A multi-class predictor based on a probabilistic model: application to gene expression profiling-based diagnosis of thyroid tumors

Naoto Yukinawa, Shigeyuki Oba, Kikuya Kato, Kazuya Taniguchi, Kyoko Iwao‐Koizumi, Yasuhiro Tamaki, Shinzaburo Noguchi, Shin IshiiBMC 2006 · DOI

Abstract

Accurate cancer diagnosis often requires moving beyond binary classifications. This study addresses the challenge of thyroid tumor differentiation—specifically between benign follicular adenoma (FA) and malignant follicular carcinoma (FC)—using gene expression profiling and a novel multi-class predictor. By employing an exhaustive combination of binary classifiers and probabilistic outputs, the proposed model achieved 85.7% accuracy on an independent test set, outperforming standard support vector machines (SVMs).

Introduction: The "Gray Zone" of Thyroid Oncology

In clinical practice, histopathology remains the gold standard for cancer diagnosis. However, thyroid oncology presents a significant "gray zone." Differentiating between Follicular Carcinoma (FC) and Follicular Adenoma (FA) is notoriously difficult because their morphological structures are nearly identical under a microscope. Fine-needle aspiration often yields indeterminate results, leading to surgical interventions that may ultimately prove unnecessary.

To address this, researchers utilized gene expression profiling to provide a molecular-level diagnostic tool. While most machine learning tools are optimized for binary (A vs. B) tasks, clinical reality involves multiple classes: Normal (N), FA, FC, and Papillary Carcinoma (PC).

Methodology: Exhaustive Binary Combinations

The study measured the expression of over 2,000 genes in 119 thyroid samples using adaptor-tagged competitive PCR (ATAC-PCR). The core innovation is the multi-class predictor strategy. Rather than relying on standard "one-vs-rest" or "one-vs-one" approaches, the team explored more exhaustive strategies:

  • One-to-Any (1A): Discriminating one class against any subset of other classes.
  • Any-to-Any (AA): Discriminating between any pair of disjoint subsets of classes.

Each binary classifier within these strategies produced a probabilistic output via logistic regression. These "opinions" were aggregated using a novel Shared Summation (SHS) method to determine the final class probability for each sample.

Discrimination power across tissue pairs. Panel (a) shows the original vs. permuted data for FA vs. Normal. Panel (b) illustrates that certain pairs, like FA vs. FC and FA vs. N, exhibit lower discrimination power, reflecting their biological similarity.

Results and Benchmarking

The 1A-SHS model achieved a test accuracy of 85.7% on an independent set of 49 samples. To validate the generalizability of the algorithm, it was also tested against public benchmarks, including the Global Cancer Map (GCM) and SRBCT datasets. In these trials, the probabilistic model consistently outperformed Multi-class SVMs and Shrunken Centroid methods.

The stability of the predictor was a key finding. As shown in the performance metrics, the accuracy of individual binary classifiers varies depending on the tissue pair and the number of genes selected (determined by p-value thresholds).

Prediction accuracy of binary classifiers across different p-value thresholds for gene selection. While pairs like FA vs. PC (red) are easily distinguished, the FA vs. N (green) pair remains a significant classification challenge.

Discussion: Biological Insights and Probability Space

The model’s internal "confusion" mirrored biological reality. The highest misclassification rates occurred between FA and Normal or FA and FC, confirming that these tissues are transcriptomically similar.

By generating probabilities for every class, the system allows for visualization in a low-dimensional "probability space." This offers clinicians more than just a label; it shows how close a sample sits to the boundaries of other classes. This nuance is vital for identifying rare hybrid tumor types or samples that may have been mislabeled during initial pathology. While the exhaustive "AA" strategy is computationally intensive, it provides a superior framework for complex diagnostic panels where subtle differences between similar tissue types are critical for patient outcomes.

Summary updated

Abstract

Abstract Background Although microscopic diagnosis has been playing the decisive role in cancer diagnostics, there have been cases in which it does not satisfy the clinical need. Differential diagnosis of malignant and benign thyroid tissues is one such case, and supplementary diagnosis such as that by gene expression profile is expected. Results With four thyroid tissue types, i.e., papillary carcinoma, follicular carcinoma, follicular adenoma, and normal thyroid, we performed gene expression profiling with adaptor-tagged competitive PCR, a high-throughput RT-PCR technique. For differential diagnosis, we applied a novel multi-class predictor, introducing probabilistic outputs. Multi-class predictors were constructed using various combinations of binary classifiers. The learning set included 119 samples, and the predictors were evaluated by strict leave-one-out cross validation. Trials included classical combinations, i.e., one-to-one, one-to-the-rest, but the predictor using more combination exhibited the better prediction accuracy. This characteristic was consistent with other gene expression data sets. The performance of the selected predictor was then tested with an independent set consisting of 49 samples. The resulting test prediction accuracy was 85.7%. Conclusion Molecular diagnosis of thyroid tissues is feasible by gene expression profiling, and the current level is promising towards the automatic diagnostic tool to complement the present medical procedures. A multi-class predictor with an exhaustive combination of binary classifiers could achieve a higher prediction accuracy than those with classical combinations and other predictors such as multi-class SVM. The probabilistic outputs of the predictor offer more detailed information for each sample, which enables visualization of each sample in low-dimensional classification spaces. These new concepts should help to improve the multi-class classification including that of cancer tissues.
BibTeX
@inproceedings{yukinawa2006multi,
  author = {Naoto Yukinawa and Shigeyuki Oba and Kikuya Kato and Kazuya Taniguchi and Kyoko Iwao‐Koizumi and Yasuhiro Tamaki and Shinzaburo Noguchi and Shin Ishii},
  title = {{A multi-class predictor based on a probabilistic model: application to gene expression profiling-based diagnosis of thyroid tumors}},
  booktitle = {BMC},
  year = {2006},
  doi = {10.1186/1471-2164-7-190},
}
Figure