Partial Least Squares Regression

Author

Aaron Meyer (based on slides from Pam Kreeger)

Explain unsupervised/supervised.
Can you bootstrap a PLSR/PCR/PCA model?

Common Challenge: Cue/Signal/Response Relationships

Number of components enough to make PCA/lasso useful
But how do we deal with components moving together?
And what we care about is some outcome

Many Methods for Relating a Signal to Response

Say we have some measurement from cells and how they respond:

\[ [1, 2, 1.5, 5, 6, 7] \sim [5, 10, 7, 24, 31, 35] \]

From the variation we can see that:

Low signal is correlated with low response
High signal is correlated with high response

If we can find a quantitative correlation between the input and output, we can predict new outcomes for measurements we haven’t yet seen.

Draw out correlation and then include new point.

Challenges with Univariate Relationships

The relationship between JNK activation and apoptosis appears to be highly context-dependent
- Univariate relationships are often insufficient
- Cells respond to an environment with multiple factors present

Notes about Methods Today

Both methods are supervised learning methods, however have a number of distinct properties from others we will discuss.
Learning about PLS is more difficult than it should be, partly because papers describing it span areas of chemistry, economics, medicine and statistics, with little agreement on terminology and notation.
These methods will show one example of where the model and algorithm are quite distinct—there are multiple algorithms for calculating a PLSR model.

Multi-Linear Regression (MLR)

In biology we often have multiple signals and multiple responses that were measured: \[ y_1 =a_1x_1 + b_1x_2 + e_1 \] \[ y_2 =a_2x_1 + b_2x_2 + e_2 \]

This can be written more concisely in matrix notation as: \[ Y = XB + E \]

Where Y is a \(n \times p\) matrix and X is a \(n \times m\) matrix; minimizing E and solving for B: \[ B = (X^tX)^{-1}X^tY \]

Underdetermined Systems

If \(n\) observations and \(m\) variables:

\(m<n\): no exact solution, least-squares solution possible
\(m=n\): one solution
\(m>n\): no unique solution unless we delete independent variables since \(X^tX\) cannot be inverted
- \(m>n\) is often the case in systems biology!

If a model is underdetermined with multiple solutions, there are two general approches we can take:

Regularization: We can use other information we know to focus on one answer
Sampling: We can look at all possible models

Regularization

Today we will use regularization.

We will assume the larger variation in the data is more meaningful.
Therefore, we will assume that smaller changes are less important.
This is a choice that must be correct for the relevant biological question at hand.

Principal Components Regression (PCR)

One solution - use the concepts from PCA to reduce dimensionality.

First step: Simply apply PCA!

Dimensionality goes from \(m\) to \(N_{comp}\).

What are the bounds on the number of components we can have?
What might be a concern as a consequence of this? (How do we determine how many components to keep?)

Principal Components Regression (PCR)

Decompose X matrix (scores T, loadings P, residuals E) \[X = TP^T + E\]
Regress Y against the scores (Scores describe observations – by using them we link X and Y for each observation)

\[Y = TB + E\]

What is the residual against here?
How do we invert the direction of the regression?

Challenge

How might we determine the number of components using our prediction?

Cross-validation with varying numbers of components
Metrics like AIC or BIC

Potential Problem

PCs for the X matrix do not necessarily capture X-variation that is important for Y
- So later PCs are going to be more important to regression
Example: the first components capture signaling that is related to another cell fate, while the signals that co-vary for this particular y are buried in later components

How might we handle this differently?

PLSR

Go over definition of covariance.
cov(X, X) = var(X)
var(Y) = cov(X, Y) + variance in Y independent of X
A factorization occurs to find max covar with Y
Forces direction of components

PLSR - NIPALs with Scores Exchanged

\(B\) is \(T (P^T T)^{-1} Q^T\)
\(P^T T\) is the approximation of \(X\)
PLSR components are not orthogonal in data space, but are orthogonal in O-PLS.

PLSR - NIPALs with Scores Exchanged

Components in PLSR and PCA Differ

Determining the Number of Components

Go over each.

Determining the Number of Components

Variants of PLSR

Discriminant PLSR

Tensor PLSR

Application

Practical Notes

PCR

sklearn does not implement PCR directly
Can be applied by chaining sklearn.decomposition.PCA and sklearn.linear_model.LinearRegression
See: https://scikit-learn.sourceforge.net/dev/auto_examples/plot_digits_pipe.html

PLSR

sklearn.cross_decomposition.PLSRegression
- Uses M.fit(X, Y) to train
- Can use M.predict(X) to get new predictions
- PLSRegression(n_components=3) to set number of components on setup
- Or M.n_components = 3 after setup

https://scikit-learn.org/stable/modules/generated/sklearn.cross_decomposition.PLSRegression.html

Summary

PLSR

Maximizes the covariance
Takes into account both the dependent (Y) and independent (X) data

PCR

Uses PCA as initial decomp. step, then is just normal linear regression
Maximizes the variance explained of the independent (X) data

Summary

Interpreting PLSR

\(R^2X\), \(R^2Y\), \(Q^2Y\) (maximum value of 1)
Using Q2Y to determine number of components Scores/loadings
DModY (lower = better predicton)
VIP (>1 indicates important)
Ultimately, these metrics are seconary to whether a model works upon crossvalidation

Summary

PLSR is amazingly well at prediction
- This is incredibly powerful
Interpreting WHY PLSR predicts something can be very challenging

Review Questions

What are three differences between PCA and PLSR in implementation and application?
What is the difference between PCR and PLSR? When does this difference matter more/less?
How might you need to prepare your data before using PLSR?
How can you determine the right number of components for a model?
What feature of biological data makes PLSR/PCR superior to direct regularization approaches (LASSO/ridge)?
What benefit does a two component model have over those with 3+ components?
Can you apply K-fold cross-validation to a PLSR model? If so, when do you scale your data?
Can you apply bootstrapping to a PLSR model? Describe what this would look like.
You use the same X data but want to predict a different Y variable. Do your X loadings change in your new model? What about for PCR?