Detection of early-stage lung cancer in sputum using automated flow cytometry and machine learning

A system for automated analysis of flow cytometry data, the system comprising:

• a computer processor in communication with a memory having stored therein flow cytometry data from a plurality of markers in a plurality of cells from a sputum sample of a subject wherein the plurality of markers include i) a plurality of cell lineage specific markers, ii) a cell viability marker and iii) a tetra (4-carboxyphenyl) porphyrin (TCPP) marker;
• a computer-program product embodied in a non-transitory computer readable medium, the computer-program product comprising instructions for causing the computer processor to:
• receive the flow cytometry data acquired from the plurality of cells from the sputum sample;
• select from the plurality of cells in the sputum sample a subpopulation of cells automatically selected based upon application of automatic gates selected from a bead size exclusion gate, a viability gate and a singlets gate;
• determine, from the subpopulation, flow cytometric values of interest for the plurality of cell lineage specific markers, the cell viability marker and the TCPP marker;
• automatically apply a machine trained classifier to the flow cytometric values of interest and a metadata of the subject;
• wherein the machine trained classifier training comprises:
• obtaining flow cytometric data from multiple sputum samples labeled with markers i-iii wherein the multiple sputum samples comprise a known cancer sputum sample and a known non-cancer sputum sample;
• creating an extracted matrix of per cell flow cytometric data obtained for a singlet population of cells for each of the multiple labeled sputum samples;
• obtaining a metadata comprising a chronological age in years of each donor for each of the multiple labeled sputum samples;
• deriving per cell flow cytometric signal intensity data for markers i-iii in a logicle scale and dividing by a log10 value of Side Scatter-Area (“SSC-Alog 10”) or log10 value of Forward Scatter-Area (“FSC-Alog 10”) to produce a signal density distribution divided into three ranges a low “[R1]”, a mid “[R2]” and a high “[R3]” identified for signal density as compared to a control unlabeled sputum sample;
• quantizing per cell flow cytometric data for signal intensity in the logicle scale for a pair of markers selected from markers i-iii to heuristically position a 4×4 grid on a dotplot of the per cell signal intensity as compared to an unstained sputum sample control wherein an x-axis of the dotplot represents a first of the plurality of cell lineage specific markers and a y-axis of the dotplot represents a second of the plurality of cell lineage specific markers to produce the per cell signal intensity in the logicle scale quantized into a low, a low-mid, a mid, and a high windows on the x-axis and the y-axis of the dotplot of signal intensity based on a combined labeled sputum samples;
• using an Akaike information criterion, evaluate generalized linear models (“GLMs”) generated by iterative step-forward and step-backward parameter inclusion using the per cell flow cytometric data and metadata from a randomly-selected training set of labelled sputum samples to produce a selected trained classifier;
• validating the selected trained classifier on a test set of labeled sputum samples not included in the training set of labeled sputum samples;
• setting a cut-off value for the selected trained classifier to classify the training set labeled sputum sample and the test set of labeled sputum samples as cancer or non-cancer wherein the cut-off value corresponds to a classification having ≥80% sensitivity and ≥80% specificity to produce the machine trained classifier;
• and
• automatically generate, by the processor, to an output display device, a likelihood value of between ≥0 and ≤1 of having lung cancer based upon the application of the machine trained classifier to an individual sputum sample flow cytometric data and metadata from the subject suspected of having lung cancer and comparing the likelihood value to the cut-off value for the machine trained classifier to assign a classification of cancer when the likelihood value is greater than the cut-off value or non-cancer when the likelihood value is less than or equal to the cut-off value.

A non-transitory computer-readable medium comprising program code that, when executed, causes processing circuitry to:

• obtain flow cytometer values from a viable singlet population of a subject's sputum sample based upon a side scatter, a forward scatter, a fluorescence from a tetra (4-carboxyphenyl) porphyrin (TCPP) marker, a fluorescence from a cell viability marker, and a fluorescence from a plurality of cell lineage specific markers;
• apply a machine trained classifier to a metadata from the subject and the flow cytometric values obtained, wherein the metadata of the subject is age of the subject in chronological years;
• and
• generate, based upon the application of the machine trained classifier, a classification for the sputum sample wherein the classification is selected from cancer and non-cancer, wherein the cell viability marker is FVS510, the plurality of cell lineage specific markers are selected from a fluorescent anti-CD206, a fluorescent anti-CD3, a fluorescent anti-CD19, and a fluorescent anti-CD66b and the machine trained classifier is
• Likelihood of
• Cancer
• ~
• b
• 0
• +
• b
• 1
• *
• age
• +
• b
• 2
• *
• CD
• ⁢
• 206
• low
• ⁢
• CD
• ⁢
• 66
• ⁢
• b
• /
• CD
• ⁢
• 3
• /
• CD
• ⁢
• 19
• mid
• +
• b
• 3
• *
• FVS
• ⁢
• 510
• FSC
• -
• A
• log
• 10
• [
• R
• ⁢
• 2
• ]
• +
• b
• 4
• *
• TCPP
• SSC
• -
• A
• log
• 10
• [
• R
• ⁢
• 3
• ]
• -
• b
• 5
• *
• age
• *
• FVS
• ⁢
• 510
• FSC
• -
• A
• log
• 10
• [
• R
• ⁢
• 2
• ]
• wherein:
• b0 is an intercept value when all variables in the trained classifier are set to zero;
• b1-b5 are weights for each variable determined by fitting the trained classifier to a training set;
• CD206lowCD66b/CD3/CD19mid is a population of cells that are labeled with the fluorescent anti-CD206 and identified as “low” on a signal intensity logicle scale dotplot and simultaneously labeled with the fluorescent anti-CD66b/anti-CD3/anti-CD19 and identified as “mid” on the signal intensity logicle scale dotplot wherein the CD206lowCD66b/CD3/CD19mid population of cells are determined from the signal intensity logicle scale dotplot partitioned with a 4×4 grid heuristically positioned on the dotplot of the per cell signal intensity as compared to an unstained sputum sample control wherein an x-axis of the dotplot represents a first of the plurality of cell lineage specific markers and a y-axis of the dotplot represents a second of the plurality of cell lineage specific markers to produce the per cell signal intensity in the logicle scale quantized into a low, a low-mid, a mid, and a high windows on the x-axis and the y-axis of the dotplot of signal intensity based on a combined labeled sputum samples;
• FVS510/FSC-Alog 10 [R2] is a proportion of cells with a FVS510 signal density greater than or equal to 0.3 but less than 0.6 (“[R2]”) where the FVS510 signal density is a per cell FVS510 fluorescent signal on a logicle scale divided by a log10 value of Forward Scatter-Area (“FSC-Alog 10”) wherein R2 is identified from a signal density distribution divided into three ranges a low “[R1]”, a mid “[R2]” and a high “[R3]” identified for a signal density as compared to a control unlabeled sputum sample;
• and
• TCPP/SSC-Alog 10 [R3] is a proportion of cells with a TCPP signal density greater than or equal to 0.6 (“[R3]”) where the TCPP signal density is a per cell TCPP fluorescent signal on a logicle scale divided by a log10 value Side Scatter-Area (“SSC-Alog 10”) to produce a signal density distribution divided into three ranges a low “[R1]”, a mid “[R2]” and a high “[R3]” identified for signal density as compared to the control unlabeled sputum sample.