Abstract
Receptor tyrosine kinases (RTKs) typically contain multiple autophosphorylation sites in their cytoplasmic domains. Once activated, these sites recruit downstream signaling proteins containing Src homology 2 (SH2) and phosphotyrosine-binding (PTB) domains. These domains recognize phosphotyrosine-containing short linear motifs (SLiMs) but exhibit polyspecific binding activities, leading to competition among signaling proteins. To investigate these effects, we developed a rule-based modeling approach for ligand-induced recruitment to the insulin-like growth factor 1 receptor (IGF1R). Using model restructuration to manage combinatorial complexity, we parameterized models with published protein copy numbers and binding affinities across 45 cell lines. Our simulations demonstrate that simple metrics like binding affinity or abundance fail to predict recruitment outcomes. Instead, recruitment is a complex network property that identifies correlated and anti-correlated binding partners, highlighting how early signaling events are fundamentally dependent on a cell's unique proteomic profile.
The Challenge of Promiscuous Signaling
In cellular signaling, Receptor Tyrosine Kinases (RTKs) like the insulin-like growth factor 1 receptor (IGF1R) act as critical docking stations. Once activated, they display multiple phosphorylation sites that recruit downstream signaling proteins. However, this recruitment is notoriously "promiscuous": many different proteins can bind to the same site, and one protein might bind to several different sites.
This creates a massive game of molecular "musical chairs." Because these proteins compete for limited space, the "winner" isn't always the one with the strongest grip (binding affinity) or the most players (abundance). Traditional models often oversimplify this by grouping proteins or ignoring the "combinatorial explosion"—the astronomical number of possible ways proteins can arrange themselves on a receptor. The authors sought to determine if recruitment could be predicted using simple rules of thumb or if detailed numerical simulations were necessary.
Methods: Rule-Based Modeling and Restructuration
The authors employed rule-based modeling using the BioNetGen language. Instead of writing an equation for every possible molecular state, they defined the rules of interaction. To make the math tractable, they introduced model restructuration. Without this, a single IGF1R dimer could exist in roughly $10^5$ unique configurations, making simulations impossible.
They used three specific strategies to shrink this "state space":
- Decoupling: Treating phosphorylation sites as independent entities.
- Bunching: Grouping ligand-binding sites into logical "pockets."
- Scaling: Adjusting abundances and rate constants to allow the model to run as a system of monomers rather than dimers while preserving the correct physics.
This reduced the model from hundreds of thousands of potential species to just 577, allowing for efficient simulation via Ordinary Differential Equations (ODEs).
Data Integration
The study integrated several high-quality proteomics and biochemical datasets:
- Binding Affinities ($K_D$): Derived from established studies characterizing SH2 and PTB domain interactions.
- Protein Copy Numbers: Proteomics data for 45 different cell lines, including HeLa S3 and the NCI-60 cancer cell panel. These provided the "starting lineup" for each cell-specific model.
- Ligand Binding Kinetics: The model was calibrated using experimental data on how IGF1 binds to its receptor over time.
Key Insights: The Intuition Gap
The most surprising finding was the "intuition gap": simple ranking methods—ranking by $K_D$, by copy number, or even by the ratio of the two—could not accurately predict the results of the full numerical simulation. Competition is non-linear; a protein might be a weak binder but "win" because its primary competitors are occupied elsewhere.
The simulations also identified pairs of proteins recruited in correlated or anti-correlated fashions. These correlations emerged purely from competition for shared sites, even though the model included no explicit "cooperativity."
Results and Cell-Type Specificity
Because every cell line has a different "proteomic profile," the same IGF1R receptor can trigger entirely different downstream pathways depending on the cell type. For example, the protein YES1 was the top-ranked binder in 38% of the cell lines but ranked much lower in others. Across the 45 cell lines, CRKL, STAT1, RASA1, and YES1 were identified as the most consistently highly recruited partners.
The restructured model maintained "exactness"—it produced the same results as the complex "natural" model but with a fraction of the computational cost. The authors showed that their mechanistic model could explain why certain proteins (like SHC1) are dominant in HeLa cells due to a combination of high copy number and a 20-fold higher affinity compared to its closest competitor (IRS1).
Limitations and Future Work
The model currently does not account for receptor trafficking (how receptors move from the cell surface to the interior), which can change signaling duration. It also stops at recruitment, ignoring downstream feedback loops. The "decoupling" strategy assumes that binding at one site doesn't physically block or help binding at a neighboring site. Future directions include incorporating mass-spectrometry-based phosphoproteomics to validate site occupancy in living cells, potentially leading to more personalized cancer treatment strategies.
Summary updated