Limits...
Fuzzy logic analysis of kinase pathway crosstalk in TNF/EGF/insulin-induced signaling.

Aldridge BB, Saez-Rodriguez J, Muhlich JL, Sorger PK, Lauffenburger DA - PLoS Comput. Biol. (2009)

Bottom Line: Simulations based on fuzzy logic recapitulate most features of the data and generate several predictions involving pathway crosstalk and regulation.We also find unexpected inhibition of IKK following EGF treatment, possibly due to down-regulation of autocrine signaling.More generally, fuzzy logic models are flexible, able to incorporate qualitative and noisy data, and powerful enough to produce quantitative predictions and new biological insights about the operation of signaling networks.

View Article: PubMed Central - PubMed

Affiliation: Center for Cell Decision Processes, Cambridge, Massachusetts, United States of America.

ABSTRACT
When modeling cell signaling networks, a balance must be struck between mechanistic detail and ease of interpretation. In this paper we apply a fuzzy logic framework to the analysis of a large, systematic dataset describing the dynamics of cell signaling downstream of TNF, EGF, and insulin receptors in human colon carcinoma cells. Simulations based on fuzzy logic recapitulate most features of the data and generate several predictions involving pathway crosstalk and regulation. We uncover a relationship between MK2 and ERK pathways that might account for the previously identified pro-survival influence of MK2. We also find unexpected inhibition of IKK following EGF treatment, possibly due to down-regulation of autocrine signaling. More generally, fuzzy logic models are flexible, able to incorporate qualitative and noisy data, and powerful enough to produce quantitative predictions and new biological insights about the operation of signaling networks.

Show MeSH

Related in: MedlinePlus

Fitting MK2 rule weights.(A) A heatmap depicts the data, untrained model (Figure 3B), and trained model time                            courses for MK2. (B) The regressed rule weights are plotted for the 12                            candidate rules. The rules are indicated in tabular format; the first                            two rows describe the state of the inputs, TNF and time, and the last                            row is the output MK2 state. L and H represent low and high states, and                            E is the state describing the early response lag. Symbols above the plot                            show whether the rules were present (✓) or not applicable () in the untrained model.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2663056&req=5

pcbi-1000340-g007: Fitting MK2 rule weights.(A) A heatmap depicts the data, untrained model (Figure 3B), and trained model time courses for MK2. (B) The regressed rule weights are plotted for the 12 candidate rules. The rules are indicated in tabular format; the first two rows describe the state of the inputs, TNF and time, and the last row is the output MK2 state. L and H represent low and high states, and E is the state describing the early response lag. Symbols above the plot show whether the rules were present (✓) or not applicable () in the untrained model.

Mentions: In the work described above, logic rules and membership functions for each gate were established manually. A better approach is to use training to optimize the weights of all possible rules in a gate by minimizing the sum of the squared differences between the experimental data and local model output (see Methods). Following optimization, logic rules that are supported by the data should have weights near 1, while poorly-supported rules should have weights near 0. We tested the fitting algorithm on the MK2 gate. For such a gate, which has two MFs each for the two inputs (TNF and time) and the output (MK2 activity), 23 = 8 explicit rules are possible. MK2 data from the 10 cytokine treatment conditions were used to optimize a vector containing the 8 rule weights. Our initial optimization attempt failed because time-dependent MFs were not parameterized so as to capture rapid increases in signals following cytokine treatment. We had implicitly ignored this discrepancy when fitting the model by hand. To improve the automated fitting procedure, an additional MF for time was included to represent immediate-early responses, increasing the number of candidate rules to 12. Optimization yielded a gate with a good fit to data using only six rules with weights near one (Figure 7A). These six rules were identical to those assembled manually with the exception of the new rule needed to represent immediate early signaling (Figure 7B). To test FL gate regression with more rules, we applied the algorithm to the same MK2 data using one additional membership function (for medium activity levels) and compared it to an untrained model using the same MFs. The training process created several rules that were nearly identical to those introduced manually as well as several new ones (Figure S1). The MK2 test case suggests that it is possible to optimize rule weights as a means to fit logic rules without bias and is a first step towards a more rigorous approach to logic-based modeling.


Fuzzy logic analysis of kinase pathway crosstalk in TNF/EGF/insulin-induced signaling.

Aldridge BB, Saez-Rodriguez J, Muhlich JL, Sorger PK, Lauffenburger DA - PLoS Comput. Biol. (2009)

Fitting MK2 rule weights.(A) A heatmap depicts the data, untrained model (Figure 3B), and trained model time                            courses for MK2. (B) The regressed rule weights are plotted for the 12                            candidate rules. The rules are indicated in tabular format; the first                            two rows describe the state of the inputs, TNF and time, and the last                            row is the output MK2 state. L and H represent low and high states, and                            E is the state describing the early response lag. Symbols above the plot                            show whether the rules were present (✓) or not applicable () in the untrained model.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2663056&req=5

pcbi-1000340-g007: Fitting MK2 rule weights.(A) A heatmap depicts the data, untrained model (Figure 3B), and trained model time courses for MK2. (B) The regressed rule weights are plotted for the 12 candidate rules. The rules are indicated in tabular format; the first two rows describe the state of the inputs, TNF and time, and the last row is the output MK2 state. L and H represent low and high states, and E is the state describing the early response lag. Symbols above the plot show whether the rules were present (✓) or not applicable () in the untrained model.
Mentions: In the work described above, logic rules and membership functions for each gate were established manually. A better approach is to use training to optimize the weights of all possible rules in a gate by minimizing the sum of the squared differences between the experimental data and local model output (see Methods). Following optimization, logic rules that are supported by the data should have weights near 1, while poorly-supported rules should have weights near 0. We tested the fitting algorithm on the MK2 gate. For such a gate, which has two MFs each for the two inputs (TNF and time) and the output (MK2 activity), 23 = 8 explicit rules are possible. MK2 data from the 10 cytokine treatment conditions were used to optimize a vector containing the 8 rule weights. Our initial optimization attempt failed because time-dependent MFs were not parameterized so as to capture rapid increases in signals following cytokine treatment. We had implicitly ignored this discrepancy when fitting the model by hand. To improve the automated fitting procedure, an additional MF for time was included to represent immediate-early responses, increasing the number of candidate rules to 12. Optimization yielded a gate with a good fit to data using only six rules with weights near one (Figure 7A). These six rules were identical to those assembled manually with the exception of the new rule needed to represent immediate early signaling (Figure 7B). To test FL gate regression with more rules, we applied the algorithm to the same MK2 data using one additional membership function (for medium activity levels) and compared it to an untrained model using the same MFs. The training process created several rules that were nearly identical to those introduced manually as well as several new ones (Figure S1). The MK2 test case suggests that it is possible to optimize rule weights as a means to fit logic rules without bias and is a first step towards a more rigorous approach to logic-based modeling.

Bottom Line: Simulations based on fuzzy logic recapitulate most features of the data and generate several predictions involving pathway crosstalk and regulation.We also find unexpected inhibition of IKK following EGF treatment, possibly due to down-regulation of autocrine signaling.More generally, fuzzy logic models are flexible, able to incorporate qualitative and noisy data, and powerful enough to produce quantitative predictions and new biological insights about the operation of signaling networks.

View Article: PubMed Central - PubMed

Affiliation: Center for Cell Decision Processes, Cambridge, Massachusetts, United States of America.

ABSTRACT
When modeling cell signaling networks, a balance must be struck between mechanistic detail and ease of interpretation. In this paper we apply a fuzzy logic framework to the analysis of a large, systematic dataset describing the dynamics of cell signaling downstream of TNF, EGF, and insulin receptors in human colon carcinoma cells. Simulations based on fuzzy logic recapitulate most features of the data and generate several predictions involving pathway crosstalk and regulation. We uncover a relationship between MK2 and ERK pathways that might account for the previously identified pro-survival influence of MK2. We also find unexpected inhibition of IKK following EGF treatment, possibly due to down-regulation of autocrine signaling. More generally, fuzzy logic models are flexible, able to incorporate qualitative and noisy data, and powerful enough to produce quantitative predictions and new biological insights about the operation of signaling networks.

Show MeSH
Related in: MedlinePlus