Deterministic and probabilistic fate decisions co-exist in a single retinal lineage

Correct nervous system development depends on the timely differentiation of progenitor cells into neurons. While the output of progenitor differentiation is well investigated at the population and clonal level, the possibilities and constraints for fate decisions of specific progenitors over development are less explored. Particularly little is known about their variability and competence plasticity. To fill this gap, we here use long-term live imaging to follow the outcome of progenitor divisions in the zebrafish retina. We find that neurogenic Atoh7 expressing progenitors produce different neuronal types over development with time-dependent probabilities. Interestingly, deterministic and probabilistic fate decisions co-exist in the same lineage. While interference with the deterministic fate affects lineage progression, interference with fate probabilities of the stochastic lineage branch results in a broader range of fate possibilities than seen in controls. When tissue development is challenged, Atoh7 expressing progenitors can produce any neuronal type, arguing against the concept of fixed competence windows. Stochastic modelling of fate probabilities in challenged conditions revealed a simple gene regulatory network able to recapitulate the observed competence changes during development. Based on our results, we postulate that fate plasticity could be involved in robust retinal development, a concept possibly applicable to other tissues.

To generate organs in a developing embryo, cells progressively become more specialized as 66 they differentiate. Cell differentiation needs to be tightly regulated to produce the correct cell 67 types at the right developmental time. Impairment of the temporal sequence of differentiation 68 can have detrimental consequences for organismal development, including incorrect organ size 69 or cellular arrangements 1,2 . It is thus important to unveil the factors that ensure the production 70 of the right cell type with the correct timing. This is particularly true for the formation of the 71 disclosed. This is different for the Drosophila nervous system, where the temporal regulation 78 of fate decisions has been more thoroughly explored for example in the CNS and the optic 79 lobe 6-9 . In these areas, a defined sequence of progenitor divisions leads to the formation of the 80 different neurons in a consecutive manner. This sequence arises during development as 81 multipotent progenitors are competent to form different neuronal types through the sequential 82 expression of defined transcription factors 10,11,6,12-14 . In the vertebrate nervous system, 83 however, it is less understood whether and how single progenitors competence changes during 84 development to give rise to different neuronal types at the right time and in the right 85 proportions. So far, some progress has been made to understand neuronal birth orders in areas 86 including the neocortex, spinal cord, olfactory bulb and the retina 15-20 but these studies have 87 been mostly performed at the clonal or the population level. This eventually led to different 88 interpretations on whether progenitors competence is pre-determined in progenitor cells, 89 resulting in fixed competence windows during development 19,21,22 , or whether it is variable and 90 influenced by stochastic processes acting on fate decision mechanisms [23][24][25][26] . 91 To better understand the variability and stereotypicity of fate decisions in the vertebrate CNS, 92 a quantitative appreciation of whether and how far the fate outcome of defined progenitor 93 divisions changes over development is needed. To date, this has been challenging due to the 94 fact that many parts of the vertebrate brain are not easily accessible for experimental 95 manipulation and due to the plethora of different progenitors and neuronal types inhabiting 96 different brain areas. An attractive system to circumvent these issues is the developing retina, 97 the part of the CNS responsible for light collection and transmission. It is located on the outside 98 precursor divisions remained unclear. We found that the majority of Crx+ cells at the apical 141 We conclude that within the same lineage, deterministic and probabilistic fate decisions can 176 co-exist. Atoh7+ progenitor divisions throughout the neurogenic window always generate one 177 PRpr and a sister cell of variable fate. The sister cell can become an RGC or an inhibitory 178 neuron. We never observed a bipolar cell (BC) or a second PRpr. Furthermore, we find that the 179 probabilities of producing RGCs or INs gradually changed over development.  We showed that Atoh7+ progenitor divisions can give rise to PRpr, RGC, HC and AC, 183 but we never observed a division producing a BC ( Figure 1). This is in line with the previous 184 notion that in zebrafish BCs arise from a different lineage that does not express Atoh7 32 . 185 Analysis of a double transgenic line expressing Atoh7 and Vsx1 (a BC marker 32,43 ) confirmed 186 that BCs do not express Atoh7 ( Supplementary Figure 2 A, B). As we previously showed that 187 divisions that produce one Atoh7+ progenitor are asymmetric and also produce one Atoh7-  Taking these findings into  216 account, we decided to use a previously established Prdm1a morpholino knockdown approach 217 to interfere specifically with the emergence of PRprs 59,60 . We found that Prdm1a morphant 218 retinas appeared smaller than controls at 48 hpf (Figure 2 A) and showed more severe 219 microphthalmia by 72 hpf (Figure 2 B, E, N = 12 embryos). While control embryos at 48 hpf 220 feature a layer of Ath5+Crx+ photoreceptors at apical positions, this layer is missing in most 221 of the Prdm1a morphants (Figure 2 A, A'). At 72 hpf a layer starts to form, but it is mostly 222 occupied by Crx-and Zpr1-negative cells, confirming a significant reduction in PR production 223 (Figure 2 B, B', N = 10 embryos). In extreme cases we found a total depletion of this cell layer 224 (3/10 embryos, Figure 2D). Further, an overall reduction of retinal thickness, mostly due to 225 shrinkage of the outer nuclear layer (ONL) and of the ganglion cell layer (GCL) was observed 226 Here, neurons that in controls were mainly produced at early stages were also born later. This 477 finding is in line with previous reports in the RP2/sib neuroblast lineage in Drosophila, where 478 it was shown that progenitors exhibit temporal plasticity and can give rise to early lineages at 479 later stages 80 . Furthermore, in double Atoh7/Ptf1a morphants, BCs were seen to arise from the 480 Atoh7 lineage, an outcome never observed in controls. In this case, these usually late-born 481 neurons appeared earlier than normally observed ( Figure 3E). This indicates that Atoh7+ 482 progenitors' can generally produce any neuronal fate at any time, challenging the idea that 483 retinal progenitor cells can give rise to certain fates only during fixed competence windows 22 Figure 2C. 871 Retinal diameter was measured on three different z planes per retina, and the average 872 measurement was plotted for each replicate. 873 Retinal thickness was measured on three different z plane of the central region of the tissue, 874 and the average measurement was plotted for each replicate. The outer nuclear layer (ONL) 875 was assigned as the distance between the apical side and the outer plexiform layer; the inner 876 nuclear layer (INL) as the distance between the inner plexiform layer and the outer plexiform 877 layer; the ganglion cell layer (GCL) was assigned as the layer between the inner plexiform 878 layer and the most basal position of the retina. 879

PH3+ cells count 881
To count the number of PH3+ cells in whole retinas, the Tg(hsp70:H2B-RFP) was used to 882 identify the boundaries of the tissue. Then, stacks covering the whole tissue were acquired 883 using the laser-scanning confocal. Images were imported in Imaris and the mask tool was used 884 to isolate the retinal tissue from the rest of the tissues in the image. The spot detection tool was 885 used to count the number of PH3+ cells with these parameters: diameter = 8 µm (x-y 886 dimensions), PSF modelling = 16 µm. Threshold for detection of the PH3+ nuclei was adjusted 887 manually to ensure that all the PH3+ nuclei were counted but was usually kept around 1800. 888 889

Statistical analysis 890
All Statistical tests used are indicated in the figure legend, as well as the definitions of error 891 bars. All test used were two-sided and 95% confidence intervals were considered. P values are 892 indicated in the figure legends, as well as sample sizes, or in Table 1 for experiments in Figure  893 3B. Data were analysed using GraphPad Prism 6 or Python 3. Statistical analysis was 894 performed using GraphPad Prism 6 and Julia 1.

Event plots 902
Event plots were produced from the raw experimental data using a custom data analysis and 903 plotting pipeline developed by the authors, which is available at https://git.mpi-904 cbg.de/nerli/lineage-analysis. The code comes together with the data files, allowing for an easy 905 and complete replication of the analysis and plots that appear in this work. 906 The data analysis consists of the following steps: Since we know from the lineage analysis data that each division produces one PRpr and 922 one sister cell whose fate is stochastically determined, we modelled only one fate decision per 923 division. Also, as the TFs Atoh7, Ptf1a and Prdm1a are responsible for the specification of the 924 different neuronal fates, we modelled their expression as random variables, with time-925 dependent expected values. Their expression levels were normalized with respect to an effective 926 threshold, i.e., a level at which they start to carry out their inhibitory function and produce a 927 downstream effect. 928 The simulation of this model relies on drawing N=100 fate decisions for each time point 929 to estimate the shares of different fates, this was then repeated for R=100 times to compute the 930 mean and confidence interval of these shares. The TF levels were modelled as normally 931 distributed, with a time-dependent mean value and a constant standard deviation. Being the 932 threshold set at the arbitrary value of 1.0, in the model we set the mean expression levels of 933 Atoh7, Prdm1a at the values of 1.1 and 1.05 respectively. The mean levels for these TFs were 934 kept constant over time. The mean expression level for Ptf1a was instead time-dependent 69 and 935 increased according to a logistic function with lower and upper asymptotes at 0.9 and 1.1 936 respectively. 937 The standard deviation was constant over time for all TFs and was computed by multiplying a 938 fixed coefficient of variation (CV=0.14) by the time-averaged expression level of each TF. 939 Finally, the GRN was modelled as a decision rule occurring at the terminal branching point of 940 the lineage (Figure 1G), taking the TF levels as input, and producing a fate choice as output. 941 Any inhibition in the GRN was translated into a precedence rule, i.e., a TF that inhibits a second 942 one is able, if its level is above the threshold, to determine the acquisition of a certain fate, 943 while the inhibited TF is ignored. The precedence rule for Scenario A ( Figure 4C,E)

949
To achieve inhibition of the target TF in additive inhibitions, it was enough that the sum of the 950 levels of the inhibitors involved was above the threshold. The precedence rule in the additive 951 inhibition case (Scenario B, Figure 4D,F)

956
In summary, the fate decision model consists of (1) drawing independent and normally 957 distributed stochastic TF expression levels, (2) applying the decision rule and recording the 958 resulting fate, (3) repeating this stochastic decision for 100 divisions taking place at any given 959 time point and (4) again repeating the whole process for 100 times, creating a synthetic dataset 960 that can then be analysed in the same way as the experimental data.