Supplementary Information for Mechanical Stress Inference for Two Dimensional Cell Arrays

Mechanical Stress Inference for Two Dimensional Cell Arrays Kevin K. Chiou, Lars Hufnagel, Boris I. Shraiman1,3,∗ 1 Department of Physics, University of California, Santa Barbara, CA 91306, USA. 2 Department of Cell Biology and Biophysics, European Molecular Biology Laboratory, Heidelberg, Germany. 3 Kavli Institute for Theoretical Physics, University of California, Santa Barbara, CA 91306,USA. ∗ E-mail: shraiman@kitp.ucsb.edu


Introduction
Genetics and biochemistry are central to all aspects of biological function. Physics is often less recognized yet also important at many levels, everywhere from intramolecular to organismal scales. For example, many important aspects of cell behavior depend directly and indirectly on its mechanical state defined by its interaction with neighboring cells and adhesion to the extracellular matrix [1][2][3]. Cytoskeletal mechanics and cell-cell adhesion determine geometric properties of cells [1,[4][5][6], as well as the dynamics of biological tissues [5,[7][8][9][10][11][12][13]. In plants, cells do not move, but the rigidity of cellulose membranes makes mechanical stress an obvious factor for cell division and proliferation [14,15]. It is known that animal cell proliferation also depends on substrate adhesion and the degree of cell confinement [2,[16][17][18][19]. It has also been demonstrated that (stem) cell differentiation is affected by substrate rigidity [20]. More speculatively, mechanical feedback interactions have been conjectured to have a role in coordination of growth during development [1,9,21,22]. Mechanical transformation of epithelial tissue is of course itself central to many morphogenetic processes: gastrulation [7] and convergent extension [1], to name a few. Understanding how mechanical changes in cells orchestrate morphological reorganization of tissues is an open problem and a subject of much current work [1,7,8] Our present understanding of the role of mechanics as one of the regulatory inputs into the cell is strongly impaired by the difficulty of quantitatively characterizing the mechanical state (i.e. stress and deformation) of the cell. Among the available techniques are laser tweezers [23] and ''traction force microscopy'' [18,24] performed on cultured cells. UV laser ablation allows the mechanical perturbation of tissues [8,25,26] on the cellular scale with the time-lapse imaging of subsequent relaxation providing information on the mechanical state of the tissue. The ablation approach is widely used on live preps, for example, in the study of Drosophila embryonic development. Yet, this technique is definitely not a ''non-destructive'' one.
On the other hand one of the major recent technical advances in developmental biology is the improvement of live fluorescent imaging. These provide high quality time lapse movies of developmental processes, including interesting morphological transformations such as gastrulation and convergent extension [7,26,27]. The purpose of the present investigation is to explore what insight into the mechanical state of cells may be gleaned from a quantitative examination of high quality images of the type shown in Fig. 1A. Our goal is to use image analysis as a nondestructive approach to obtaining quantitative measures of stress in these systems. A similar strategy has been pursued by the recently proposed ''Video Force Microscopy'' (VFM) approach by Brodland et al [28]. Our approach will differ from VFM in its assumptions about mechanical state of tissue, in the parameterization of forces and in the way imaging data is utilized.
Below we shall define a general model parameterizing the mechanical state of cells in two dimensional epithelial tissue and provide a computational method for inferring these parameters from the observed geometry of the cell array. We shall study the sensitivity of the proposed Mechanical Inverse (MI) method to errors in measured cell geometry and identify conditions under which robust inference is possible. We then illustrate the proposed MI method by applying it to the analysis of two different biological processes: cochlear neurogenesis [29] and ventral furrow formation [27].

Model of epithelial tissue mechanics
Our approach is based on the assumption that epithelial monolayers are in an instantaneous mechanical equilibrium, characterized by a static balance of the forces acting at intercellular junctions. The second important assumption is that epithelial mechanics is dominated by the actomyosin cortices and inter-cellular Adherens Junctions [1] both localized at cell boundaries, which form a visible two-dimensional web, as shown in Fig. 1A. Thus we assume that the mechanical state of a cell can be described by an effectively two-dimensional model with tension at the interface and the hydrostatic pressure in the cell interior. Yet, because cells can independently regulate their mechanical state, e.g. by modulating myosin activity or cell-cell adhesion, we allow for the possibility of each intercellular interface to have a different effective tension, T ab , and for each cell to have a different internal pressure P a (where a labels cells and ab labels the interface between cells a and b), as shown in Fig. 1D. Mechanical equilibrium then corresponds to the condition that the forces acting on each ''vertex''r r i (defined as a junction of three cells and therefore of three interfaces) add up to zero.
Letr r i andr r j be the vertices belonging to the interface ab and let r r ij :r r j {r r i be the vector from vertex i to j. The force exerted by this interface on vertex i is where a labels vector components in the xy plane and e ba is the anti-symmetric tensor (e xy~{ e yx~1 and e xx~eyy~0 ). As shown in Fig. 2, this expression accurately represents the Young-Laplace balance between interfacial tension and the pressure differential across the interface P a {P b~kab T ab , as long as the interfacial curvature k ab is small (see Supplementary Text S1). This fact enables us to formulate all mechanical balance conditions in terms of a polygonal approximation of the cell array, thus allowing us to reduce the problem to a generalized ''vertex model'' [8,9]. Remarkably, the forces given by (1) correspond to the mechanical energy in the form of the following simple Hamiltonian where A a is the area of cell a, ' ab~D r ij D is the length of the interface between cells a and b and f' ab g denotes the set of interfaces belonging to cell a. Both A a and ' ab 's are defined in the polygonal approximation. This Hamiltonian is a generalization of the vertex models often used to describe epithelial sheet mechanics [8,9,26]. Pressure and tension are defined by considering the differential form of H: where we define T ab : LH L' ab and P a :{ LH LAa . The SabT sum runs over all edges, i.e. pairs of neighboring cells a, b. This tangent representation of mechanical energy expresses interfacial tension T and intracellular pressure P as conjugate variables to edge lengths and cell areas, respectively. (The reader will notice that strictly speaking our P a refers to a two-dimensional pressure which relates to the hydrostatic pressure only with the additional assumption that dA a entails a change of cell volume. Alternatively  [29] for details) at the E9 stage of development just following the onset of differentiation of cells into hair cell precursors and surrounding support cells. Panel (A) visualizes cell boundaries using and anti-cingulin (a tight junction protein) staning. Panel (B) shows the same tissue with pro-neural cells [29] stained via an anti-hair cell antigen. (C) is a computer generated segmentation of the raw image in (A) as a polygonal tiling which approximates cell geometry. The zoomed-in image (D) defines our parametrization of cell geometry in terms of vertex coordinates r i and of the mechanical state of the cell in terms of interfacial tensions T ab and hydrostatic pressures P a . doi:10.1371/journal.pcbi.1002512.g001

Author Summary
Mechanical forces play many important roles in cell biology and animal and plant development. In contrast to inanimate matter, forces in living matter are generated by active and highly regulated processes within and between cells. The ability to directly measure forces and mechanical stress on the cellular scale within living tissues is critically important for understanding many morphogenetic processes but is a serious experimental challenge. The present work proposes an alternative approach based on the analysis of images that provide a visualization of cell boundaries in two dimensional epithelial tissues. The method uses the assumption of force balance within the epithelial layer to infer, on the basis of image-derived geometric data, the mechanical state of each cell. The proposed Mechanical Inverse method is illustrated on the analysis of two examples: the initial step of the gastrulation process in the Drosophila embryo, and the process of neurogenesis in the developing avian cochlea.
P may be thought of as the axial component of the threedimensional stress tensor.) Mechanical equilibrium means that H is minimized with the respect to vertex positions which defines the static force balance constraints (the sum is over the vertices j(i) neighboring i). Our analysis will be based on the assumption that the cell layer is close to mechanical equilibrium in the sense of the magnitude of the net resultant force acting on vertices being much smaller than the average magnitude of the component forces that vectorially add to the resultant Stated in other words, we assume that the internal forces that balance each other in the (approximate) instantaneous mechanical equilibrium state are much larger than the unbalanced residual force that drives residual physical motion and (through viscous effects) defines its velocity. More generally, the dynamics of passive relaxation towards this mechanical equilibrium would be described by n P and the sum is again over the vertices j(i) that neighbor i. In principle, given that vertex velocities can be directly measured by time-lapse microscopy in live tissues, d dtr r i {r r j(i) Â Ã can be obtained directly from the experiment, allowing a straightforward extension of the method described below toward the VFM method [28] (where viscous forces were assumed to dominate).

The mechanical inverse problem
We can now inquire to what extent the knowledge that a given cell array geometry is in a mechanical equilibrium constrains the parameters P a , T ab describing the mechanical state of cells. We proceed by a simple count of mechanical constraints and of the free parameters for two cases i) a closed cell array, shown in Fig. 3A and ii) an open cell array, shown in Fig. 3B.
Let us begin with the closed cell array where v,e, and c are the total number of vertices, edges, and cells, respectively. For v vertices in two dimensions, we have exactly n c~2 v{3 mechanical constraints, where the extra three degrees of freedom are associated with global translation and rotation symmetries (alternatively, three constraints are redundant because the total force and total torque in the closed system are equal to zero). On the other hand, the number of unknown tension parameters is e, and the number of unknown pressures is c, so that the total number of parameters is n p~e zc. Our closed system, if we count the exterior as an additional ''cell'', is topologically equivalent to a sphere so that Euler's theorem reads Combining this relation with the condition that vertices are points where three edges meet and each edge impinges on two vertices, that is 3v~2e, we obtain the result

ezc~2vz1: ð6Þ
This implies n p~nc z4, which means that our unknown parameters can be determined up to four free constants. One of the latter is the arbitrary overall scale of T ab and P a which cannot be constrained by the force balance conditions (note that since P a is only defined up to an additive constant, one can set the pressure in the exterior of the domain to zero). Yet the good news is that the number of free constants is finite, while the number of nontrivial constraints scales with the number of cells! This counting argument can be readily generalized (see SI) to the case where a fraction of vertices has more than three incoming edges: the so called ''rosettes'' that can be quite common in certain tissues [30].
Repeating the counting procedure for the open system, one finds that ezc~2vzbz1, where b is the number of cells at the boundary of the domain. It follows that n p~nc zbz1. Thus mechanical parameters are determined up to bz1 free constants: we can still choose the overall scale while the additional b degrees of freedom may be regarded as the boundary conditions such as P a 's of the cells at the edge of the domain. Again, for a large array, because b* ffiffi ffi c p while n p *c, the number of parameters and constraints is much larger than the number of free constants.  To actually determine the T ab and P a parameters, we use the fact that they appear only linearly in the force balance equations (4). This results in a linear system for with M being an n p |(n c z1) matrix where the 1st n c rows impose the force balance conditions from equation (4), and the additional row imposes the scale. This is performed by constraining the average tension to be equal to one. Correspondingly the top n c entries of the column vector C are zero, while the bottom row C ncz1~e .
The rectangular system (8) is solved via a pseudo-inverse [31] with the general solution of the form Mw n~0 ð11Þ whereM M {1 is the pseudo-inverse of the rectangular matrix M and the free parameters, A n , are the amplitudes of the n z~np {n c {1 ''zero modes'' w n . Additional details regarding the formulation and solution of the inverse problem are provided in the Supplementary Text S1. Fixing the remaining n z degrees of freedom requires introducing additional constraints: e.g. one may have reasons to seek a solution that minimizes variation of P a 's or T ab 's. In choosing such additional assumptions one may want to use all the information that one has for specific applications, as we shall do below. However, before proceeding to the applications we must consider the issue of error sensitivity.

Sensitivity of the inverse
Our approach to mechanical parameter inference is based on the observed geometry of the cell array. How sensitive are the results to the inaccuracy of vertex positions fr r i g? Such inaccuracies will inevitably arise in the process of imaging, image segmentation, and more importantly from the fact that the cells themselves fluctuate. (These fluctuations are of course related to the fact that mechanical equilibrium is itself at best approximate.) To quantify the stability of the inverse we consider the effect of an arbitrary small perturbation in vertex positions, fdr r i g. Since the inhomogeneous term C in (8) is independent of cell geometry, the first order response of the parameters dy to positional error is given by Ideally the error response matrix L would have small eigenvalues providing a relatively robust inverse. On the other hand, large eigenvalues of L would indicate high error sensitivity. These sensitive modes appear via the pseudoinverse matrixM M {1 . A histogram of singular values of the matrixM M {1 is shown in blue in Fig. 4 (for a closed system with &800 cells). One notes that a substantial fraction of modes have eigenvalues larger than one. As a result, small errors in positions can result in large error in inferred parameters.
The simplest way to solve the sensitivity problem is to reduce the number of parameters. For example, as we shall argue below, in some contexts it may be reasonable to neglect variation in cell pressure and set P a~P0 which eliminates c parameters, reducing n p from 4c to 3c. In that case the mechanical constraint system given by (8) becomes overdetermined and can be solved only in the sense of least square minimization: i.e. minimization of The solution of the minimization problem is still given by the pseudo-inverse of the rectangular matrix M' which extends the force balance matrix M by including additional (linear) equations that constrain y. Fig. 4 shows (in red) the distribution of singular values governing the sensitivity of the reduced or partial inverse problem. We note a substantial reduction in sensitivity.
The partial inverse approach is then tested in silico. To that end we consider a closed array of cells and define cell geometry by minimizing elastic energy given by with uniformly distributed k ab [½0:7,1:3. The absence of area terms imposes a constant pressure. (The closed cell array is relaxed under toroidal boundary conditions to prevent a collapse into the zero tension ground state.) The vertex model parameters are computed via equation (2). These quantities are then compared to values obtained by applying the partial inverse algorithm to the vertex ''data'' fr r i g corrupted by random noise fdr r i g with an r.m.s. variation of 5% of the average length of cell edge (see Fig. 5). The correlation coefficient between inferred and computed parameters is 0.85, which confirms the ability of our method to extract information from noisy data. The Supplementary Figure S4 shows results under a 10% random corruption in vertex positions. In that case the correlation coefficient is reduced to 0.65, which remains serviceable. We note that the''soft modes'' that contribute to the sensitivity of the full inverse problem are quite interesting. The formulation of the minimally constrained problem is analogous to the isostatic systems studied in jamming transitions of amorphous solids [32]. These isostatic systems live on the boundary of Maxwell's criterion for rigidity, and much like amorphous solids they must satisfy both the local and global rigidity conditions. In our mechanical inverse formulation, ''rigidity'' corresponds to a fully constrained set of mechanical (T ab and P a ) parameters. Amusingly, local soft modes for the MI problem correspond to special local geometries: specifically, polygons that can be inscribed into circles (i.e. a generalization of regular polygons) -a category which includes triangles of any shape. These interesting mathematical aspects of the problem will be discussed in a separate publication.

Mechanical differentiation of cells in the developing avian cochlea
In cochlear development, which takes place during the first two weeks of chick embryonic development, cells in an initially homogeneous two dimensional epithelial layer differentiate into pro-neural (hair-cell) and support cell fates [29]. The process is driven by Delta/Notch-mediated cell-contact signaling [33], which causes lateral inhibition: cells which express Delta ligand on their surface prevent their immediate neighbors from doing the same. Expression of Delta is an early marker of the pro-neural fate of cells. Fig. 1A,B presents a micrograph of the cochlea epithelium, obtained by Goodyear and Richardson [29] at the stage of development shortly after the onset of differentiation. The images in Fig. 1A,B were obtained as described in [29] using a double fluorescent antibody labeling: antibody to the tight junction protein cingulin allowing visualization of cell boundaries and 275 kDa hair-cell antigen staining labeling pro-neural cells. Note that the two cell types already have discernibly different morphology: pro-neural cells are somewhat smaller and have curved edges. This dimorphism is supported by direct labeling of specific pro-neural markers, shown in Fig. 1B and demonstrated in [29].
Our goal is to infer, based on the analysis of the image in Fig. 1A, the variation in the mechanical parameters between cells. The visible positive curvature associated with pro-neural cells suggests that they are under higher internal pressure. Can the Mechanical Inverse method determine pressure differentials between cells? Because our approach requires only positions of cellular vertices, it does not use the information provided by the interfacial curvatures which are readily measurable on the image. This additional information will be used as an a posteriori validation of the inferred results.
To reduce the number of parameters, we assume that interfacial tensions can be expressed as T ab~ha zh b in terms of constant cortical tensions h a , h b of adjacent cells which reduces the number of parameters by e{c~2c. This is sufficient to render a robust partial inverse (in the sense of least squares), yielding h a and P a for every cell. Fig. 6 shows the distribution of inferred intracellular pressures and cortical tensions for the two cell types. We see that pro-neural cells have on average higher tension and pressure. While pressure shows some correlation with cell area, there is no correlation between interfacial tension and its length. However, there is no reason to expect any specific correlation between these quantities. On the other hand, Laplace's Law predicts P a {P b~kab T ab which we are in a position to check directly, thanks to the fact that interfacial curvatures k ab are directly measurable on the images such Fig. 1A. Fig. 7 presents the ''empirical'' Laplace's Law obtained on the basis of the inferred P a {P b and T ab . Because the Mechanical Inverse algorithm did not in any way use the interfacial curvature information, the fact that inferred parameters approximately obey the Laplace's Law provides a validation of the inverse method.

Mechanical anisotropy at the onset of the ventral furrow formation in Drosophila
During the initial stage of development a Drosophila embryo is comprised of an ellipsoidal monolayer of cells. The first step  toward more complex morphology that is achieved after gastrulation, is the formation of a ventral furrow that begins with the contraction of the apical surfaces of cells along the ventral midline [27,34,35]. Fig. 8A presents the ventral view of a Drosophila embryo at the beginning of this mechanical transformation. The high quality of these images (kindly provided by the Weischaus lab [27]) makes it possible to attempt the Mechanical Inverse analysis. Since the process begins even before cellularization is completed it is reasonable to assume that cells have the same internal pressure P a~P0 , allowing us to reduce the number of parameters enough to achieve a robust partial inverse and infer T ab for every cell boundary. Note that in contrast to the developing avian cochlea, cell-cell interfaces exhibit little curvature and (apical surfaces of) cells are well approximated by polygons (see in Fig. 8A), which is consistent with pressure differentials being weak compared to interfacial tensions.
Interestingly, comparing images separated by merely two minutes ( Fig. 9) we found that the inferred T ab at the later time-slice exhibited statistically significant anisotropy with estimated tensions of cell interfaces along the AP axis being on average about 15% higher than those along the DV axis. The inferred increase in AP tension (relative to DV) is consistent with the laser ablation measurements made in the Wieschaus lab [7,27]. Yet, mechanical inverse inference gives information not only on the global, tissuewide level, but also on the scale of a single cell and interface. The analysis also clearly demonstrates the ability to make specific predictions (for interfacial tensions) that can be directly tested by combining high quality live imaging with UV pulsed laser ablation.

Intercellular traction forces
The variation of tension from one interface to another implies the existence of traction forces acting between cells. This traction, or shear stress, must be entirely borne by the cadherins and other cell adhesion molecules which bridge cellular membranes and connect actomyosin cortices of apposing cells [1]. In Fig. 10 we zoom in on an interface decomposing interfacial tension into the cortical tensions on the opposite sides of the interface T ab~ha (x)zh b (x), now allowing for the possibility that the latter are not constant along the interface and vary as a function of position along the edge x[½0,' ab . This transfer of tension from the cortical bundle in one cell to the other is possible because of cadherin mediated traction forces acting between cells. The total shear stress on the interface is t ab~½ h a (0){h a (' ab )=' ab~½ h b (' ab ){h b (0)=' ab . In the Supplementary Text S1 we show that because cortical tensions are constrained by the continuity conditions at cell ''corners'' they can be   readily expressed in terms of interfacial tensions leading to the following simple expression for the traction force acting between cells a and b.
Fig . 8B shows inferred tractions calculated for the ventral furrow data taken two minutes prior to invagination. We observe a significant variability in tractions at different interfaces. Because traction forces stretch trans-cellular cadherin dimers, they may be physiologically important. Since at present there is no way of measuring them directly the possibility of indirect inference is particularly interesting.

Discussion
We have demonstrated that the readily visualized two dimensional network of cellular interfaces in an epithelial tissue holds, potentially, a wealth of information on the relative strength of mechanical stresses acting in the tissue. The main precondition is that the tissue is close to the mechanical equilibrium in which internal cytoskeletal forces are balanced by intercellular interactions. Any imbalance of forces corresponding to directed or fluctuating motion must be small in comparison to the magnitude of internal forces that balance each other in mechanical equilibrium. Force balance is achieved by the suitable adjustment of cell geometries (parameterized by the positions of vertices). Conversely we envision changes in tissue geometry to be driven adiabatically -i.e. without disruption of the mechanical equilibrium -by changes in cytoskeletal forces within cells. This picture is at once similar and dissimilar to the case of soap froths. The geometry of a soap froth [36][37][38] is also defined by the instantaneous force balance and changes adiabatically (when gas diffuses out of cells with higher internal pressure). Yet epithelial cells, in contrast to soap bubbles, can control interfacial tension by regulating myosin activity within actomyosin cortices and therefore can generate variation in tension on sub-cellular scale, even between different interfaces of the same cell.
Our Mechanical Inverse method is fundamentally different from the Video Force Microscopy [28]. In contrast to our assumption that cytoskeletal forces are in an approximate instantaneous balance, VFM is based on the assumption that bulk forces acting within the tissue are balanced by viscosity. It is therefore based on the observed velocity of tissue motion and employs finite element methods to define forces on a computational grid rather than the underlying cells. The two methods are complementary in the sense that VFM provides information about the distribution of unbalanced bulk force which drives motion on the scale of the embryo, while our Mechanical Inverse is focused on the internal balance of forces in relation to cell geometry and its local changes. Our approach can be extended to include measured velocities which can be used to define net forces on the vertices, as explained below Eq. (4), leading to a modified inverse problem. This generalization would bridge the static inference presented here with the VFM approach. Yet, to the extent that the dynamics of normal epithelial cell rearrangement unfolds relatively slowly (on the time scale of minutes) compared to the rapid (time scale of seconds) viscosity limited retraction of laser ablated interfaces, it is reasonable to assume that the contribution of viscous forces during slow normal developmental dynamics is small compared to the balancing internal forces, which is the assumption underlying our Eqn. (4).
Recent experiments have demonstrated that actomyosin structures transiently assembling on the apical or basal surfaces of the cell, play an active role in defining its mechanical state [7,26,39]. In particular, [7] and [26] argue that coalescing pulses of ''medial myosin'' on the apical surface drive a ratchet of apical surface contraction. Presently, our mechanical model does not explicitly incorporate such effects, which in full generality would require introduction of many more parameters (characterizing intracellular heterogeneity and anisotropy). On the other hand, these effects are not observed in all epithelial tissues at all times, leaving the present approach with many possible applications. Furthermore it may be possible to generalize our approach to model medial myosin as well, especially if additional information from cell imaging is used. For example, during convergent extension investigated in [26] one often observes intracellular medial myosin filaments attaching to the lateral cortex and causing measurable deformation of cell-cell boundary. It that case it may be possible to define an additional ''vertex'' corresponding to the attachment point, apply considerations of mechanical balance discussed above and obtain an estimate of the force applied by the medial myosin as compared to the cortical tension. Alternatively, when medial actomyosin structures appear to be isotropic, their effect may be well approximated by a uniaxial stress which is already parameterized already by our existing model. Studying the effect of medial myosin would be an interesting direction for future work.
The proposed Mechanical Inverse method converts clearly stated assumptions about the nature of cellular stresses into readily falsifiable predictions. Using the example of avian cochlea, we were able to demonstrate that mechanical parameters inferred via the Mechanical Inverse satisfy non-trivial cross-checks provided by independent additional information (interfacial curvature measurements) read off the tissue images. Thus our approach is capable, in realistic applications, of inferring mechanical parameters and to uncover interesting aspects of the internal state of the cell. By combining high quality live imaging with UV pulsed laser ablation, one will be able to put predictions for local interfacial tensions obtained via the Mechanical Inverse, to a rigorous experimental test. We note however, that the predictions do not have to be very accurate to be useful. Even if inferred tensions each carry only a single bit of information -i.e. identify interfaces with high or low tension -correlating tension with the observed level of myosin, cadherin and/or other proteins involved in regulation of cell mechanics could be extremely informative (in addition, since a large number of cells can be imaged and analyzed, the method is effectively ''high throughput''!). Finally, our approach allows for inference of quantities such as intercellular traction forces (or shear stress), which may be important for the stability of Adherens Junctions but cannot be directly measured by any means presently available. Future development of FRET based molecular sensors of stress [40] may nevertheless make such measurements possible in the future. Hence we expect that further development, validation, and application of the Mechanical Inverse method will lead to new insights into the molecular biology of epithelial cells and tissues. Figure S1 A comparison of inferred tensions between two optimization schemes: linear least squares (i.e. the pseudo-inverse) and linear least squares with a tension positivity constraint (i.e. quadratic programming). Note the small tail of negative tensions predicted (by the pseudo-inverse) when positivity is not imposed.  Figure S3 The ''collapse'' of two three-fold coordinated vertices into a four-fold coordinated vertex that is shared by cells a, b, c, and d (also known as a ''rosette''). This process can be performed iteratively by collapsing more three-fold coordinated vertices to obtain vertices of arbitrary order. Each ''collapse'' creates a new vertex that is of one coordination higher than the previous, as well as subtracting one vertex and one edge from the total count. (TIF) Figure S4 Tension inference of pressure-constrained simulated tissue under artificially induced 10% error in vertex positions. The correlation coefficient here &0:6. While individual tensions may deviate considerably from their simulated values, the correlation between inferred and simulated tensions is still significant. (TIF) Figure S5 Scatter plot of edge length against mechanically inferred tension of two time points during Drosophila ventral furrow formation. Panel A and B respectively show results from 4 and 2 minutes prior to invagination. Analysis indicates only a very weak negative correlation between inferred tension and edge length. (TIF)

Supporting Information
Text S1 Additional technical details on the a) vertex model; b) computational Implementation of the Mechanical Inverse; c) interfacial traction; d) tension anisotropy; e) counting argument with higher order vertices. (PDF)