$ cat projects/crispr-quantum/method.md
CRISPR networks — method
What the paper did, in its own terms: the data, the three architectures, and how they were trained and scored. Each section cites the pages of the paper it comes from.
datasets and systematic screening
The datasets were extracted from the DepMap project and include genetic dependency data (CRISPRGeneDependency) and gene copy-number data (OmicsCNGene), representing the genetic dependencies of hundreds of cancer cell lines. The data were analysed so that there were no redundant records, and the training and test sets were adjusted to allow experiments at different scales. Feature engineering selected the most relevant variables, reducing the dimensionality of the dataset through correlation-based filtering. Figure 2 shows the pipeline of the systematic screening of the relationship between CRISPR gene dependency (Dep) and gene copy number (CNGene) in the CCLE (Cancer Cell Line Encyclopedia) project.
Preprocessing and exploratory data analysis used Pandas, Matplotlib, Seaborn, NumPy, and SciPy. Scatter plots showed the correlations between CRISPR knock-out effects and gene copy-number levels across cell lines, and a Spearman correlation matrix between CRISPR genetic dependency and gene copy number was used to compare diagonal and off-diagonal values. The paper reports a Wilcoxon test with a mean test statistic of 112,246.404 and a mean p-value of 2.64e-43 (printed as 112.246,404 and 2,64e-43), which it interprets as a strong and highly significant association between CRISPR genetic dependency and gene copy number that merits further investigation.
model architectures (classical vs hybrid)
Three neural network architectures were built using TensorFlow, Keras, and PennyLane: Classical NNW (14,814 total parameters), Partially Quantum Hybrid NNW (14,790 total parameters), and Fully Quantum Intermediate NNW (14,814 total parameters). All models feature an 8-neuron input layer with ReLU activation and a 6-neuron densely connected output layer.
The Classical NNW consists of an 8-neuron input layer, three hidden dense layers of 8 neurons with ReLU, and a 6-neuron output layer (parameters: 14,544; 72; 72; 72; 54).
The Partially Quantum Hybrid NNW (H-NNW) integrates one quantum intermediate layer built with PennyLane between classical layers (Dense 8, KerasLayer 8 with 48 parameters, Dense 8 with 72 parameters, Dense 8 with 72 parameters, Dense 6 with 54 parameters). The quantum circuit is defined with the qnode function and consists of an AngleEmbedding layer followed by StronglyEntanglingLayers; in the description of Figure 5, the inputs correspond to the RxEntrada qubits, the AngleEmbedding refers to the first three columns (rz, ry and rz theta) with a randomly generated initial theta, the StronglyEntanglingLayers represent the intermediate values (Rz, Ry and Rz theta) associated with the quantum layer's weights, and the final value is obtained by measuring all qubits with the expectation of the Pauli-Z operator.
The Fully Quantum Intermediate NNW (HQNNW) contains three consecutive quantum layers (Dense 8 with 14,544 parameters, followed by three KerasLayer QNodes each of shape (None, 8) and 72 parameters, ending in Dense 6 with 54 parameters).
training and hardware simulation setup
Experiments were developed in Python Notebook using IBM Quantum, PennyLane, and Qiskit. Quantum circuits were executed locally using classical hardware simulation via the PennyLane API to emulate quantum circuit behavior across 20 training epochs.
Tools named in the paper: Python · Python Notebook · IBM Quantum · PennyLane · Qiskit · TensorFlow · Keras · Pandas · NumPy · SciPy · Matplotlib · Seaborn.
evaluation metrics
Model performance was evaluated using execution time (reported as CPU Time and Wall Time), Accuracy, Loss, Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), Mean Squared Logarithmic Error (MSLE), Cosine Similarity, Log-cosh, and R2.
source
Marcus Navarro Gabrich, Henrique Cota de Freitas, Matheus Alcântara Souza. Análise de Redes Neurais para CRISPR: Uma Abordagem com Computação Quântica (Neural Network Analysis for CRISPR: A Quantum Computing Approach). Anais do XXV Simpósio em Sistemas Computacionais de Alto Desempenho (SSCAD 2024). doi:10.5753/sscad.2024.244778