To run, install the conda environment from env.yml
Jupyter notebook demo: Assignment 4/main.ipynb
We chose to use a 1D convolutional neural network, followed by a single dense layer, to learn meaningful embeddings for protein sequences. Namely, we chose to predict the next amino acid from contexts of length 20 amino acids, extracted from known protein sequences. Even though we had access to a GPU, we limited the training set size signficantly and only trained for 10 epochs. The training loss seems to converge, although it is not very low, and the model only does somewhat better than randomly guessing (as calculating by the log likelihood of the cross entropy loss).
The model could be improved by making it more expressive (increasing the hidden dimension, adding layers, training for more epochs). A transformer archiecture may also work better than a 1D CNN. However, for the sake of this assignment, we did not explore these more computationally costly approaches. At a larger scale, it will be more important to consider memory requirements and make the processing of protein sequences to context encodings more efficient.
Generally, similar amino acids are grouped together. For example, G is the only amino acid without a side chain, and it is isolated in the emebedding space. Aromatic amino acids such as F and Y are close to one another, as are those with polar hydroxyl groups (S and T). Negatively charged and hydrophobic amino acids are also generally closer to one another. Thus, even with a limited amount of training on a small dataset, the learned embeddings capture differences in amino acids.
Weights saved from the logistic regression model: Assignment 1/model.sav
Function to test code (used for grading): run_test_code from Assignment 1/util.py
Jupyter notebook demo: Assignment 1/main.ipynb
We performed multiclass classification of individual cells into one of 18 cell types, based on the mean expression profile across the cell, for 51 different markers. From marker expression panel analysis, some of the markers with the biggest impact include Au; Na; and the sum of H3K27m3, H3K9ac, and dsDNA. Most of the markers have relatively low impact.
We then trained a logisitic regression model to predict the classification of each individual cell, based on the mean expression profile. While this is a very basic model, as it is linear, it still performed decently. For most cell types, the the correct cell type was predicted most often. However, notably, the model struggles to correctly classify endothelial, Nk, and Mono-Neu cells, often getting them confused for other cells.
We suggest that the performance of the model could be improved by introducing a more expressive model, such as a neural network. Furthermore, regularization could be used to reduce the weights in some of the less important marker features, reducing overfitting.