DNA Molecular Computing with Weighted Signal Amplification for Cancer miRNA Biomarker Diagnostics

Abstract

The expression levels of microRNAs (miRNAs) are strongly linked to cancer progression, making them promising biomarkers for cancer detection. Enzyme-free signal amplification DNA circuits have facilitated the detection of low-abundance miRNAs. However, these methods may neglect the diagnostic value (or weight) of different miRNAs. Here, a molecular computing approach with weighted signal amplification is presented. Polymerase-mediated strand displacement is employed to assign weights to target miRNAs, reflecting the miRNAs' diagnostic values, followed by amplification of the weighted signals using localized DNA catalytic hairpin assembly. This method is applied to diagnose miRNAs for non-small cell lung cancer (NSCLC). Machine learning is used to identify NSCLC-specific miRNAs and assign corresponding weights for optimum classification of healthy and lung cancer individuals. With the molecular computing of the miRNAs, the diagnostic output is simplified as a single channel of fluorescence intensity. Cancer tissues (n = 18) and adjacent cancer tissues (n = 10) are successfully classified within 2.5 h (sample-to-result) with an accuracy of 92.86%. The weighted amplification strategy has the potential to extend to the digital detection of multidimensional biomarkers, advancing personalized disease diagnostics in point-of-care settings.

Keywords: DNA computing; NSCLC diagnostics; machine learning; miRNA; signal amplification.

Figure 1

Weighted amplification molecular computing strategy for cancer diagnosis. A) Traditional isothermal signal amplification techniques achieve signal enhancement by circulating targets. However, these techniques only focus on amplifying a single miRNA target and the output results are difficult to be used directly for diagnosis. B) We integrated PMSD and LCHA reactions to the biological information provided by the differential expression of multiple miRNAs, finally realizing direct diagnosis. According to the machine learning screened miRNAs and their weight information, PMSD assigns weights for multiple miRNA biomarkers by releasing weight DNA strands, and LCHA amplifies these weight DNA strands to achieve a detectable fluorescence signal. Finally, this strategy was used for NSCLC tissue sample diagnosis.

Figure 2

miRNA screening and weight assigning for classification performance assessment. A) Workflow of miRNA screening and weight assigning. miRNA expression profiles of publicly available healthy and NSCLC individuals from the TCGA database, candidate miRNAs were preliminarily screened, miRNA combinations were selected from candidates and assigned with weights by SVM algorithm, evaluation of classification performance. Six miRNAs and their weights were obtained. B) The screened miRNAs and their corresponding weights were used to classify the training set with good classification results, AUC = 0.97. C) Confusion matrix analysis of training set (n = 869), with an accuracy of 97.46%. D) The screened miRNAs and weight combinations were used to classify the validation set with good classification results, AUC = 0.99. E) Confusion matrix analysis of the validation set (n = 218), with an accuracy of 99.08%.

Figure 3

Weighted amplification molecular computing strategy construction and characterization. A) Schematic of the weighted amplification molecular computing strategy. PMSD assigns weights for different miRNA biomarkers by releasing weight DNA strands. LCHA amplifies these weight DNA strands to achieve a detectable fluorescence signal. B) Fluorescence kinetic curves generated by the LCHA were triggered by five target concentrations. C) Fluorescence‐concentration linearity every 5 min over 30 min with R² > 0.98. D) Fluorescence kinetic curves generated by the weighted amplification molecular computing strategy with weights 1–5 at 10 pm. N denotes the weight of the target. E) The linear relationships of fluorescence intensity at 15 min and concentration, R² > 0.96. The fluorescence‐weight linear relationship is shown in Figure S7B (Supporting Information). ΔF was calculated as the difference in fluorescence intensity between the experimental and control groups at 15 min. Condition: 10 nm PMSD, 100 nm LCHA, 3.98 nm Bst polymerase. 1×ThermoPol reaction buffer (20 mm Tris‐HCl, 10 mm (NH₄)₂SO₄, 10 mm KCl, 4 mm MgSO₄, 0.1% Triton X‐100, pH 8.8) at 25 °C. Data are mean ± S.D. (n = 3 independent experiments).

Figure 4

Weighted amplification molecular computing for miRNAs. A) The molecular computing workflow. Three positive and three negative weights PMSDs, each running an additive operation and an artificial subtractive operation. B) Fluorescence characterization of the three positive weights at a miRNA concentration of 10 pm. C) Fluorescence characterization of the three negative weights at an input concentration of 10 pm. The double Y‐axis shows the one‐to‐one correspondence between ΔF and weight (D) The extracted fluorescence intensities of the positive weight summation were linearly fit with weight values. E) The extracted fluorescence intensities of the negative weight summation were linearly fit with weight values. ΔF was calculated as the difference in fluorescence intensities between the experimental and control groups at 15 min. F) Diagnostic results for 10 NSCLC synthetic samples and 10 healthy synthetic samples, orange for NSCLC and blue for healthy. Data are mean ± S.D. (n = 3 independent experiments). G) Confusion matrix analysis of the diagnostic performance of synthetic samples. Condition: 10 nm PMSD, 100 nm LCHA, 3.98 nm Bst polymerase. 1×ThermoPol reaction buffer (20 mm Tris‐HCl, 10 mm (NH₄)₂SO₄, 10 mm KCl, 4 mm MgSO₄, 0.1% Triton X‐100, pH 8.8) at 25 °C.

Figure 5

NSCLC cell line and clinical sample diagnostics. A) The entire procedure from sample to result. B) Cell line results. All NSCLC cell results [FAM‐HEX] > 0, normal cell results [FAM‐HEX] < 0. C) Diagnostic results of clinical tissues, [FAM‐HEX] > 0 indicates NSCLC, and [FAM‐HEX] < 0 indicates adjacent. D) Confusion matrix for diagnosis of 18 NSCLC samples and 10 adjacent cancer samples with 92.86% accuracy. Condition: 10 nm PMSD, 100 nm LCHA, 3.98 nm Bst polymerase. 1×ThermoPol reaction buffer (20 mm Tris‐HCl, 10 mm (NH₄)₂SO₄, 10 mm KCl, 4 mm MgSO₄, 0.1% Triton X‐100, pH 8.8) at 25 °C. Data are mean ± S.D. (n = 3 independent experiments).