Molecule Docking

1. We firstly used AlphaFold3^[1] to predict the nucleic acid structure. Since the directly downloaded file format was .cif, we employed Open Babel^[2] to convert the .cif file into a .pdb file for molecular docking.

2. We retrieved the experimentally validated structure of vancomycin (PDB ID: 1SHO) from the PDB database.
   According to the original publication, this file contains an antibiotic dimer, two chloride ions, one acetate ion, and 105 solvent water molecules^[3].

Using PyMOL, we removed the acetate ion and chloride ions to obtain a vancomycin monomer .pdb file containing water molecules.

3. Subsequently, LBO1 was docked with vancomycin using AutoDock4^[4] (50 runs). The results are presented below.

Analysis of the docking results revealed that:

• Vancomycin tends to form an asymmetric dimer, with a binding energy of –3.26 kcal/mol.
• The binding energy between LBO1 and vancomycin was –5.79 kcal/mol, indicating a more stable interaction due to lower (more negative) binding energy.
• Our docking results are consistent with the original publication^[5], showing that the binding site of vancomycin lies within the central loop region of the aptamer.

Machine Learning Model

Model Name — DSAPred

Full name: DNA–Small molecule binding Affinity Predictor
Abbreviation: DSAPred

Project Motivation and Background

Antibiotic contamination poses a growing threat to food safety, medical diagnostics, and public health. Our project focuses on developing a modular aptamer-based biosensing platform that can rapidly detect antibiotics through specific molecular recognition. Within this system, DNA aptamers act as the key sensing components, and their binding affinity (Kd) toward target antibiotics fundamentally determines the sensor's sensitivity and response reliability.

To improve sensor performance, identifying or designing aptamers with higher affinity and specificity is essential. However, experimental screening for such aptamers is labor-intensive and time-consuming. To overcome this challenge, we developed a machine learning–based prediction tool, DSAPred (DNA–Small molecule binding Affinity Predictor), which quantitatively estimates the apparent binding affinity (pKd) between DNA aptamers and small organic molecules.

By systematically analyzing existing aptamer–small molecule datasets, DSAPred allows us to computationally discover and prioritize promising aptamer candidates, thereby building a modular component library that supports the rapid assembly of antibiotic-sensing systems.

Although time constraints limited full optimization, our modeling work reflects an ongoing effort toward rational parameter tuning—for example, accounting for reaction conditions close to real biosensing assays, such as room temperature, moderate ionic strength, and physiological ion compositions (Mg²⁺, K⁺). These conditions strongly influence aptamer folding stability and ligand interaction, and incorporating them enhances the robustness and real-world applicability of the predictions.

Importantly, DSAPred is not limited to antibiotics. Since its input features represent general nucleic acid–small molecule physicochemical relationships, the same framework can be applied to predict interactions with a broad range of targets. This universality makes DSAPred a valuable computational foundation not only for antibiotic biosensor development but also for expanding our modular sensing platform to diverse chemical analytes in the future.

In summary, this model bridges computational prediction with experimental design, offering both a rational guide for aptamer selection and a scalable library of functional components that advance the engineering of responsive biosensors across multiple application domains.

Description

DSAPred (DNA–Small molecule binding Affinity Predictor) is a predictive model developed to estimate the binding affinity (Kd/pKd) between DNA aptamers and small organic molecules.

Integrating sequence-derived descriptors, predicted secondary structure features, and physicochemical properties of small molecules, DSAPred establishes a quantitative mapping between aptamer composition and molecular characteristics.

The model was trained on a curated dataset of 759 aptamer–small molecule pairs, cleaned and standardized through rigorous preprocessing (including feature screening, missing-value handling, and duplication removal).

Using a forward feature selection strategy and cross-validation verification, DSAPred achieved strong generalization performance (CV r = 0.705± 0.049; MAE = 0.572 ± 0.055; Test r = 0.636; MAE = 0.573).

This model serves as the computational foundation for our aptamer-based antibiotic detection biosensor, guiding rational probe design and sensitivity optimization.

Data Collection & Processing

1. Data Source

Our dataset was compiled from multiple curated aptamer databases and literature reports, focusing on DNA aptamers targeting small organic molecules.

The main sources include:

(1) AptaDB

A comprehensive literature-curated database containing experimentally validated aptamer–target pairs with binding constants (Kd), sequences, target categories, and selection methods (SELEX variants).

We extracted DNA aptamers binding to small organic molecules, collecting sequence, Kd, and target details as base data.

(2) AptaIndex (Aptagen, USA)

Primary source of aptamer–target pairs, sequences, and binding constants (Kd).

(3) UTexas Aptamer Database (Sept 2023 version)

Provided additional experimental data with measured Kd values and binding conditions.

(4)Alkhamis et al., 2024

Exploring the relationship between aptamer binding thermodynamics, affinity, and specificity (Department of Chemistry, North Carolina State University, USA) — This dataset provides detailed thermodynamic characterizations (ΔG, ΔH, TΔS) and specificity profiles of over 200 aptamers toward multiple small molecules, serving as a crucial benchmark for integrating binding thermodynamics into our predictive framework.

From these sources, we collected ~860 unique DNA aptamer–small molecule binding pairs, each containing the following essential attributes:

2. Data Standardization

To ensure consistent analysis, we conducted a multi-step data cleaning process:

• Unit normalization — All Kd values were converted into molar units (M), and then transformed to logarithmic scale (pKd = –log₁₀(Kd)).
• Sequence cleaning — All aptamer sequences were standardized to uppercase A/C/G/T (replacing U→T).
• Target harmonization — Each target molecule was mapped to a unique PubChem CID using PubChemPy API; if successful, its isomeric SMILES and connectivity SMILES were recorded.
• Filtering — Only records with:
  • Type of Nucleic Acid = DNA (including ssDNA)
  • Target Category = Small Organic Molecule
  were retained for modeling.

Result: 859 valid DNA–small molecule pairs were obtained.

3. Secondary Structure Prediction

Each DNA sequence was folded using RNAstructure v6.5 (Fold.exe) in DNA mode (-d) at 25°C.
The minimum free energy (MFE) structure was exported as dot–bracket notation, together with the corresponding ΔG (kcal/mol) value.

This provides a thermodynamic and structural descriptor of each aptamer, forming the foundation for structure-based feature extraction.

4. Feature Extraction

We extracted 502 DNA aptamer features following the QSAR (Quantitative Structure–Activity Relationship) principle, combining sequence, structure, and physicochemical properties:

All features were standardized using the parameters derived from the training set.

5. Small Molecule Descriptors

Each small molecule was processed through the following pipeline:

• SMILES conversion — Retrieved via PubChem CID → SMILES.
• 3D structure generation — Constructed using RDKit, hydrogens were explicitly added, and energy minimized using the MMFF94 force field.
• Descriptor calculation — Extracted using the Mordred package, including 1D, 2D, and 3D descriptors (1826 features).

These features describe molecular topology, charge distribution, and geometric conformation.

6. Final Feature Set

After merging and standardizing both aptamer and small molecule features, we obtained:

• Aptamer features: 502 dimensions
• Molecule features: 1826 dimensions
• Combined total: 2328 descriptors per sample

Data Filtering and Quality Control

Before model training, the combined dataset underwent rigorous filtering and preprocessing to ensure both data integrity and statistical reliability.

After these steps, the dataset was reduced to 759 high-quality DNA–ligand binding pairs, ready for feature standardization and modeling.

Modeling and Evaluation

1. Modeling Overview

To quantitatively predict the binding affinity (pKd) between DNA aptamers and small organic molecules, we constructed a polynomial-extended linear regression model.

The approach follows the principle of Quantitative Structure–Activity Relationship (QSAR), integrating both aptamer sequence–structure descriptors and molecular physicochemical features.

2. Feature Preparation

After merging aptamer and small-molecule descriptors and data filtering, the initial feature space contained approximately 1225 features.

To enhance the model's ability to capture nonlinear relationships, we introduced second-order (squared) terms for all features, effectively expanding the descriptor set while retaining interpretability.

3. Data Splitting and Standardization

To ensure objective and reproducible evaluation, we split the dataset into training (80%) and test (20%) with random_state=41.

All features were standardized using Z-score normalization, where the mean and standard deviation were computed from the training set only.

4. Feature Selection — Forward Feature Selection (FFS)

Given the expanded feature space (original + squared terms), we employed Forward Feature Selection (FFS) to iteratively identify the most informative and uncorrelated subset.

At each iteration:

1. One candidate feature is added only if it improves training performance (Pearson correlation coefficient(*r)*↑, Mean Absolute Error(*MAE)*↓);
2. Candidate evaluation uses 5-fold cross-validation (CV) with random shuffling for stability;
3. To ensure feature independence, any candidate feature with a pairwise Pearson correlation coefficient above |r| > 0.90 with already-selected features was excluded from consideration. This avoids redundancy and ensures interpretability of coefficients.
4. The process stops when performance improvement (Δr < 0.001, ΔMAE < 0.001) becomes insignificant.

Finally, the elbow point of the FFS curve was used to determine the optimal number of features — 20 selected features were retained for the final model.

Bias = 6.209472309338183

5. Regression Model

The final model was a multiple linear regression (OLS) model trained on 20 selected features (including their squared counterparts).

This model maintains both interpretability and computational simplicity, allowing visualization of each feature's contribution via its regression coefficient.

6. Model Performance

Cross-Validation (5-fold, shuffled)

The CV results demonstrate stable internal generalization, suggesting that the selected 20-feature model captures consistent patterns across resampled folds.

Independent Test Set

The model performed comparably on the independent test set, confirming that it generalizes well to unseen aptamer–ligand pairs.

Training set: r = [0.749]; MAE = [0.535]; R² = [0.561]. Most points fall within ± 0.5 pKd, showing stable fitting for moderate-affinity samples (pKd ≈ 4–7).

Test set: r = [0.636]; MAE = [0.573]; R² = [0.352]. The similar distribution pattern demonstrates robust generalization across unseen aptamer–ligand pairs.

7. Model Interpretability — Coefficients and SHAP Analysis

We employed SHAP to quantify sample-level positive/negative contributions of each feature.

Key findings:

• Polarizability and charge-distribution descriptors (e.g., BCUTs-1h_sq, JGI7, MATS4Z) dominate and generally increase pKd when high—consistent with chemical intuition that polar/electrostatic matching stabilizes DNA binding.
• Several 3D shape/volume Moran descriptors (Mor14v_sq, Mor17p, Mor30m, Mor24se) contribute strongly, highlighting the role of spatial complementarity.
• Certain sequence/structural features (k3_CGA_sq, u_3mer_GAA_sq) show sparse but strong impacts, indicating context-specific effects (e.g., loop/stack configurations).

8. Summary

This regression-based model successfully integrates molecular and nucleic acid descriptors to predict aptamer–ligand affinity quantitatively.

Despite its simplicity, it achieves consistent performance across CV and test sets and maintains excellent interpretability, making it suitable for rational aptamer screening and modular biosensor design.

9. Model Input/Output

All features were normalized using the stored mean and standard deviation from the training dataset before prediction.

Squared terms (feature²) were automatically generated for nonlinear correction.

Case Study

To evaluate the real-world applicability of DSAPred in antibiotic recognition systems, we conducted an independent case study using a vancomycin-specific DNA aptamer reported by:

Xiaona Fang, Tian Gao, Zhaofeng Luo, and Renjun Pei, "Efficient selection of vancomycin-specific aptamers via particle display and development of a high-sensitivity fluorescent apta-sensor for vancomycin detection," Sensors and Actuators B: Chemical, 436 (2025), 137681.

The aptamer sequence used for validation was:

5′-CGACCGAGGGTACCGCAATAGTACTTATTGTTCGCCTATTGTGGGTCGGGTCG-3′

This sequence was not included in the training dataset, making it an ideal external sample for testing model generalization.

The predicted apparent binding affinity from DSAPred was pKd = 6.83, which agrees closely with the experimentally measured value of pKd = 6.85, showing a deviation of only 0.02 (≈4.7%). This remarkable agreement demonstrates the model's robust predictive capability for real antibiotic-binding systems.

SHAP value analysis further highlighted the underlying feature contributions:

• Trinucleotide motifs "CGA" strongly increased the predicted pKd, suggesting favorable sequence-context effects in vancomycin binding.
• High values of molecular descriptors MATS4Z and BCUTs-1h_sq positively contributed to affinity, indicating the significance of polarizability and electrostatic complementarity between the aptamer and the ligand.
• Conversely, descriptors reflecting molecular shape heterogeneity (e.g., Mor24se) had minor negative contributions.

Collectively, this case study confirms that DSAPred can accurately predict binding affinities for unseen aptamer–antibiotic pairs while providing mechanistic interpretability through feature contribution analysis. It demonstrates DSAPred's strong potential as a design and optimization platform for antibiotic-sensing aptamer elements.

NUPACK Simulation

We used NUPACK to predict the binding behavior of DNA components before starting experiments, achieving the goal of guiding design.

NUPACK (Nucleic Acid Package) is a software package developed by institutions such as Caltech for structural modeling, analysis, and design of DNA/RNA. It is particularly suitable for multi-chain systems and widely used in DNA nanotechnology, synthetic biology, and molecular diagnostics.

By rationally using NUPACK, we can:

• Design nucleic acid strands
• Pre-screen for non-specific secondary structures
• Predict hybridization efficiency
• Save manpower and resources in experimental validation

Simulation of T7 System's Sequences

Designed 5TS and 3TS Constructs

Cas System Simulation