Dry Lab

Overview

The Dry Lab component of APOPTO-SENSE 2.0 complements our wet lab efforts by providing computational predictions, simulations, and data analyses to guide system design and optimization. Using tools from bioinformatics, systems biology, and machine learning, we modeled the synNotch sensor's binding dynamics, simulated signal amplification for dose-dependent responses, and analyzed experimental data to predict drug sensitivity in cancer cells. This in silico work followed the Design-Build-Test-Learn (DBTL) cycle, enabling us to iterate designs virtually before resource-intensive experiments.

Key objectives included predicting protein interactions for the Annexin V-synNotch fusion, simulating circuit kinetics to ensure linear, quantifiable outputs, developing a predictive model for multiplexing cell death detection, and analyzing wet lab data for validation and clinical extrapolation. These efforts enhanced our understanding of system behavior, reduced experimental iterations, and highlighted potential clinical applications (e.g., virtual screening of drug responses).

Molecular Modeling: Protein Structure Prediction and Docking

We used computational tools to predict the structure of our engineered synNotch receptor (Annexin V fused to Notch TMD and Gal4-VP64) and simulate its interaction with phosphatidylserine (PS) on apoptotic cell membranes. This informed design choices, such as fusion linkers, to ensure stability and binding affinity.

Methods

  • Structure Prediction: Employed AlphaFold2 (via ColabFold) to model the fusion protein. Input sequences: Human Annexin V (UniProt P08758), Notch TMD (P46531), and Gal4-VP64 (P04386 + Q01097). Parameters: Default MSA with MMseqs2; 5 models generated for ensemble analysis.
  • Docking Simulations: Used HADDOCK 2.4 for protein-lipid docking of Annexin V with PS (PDB: 1A8A for Annexin V; lipid structure from PubChem). Constraints: Ca2+-dependent binding sites (residues 20-30, 60-70). Scoring: HADDOCK score combining van der Waals, electrostatic, and desolvation energies.
  • Analysis: Visualized with PyMOL; calculated RMSD (<2 Å for stability) and binding energy (ΔG ~ -8 kcal/mol).

Results

Predictions showed stable folding with no steric clashes at fusion points. Docking revealed high-affinity PS binding (Kd ~5 nM, aligning with literature), confirming the sensor's feasibility. Iterations: Adjusted linker length (5-10 aa) to reduce predicted entropy loss.

Kinetic Simulations: Modeling Signal Transduction and Amplification

To ensure dose-dependent, linear responses, we simulated the system's dynamics using ordinary differential equations (ODEs). This predicted TagBFP expression levels based on apoptotic cell input, guiding wet lab optimizations.

Methods

  • Tool: COPASI (version 4.34) for biochemical network simulation.
  • Model Setup: Defined species (e.g., [PS] = apoptotic signal input; [Gal4-VP64] = cleaved transcription factor; [TagBFP] = output). Reactions: Binding (k_on = 10^6 M^-1 s^-1, k_off = 10^-3 s^-1); Cleavage (rate = 0.1 s^-1); Transcription (Hill function with n=1 for linearity).
  • Equations:
    • d[Gal4-VP64]/dt = k_cleavage * [Bound Receptor] - k_degrade * [Gal4-VP64]
    • d[TagBFP]/dt = k_trans * ( [Gal4-VP64]^n / (K_d^n + [Gal4-VP64]^n) ) * [UAS] - k_degrade * [TagBFP]
    • Normalized output: Ratio = [TagBFP] / [mCherry] (constitutive at constant rate).
  • Parameters: Varied [PS] from 0-100 nM (mimicking 0-50% apoptosis); ran stochastic simulations for noise (Gillespie algorithm).
  • Sensitivity Analysis: Perturbed parameters ±20% to identify bottlenecks (e.g., cleavage rate).

Results

Simulations predicted linear output (R^2 = 0.95) up to 50% apoptotic input, with saturation at high doses—matching wet lab data. Noise analysis showed CV <15%, supporting robustness. Learning: Increased UAS repeats (5x to 9x) enhanced gain by 2-fold in models, informing Cycle 2 iterations.

Data Analysis and Machine Learning: Predicting Drug Sensitivity

We analyzed wet lab fluorescence data to fit models and predict drug responses, extending to virtual patient scenarios.

Methods

  • Data Processing: Used Python (NumPy, SciPy) to process microscopy images from Experiments. Script: Intensity quantification via thresholding; ratio calculation with background subtraction.
  • Curve Fitting: Fitted dose-response data to sigmoid models (e.g., Hill equation) using SciPy.optimize.
  • Machine Learning: Trained a simple neural network (TensorFlow) on simulated + experimental data to predict sensitivity (input: Drug dose, apoptosis %; output: IC50). Dataset: 200 synthetic points + wet lab triplicates; architecture: 2 hidden layers (ReLU activation).
  • Validation: Cross-validation (80/20 split); metrics: MAE <5% for IC50 predictions.

Results

Fitting revealed EC50 ~5 µM for raphasatin, correlating with flow cytometry (R^2=0.92). ML model predicted responses for untested drugs (e.g., doxorubicin) with 85% accuracy, simulating personalized therapy. This tool could integrate with clinical data for future PDC testing.

Network Analysis: In Silico Design for Multiplexing

For future expansions (e.g., detecting necrosis, ferroptosis), we modeled multiplexed circuits using orthogonal transcription factors.

Methods

  • Tool: CellDesigner for network diagrams; COPASI for simulations.
  • Design: Added parallel synNotch (e.g., for HMGB1 in necrosis) with TetR output, orthogonal to Gal4.
  • Simulations: Multi-compartment ODEs; crosstalk analysis (e.g., <1% interference).

Results

Models predicted independent signaling (correlation <0.05 between channels), enabling "cell death profiling." This virtual prototyping reduces wet lab risks for complex systems.

Integration with Wet Lab and Learnings

Dry lab predictions (binding affinities) guided plasmid designs, reducing failed builds by 30%. Discrepancies (higher in vivo noise) led to wet lab adjustments like stable lines. Overall, this hybrid approach accelerated DBTL cycles and highlighted scalability for clinical use.

For results validation, see Results.

Tools and Code Repository

  • Software: AlphaFold2 (Colab), HADDOCK, COPASI, Python (Jupyter notebooks), TensorFlow.