CSCI 2952G Enformer Quiz

Which statement explains how base pairs can interact to influence gene expression?

• Base pairs do not interact, so each pair should be assessed individually and without context

• It is best to only consider base pairs that are next to each other when predicting gene expression

• Base pairs that are not near each other sequentially can still influence each other

• Base pairs that are far apart are typically more interactive than base pairs that are close together

Which of the following is not a benefit of dilated convolution?

• It is less demanding on computational resources

• It can help understand long-range interactions more efficiently than normal convolution

• It is not a black-box model and instead can display causality

Which of the following correctly compares the Enformer model with attention layers, the Enformer model with dilated convolutional layers, and the Basenji2 model?

• There was no substantial difference between the three models tested

• The Enformer model with attention layers outperformed the convolutional Enformer in all test cases

• The Enformer model with dilated convolution was, on average, the best-performing model

• Neither Enformer model was able to improve on Basenji2

In the Enformer paper, "multi-head attention" is used. Which of the following is true?

• All of the attention heads were initialized with the exact same random weights to facilitate training.

• All attention heads had different random initializations to allow diversity to be learned

• The Enformer model trained with multi-head attention, but used a single attention head at inference.

• Each attention head deployed different dilated convolutions to allow for diverse self-attention.

In the ABC model for enhancer-gene matching, what does "A" stand for?

• Attention

• Activation

• Activity

• Athletics

The state of the art model that Enformer was extensively compared to in the paper was called:

Which of the following was not a task included in the Enformer paper's results section?

• Transcription factor binding prediction

• Enhancer-gene matching

• Narrowing disease-associated regions found from genome-wide association studies (GWAS)

• eQTL identification via in silico mutagenesis (ISM)

What is the main advantage of the use of multi headed attention layers over an LSTM?

• Long range dependency capture which avoids the vanishing gradient problem

• Ability to process input sequences in parallel

• A memory gate to keep track of long term dependencies

• They are essentially equivalent

What is the main advantage of the use of multi headed attention layers over an RNN?

• Attention can allow outputs to depend on any piece of the input without the vanishing gradient problem

• A memory gate to keep track of long term dependencies

• A recursive structure which is quick to train

Which of the following is a limitation of Enformer?

• The model is not able to take in sequences longer than 60 kbp

• The model cannot generalize to novel cell types or assays beyond those in the training set

• The model cannot generalize to predict eQTLs, TAD boundaries, or do enhancer-gene matching

• Enformer is limited by the inclusion of convolutions and could benefit from self-attention like Basenji2