Algorithms
This section aims to explain briefly some of the algorithms used in pixelator. For extended explanations, if relevant, we try to include references in each sub-section.
Edge Collapsing
From the sampled NGS reads, pixelator tries to reconstruct the original molecules that have been assayed by MPX. In order to achieve this, we use the combination of the antibody barcode, the UMI and UPIA to uniquely associate all UPIBs that show that fragment combination in the reads. Once that is done, after translating each nucleotide sequence to a binary sequence (using 2-bit encoding), we use Annoy (Approximate Nearest Neighbors Oh Yeah) to cluster those fragments at a certain Hamming distance from each other (default: 2 mismatches). The algorithm will then group a set of Ab/UMI/UPIA fragments into a different groups where all fragments are separated just by less than the given Hamming distance from each other.
The reason to use Annoy is that many to many comparisons takes a lot time. Performance in a worst case scenario is O(N^2) algorithm. On the other hand, Annoy is very quick and reduces the time complexity to O(log N).
Along the groups of fragments, pixelator creates a frequency table and we pick the most frequent UPIB and associate that with the Ab/UMI/UPIA fragment. That fragment combination with most frequent UPIB becomes the representative molecule or edge from each of the groups. Pixelator stores all edges in a large table, and for each edge it computes the total count of UPIBs (associated to the number of reads of a certain molecule), the number of fragments, and the total number of unique UPIBs.
The number of resulting edges is dependant on many factors: the origin of the cells, the experimental conditions (e.g. any stimulation of the cells), the number or reads sequenced for the library, and the number of cells in the experiment.
Multiplet recovery
From the edge list, pixelator generates an undirected bipartite graph where each Ab/UMI combination forms an edge, and each UPIA and UPIB form a node. Each graph is a representation of a putative cell but at this stage of the pipeline we use the concept of a component to refer to them.
Due to the complex nature of the assay biochemistry and the fact that the reaction is carried on a single tube, MPX edge lists may have a proportion of undesired artefactual edges. Those few edges (only a small proportion of the total edge number of the experiment) are removed in a step we call multiplet recovery in the graph
pipeline stage if the --multiplet-recovery
removal flag is activated (by default it is on).
This removal is achieved by using community detection algorithms that operate on the graph. Densely connected communities are assumed to represent cells and pixelator removes any spurious edges connecting them. Each component of the resulting edge list is assigned a recovered ID and pixelator saves a table of all the recovered IDs in relation to original IDs.
Edge Rank Plot and Cell Calling
The edge rank plot is a useful tool to visualize the range and distribution of antibody counts (edges) across cells called by pixelator, and to help you to assess the data quality. The edge rank plot is constructed by visualizing all graph components (cells) ranked from highest to lowest by the number of unique antibodies (edges). Pixelator uses the edge rank plot internally to perform an automatic cell calling by finding the threshold point which distinguishes real cells from a background of small spurious components. The threshold is recognized as the elbow point, where the component size (edges) rapidly declines in relation to the rank. Although this process is automatic, it is advisable to plot the edge rank plot and set a manual cutoff to refine the selection of cells. The cutoff should be selected to approximate the elbow point where the component size is sharply declining in relation to the size rank. The number of edges that corresponds to an appropriate cutoff will vary depending on the dataset, as the number of molecules detected varies by multiple factors, such as sequencing depth, cell type, cell size, and cell states, such as whether the cell has undergone any type of stimulation or not.
Antibody Count Distribution Outliers
Cell components might rarely have a diverging distribution of counts across proteins, such that the counts might be skewed to a single
protein or evenly distributed to most proteins. This can be an indication that the component is not originating from
a cell, and could instead originate from debris or from an antibody aggregate. Antibody aggregates form rarily,
and usually have either high complexity, consisting of wide range of different antibodies, or low complexity, mainly constituted
by a single antibody. Pixelator detects these by using the metric Tau (tau
), adjusted from Yanai et al.,
which measures the skewness of antibody counts across different antibodies for a given component. Tau is a numerical value
ranging from 0 and 1 describing the degree to which skewness the distribution of counts across markers. A Tau of 0 indicates
equal distribution of counts across all antibodies and a Tau of 1 indicates that all counts are distributed to a single antibody.
Pixelator classifies components by their Tau score in tau_type
as high
if the Tau score is above 0.995 or deviates from the
population median with more than 2 interquartile ranges (IQRs), and as low
if the Tau score is lower than 5 IQRs from the
population median. If the tau_type
is neither high
or low
, the tau_type
is set to normal
. A typical down-stream
quality control will involve removing components marked as high
or low
tau_type
.
Additionally, aggregates often have dense pixels with higher levels of antibodies detected per each pixel, which can be measured
using the mean_umi_per_upia
metric stored in the component meta data.
MPX Polarity Scores
The polarity score aims to describe the degree of spatial clustering a protein exhibits upon the surface of a cell. It
is calculated as Moran's I (morans_i
), a statistic of spatial autocorrelation originally developed by Patrick Alfred Pierce Moran.
In the context of MPX, morans_i
measures the degree of spatial clustering of a protein within the graph component (the cell),
and ranges from -1 (perfectly dispersed), to 1 (perfectly clustered), where 0 represents no spatial autocorrelation. morans_i
can be represented as a z-score (morans_z
), which is calculated as the number of standard deviations the Moran's I statistics
deviates from the expected mean of an analytical random distribution. The Z-score aims to normalize the polarity score to make the metric comparable between cells,
but the two metrics are correlated in MPX data, and often yields similar result.
MPX Colocalization Scores
The MPX colocalization score is a metric to assess the degree of cooccurence of a pair of proteins within small
neighborhoods upon a cell. The scores are calculated as 1) the Pearson correlation (pearson
) and 2) Jaccard index (jaccard
)
between counts of two markers within a neighborhood of one A pixel and its immediate neighboring A pixels. Pixelator
precalculates colocalization scores and performs a Monte Carlo simulation, to recast the score to reflect its deviation
from the degree of colocalization that would be expected by random chance. The resulting score is a Z-score where values
above zero indicate higher degree of colocalization than would be expected by random chance, while lower than zero indicates
less colocalization than expected by random chance.
Local
Local ("gi"), also known as Getis-Ord statistic, is a statistic used to find regions of a given study area where some measured variable of interest clusters. A classical application of this statistic is to identify local spatial autocorrelation patterns or "hot spots" where the variable of interest takes higher or lower values than expected. The study area is typically represented as a map, with measurements taken at specific sites. Using information about the relative distances between the sites, we can define local neighborhoods around each site. The local statistic then quantifies the degree to which the variable of interest clusters within these neighborhoods relative to the entire study area.
In MPX data analysis, the study area is a graph and the measured values are protein marker counts. Here, the marker counts are found in the nodes of the graph, and while we do not know the exact spatial positions of the nodes, we can define neighborhoods using the edges of the graph instead. One useful application of local in this context is to identify regions of a cell component graph where a marker of interest has a higher abundance than expected which is the case for polarization events.
Low-abundance markers can be of significant interest but are typically sparsely distributed in a cell component graph, making it challenging to identify and visualize spatial patterns such as polarization events. By summarizing abundance information across neighborhoods, the local statistic reduces the limitations of sparse data and becomes a valuable tool for creating more interpretable 3D visualizations of polarized marker expression.
Local is a Z-score that quantifies the deviation of the observed local marker expression from the expected expression under the null hypothesis of no spatial clustering. The sign of the score indicates whether the observed marker counts are higher or lower than expected, thus identifying hot and/or cold spots.
The local metric is calculated using the following formula (Ord and Getis, 1995, equation 6):
where
and
In this equation for , the condition that is central. An alternative definition, denoted ("gstari") relaxes this constraint, by including as a neighbor of itself. This statistic is expressed as (Ord and Getis, 1995, equation 7):
where
and
Both methods ("gi" and "gstari") are implemented in the pixelator (Python) and pixelatorR (R) analysis tool kits.
If we apply local or local on cell component data, corresponds to node counts for a selected protein marker. The statistic is calculated for a single node , using information about protein marker abundance () in the local neighborhood of . This neighborhood could for example consist of the direct neighbors to in the graph. We can also expand the neighborhood around to include nodes that are up to steps away. The effect of increasing the neighborhood size is that the spatial scale is increased, and we can identify larger spatial patterns in the graph. For visualization purposes, increasing can be an effective strategy to highlight large spatial patterns such as polarization events.
Local G scores reveal a polarization pattern in a cell component graph. The figure displays data for three protein markers: CD37, CD50, and CD162. The top row presents the log1p-transformed counts for each marker, while the bottom row illustrates the corresponding local G scores. Each point corresponds to a node in the graph.
Weights in Local
The marker counts in nodes of a component graph are correlated with the degree of the nodes. If this is not accounted for, high degree nodes will contribute more to the statistic. Additionally, if we set to a value larger than 1 to include neighbors that are more than 1 step away from , we need to assign lower weights to nodes that are further away. For these reasons, the Local G algorithm that we provide in pixelator (Python) and pixelatorR (R) also computes edge weights between and its neighbors (). These edge weights are based on transition probabilities between nodes, and aim to correct for the degree bias and the distance between nodes within the neighborhood.
MPX layouts
Cell component graphs can be visualized in 3D layouts using graph drawing techniques. These often aim to create layouts that are visually intuitive, making it easier to identify and interpret group structures and relationships among nodes, highlighting key features such as clusters, hierarchies, and pathways within the data. For instance, in social network analysis, it is common to draw a 2D layout of the network that highlights community structure. However, MPX graphs are distinct from many other types of graphs because they contain information about spatial relationships. Here, the goal is to find a 3D layout of the graph that gives an abstract representation of the structure of the cell.
Pixelator offers options to compute layouts for each MPX component in a data set using three different algorithms: Fruchterman-Reingold (FR), Kamada-Kawai (KK), and pivot Multidimensional Scaling (pMDS). The FR and KK algorithms are force-directed methods that assign forces to edges and nodes, and then allow these forces to repel and attract each other until the nodes reach a mechanical equilibrium. A major limitation with force-directed algorithms is the high running time, which in general is considered to run in cubic time O(N^3). This limitation presents a computational challenge as an MPX data set consists of hundreds to thousands of relatively large graphs. For additional details on these methods, please refer to the following articles:
The pMDS technique, which is the default graph drawing method in Pixelator, overcomes the running time limitations with force-directed algorithms. It builds on the same principles as classical multidimensional scaling (MDS), a widely studied method for graph drawing that converts distances between pairs of nodes into a configuration of points in an abstract Euclidean space. In brief, given the shortest path distances between each pair of nodes in the graph, the MDS technique attempts to place each node in a low-dimensional space (typically 2D or 3D) such that the distances between nodes in this space reflect the original distances as closely as possible. If we consider an MPX graph, we know that nodes that are connected by an edge should in theory correspond to DNA pixels that are positioned close to each other on the cell surface. Knowing that edges correspond to proximity relationships make MDS a good candidate for graph drawing of MPX component graphs.
While classical MDS also suffers from a high running time, this is mitigated with pMDS (Brandes et al.). Instead of computing all pairwise distances, pMDS only considers the shortest path distances from all nodes to a set of randomly selected "pivot" nodes. Below, we outline the general steps of the algorithm for computing a 3D graph layout:
Pivot MDS Input: an undirected graph , pivot nodes where
Steps:
- Select random pivot nodes from
- Compute the shortest path distances for nodes to pivot nodes to yield a distance matrix . For unweighted graphs, we can use a Breadth-First Search (BFS) to compute the distances.
- Create a transformed distance matrix , which is double-centered
- Decompose with Single Value Decomposition (SVD) and select the three singular vectors with the highest singular values
- The layout coordinates are obtained by computing: