in generating embryoids with correct amniotic cavity formation. Consistent with established size checkpoints in mouse development7–9, and with the demonstrated role of proamniotic cavity formation in regulating downstream embryonic patterning82, we define single lumen formation as a proxy for viability in a singleton pregnancy. Multi-lumen formation reflects failure to achieve the correct morphogenetic outcome at this size checkpoint, analogous to defects observed in oversized mouse embryos9. This definition is consistent with viability metrics used in other human post-implantation stem cell models38,42. Thus, we consider embryoids that form a single lumen to be successful (which we define as “viable” structures), while those containing multiple lumens are considered to be unsuccessful (defined as “nonviable” structures; Fig. 3b). The high-throughput nature of the heX-embryoid platform offers an opportunity to explore how variability in early embryo size impacts cell behaviour and vice versa. Aligned with the EGT model, embryoid size is defined by the number of cells in the structure. We use the projected two-dimensional area of heX-embryoids as an accessible proxy for measuring embryoid size in our high-throughput studies, owing to the fact that it is a well-correlated metric for total cell number (Supplementary Fig. S6a). The stochasticity of hiPSC cluster formation results in a large degree of variability in embryoid sizes (Supplementary Fig. S6b). This reveals a relationship between size and single-lumen viability. Intriguingly, the target size range that supports viability corresponds to the size of the human embryo at the stage of lumen formation (Fig. 3c, control experiment; see “Methods”)83. Nonviable embryoids are generally larger than the human embryo size range (Fig. 3c, d, control experiment). To indirectly explore the effect of killing, we inhibit cell death by treating the heX-embryoid culture with a pan-caspase inhibitor (Z-VAD-FMK). Caspase inhibition does not impact the ability to form lumens (Fig. 3e and Supplementary Fig. S6b, c), but it causes an overall increase in embryoid size, pushing more structures outside of the target size range compared to control (Fig. 3c, e, f and Supplementary Fig. S6d). As a result, a greater proportion of embryoids are considered nonviable following caspase inhibition (Fig. 3e, f and Supplementary Fig. S6e). However, caspase inhibition does not affect the proportion of structures that form a single lumen for structures within the target size range (Fig. 3f and Supplementary Fig. S6d), in line with a previous report in mouse9. Overall, these results suggest a relationship between cell death, embryoid size, fates and amniotic cavity formation (used here as a proxy for viability). Furthermore, the increase in nonviable structures when death is inhibited hints that it is a responsive behaviour. Killing and contributing are conditional strategies To explore how responsive behaviour influences the maintenance of killing, we introduce conditional strategies to the game. So far, we considered “unconditional” strategies; these are predetermined and do not allow players to respond to the behaviours of others. Conditional strategies, on the other hand, capture the plastic nature of cell behaviour. Cellular plasticity is inherent to mammalian development, where embryonic cells make different fate decisions in response to cues in their microenvironment84. In our case, we want to capture cells that decide to contribute or kill in each round in response to embryoid size. We model a conditional strategy by allowing a player to choose between two actions, a or b, in each round of the game by comparing the total embryoid size against a threshold77, denoted (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b). If the size is below \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ, the player chooses action a, otherwise, it chooses action b (see example in Fig. 4a). The strategy space increases with this setup; for example, the two-round game we presented had 6 unconditional strategies, but the new setup generates 81 possible strategies (see “Methods”).Fig. 4Killing is a condition strategy for embryo size correction.a Schematic of a two-player, two-round CRD embryo game with conditional strategies, high difficulty of achieving the target and low survival probability (p). Player 2 plays the same conditional strategy in both games (player 2 versus player 1 and player 2 versus player 3). In the first game (left), player 2 contributes in round 1 and kills in round 2 because the total cell number exceeds a threshold of one after round 1. In the second game (right), player 2 contributes in both rounds because player 3 does not contribute, and the total cell number does exceed the threshold after round 1. b Frequency of strategies that evolve over generations of the game as a function of p for hard and moderate difficulty. c Frequency of cell death (cleaved caspase-3 positive cells) in heX-embryoids that are below, within or above the target size range. Data shown includes: in “control” n = 55 (below), n = 175 (within), n = 129 (above); and in “caspase inhibitor” n = 22 (below), n = 117 (within), n = 136 (above). Data from the heX-embryoids is combined across n = 2 biological replicates. Exact p-values: control below versus above, p = 0.000416; control below versus within, p = 0.000416; control versus caspase inhibitor below, p = 2.07 × 10−110; control versus caspase inhibitor within, p = 1.23 × 10−176; control versus caspase inhibitor above, p = 5.64 × 10−187. d Degree of cell proliferation (Ki67 intensity/area) in heX-embryoids that are below, within or above the target size range. Data shown includes: in “control” n = 67 (below), n = 215 (within), n = 205 (above); and in “caspase inhibitor” n = 52 (below), n = 139 (within), n = 250 (above). Data from the heX-embryoids is combined across n = 2 biological replicates. Box-plot reports median as centre line, upper and lower quartiles and error bars indicate standard error. Exact p-values: control below versus above, p = 0.005; control within versus above, p = 0.01; caspase inhibitor below versus within, p = 3.72 × 10−6; caspase inhibitor below versus above, p = 0.007; caspase inhibitor within versus above, p = 0.002; control versus caspase inhibitor below, p = 1.98 × 10−25; control versus caspase inhibitor within, p = 3.75 × 10−40; control versus caspase inhibitor above, p = 8.73 × 10−50. Pairwise comparisons were made using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons in (c, d). **p < 0.01, ****p < 0.0001. e, f Payoff matrix and simplex plot with streamlines showing the evolutionary direction of the system towards stable states in a limited strategy space. The target size difficulty is changed between conditions shown. Blue and yellow indicate regions of low and high velocity, respectively. Source data are provided as a Source Data file. Abbreviations indicate: C = contribute, K = kill, N = do nothing. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing; = degree of game hardness. The target icon was taken from Flaticon.com, with attribution to Good Ware. Killing is a condition strategy for embryo size correction. a Schematic of a two-player, two-round CRD embryo game with conditional strategies, high difficulty of achieving the target and low survival probability (p). Player 2 plays the same conditional strategy in both games (player 2 versus player 1 and player 2 versus player 3). In the first game (left), player 2 contributes in round 1 and kills in round 2 because the total cell number exceeds a threshold of one after round 1. In the second game (right), player 2 contributes in both rounds because player 3 does not contribute, and the total cell number does exceed the threshold after round 1. b Frequency of strategies that evolve over generations of the game as a function of p for hard and moderate difficulty. c Frequency of cell death (cleaved caspase-3 positive cells) in heX-embryoids that are below, within or above the target size range. Data shown includes: in “control” n = 55 (below), n = 175 (within), n = 129 (above); and in “caspase inhibitor” n = 22 (below), n = 117 (within), n = 136 (above). Data from the heX-embryoids is combined across n = 2 biological replicates. Exact p-values: control below versus above, p = 0.000416; control below versus within, p = 0.000416; control versus caspase inhibitor below, p = 2.07 × 10−110; control versus caspase inhibitor within, p = 1.23 × 10−176; control versus caspase inhibitor above, p = 5.64 × 10−187. d Degree of cell proliferation (Ki67 intensity/area) in heX-embryoids that are below, within or above the target size range. Data shown includes: in “control” n = 67 (below), n = 215 (within), n = 205 (above); and in “caspase inhibitor” n = 52 (below), n = 139 (within), n = 250 (above). Data from the heX-embryoids is combined across n = 2 biological replicates. Box-plot reports median as centre line, upper and lower quartiles and error bars indicate standard error. Exact p-values: control below versus above, p = 0.005; control within versus above, p = 0.01; caspase inhibitor below versus within, p = 3.72 × 10−6; caspase inhibitor below versus above, p = 0.007; caspase inhibitor within versus above, p = 0.002; control versus caspase inhibitor below, p = 1.98 × 10−25; control versus caspase inhibitor within, p = 3.75 × 10−40; control versus caspase inhibitor above, p = 8.73 × 10−50. Pairwise comparisons were made using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons in (c, d). **p < 0.01, ****p < 0.0001. e, f Payoff matrix and simplex plot with streamlines showing the evolutionary direction of the system towards stable states in a limited strategy space. The target size difficulty is changed between conditions shown. Blue and yellow indicate regions of low and high velocity, respectively. Source data are provided as a Source Data file. Abbreviations indicate: C = contribute, K = kill, N = do nothing. Symbols indicate: = target size; = viable, = nonviable; = contribute, = kill, = do nothing; = degree of game hardness. The target icon was taken from Flaticon.com, with attribution to Good Ware. Overall, conditional games yield the same trends: killers emerge when the probability of survival is low, and the target is hard to achieve (Fig. 4b). However, conditional killers are more abundant than unconditional killers and may be either over- (C C|C K) or fair contributors (C N|C K) themselves. Conditional over-contribution (C C|C K) is a strategy that over-contributes in one condition and kills in another, whereas conditional fair-contributors (C N|C K) contribute early in all situations and only kill in some. In addition to conditional killers, we see the emergence of conditional contributors (C C|C N), which switch between over and fair contribution depending on embryo size. In unconditional games, when the target is hard to achieve, and survival is low, we see late contributors in high abundance; however, in conditional games for the same parameter set, we see an increase in early compared to late contributors. Thus, our model predicts that conditional behaviours (both killing and contributing) are beneficial in reaching the correct target size for viability. We also find that conditional killers are more abundant than their unconditional counterparts when the target is hard to achieve. These predictions are further supported by results from the heX-embryoids. Structures below the target size range contain lower proportions of dying (cleaved caspase-3 positive) cells (Fig. 4c, Supplementary Figs. S7–10 and Supplementary Movie S1–4). This suggests that cell death is a conditioned response to embryoid size. Analogously, cell proliferation (marked by Ki67) decreases with increasing heX-embryoid size, with reduced levels in structures that are above the target size range (Fig. 4d, Supplementary Figs. S7–10 and Supplementary Movie S1–4). Proliferation is lower following caspase inhibition (Fig. 4d, Supplementary Figs. S11–13 and Supplementary Movie S5–7), suggesting that this is also a conditioned response and may help support size correction, consistent with our model and a previous report in mouse8. Further, confocal imaging results suggest an enrichment of dying cells between separated lumens. These lumen-associated areas of cell death are concentrated close to the bottom of the structures, where amnion is found to be specified in heX-embryoids41 (Supplementary Figs. S8, S10). This aligns with the scRNA-seq results, and further suggests that extraembryonic amnion cells may preferentially be sacrificed to support viability by merging lumens. What does this mean? Is killing needed or not? Both models suggest killing is beneficial when targets are hard. However, it is clear that players have more than one strategy available to them, and these divergent strategies may be present at the same time. To understand which strategies can coordinate or co-exist, and since the embryoids, like real embryos, have a limited number of initial cells, we run simulations for three-strategy systems. We consider the case where killing cannot happen and find that target size drives whether conditional contributors co-exist with fair and over-contributors (Fig. 4e). On the other hand, conditional killers (C C|C K) co-exist with conditional (C C|C N) or unconditional over-contributors (C C) for a hard target (Fig. 4f). Increasing target size reduces the number of killers, leaving conditional as well as fair-contributors to support correct embryo size. These conditional killers may be either over- or fair-contributors themselves (Fig. 4f and Supplementary Fig. S14). But when we consider successful games, conditional killers (C C|C K) fair better with conditional contributors (C C|C N) (Fig. 4f). In line with this idea, the lower rates of proliferation under caspase inhibition in heX-embryoids suggest that cells are responding to their environment and implies that conditional strategies may co-exist. While the embryoid model cannot explore a mixed strategy space, it can show that in the absence of killing, conditional proliferators may still switch their behaviour from (C C) to (C N), resulting in lower overall proliferation. These observations suggest that cells interact in a heterogeneous system, where behaviours change depending on the player and its circumstances. Altogether, the predictions of our CRD embryo model and data from human embryoids suggest that a cell can respond as a killer or a proliferator. They also suggest that killing is maintained in these complex systems because it is beneficial. Killing is useful when the number of cells exceeds the target size. However, the model also suggests that the threat of killing (being a conditional killer) is enough to deter the emergence of over-contributors, which exist at a lower frequency compared to games containing unconditional killers (Fig. 4b versus 2b). This is observed even as the target size is increased, where maintaining the presence of killers in the system limits over-contributors to ensure cooperation. Discussion Cell-cell interactions are essential for coordinating the diverse cell types emerging during development, facilitating embryo growth, differentiation, tissue formation, and organ development. Despite their essential role in ensuring proper function and survival, most cell behaviours are technically challenging to probe, leaving them hidden. We often assume that normal cells are exclusively cooperative in healthy tissues, whereas defecting cells emerge only in disease85. However, the newly observed phenomenon of “cell competition”, in which cells engage in contact-mediated killing of neighbours in the same tissue, urges us to reconsider this paradigm. The first example of cell competition was reported in the Drosophila embryo86. The last decade has uncovered instances of competition in mouse18–20,25,80,81,87–91 and zebrafish embryos27,92 as well as cultured human embryonic stem cells93,94, revealing that killing is a conserved behaviour across species. Emerging examples have suggested a role for cell competition as a mechanism to remove unwanted cells during Drosophila wing development95–98, incorrectly localised cells during tissue patterning in zebrafish development27,92, cells stuck between lineages in early mouse embryonic development89, and low-quality cells in mouse epidermis25. It has been shown that cell competition optimises mouse skin fitness25 and elongates adult Drosophila lifespan29, implicating killing as a mechanism to remove undesirable cells in tissue development and homoeostasis34. However, few studies to date have explored the link between cell killing and organism success. Why is killing maintained in a developing system that may depend on cooperative rather than competitive behaviour? Cooperation is expected to increase survival, as it allows groups to work together for mutual benefit, as in the embryo. How do aggressive behaviours, such as competitive killing, benefit the collective? Is killing random or strategic? To answer these questions, we need to capture information flow across biological scales, linking cell behaviour to multicellular dynamics and organism outcomes. Here, we combine an in silico EGT model (the CRD embryo game) and an in vitro human embryo model (the heX-embryoid41) to explore the basis of killing in a broad context, as well as its relevance to human development (Fig. 5). We find that killing is beneficial for the good of the collective embryo but costly for the individual player. While killing is a behaviour typically associated with destruction or defection, our analysis suggests that it can be beneficial and strategic. Killers reduce their own payoff for the benefit of the collective, removing extra cells to hit the target size. We predict killing is beneficial when the stakes are high and the target size is limited or hard to achieve relative to the proliferation of cells. The insight that an individually costly behaviour (cell killing) can be favoured by evolutionary selection when it benefits the collective is a signature prediction of game-theoretic models and would not readily emerge from statistical analyses of cell behaviour alone. Thus, our findings suggest that although counter-intuitive, competition underpins cooperation amongst cells for embryonic success.Fig. 5A dual EGT and embryoid modelling framework to link cellular behaviours to embryo outcomes.Schematic summary of the combined EGT and embryoid modelling framework, mapping underlying cellular behaviours to embryo viability. The model is used to uncover the role of cell killing in removing extra cells, pushing embryo size to within a target size to maximise the probability of survival. This framework not only addresses fundamental open questions in developmental biology but can provide insights for embryo selection for assisted reproduction. Symbols indicate: = viable, = nonviable. A dual EGT and embryoid modelling framework to link cellular behaviours to embryo outcomes. Schematic summary of the combined EGT and embryoid modelling framework, mapping underlying cellular behaviours to embryo viability. The model is used to uncover the role of cell killing in removing extra cells, pushing embryo size to within a target size to maximise the probability of survival. This framework not only addresses fundamental open questions in developmental biology but can provide insights for embryo selection for assisted reproduction. Symbols indicate: = viable, = nonviable. Our CRD embryo model not only shows that killing is beneficial but that it is responsive, allowing cells to switch between behaviours: killing when there are too many cells and contributing when there are too few. Mirroring this observation, in the heX-embryoid we find that cell proliferation and death are responsive; embryoids below the expected size of the native embryo83 contain lower proportions of dying cells, and proliferation decreases with increasing embryoid size. The observation that proliferation is also lowered following caspase inhibition is consistent with this interpretation because in the absence of killing, cells are forced to decrease their contributions to hit the target. These results suggest that cells are responding to their microenvironment, switching their behaviours as needed. Responsive behaviours provide the embryo with adaptiveness and resilience to error, adjusting strategies to balance contribution and killing to hit the target. Aligned with our results, early landmark studies involving over-sized mouse embryos revealed that cells exhibit responsive behaviours: conditional proliferators contribute fewer progeny when embryo size is too large shortly after implantation to the uterine wall7,8. Responsive behaviours capture phenotypic switching, a phenomenon well-known in the context of adaptive strategies in ecology99 and cancer100,101. These behaviours allow for quick cell responses to the microenvironment and are ideal in uncertain conditions99,102, making their role in development unsurprising103,104. Our results suggest that a responsive strategy is beneficial during the early stages of development when cells are few, and the viability of the embryo depends on their interactions. Our results also shed new light on early human development, which has largely remained a black box. We observe that embryoids closely mirror the size of the native human embryonic disc at the stage of amniotic cavity formation, and those that hit this target size are able to form single lumen structures. When cell death is inhibited, embryoids increase in size. Lumen formation is known to be a size checkpoint in mice as it is sensitive to cell numbers and exhibits concentrated cell death8,21,82. Double-sized mouse embryos exhibit defects in lumen formation, producing multiple rather than single lumens9. Analogously, multi-lumen formation increases in human embryoids that exceed their expected size range, suggesting a link among cell death, target size and embryoid success. The molecular mechanisms that mediate cell killing remain largely unknown, particularly in the context of human development, where studies of cell competition remain in their infancy93. As a result, our ability to distinguish the various cellular strategies that drive death (self-killing or competitive killing) is technically challenging. Nevertheless, our results provide the first link between cell death and collective success in human development. Leveraging these embryoids, we find that proliferation is reduced when death is inhibited to regulate embryoid size and facilitate successful lumenogenesis. Analogously, proliferation is reduced in larger embryoids. This suggests that cell proliferation and death are cellular behaviours used to regulate embryo size in human embryoid development, a phenomenon that has been reported in mouse7–9. To our knowledge, this is the first report of the conserved nature of this phenomenon in the context of human development. Further, we uncover molecular signatures similar to mouse epiblast winner fate in heX-embryoids. Though we do not directly assess the presence of contact-mediated elimination of losers in heX-embryoids, these observations support the possibility that cell competition may be conserved in early human development. Together with confocal imaging analysis, we find an enrichment of dying cells within the lower cell layer (the putative amnion compartment), particularly at the junctions between lumens in multi-lumen embryoids. This suggests that the embryo may preferentially sacrifice extraembryonic cells during error-correction to promote viability while preserving the fittest cells for the posterior axis, which ultimately undergoes gastrulation to give rise to the body axis. Alternatively, less fit cells may show a bias towards an amniotic fate with a higher susceptibility for subsequent elimination. Future work can untangle the underlying mechanisms at play in mediating this process. Classical examples of cell competition, such as in the Drosophila wing disc95–98, identify the trigger for killing as a relative fitness difference between neighbouring cells. Our model suggests a distinct but complementary mechanism: killing can also be triggered by a collective property (deviation from a target size). The emergence of killing in largely isogenic heX-embryoids supports this interpretation, pointing to size-based regulation as a potentially distinct mode of cell competition that may be particularly relevant to early embryogenesis. We also predict an intriguing effect that results from the responsive nature of killing: the presence of conditional killers protects the embryo from the emergence of over-contributors. The “threat” of killing is also hypothesised to occur in cancer and other ecological niches – cells behave “properly” to avoid retaliation105–107. Cancer cells exploit other cells to do their bidding108,109. Models that explore manipulative interactions propose cyclical dynamics that lead to compliant behaviours110. Whether conditional strategies enforce cooperative proliferation is beyond the scope of our model system. Future work can disentangle the beneficial role of killing and whether killers remain in the body, asking if these strategies provide another safeguarding mechanism to enforce cooperation, as in the specialised role of the immune system. Although the EGT and heX-embryoid models intersected here have provided valuable insights into the beneficial role of killing for embryo viability, they also have several important limitations to consider. First, the CRD embryo game introduced here is intentionally abstracted and restricted to relatively simple games (two players, up to two rounds of play, a restricted strategy space). It captures the evolutionary logic of killing and contributing but does not incorporate other molecular or cellular mechanisms at play in embryogenesis (spatial organisation, signalling gradients, mechanical forces that govern lumen formation, divergent cell states, etc). Future iterations of the CRD embryo model that integrate these features may reveal additional constraints on when killing strategies are favoured, as well as the underlying biological mechanisms that drive these behaviours. Follow-up studies can explore more complex games with biologically relevant parameters, such as larger groups of players, longer game durations, and more complex strategy spaces that more closely recapitulate the growth/death rates of cells in the embryo and capture the role of different cell types in the system. Second, our heX-embryoid data cannot currently distinguish between apoptosis driven by competitive cell interactions (contact-mediated killing) and cell-autonomous death triggered by intrinsic stress or developmental signals. The molecular mechanisms underlying the cell death we observe in heX-embryoids remain to be fully characterised. Third, while the correlation between posterior fate and winner-like transcriptional signatures is intriguing, we cannot exclude the possibility that this reflects differential transcriptional programmes associated with lineage specification rather than competitive fitness. Finally, heX-embryoids, while recapitulating key aspects of early human development, are an in vitro model. The extent to which these findings translate to native human embryogenesis will require further exploration, as will the requirement to identify the timing and context in which competition may occur in development. In providing a predictive link between cell behaviours that are necessary for the formation of viable embryos, our joint modelling scheme can be leveraged to inform in vitro fertilisation (IVF) screening practices. Specifically, our results point to early proliferation as a beneficial cell behaviour for achieving a viable target size. This is mirrored in embryo grading practices in IVF clinics, where the cell cycle dynamics of early embryonic cells are used to establish morphokinetic metrics111,112. Healthy embryos with high implantation potential are identified based on faster cleavage from the 2- to 8-cell stage and can be predicted in the first 48 hours of imaging, where successful embryos experience a short duration of the first few divisions113. This has also been suggested to serve as a proxy for identifying genetically mosaic embryos that contain aneuploidic cells, where these embryos reportedly display prolonged or irregular cell cycle transitions114,115 and time to reach requisite cell numbers to enter the blastocyst stage116. Building on these existing best practices, our model provides insights into how early cell-cell interactions shape embryo viability by regulating size, providing new insights for IVF best practices. For example, we predict that responsive proliferators and killers are the drivers of embryo size correction. Markers for these subpopulations could serve as indicators for embryo robustness to error, including deviations to embryo size but perhaps also the emergence of genetically aberrant cells97,117 and abnormal morphogenesis9. These possibilities could seed follow-up studies to explore the survival probability of human embryos for IVF, exploring the effect of a larger strategy space that captures biologically relevant embryo grading metrics, such as proliferation rates. Further, while we modelled an “exact-or-nothing” game that assumes viability in a narrow range of target size, data on size variation in IVF embryos can be used to explore the impact of survival probability functions53 on cell behaviour. By bridging mathematical models and engineered stem cell models, we can unlock a deeper understanding of multicellular systems, including the native embryo, embryoids, and organoids. We can also catalyse the development of predictive models that link hidden cellular behaviours with emergent properties of multicellular populations. By capturing information flow across biological scales, we can simulate engineered cellular systems, accelerating discovery and therapeutic development. Methods Collective risk dilemma of embryonic development We model embryonic development as a collective risk dilemma (CRD) with multiple rounds (Fig. 1a). Multiple rounds mean players interact multiple times before the final payoff is calculated. Players can have a different action for each round. Their strategy is predefined at the beginning of the game. There are three basic actions: contributing progeny to the embryo pot (C), killing cells contributed to the embryo pot in previous rounds (K), or doing nothing (N). Only one action can be used in a given round.Contributing: the action of adding cells to the embryo pot. C represents one cell. It increases the payoff by 1.Killing: the action of removing a random cell from the embryo pot. Killing does not occur if the embryo pot is empty. Killing decreases the payoff of the player who lost a cell by 1.Doing nothing: the action of choosing to skip the round. The player misses a chance to increase their payoff or decrease someone else’s. Doing nothing can be altruistic, as it may indirectly help ensure the embryo does not exceed the target. Contributing: the action of adding cells to the embryo pot. C represents one cell. It increases the payoff by 1. Contributing: the action of adding cells to the embryo pot. C represents one cell. It increases the payoff by 1. Killing: the action of removing a random cell from the embryo pot. Killing does not occur if the embryo pot is empty. Killing decreases the payoff of the player who lost a cell by 1. Killing: the action of removing a random cell from the embryo pot. Killing does not occur if the embryo pot is empty. Killing decreases the payoff of the player who lost a cell by 1. Doing nothing: the action of choosing to skip the round. The player misses a chance to increase their payoff or decrease someone else’s. Doing nothing can be altruistic, as it may indirectly help ensure the embryo does not exceed the target. Doing nothing: the action of choosing to skip the round. The player misses a chance to increase their payoff or decrease someone else’s. Doing nothing can be altruistic, as it may indirectly help ensure the embryo does not exceed the target. The embryo is empty at the beginning of each game. Players contribute to the embryo but are not part of the embryo. The target is defined as the total number of cells in the embryo pot needed to be achieved at the end of the game, after r rounds. If the players collectively reach this set target size, all players obtain their individual payoff (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π). When the size of the embryo fails to reach the target or exceeds it, all players keep their payoff with a probability, p (Fig. 1b). This corresponds to the probability of embryo survival, which determines the outcome of the collective behaviour at the end of the game with an “exact-or-nothing” probability function:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P\left[p,{target},{embryo}\,{pot}\,{size}\right]=\left\{\begin{array}{c}{embryo}\,{pot}\,{size}-{target}=0\to 1\\ {embryo}\,{pot}\,{size}-{target}\ne 0\to \,p\end{array}\right.$$\end{document}Pp,target,embryopotsize=embryopotsize−target=0→1embryopotsize−target≠0→p The payoff of each player is given by:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi= \left({total}\,{cells}\,{the}\,{player}\,{contributed}-{total}\,{progeny}\,{killed}\right) \\ \times P[p,{target},{embryo}\,{pot}\,{size}]$$\end{document}π=totalcellstheplayercontributed−totalprogenykilled×P[p,target,embryopotsize] Throughout the game, players face a social dilemma where they are tempted to contribute as many cells as possible to increase their individual payoff, but where doing so risks the overall viability of the organism, and with it their payoff. Replicator dynamics To understand the dynamics of strategies in the game, we utilise replicator dynamics. This concept is based on the evolutionary principle that suggests beneficial strategies spread quickly within a population while less beneficial ones are rapidly eliminated. Fitness is calculated using the average payoff. We model an infinite population of player strategies determined by the number of rounds in the game. In a one-round game, three strategies can be employed: a non-contributor (plays N, doing nothing), a contributor (plays C, adding one cell), an over-contributor (plays 2 C, adding two cells), and a killer (plays K, removing one cell). As the number of rounds increases, the available strategies also expand as a function of the round strategies. For example, a two-player, two-round game with three actions (C, K, or N) generates 9 possible strategies since players can choose a different strategy per round (R1 R2), such as: contribute in both rounds (C C) or contribute in the first and kill in the second (C K). We describe over-contributors as players that contribute in both rounds, (C C); fair-contributors as players that contribute in one round and do nothing in the other, including early contributors (C N) and late contributors (N C); killers as those that either contribute (C K) or do not contribute (N K); and non-contributors as players that do nothing for both rounds (N N). Note that strategies where players kill in the first round of the game such as (K N), (K K) and (K C) are omitted since the killers in the first round are not able to remove cells from the empty embryo pot and thus behave equivalently to N. Two strategies are chosen from the population to play a game; each player decides whether to contribute or not to the embryo pot. The number of cells contributed to the embryo pot can grow or decrease across rounds, depending on the strategy used. In our setup, the player can contribute at most one cell in each round, which is added into the embryo pot. The number of cells contributed by each player directly influences the payoff of their strategy as well as the number of cells in the collective embryo pot (it does not influence the population size, which is fixed). Using a payoff matrix represented as (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}c, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d$$\end{document}d) in a two-player game and a two-strategy system (Strategy 1, Strategy 2), the following definitions apply: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a represents the payoff when two players of Strategy 1 interact; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b represents the payoff when a Strategy 1 player interacts with a Strategy 2 player; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}c represents the payoff when a Strategy 2 player interacts with a Strategy 1 player; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d$$\end{document}d represents the payoff when two Strategy 2 players interact. The fitness of each strategy is calculated as the average payoff for that strategy based on the distribution of the population. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document}x be the frequency of Strategy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1$$\end{document}1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1-x$$\end{document}1−x be the frequency of Strategy 2. The payoffs, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi$$\end{document}π, for the players can be defined as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\pi }_{1}={a\; x}+b\left(1-x\right)$$\end{document}π1=ax+b1−x4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\pi }_{2}={c\; x}+d(1-x)$$\end{document}π2=cx+d(1−x) This framework allows us to analyse how strategies evolve over time based on their relative payoffs. Assumptions and limitations of the model Our model operates under a set of assumptions that simplify the evolutionary dynamics of player strategies:Single action per round: In each round of the game, each player can only make a single decision to “do nothing”, “contribute” or “kill”.Competitive killing of contributed cells: Our strategy space captures killing of cells that were contributed in the previous rounds, focusing on the evolution of strategies where killing is a response to previous interactions. A player cannot eliminate themselves from the game, but rather can only eliminate cells that have been contributed to the embryo pot, including their own contributions. This may not fully represent all forms of killing; for instance, in the naive embryo, apoptotic self-killing is possible and can be triggered by genetic abnormalities or signals from the microenvironment, such as a lack of nutrients. The model presented here, on the other hand, focuses on competitive cell killing strategies, which are represented by a player killing cells present in the embryo pot, regardless of which player contributed the cell. Despite its limitations, this model allows us to investigate whether killing can evolve in a multicellular system and whether it is beneficial.Thresholding conditional strategies: The model involves thresholding, where players may only decide to contribute or kill if certain conditions are met (e.g., the number of cells exceeds the threshold). This is a limitation of the strategy space, as it still simplifies decision-making to a binary action rather than a more nuanced strategy based on complex conditions or interactions.Contributed cells do not interact: Only the players make decisions about the embryo pot, the contributed cells do not interact or contribute to the game. Each player is defined by a strategy, and that strategy represents heritable information. Strategies capture behavioural rules that evolved over time, and we expect that these strategies helped achieve the correct embryo size. These strategies influence the number of progeny in each round. This allows us to capture and quantify how cells with a particular strategy would behave in a temporal window (rounds). Single action per round: In each round of the game, each player can only make a single decision to “do nothing”, “contribute” or “kill”. Single action per round: In each round of the game, each player can only make a single decision to “do nothing”, “contribute” or “kill”. Competitive killing of contributed cells: Our strategy space captures killing of cells that were contributed in the previous rounds, focusing on the evolution of strategies where killing is a response to previous interactions. A player cannot eliminate themselves from the game, but rather can only eliminate cells that have been contributed to the embryo pot, including their own contributions. This may not fully represent all forms of killing; for instance, in the naive embryo, apoptotic self-killing is possible and can be triggered by genetic abnormalities or signals from the microenvironment, such as a lack of nutrients. The model presented here, on the other hand, focuses on competitive cell killing strategies, which are represented by a player killing cells present in the embryo pot, regardless of which player contributed the cell. Despite its limitations, this model allows us to investigate whether killing can evolve in a multicellular system and whether it is beneficial. Competitive killing of contributed cells: Our strategy space captures killing of cells that were contributed in the previous rounds, focusing on the evolution of strategies where killing is a response to previous interactions. A player cannot eliminate themselves from the game, but rather can only eliminate cells that have been contributed to the embryo pot, including their own contributions. This may not fully represent all forms of killing; for instance, in the naive embryo, apoptotic self-killing is possible and can be triggered by genetic abnormalities or signals from the microenvironment, such as a lack of nutrients. The model presented here, on the other hand, focuses on competitive cell killing strategies, which are represented by a player killing cells present in the embryo pot, regardless of which player contributed the cell. Despite its limitations, this model allows us to investigate whether killing can evolve in a multicellular system and whether it is beneficial. Thresholding conditional strategies: The model involves thresholding, where players may only decide to contribute or kill if certain conditions are met (e.g., the number of cells exceeds the threshold). This is a limitation of the strategy space, as it still simplifies decision-making to a binary action rather than a more nuanced strategy based on complex conditions or interactions. Thresholding conditional strategies: The model involves thresholding, where players may only decide to contribute or kill if certain conditions are met (e.g., the number of cells exceeds the threshold). This is a limitation of the strategy space, as it still simplifies decision-making to a binary action rather than a more nuanced strategy based on complex conditions or interactions. Contributed cells do not interact: Only the players make decisions about the embryo pot, the contributed cells do not interact or contribute to the game. Each player is defined by a strategy, and that strategy represents heritable information. Strategies capture behavioural rules that evolved over time, and we expect that these strategies helped achieve the correct embryo size. These strategies influence the number of progeny in each round. This allows us to capture and quantify how cells with a particular strategy would behave in a temporal window (rounds). Contributed cells do not interact: Only the players make decisions about the embryo pot, the contributed cells do not interact or contribute to the game. Each player is defined by a strategy, and that strategy represents heritable information. Strategies capture behavioural rules that evolved over time, and we expect that these strategies helped achieve the correct embryo size. These strategies influence the number of progeny in each round. This allows us to capture and quantify how cells with a particular strategy would behave in a temporal window (rounds). CRD difficulty assessment We categorise games as easy, medium, or hard based on the number of Nash equilibria present and the type of strategy coordination required. In the CRD game, all strategy combinations that achieve the targets represent a Nash equilibrium. In addition, we introduce a weighting and frequency element: strategies that are similar are considered easier than those that are different. The greater the number of strategies leading to a Nash equilibrium, the more difficult it is to coordinate, thus increasing the overall difficulty of the game. The combination of these elements results in the designation shown in the lookup matrix in Supplementary Fig. S1a. Conditional strategies We model a conditional strategy by using a threshold that determines the action a player will choose between two options, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b. A player carries a threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ for each round, r, of the game, it compares the total number of cells contributed to the embryo pot so far (Τ) against \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θr. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{T}}}$$\end{document}T ≤ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ, the player chooses action a, otherwise, if Τ > \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ, the player chooses action b. The strategy for each round is denoted (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b) and increases with round number, (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b1)1, (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θ2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}a2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b2)2 … (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}θr, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a$$\end{document}ar, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}br)r. Evolutionary dynamics of CRD embryonic development The order of events of the evolutionary process of the CRD embryo game is as follows (Supplementary Fig. S15):The model is initialised with target size (number of cells), total round number (length of the game), survival probability (a number between 0-1) and types of strategies (conditional or unconditional).Using the parameters defined in step 1, the model then defines a strategy space, S. For example, in a one-round game, S = {N, C, K}.Payoff matrices are calculated for a two-player game using S.The expected payoffs are used in the ordinary differential equation of the replicator dynamics.Replicators are also initialised with generation time and with random frequencies (this is repeated 20,000 times).We analyse the average frequencies across all simulations. The model is initialised with target size (number of cells), total round number (length of the game), survival probability (a number between 0-1) and types of strategies (conditional or unconditional). The model is initialised with target size (number of cells), total round number (length of the game), survival probability (a number between 0-1) and types of strategies (conditional or unconditional). Using the parameters defined in step 1, the model then defines a strategy space, S. For example, in a one-round game, S = {N, C, K}. Using the parameters defined in step 1, the model then defines a strategy space, S. For example, in a one-round game, S = {N, C, K}. Payoff matrices are calculated for a two-player game using S. Payoff matrices are calculated for a two-player game using S. The expected payoffs are used in the ordinary differential equation of the replicator dynamics. The expected payoffs are used in the ordinary differential equation of the replicator dynamics. Replicators are also initialised with generation time and with random frequencies (this is repeated 20,000 times). Replicators are also initialised with generation time and with random frequencies (this is repeated 20,000 times). We analyse the average frequencies across all simulations. We analyse the average frequencies across all simulations. heX-embryoid culture The rtTA expressing PGP1 (male; Coriell GM23338) and PGP9 hiPSCs (female; Coriell GM29087) were transfected using Lipofectamine 3000 (Thermo Fisher Scientific) with Super PiggyBac Transposase (System Biosciences) and the PiggyBac transposon vector with human GATA6-2A-EGFP under the control of the tetracycline responsive element promoter41,118. Transfected cells were selected by adding 0.5 mg ml−1 puromycin to the mTeSR-1 maintenance medium. As before41, the GATA6-engineered hiPSCs were seeded at either a ratio of 81/5 or 4/1 with rtTA-expressing PGP1 or PGP9 hiPSCs either containing or lacking an mKate reporter gene at a total density of 54,000 cells per cm2 in mTeSR-1 supplemented with 10 µM Y-27632. The next day, the medium was changed to mTeSR-1 with 1 μg/ml doxycycline to induce expression of the GATA6 transgene, and this medium was used for daily replacement for up to 5 days. Cell death was inhibited by application of 100 uM Z-VAD-FMK into the normal media at day 0 of the heX-embryoid culture, at the time of induction with doxycycline9,19. This concentration of Z-VAD-FMK was maintained during media changes on subsequent days until the termination of culture. Ethics about the development of heX-embryoids This research was approved and performed under the oversight of the University of Pittsburgh Human Stem Cell Research Oversight Committee to generate human embryo models from human iPSC lines (approval number PROTO202300020). As before41, the work that has been reported in this study followed the 2016 Guidelines for Stem Cell Research and Clinical Translation as well as 2021 ISSCR Guidelines on Ethical Standards for Stem Cell Embryo Model119, including the 2025 update. The heX-embryoids reported in this study are generated from human iPSCs that are derived from human somatic cells, such as fibroblasts and can be purchased from Corriel. Consent was obtained within the Personal Genome Project (https://www.personalgenomes.org/). The heX-embryoids are attached to a cell culture dish, lacking an extra-embryonic trophectoderm tissue critical for full integrated embryo development and implantation in the uterine cavity. The yolk sac-like cavity in these embryoids is not closed, and the tissues cannot be harvested for any implantation without substantial disruption of their structures. These features collectively restrict the ability of this model to undergo the full integrated development of a human embryo in vitro and/or implant to support further development of a conceptus in vivo. This study does not use human blastocysts, does not involve the derivation of any human embryonic stem cell lines and does not use samples obtained from foetal abortions. This study does not involve in utero transfer of any human cells or structures into any other species. Antibodies The following antibodies were used in this study. All primary antibodies for immunofluorescence were used at a dilution of 1:200. All secondary antibodies were used at a dilution of 1:400. The specific clone for monoclonal antibodies is given in parentheses when applicable. GFP Aves Labs (Polyclonal) GFP-1020 PODXL Invitrogen (AB_2532205) 433140 Hoechst 33342 Thermo Fisher H3570 Phalloidin-iFluor 405 AB176752-1001 Ki-67 R&D (Polyclonal) AF7617 Cleaved Caspase-3 Cell Signalling (5A1E) Asp175 Immunofluorescence staining Cells were grown on 14 mm coverslip bottom dishes (Mattek Corporation), referred to as “coverslips” in this protocol. Cultures were fixed for 20 min in 4% paraformaldehyde (Electron Microscopy Sciences) at room temperature. Coverslips were then washed three times with PBS, followed by 15 min permeabilisation with 0.2% Triton X-100 in PBS. Subsequently, the coverslips were washed three times in wash buffer (0.05% Tween-20 in PBS) for 5 min and blocked for 20 min in 200 ml of wash buffer plus 5% normal donkey serum (Jackson ImmunoResearch Laboratories). The primary antibodies were diluted in 5% normal donkey serum in PBS and incubated with the tissues for 1 h at room temperature, followed by three washes in wash buffer for 5 min each. The secondary antibodies were diluted in 5% normal donkey serum in PBS and incubated with the tissues for 1 h at room temperature; if phalloidin was used to mark F-actin, this was diluted with the secondary antibodies at a concentration of 1:1000. Following this, coverslips that were not stained using phalloidin were incubated for 15 min in Hoechst 33342 (10 μl/mL in wash buffer, Thermo Fisher Scientific) to counter stain. Either step was followed by two washes in wash buffer for 5 min each. Afterwards, coverslips were stored in PBS at 4 °C in the dark prior to imaging. Image acquisition and processing Images were acquired using the EVOS M700 automated scanning microscope (M7000 Software Revision v.2.0.2094.0) or Nikon A1 confocal microscope, and processed using Fiji/ImageJ software (National Institutes of Health, NIH). Any contrast adjustments were made in individual channels and applied evenly across the whole image in that channel. Contrast and colour balance for colour images were applied evenly across the whole image. Analysis of wild-type cluster areas Tiled whole-coverslip images were cropped to a central circumscribed square (typically 9000 μm2) image for analysis. Wild-type compartment analysis was performed using an in-house MATLAB (Mathworks) pipeline (built in MATLAB v.R2020a and run in v.R2023b). In brief, wild-type compartments were detected programmatically by thresholding the nuclear dye or F-actin stain for areas of high signal intensity. Maximum individual compartment area and average cell size were defined manually. Compartments were filtered to remove those with areas close to the defined individual cell area by using the following equation:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${lower}\,{size}\,{limit}\,=\,\frac{\max {compartment}\,{area}}{{10}^{\left(\frac{\max {compartment}\,{area}-{single}\,{cell}\,{area}}{2}\right)}}$$\end{document}lowersizelimit=maxcompartmentarea10maxcompartmentarea−singlecellarea2 For compartment counting and wild-type area analysis, characteristics of each compartment were recorded using the regionprops function in MATLAB. Furthermore, the distance to the nearest wild-type compartment was recorded. The Biovoxxel 3D Box Fiji/ImageJ plugin was used for determining the percentage cleaved caspase-3 areas120. The areas of cleaved caspase expression were converted to a binary image using the Threshold function in Fiji/ImageJ, with a consistent threshold used across images of each experiment. Wild-type regions identified via the Matlab script above were converted to binary objects. These images were then used as inputs for the “Overlap Extractor” function from the Biovoxxel plugin; the outputs of this function were used as estimates of relative cleaved caspase 3 expression in each wild-type structure. Calculating human embryo size To measure the size of human embryos at the stage of lumen formation, we referenced the Carnegie dataset83. We estimated the cross-sectional area of embryos, assuming they are an elliptical shape, using measurements of the two elliptical axes provided in the original publications as the dimensions of the “embryonic disc”. Day 8 (CS5a-2) embryos begin amniotic cavity formation and have embryonic disc sizes with cross-sectional axes of 50 μm and 90 μm length, with an estimated area of 14,137 μm2. Day 9 (CS5b) embryos have formed a clear amniotic cavity and have embryonic disc sizes with cross-sectional axes of 100 μm and 130 μm length, with an estimated area of 41,469 μm2. Statistical analysis Statistical analyses were conducted using R (version 4.4.2)121. Non-parametric hypothesis testing was performed using the Wilcoxon rank-sum test, as the data did not meet the assumption of normality. Pairwise comparisons between groups were carried out with the ggpubr (version 0.6.3)122 and rstatix (version 0.7.2)123 packages. To control for multiple comparisons and reduce the likelihood of false positives, we applied the Benjamini-Hochberg correction method. Pairwise comparisons of categorical data were conducted in Python (version 3.12.3) with the SciPy library (scipy.stats, version 1.17.1), using chi-squared tests of independence (chi2_contingency) or Fisher’s exact test (where expected counts were below 5) with Holm-Bonferroni correction for multiple comparisons. scRNA-seq processing and quality control Seurat V5 (version 5.4.0) was used for downstream processing of the scRNA-seq data124. Single-cell data were excluded on the basis of a high mitochondrial genome transcript ratio (> 10%) and either high or low feature or UMI counts. Genes with UMI counts in fewer than five cells were removed from consideration. For scRNA-seq data processing and cluster analysis using Seurat, we used the following general standardised pipeline for processing of a CellRanger output: first, previously clustered datasets from day 0–5 heX-embryoid published previously41 were loaded into R. Cluster 10 from the original dataset was omitted, as many of the top differentially expressed genes were mitochondrial genes, indicating primarily being comprised of low-quality cells. Clusters 3, 9, 15, and 16 from the original dataset were used, corresponding to day 3–5 epiblast, day 0-2 epiblast, amnion-like, and posterior-like clusters in this dataset respectively Visualisation was achieved by the use of uniform manifold approximation and projection (UMAP) plots identifying cells, clusters and selected gene expression in each cell, as well as heatmaps and violin plots showing the expression level of genes by cluster. Visual assessments of the percentage of apoptosis genes and mitochondrial genes were produced using the appropriate quality control metrics from the scCustomize package in R (samuel-marsh/scCustomize: Version 1.1.3)125. Calculating similarity between marker gene lists For the similarity quantification between lists of marker genes, we quantify the similarity between two lists of marker genes A and B by using the hypergeometric test for overrepresentation. This is equivalent to a one-tailed Fisher’s exact test. Bonferroni correction for multiple comparisons was applied. Data obtained for comparison were taken from Lima et al., 202179. Conversion of mouse genes to human orthologs was performed using the orthogene package (version 1.16.1) in R126. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Peer Review file Description of Additional Supplementary Files Supplementary Data 1 Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Supplementary Movie 5 Supplementary Movie 6 Supplementary Movie 7 Reporting Summary Supplementary Information Peer Review file Description of Additional Supplementary Files Supplementary Data 1 Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Supplementary Movie 5 Supplementary Movie 6 Supplementary Movie 7 Reporting Summary Source data Source Data 1 Source Data 2 Source Data 3 Source Data 4 Source Data 5 Source Data 6 Source Data 7 Source Data 8 Source Data 1 Source Data 2 Source Data 3 Source Data 4 Source Data 5 Source Data 6 Source Data 7 Source Data 8