Multi-Input RNAi-Based Logic Circuit for Identification of Specific Cancer Cells

See allHide authors and affiliations

Science  02 Sep 2011:
Vol. 333, Issue 6047, pp. 1307-1311
DOI: 10.1126/science.1205527


Engineered biological systems that integrate multi-input sensing, sophisticated information processing, and precisely regulated actuation in living cells could be useful in a variety of applications. For example, anticancer therapies could be engineered to detect and respond to complex cellular conditions in individual cells with high specificity. Here, we show a scalable transcriptional/posttranscriptional synthetic regulatory circuit—a cell-type “classifier”—that senses expression levels of a customizable set of endogenous microRNAs and triggers a cellular response only if the expression levels match a predetermined profile of interest. We demonstrate that a HeLa cancer cell classifier selectively identifies HeLa cells and triggers apoptosis without affecting non-HeLa cell types. This approach also provides a general platform for programmed responses to other complex cell states.

Synthetic biomolecular pathways with elaborate information processing capability will enable in situ response to complex physiological conditions (1). For example, multiple cancer-specific biomarkers (2) can be sensed and integrated in such pathways, providing precise control of therapeutic agents (3). A combination of up to two tissue-specific signals, including promoter and/or microRNA (miRNA) activity, mRNA, and protein levels, have been used to partially restrict therapeutic action to cancer cells (49). In parallel, research in synthetic biology has demonstrated multi-input information processing in living cells (1018), but the interaction between these systems and the cellular context has been limited. Yet, general-purpose mechanisms for programmable integration of multiple markers are required to detect cell state and precisely regulate therapeutic actuation. Here, we describe such a mechanism, a “classifier” gene circuit that integrates sensory information from a large number of molecular markers to determine whether a cell is in a specific state and, if so, produces a biologically active protein output. Specifically, when transiently expressed inside a cell our classifier ascertains whether the expression profile of six endogenous miRNAs (19) matches a predetermined reference profile characteristic of the HeLa cervical cancer cell line. A match identifies the cell as HeLa and triggers apoptosis (Fig. 1A).

Fig. 1

High-level architecture of a cell type classifier. (A) Schematic representation of a HeLa-specific classifier circuit operation. Gray circle, healthy cells; light green, HeLa cells. (B) High-level and detailed description of double-inversion module for sensing HeLa-high miRNAs. Act, activator; R, repressor. (C) High-level and detailed description of HeLa-low miRNA sensor. (D) Schematic representation of an integrated multi-input classifier. Lines with bars indicate down-regulation. R1 and R2 represent double-inversion modules. The entire network implements a multi-input AND-like logic function for identification and selective killing of HeLa cells through regulated expression of hBax.

A HeLa reference profile was constructed by choosing a set of HeLa-high and HeLa-low markers so that HeLa cells express high or low levels of these, respectively, whereas expression in any healthy cell differs substantially from that of a typical HeLa cell for at least one of these markers. Multiple reference profiles are possible, and these profiles do not need to differentiate HeLa from other cancers. We designed a sensor motif for HeLa-high markers comprising a “double-inversion” module that allows output expression only if the marker is present at or above its level in HeLa cells but efficiently represses the output if the marker’s level is low. The design is based on our previously described module that consists of the small interfering RNA (siRNA)–targeted transcriptional Lac repressor (LacI) and a LacI-controlled promoter CAGop (CAG promoter followed by an intron with two LacO sites) (17). We improved the ON:OFF ratio of our original design to approximately 8- to 10-fold by introducing a reverse tetracycline-controlled transactivator (rtTA) to control LacI and by targeting both the repressor and the activator using miRNA in a feed-forward loop (Fig. 1B and fig. S1) (20, 21). A HeLa-low marker sensor was implemented by fusing four repeats (5) of fully complementary target sites into the output’s 3′-untranslated region (3′-UTR) (Fig. 1C). A complete classifier circuit consists of a set of HeLa-high and HeLa-low marker sensors all arranged to target the same output (Fig. 1D). Multiple HeLa-low sensors are combined by fusing their corresponding miRNA targets in the output’s 3′-UTR (17). Multiple HeLa-high sensors are integrated by regulating the same output in parallel. Accordingly, all HeLa-high markers must be present at the same time to enable output expression from CAGop because any individual double-inversion module can efficiently repress the output by itself (22).

To choose HeLa markers for the reference profile and analyze expected circuit performance, we created a mathematical model consisting of a multi-variable circuit response function that uses experimentally derived responses (23) of individual sensors to their miRNA inputs [fig. S2 and supporting online material (SOM) text]. We first evaluated HeLa-high marker combinations from the set of the top 12 candidates determined by analyzing expression data from the MicroRNA Atlas (fig. S3) (24). On the basis of reasonable assumptions regarding sensor response parameters (22), we found that using miR-21 together with a composite marker that includes both miR-17 and miR-30a (miR-17-30a) results in at least a fivefold difference between circuit output in HeLa cells and the output in all but a few healthy cell types profiled in the MicroRNA Atlas (Fig. 2A and fig. S2D) (24). We then searched for HeLa-low markers and found that miR-141, miR-142(3p), and miR-146a are highly expressed in the potentially misclassified cell types but unexpressed in HeLa (Fig. 2B). With inclusion of these inputs in the response function, the model predicts at least a sixfold output difference in HeLa cells relative to the closest other cell type USSC-7d and on average about a 160-fold difference relative to the rest of the cells (fig. S2E). We also computed selectivity of all possible marker subsets and found that it steadily increases as the number of markers goes from one to five (Fig. 2C). With all inputs included, the response function is well approximated by a Boolean expression (fig. S2F): Embedded ImageIn addition to HeLa, we tested circuit operation in six control cell lines. For practical reasons, most of these are cancer cell lines that can be handled with ease and that are predicted by the model to produce less than 1% of HeLa output (22, 24). First, we measured all miRNA activities and HeLa-high sensor responses (Fig. 2D and fig. S4). For the HeLa-high sensors, we found a qualitatively correct relationship between miRNA activity and the output of the double-inversion modules (fig. S5A). We also observed that the combined effect of all three HeLa-low sensors is generally stronger than a simple prediction according to individually measured miRNA activities (fig. S5B).

Fig. 2

Design of HeLa classifier. (A) Simulated output levels of a two-input circuit with miR-21 and miR-17-30a inputs. Each dot represents a particular cell type, and the contour line shows input combinations that result in 20% output relative to HeLa cells (red dot). Dots above the contour line represent “false-positive” cell types for this specific circuit configuration. The numbers on the axes are given in cloning frequency (CF) units. (B) CF of selected HeLa-low markers in false-positive cell types. (C) Circuit selectivity as a function of the number of markers. For a given number of markers (x axis), we computationally generate all possible classifier circuits by choosing subsets of markers from the set of six chosen markers and evaluate circuit output in HeLa cells and in the profiled healthy cell types. We then choose a circuit that maximizes the worst output ratio between HeLa and all non-HeLa cell types (“minimal”) and also calculate the average output ratio for this circuit (“average”). (D) miRNA marker sensor responses in different cell lines (tables S1 and S3). Bars show mean ± SD of normalized sensor output measured in three independent flow cytometry experiments (fig. S4). (E) Schematics of the complete classifier circuit. For simplicity, four adjacent miRNA target sites are shown as a wider box, and DNA/RNA species are lumped together. Two double-inversion modules for HeLa-high markers and three sensors for HeLa-low markers are shown. rtTA crosstalk is indicated with dotted lines. The logic “computed” by the circuit is shown on top. (F) Optimized sensor configuration for HeLa-high markers. (Inset) Simplified network diagram of the sensing process. miR-FF4 is a synthetic microRNA co-transcribed with LacI.

Next, we validated the response of the complete circuit (Fig. 2E) in HeLa cells by systematically modulating all inputs and measuring the output for a total of 25 = 32 combinations of ON and OFF inputs (fig. S6). Although the results generally agree with expectation in all 32 cases, we detected undesirably high output levels in the three conditions in which one of the HeLa-high markers and all HeLa-low markers were set to OFF. To reduce this leakiness, we introduced in a HeLa-high sensor an additional posttranscriptional repression element by engineering intronic miRNA to coexpress with LacI (25, 26), improving the worst case ON:OFF ratio of the circuit from approximately 4- to ~10-fold (Fig. 2F and fig. S7).

Subsequently, we tested subcircuits of increasing complexity in the different cell lines (Fig. 3A and figs. S8 and S9). In any cell line, the upper bound on circuit output is obtained when a red fluorescent protein DsRed is constitutively expressed from unrepressed CAGop promoter, whereas the lower bound is obtained when DsRed is repressed by fully active double-inversion modules and endogenous HeLa-low markers. Constitutive CAGop expression varies widely among the cell lines (Fig. 3A, top row), but repression was uniformly efficient (Fig. 3A, middle row). For CAGop-driven DsRed with the three HeLa-low sensors (“all-low” circuit), and an “all-high” circuit with only HeLa-high sensors, output expression levels normalized to the output upper bound in each cell line (fig. S8) were generally consistent with expectation. In the all-low circuits, in many cases we actually observed an improvement over a simple prediction based on individual miRNA activities (fig. S5C). In all-high circuits, the performance seems to reflect an additive function of the individual sensor responses (fig. S5D), suggesting that the circuit operates in a gradual fashion in at least some of the cell lines. Despite occasional deviations of subcircuit behavior from simple computational predictions, the complete circuit performs as desired experimentally, resulting in an output that is substantially higher in HeLa cells both in absolute and relative units as compared with the other cell lines in the set (Fig. 3A, bottom row, and fig. S5E). Analysis reveals that for the cell types assayed, the circuit operates well because of one of two conditions: At least one sensor exhibits a response in a non-HeLa cell that is an inverse of that sensor’s response in a HeLa cell [human embryonic kidney (HEK) 293, MCF-7, and SH-SY5Y], or a sufficient number of responses deviate moderately from the ones observed in HeLa cells (DAOY, SKBR3, and T47D). The classifier’s output in T47D is much better than the prediction, suggesting that further model refinement may be necessary.

Fig. 3

The classifier circuit identifies and specifically kills HeLa cells (tables S3A and S3B). (A) Fluorescence reporter assays. (Left to right) Schematics of tested circuits, representative flow cytometry scatter plots, and absolute and relative output levels. Each bar shows mean ± SD from at least three independent replicates (22). (B) Modifications to the circuit for apoptotic protein output, in which DsRed is replaced with a gene for hBax protein. LacI protein is cotranslated with an hBax inhibitor bcl2 by using a 2A linker (only one LacI construct is shown). (C) Apoptosis assays in HeLa and HEK293 cell lines. The CAG-AmCyan transfection marker indicates cell survival. AmCyan+ fraction—percentage of AmCyan-positive cells gated using untransfected cells as a reference (fig. S8)—was measured 4 days after transfection by means of fluorescence-activated cell sorting (FACS). “No cell death” indicates cells cotransfected with an all-low circuit with a DsRed output. hBax-Tgts, cells cotransfected with an all-low circuit with a hBax output; circuit, complete classifier circuit. Each bar represents mean ± SD of the AmCyan+ fraction from three independent replicates. (Right) Comparison of the circuit-killing efficiency for HEK293 versus HeLa cells (22).

Lastly, we tested whether the circuit can selectively trigger apoptosis by regulating expression of human Bcl-2–associated X protein hBax (27). Because constitutive expression of hBax resulted in efficient killing only in HeLa and HEK293 cells, we chose to focus on this pair. To quantify circuit-induced cell death (5), constitutively expressed cyan fluorescent protein (AmCyan) transfection/survival reporter and the apoptosis-inducing classifier (Fig. 3B) were cotransfected into cells. We observed that among transfected cells, apoptosis-inducing classifier circuit inflicts almost the same degree of cell death in HeLa cells as does constitutively expressed hBax but causes little cell death in HEK293 cells, indicating selective actuation (Fig. 3C).

An important measure of circuit performance is specificity and selectivity in heterogeneous cell populations. Cocultured fluorescent HEK293-Cerulean cells (22) and unmodified HeLa cells were transfected with various circuits expressing DsRed output (Fig. 4A and figs. S10 and S11A). In the mixture transfected with the complete circuit, the DsRed+ population was strongly biased toward HeLa cells relative to the control, as expected. To test selective induction of HeLa cell death in a cell mixture, CAG-driven transfection/survival reporter DsRed was cotransfected with the apoptosis-inducing classifier to cocultured HEK293-Cerulean and HeLa-EYFP cells (Fig. 4B and fig. S11B) (22). The classifier with hBax results in markedly stronger apoptosis of transfected HeLa-EYFP cells compared with HEK293-Cerulean cells. Taken together, the results confirm that programmed detection and apoptosis of HeLa cells operate correctly in a cell mixture. However, the observed degree of false-positive cell detection and cell death as well as false-negative cell survival warrants continuing circuit optimization. In part, it can be explained by high circuit output in HEK293 cells due to high CAGop activity in these cells relative to HeLa that could be alleviated by using a constitutive promoter with more uniform activity in different cell lines.

Fig. 4

Fluorescent reporter assays and killing experiments in cell mixtures (table S4). (A) Fluorescent reporter assays. The histograms show the fractions of HeLa and HEK-Cerulean in the DsRed+ cell population. The bar chart shows the absolute numbers of DsRed+ cells (mean ± SD) under the indicated conditions. (B) Apoptotic assays in cell mixture. The scatter plots show the fractions of HeLa-EYFP and HEK293-Cerulean cells among the DsRed+ cell population interpreted as surviving transfected cells. The bar chart shows the absolute number of surviving HEK293 and HeLa DsRed+ cells (mean ± SD) under the indicated conditions.

We show that synthetic biological networks and potentially their combinations (fig. S12) can trigger programmed biological actuation when complex intracellular conditions are detected. The results suggest potential therapeutic usage for the circuit, provided that challenges such as efficient in vivo DNA delivery to cells (28) are overcome. At the same time, in vitro applications that include drug screening or monitoring of developmental processes may be enabled in the near term by means of stable integration of classifier circuit into cultured cells. In addition, our engineered convergent and divergent control by miRNA together with complex interactions between miRNA and transcription factors validates recently suggested endogenous roles of miRNA in regulatory pathways (29, 30). Such synthetic transcriptional/posttranscriptional gene circuits will continue to be complementary to existing tools for elucidating mechanisms of natural biological processes.

Supporting Online Material

Materials and Methods

SOM Text

Figs. S1 to S13

Tables S1 to S6

References (31–37)

References and Notes

  1. Materials and methods are available as supporting material on Science Online.
  2. Acknowledgments: We thank M. Leisner, J. Lohmueller, L. Bleris, and the members of FAS Center for Systems Biology for discussions; G. Winer for critical reading of the manuscript; and M. Leisner for CAGop-ZsYellow plasmid. We thank B. Tilton, P. Rogers, and G. Paradis for assistance with flow cytometry. The research was funded by the Bauer Fellows program, the National Institute of General Medical Sciences grant GM068763 for National Centers of Systems Biology, U.S. Army Congressionally Directed Medical Research Programs grant W81XWH-09-1-0240 BC085163, National Cancer Institute grant 1R01CA155320-01, and ETH Zürich core funding. Y.B., Z.X., and R.W. conceived of the ideas implemented in this study. Z.X. performed most of the experiments and data analysis. L.W. constructed stable, cell lines and performed experiments with MCF-7 cells and initial testing of the hBax protein and the 2A linker. L.P. performed experiments and data analysis in HEK293, HeLa, and T47D cells. Y.B. and R.W. supervised the project and performed data analysis. Y.B., Z.X., and R.W. wrote the paper. Raw and processed data are stored on our laboratory servers and are available upon request. A U.S. patent concerning the technology described in this paper has been filed by Harvard University and MIT entitled “Multiple Input Biologic Classifier Circuits for Cells.”
View Abstract

Navigate This Article