Quantum Cognition Machine Learning for Forecasting Chromosomal Instability
Authors: Giuseppe Di Caro, Vahagn Kirakosyan, Alexander G. Abanov, Jerome R. Busemeyer, Luca Candelori, Nadine Hartmann, Ernest T. Lam, Kharen Musaelian, Ryan Samson, Harold Steinacker, Dario Villani, Martin T. Wells, Richard J. Wenstrup, Mengjia Xu
Organizations: Qognitive, Inc., Miami Beach, FL, USA · Epic Sciences, San Diego, CA, USA · Stony Brook University, Department of Physics and Astronomy, Stony Brook, NY, USA · Indiana University, Department of Psychological and Brain Sciences, Bloomington, IN, USA · Wayne State University, Department of Mathematics, Detroit, MI, USA · University of Vienna, Department of Physics, Vienna, Austria · King’s College London, Department of Mathematics, London, UK · Cornell University, Department of Statistics and Data Science, Ithaca, NY, USA · New Jersey Institute of Technology, Department of Data Science, Newark, NJ, USA
The accurate prediction of chromosomal instability from the morphology of circulating tumor cells (CTCs) enables real-time detection of CTCs with high metastatic potential in the context of liquid biopsy diagnostics. However, it presents a significant challenge due to the high dimensionality and complexity of single-cell digital pathology data. Here, we introduce the application of Quantum Cognition Machine Learning (QCML), a quantum-inspired computational framework, to estimate morphology-predicted chromosomal instability in CTCs from patients with metastatic breast cancer. QCML leverages quantum mechanical principles to represent data as state vectors in a Hilbert space, enabling context-aware feature modeling, dimensionality reduction, and enhanced generalization without requiring curated feature selection. QCML outperforms conventional machine learning methods when tested on out of sample verification CTCs, achieving higher accuracy in identifying predicted large-scale state transitions (pLST) status from CTC-derived morphology features. These preliminary findings support the application of QCML as a novel machine learning tool with superior performance in high-dimensional, low-sample-size biomedical contexts. QCML enables the simulation of cognition-like learning for the identification of biologically meaningful prediction of chromosomal instability from CTC morphology, offering a novel tool for CTC classification in liquid biopsy.
Figures & tables
Figure 1: Motivation and study workflow
Figure 2: Distribution of genomically determined LST values across all cases. Scatter plots show the genomically determined LST ground truth values which are shown per CTC, sub-grouped by case ID number and ranked from left to right by increasing LST values. We show the visual representation of the heterogeneity in the distribution of LST values across each case ID.
Figure 3: Count of CTCs with specific LST values. The bar plot comparison illustrates the number of CTCs ranked from left to right by increasing LST values. The bar plot at the top displays the LST values ground truth, which is the genomically determined LST per CTC, while the bottom plot shows the morphology-predicted LST values computed by QCML. The count of CTCs for the ground truth genomically determined LST follows the same bimodal trend as that of the morphology-predicted LST.
Figure 4: Linear correlation of cellular and nuclear morphology and protein intensity features with the extent of LST as determined by genome sequencing. In the scatter plots each dot is an individual CTC and graphs show the best-fit line between the extent of LST as a measure of chromosomal instability and a) nuclear size, b) cellular size, c) nuclear fractal, d) cellular LQI CD45/CD31, e) nuclear MFI DAPI, f) cellular IQR DAPI, g) LBP DAPI-CK, h) LBP CK-CD45/CD31.
Models
Fold training
Fold validation
Sensitivity
Specificity
Balanced Accuracy
Sensitivity
Specificity
Balanced Accuracy
QCML
91% ± 3%
67% ± 6%
79% ± 3%
81% ± 10%
53% ± 16%
67% ± 9%
TabPFN [ 22 ]
100% ± 0%
99% ± 2%
99% ± 1%
90% ± 8%
41% ± 16%
66% ± 8%
CatBoost [ 45 ]
100% ± 0%
100% ± 1%
100% ± 1%
93% ± 7%
36% ± 17%
64% ± 8%
SVM Gaussian Kernel
98% ± 1%
73% ± 5%
85% ± 2%
90% ± 7%
39% ± 16%
64% ± 8%
Elastic Net
95% ± 2%
52% ± 8%
73% ± 4%
86% ± 8%
40% ± 16%
63% ± 8%
Table 1: Fold-training and fold-validation performance of QCML and classical ML models forecasting LST+. Showing the average and standard deviation of sensitivity, specificity and balanced accuracy across 25 folds (5-fold with 5 repeats).
Fold training
Fold validation
Model
Sensitivity
Specificity
Balanced Accuracy
ROC AUC
Sensitivity
Specificity
Balanced Accuracy
ROC AUC
QCML POVM
92% ± 3%
83% ± 6%
87% ± 3%
0.946
76% ± 9%
58% ± 18%
67% ± 9%
0.739
QCML
81% ± 5%
85% ± 4%
83% ± 2%
0.922
67% ± 13%
62% ± 18%
65% ± 9%
0.710
CatBoost
100% ± 0%
100% ± 0%
100% ± 0%
1.000
80% ± 10%
55% ± 18%
68% ± 8%
0.724
TabPFN
100% ± 0%
100% ± 0%
100% ± 0%
1.000
76% ± 10%
57% ± 18%
66% ± 6%
0.714
XGBoost
100% ± 0%
100% ± 0%
100% ± 0%
1.000
76% ± 11%
55% ± 18%
66% ± 7%
0.715
Table 2: Fold-training and fold-validation performance of QCML POVM and classical ML classification models forecasting LST+. Showing the average and standard deviation of sensitivity, specificity and balanced accuracy across 25 folds (5-fold with 5 repeats). Also showing the average ROC AUC score per model.
Figure 5: ROC curves for the top three performing models : QCML POVM (blue), CatBoost (red), and XGBoost (green). At ∼ 70% specificity all three models achieve similar sensitivity. At lower specificity, QCML POVM is outperforming both CatBoost and XGBoost (i.e., it can achieve higher sensitivity for a fixed specificity). At higher specificity, QCML POVM is on par with XGBoost and outperforms CatBoost.
Figure 6: QCML POVM produces full probability distribution of LST for all observations. Showing predicted LST distribution for a sample of 3 CTCs. Dotted vertical lines show the actual LST for each CTC. Having probability distribution allows one to forecast both LST quantiles as well as probabilities for LST being above/below a threshold.
HS Dims
Balanced Accuracy : Fold validation
4
67.3%
8
67.6%
16
67.3%
32
66.7%
Table 3: QCML performance across Hilbert Space dimensions.
Runtime Comparison
Model
Train Runtime
Inference Runtime (100,000 cells)
QCML
120s
0.01s
CatBoost
30s
0.05s
TabPFN
26s
59s
Table 4: Training and inference runtime comparison. QCML required more training time but had the shortest measured inference time in this benchmark.
Table 5: Regression model performance at different LST thresholds Showing the fold-validation balanced accuracy across varying LST thresholds of 10, 12, 14. QCML achieved the highest mean balanced accuracy at each tested threshold.
Figure 7: QCML gradient-based feature importance. Left: Top morphological and protein intensity features associated with high LST (blue) and low LST (red) predictions. Right: Top feature names based on absolute importance, absolute importance score values, and model predictions (i.e., High LST or Low LST).
Figure 8: Multi-dimensional scaling visualization based on a distance matrix of CTCs. Red color represents LST+ CTCs, while blue color represents stable CTCs. Using QCML distance one can achieve a visual separation between LST+ and stable CTCs.
School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA · SenSIP Center, Arizona State University, Tempe, AZ, USA · School of Computing and Augmented Intelligence, Arizona State University, Tempe, AZ, USA
Department of Computational Life Sciences, Cleveland Clinic Research, Cleveland, Ohio, USA. · Department of Quantitative Health Sciences, Cleveland Clinic Research, Cleveland, Ohio, USA. · IBM Quantum, IBM Thomas J Watson Research Center, Yorktown Heights, NY, USA. +1