Aptamers are short, single-stranded oligonucleotides (DNA or RNA) that fold into 3D structures, enabling them to bind to molecular targets with high specificity and affinity. Compared with traditional antibody-based sensors, DNA aptamers in particular offer unique advantages. Aptamers used in vitro are chemically prepared oligonucleotides, and no animals or in vivo models are involved in their isolation, synthesis, or later identification. This makes the purification and reproducibility of aptamers much more efficient than the generation of antibodies. Next, because they are synthetically obtained, the potential for increased diversity in the oligonucleotide library pool can be achieved through simple chemical modifications using standard principles of nucleic acid chemistry. Finally, oligonucleotides are more stable than antibodies, making them ideal candidates for many biomedical applications [1]. These characteristics make DNA aptamers suitable for designing biosensors targeting various targets, especially for point-of-care diagnostics [2]. The ligands of aptamers may include but are not limited to ions, peptides, proteins, microbes (viruses and bacteria), and even entire cells [3,4]. In Figure 1, a 15-mer single-stranded DNA aptamer that folds into a 2-G-quartet structure and binds thrombin is shown. This interaction prevents thrombin from binding to fibrinogen and the subsequent coagulation cascade, thereby preventing blood coagulation [5]. The increase in nucleotide count in the aptamer allows greater variability in molecular structure, ultimately enabling it to adopt a unique fold [4,6].

An aptamer’s linear nucleotide sequence, also called the primary structure, determines its ability to fold into secondary and tertiary structures (Figure 2). This is based on the ability of nucleotide bases to interact with each other through Watson-Crick base pairing. These higher-order structures create unique 3D pockets that dictate target recognition sites and target specificity. Changes in these primary sequences can significantly alter folding patterns, which in turn change target selectivity and affinity to a specific molecular target. Hydrogen bonding, electrostatic interactions, π-π stacking, and van der Waals interactions contribute to the molecular recognition of a target as well as to an aptamer’s structural stability [8,9].

The traditional method of aptamer discovery, SELEX (systematic evolution of ligands by exponential enrichment), consists of several iterative rounds of binding and washing to select aptamers from a broad aptamer library, identifying high-affinity sequences for a specific target. Using the SELEX technique, aptamers have been selected for the amino acid tryptophan [10], the antibiotic chloramphenicol [11], the natural product theophylline [12], the neurotransmitter dopamine [13], and many other small-molecule targets. However, this process is time-consuming, expensive, and requires specialized equipment, and there are still many small molecules of interest for which isolating high-affinity aptamers remains difficult [10]. Under typical SELEX conditions, a target of interest is exposed to a library pool of single-stranded DNA oligonucleotides (20‐80 nucleotides in length), and sequences with affinity to bind the target noncovalently will be collected. In contrast, the nonbinding oligonucleotides will be eluted. Multiple cycles of polymerase chain reaction–based selection and amplification will promote competition among active compounds. They should eventually result in the purification of the few molecular species with the highest affinity or efficacy for a given target (Figure 3).

Although methods such as high-throughput SELEX [14] and particle display [15] have significantly increased SELEX throughput, they are still optimized for single targets and can be resource-intensive. Therefore, there is a need to identify additional methods to reduce the complexity of SELEX by prior identification of aptamer candidates for the target of interest.

In the past few years, computational strategies have become popular to facilitate faster aptamer discovery and reduce the burden of traditional SELEX. The early in silico techniques for aptamer discovery mostly relied on structure-based modeling. For example, tools such as MFold (RNA Institute at the University of Albany) and NUPACK (Nucleic Acid Package; California Institute of Technology [Caltech]) use thermodynamic free energy minimization algorithms. All possible folding configurations of a single nucleic acid strand are evaluated, and the most thermodynamically stable structure is identified. Figure 4 shows examples of predicted secondary structures for aldosterone (ALDO) aptamer 14 using MFold (left) and NUPACK (right). These secondary-structure predictions can be used as inputs for tertiary-structure modeling on platforms such as RNAComposer or 3DRNA.

While these methods provide valuable insights into aptamer-target interactions, they are computationally expensive and difficult to scale to large aptamer libraries. They can fail for new targets, limiting their practical applications. Therefore, other data-driven methods were introduced, such as combining machine learning (ML) with genetic algorithms for finding novel RNA aptamers [11]. Other deep-learning methods can bypass explicit structural modeling and learn binding patterns directly from the primary sequence data. For example, AptaBLE (Aptamer Binding Language) uses pretrained encoders for aptamer and protein-pair inputs. The pairs are transformed into vectors representing the aptamer-molecule interactions, which in turn are used to predict binding strength or even generate new aptamer candidates [12]. However, these methods are target-specific and require retraining for each new target. This gap inspired the current work and is discussed in more detail in the following section.

Despite significant advancements from thermodynamics-based folding prediction tools (such as MFold and NUPACK) to deep-learning methods based on raw sequence data, many limitations remain in this area. First, much of the existing research, as well as available datasets, focus on RNA aptamers rather than DNA aptamers. DNA aptamers are highly preferred for biosensor fabrication because they offer greater chemical stability and lower synthesis costs [13]. Second, most ML models are target-specific and trained for a single molecule. For example, AptaBLE, discussed above, needs to be retrained each time a new target is introduced. Third, several of these models rely on accurate folding predictions, so an error at this critical step can propagate and amplify throughout the model. In summary, current research lacks a generalized, sequence-only framework for classifying DNA aptamers across multiple target categories based purely on primary sequence information. Addressing this gap enables rapid prescreening of aptamer candidates before experimental SELEX, ultimately reducing time and cost while supporting large-scale sensor design and discovery.

In this work, we present preliminary results of a sequence-based ML framework for classifying DNA aptamers by target type across diverse molecular classes. In this study, cells and steroids were selected as the targets of interest from the available UTexas database. Each aptamer sequence is represented using term frequency (TF)–inverse document frequency (IDF) rather than the raw sequence, and the resulting vector is then used as direct input to the classifier model. TF-IDF is a text-mining method that scores short sequence motifs (k-mers) according to how frequently they occur within an individual sequence (TF) relative to how commonly they occur across the entire dataset (IDF). Motifs that occur within one target class but are rare elsewhere are graded with a higher TF-IDF weight, tagging them as candidate class-discriminating features. TF-IDF was selected for this work because, unlike structure-based or deep-learning representations, it requires no folding prediction or a large target-specific pretraining set or per-target retraining, thereby directly addressing the limitations of prior approaches (discussed in the Identified Research Gap section).

This study aims to develop and evaluate a sequence-only framework for classifying DNA aptamers by target type without relying on structural prediction models and to assess whether the resulting classifier generalizes to independent, laboratory-generated sequences.

For this study, we used curated entries from the UTexas Aptamer Database [16], which contains experimentally validated DNA aptamer sequences and their corresponding molecular targets. Only records with complete sequences and a clearly defined target category were retained. The categories with sparse representation (eg, proteins or peptides with few sequences) were excluded to reduce class imbalance and improve model robustness. After filtering, 2 dominant target classes were identified: cell-target and steroid-target aptamers. “Cell” entries in the dataset were defined as macroscale, whole-cell SELEX targets, and “Steroid” entries were classified as small, low-molecular-weight hydrophobic ligands. The resulting datasets contained unique single-stranded DNA aptamer sequences, each labeled according to its target class.

Each sequence was standardized by converting to uppercase, removing ambiguous characters, and trimming whitespace to ensure fair class representation and balanced training. We applied class-weighting and random oversampling of the minority class (steroid) before training. Finally, a separate, independent dataset of 13 newly discovered aptamer sequences from our laboratory was prepared in the same format for testing. These sequences, experimentally verified to bind to steroid targets, were used exclusively to assess the model’s generalization capability beyond the training database.

Each DNA aptamer sequence underwent a 2-step feature-engineering process to transform it into a format better suited as input to the model. To capture local sequence motifs that could influence binding affinity, each aptamer sequence was decomposed into overlapping subsequences ranging from 3 to 6 nucleotides in length (character-level n-grams, n=3‐6). For example, the sequence “ACGTACGT” will generate 6-mer motifs such as “ACGTAC” and “CGTACG,” 4-mer motifs such as “ACGT,” and “CGTA,” and 3-mers such as “ACG” and “CGT.” These variable-length subsequences were then vectorized using a TF-IDF [17]. TF measures how often a k-mer motif occurs in a single aptamer sequence (Equation 1).

IDF measures how common the same k-mer motif is across all sequences in the entire dataset (Equation 2). A low IDF is a sign of a k-mer being common across multiple target groups, and is therefore potentially irrelevant for specific binding to a target class. A high IDF value is therefore preferred.

The TF-IDF factor (Equation 3) was then calculated. This helps identify motifs that frequently occur in a target set but are rare across the entire dataset, suggesting that those motifs may contribute toward binding with the target.

In addition to the TF-IDF, 2 global sequence-level features were also computed: guanine-cytosine (GC) content and total sequence length, which were also standardized and concatenated with the TF-IDF vector to form the final feature matrix used to train the classifier. The vector, therefore, now contains both local motif-level and global sequence-level properties solely derived from the primary DNA sequence, without requiring any structural predictions or folding models. To understand which sequence properties most influenced classification, feature importance was also determined for each of the 9 bagged models using gain (the improvement in model loss attributable to a feature across all splits of a model). Importance values were averaged across the ensemble to identify k-mer motifs and descriptors globally relevant to classification.

The classification model was built on an ensemble learning framework using extreme gradient boosting (XGBoost) decision trees. XGBoost was selected for this evaluation based on three primary reasons: (1) its performance is well established on structured, high-dimensional tabular data such as TF-IDF feature vectors; (2) it provides native support for handling class imbalance by assigning class weights; and (3) considering the modest size of the dataset, XGBoost has a lower chance of overfitting compared to more complex model architectures. As this study is intended primarily as a proof-of-concept for sequence-only target-type classification rather than comparative benchmarking of ML algorithms, a systematic comparison against alternative classifiers was considered outside the scope of this evaluation. Comparative model benchmarking is noted as a direction for future work (see the Future Work section).

A bagged 9-ensemble model strategy was used to reduce variance and improve generalization stability on a relatively sparse dataset. Each subdataset was split into 70% training, 20% testing, and 10% validation subsets. The validation subset was used for early stopping during training. Each model was trained on a unique bootstrapped subsample of the entire training set using a distinct random seed. The XGBoost classifier was trained with a multinomial softmax objective, using regularization parameters to prevent overfitting, including constraints on tree depth, minimum child weight, and subsampling rates. The number of boosting rounds was controlled by early stopping based on the fixed validation set (with multiclass log-loss).

Model performance was evaluated using accuracy, macroaveraged F₁-score, and per-class precision/recall to account for the imbalance between cell- and steroid-binding aptamers. Confusion matrices were used to visualize class-level classification patterns. Following training, the output probabilities from all 9 ensemble models were averaged at the soft-probability level. These averaged probabilities were then subjected to a post hoc calibration step using a logistic regression model trained on the same validation set, thereby improving the interpretability and reliability of the predicted confidence values. Finally, we implemented a per-class threshold optimization to identify decision thresholds that maximize the macroaverage. The trained, calibrated ensemble was then used to generate predictions for the held-out test set. To quantify the stability of reported performance metrics, given the small test set size, 95% CIs for accuracy and macro-F₁ were estimated via the nonparametric bootstrap resampling of test-set predictions for both the argmax and threshold-tuned decision rules.

In addition to the model’s performance on the curated dataset, we evaluated its generalization to an independent set of experimentally generated DNA aptamer sequences, which is discussed in detail in the In Vitro Aptamer Selection and Validation section. These sequences were discovered through laboratory screening using SELEX, targeting steroidal molecules, and were not part of the original aptamer database or any training/calibration subsets used during model development. The external dataset served as a true out-of-domain validation benchmark, representing the type of novel aptamers encountered during early-stage discovery. To ensure consistency, each sequence from the external set underwent the same preprocessing and feature transformation pipeline as the training data, including standardization and TF-IDF feature extraction, without retraining or refitting any part of the model. Moreover, to ensure that memorization of near-duplicate training sequences is not a factor, sequence novelty was computed for each of the 13 new sequences as the normalized Levenshtein distance to its nearest neighbor in the training set. Finally, to assess whether calibrated model confidence tracked binding strength, the Spearman rank correlation between each aptamer’s calibrated steroid-class probability and its measured K_d was computed. The graphical overview of the process is presented in Figures 5 and 6.

Full hyperparameter values and training configurations are available in the code repository, mentioned in the Data Availability section. Dataset preprocessing and model training were performed in Python. Secondary-structure analyses of the newly isolated aptamers in the laboratory were performed using MFold and NUPACK. Binding confirmation assays used Thioflavin T (ThT) fluorescence measurements, with curve fitting on GraphPad Prism 10 (described in detail in the In Vitro Aptamer Selection and Validation section).

Unless otherwise stated, all reagents were used without further purification. All oligonucleotides and modified oligonucleotides (biotinylated, fluorescein, and dabcyl-tagged) were purchased and purified via reverse-phase high-performance liquid chromatography from Sigma Aldrich. ALDO was purchased from Sigma Aldrich. Buffers, 10× concentrated phosphate-buffered saline (PBS) and 10× concentrate tris-borate-ethylenediaminetetraacetic acid, were obtained from Corning. Streptavidin agarose resin was obtained from Thermo Scientific. Mini gravity columns were purchased from Biorad. Ultracentrifugal filter units (3K and 10K) were purchased from Millipore. DNase/RNase-free water was used for all purposes unless otherwise noted. All cloning materials were purchased from Fisher Scientific. For this selection, the oligonucleotides were obtained and used following our previously published library design shown in Textbox 1.

The SELEX reaction buffer used for this selection procedure was 1× PBS concentrate with 2 mM of MgCl2 and pH adjusted to 7.4. The ALDO stock solution was prepared at 10 mM in 200-proof ethanol and stored in a refrigerator. A fresh solution of ALDO at the desired concentration was prepared before each round of SELEX by dissolving the ALDO stock solution in the SELEX buffer. The 0.5× concentrate tris-borate-ethylenediaminetetraacetic acid buffer was used as an electrophoresis buffer. The strand separation buffer was 1× PBS concentrate without MgCl₂. All oligonucleotides were dissolved separately in DNase/RNase-free water to prepare 100 μM stock solutions of each, which were divided into small aliquots sufficient for one SELEX round, and stored at −20 °C. The selection process has been described previously by us [14,15,18].

A 750 μL aliquot of 100 μM ThT solution was mixed with an equal volume of 10 μM aptamer solution and incubated for 40 minutes at room temperature, protected from light. Ten tubes were labeled, and 125 μL of SELEX buffer with 2 mM MgCl₂ was added to tubes 2 to 10. In tube 1, 250 μL of 200 μM ALDO was added. Serial dilutions were made up to tube labeled 9. The tube labeled 10 did not contain ALDO. This was followed by adding 125 μL of ThT/aptamer solution to each tube. The mixtures were incubated for 40 minutes at 30 °C, protected from light. Measurements were conducted in triplicate in 384-well black plates (Costar) using a microplate reader M2 (Molecular Devices) to record ThT fluorescence spectra at excitation and emission wavelengths of 425 and 490 nm, respectively. The results were analyzed using GraphPad Prism 10 using nonlinear regression curve: curve fit, exponential, one phase decay [19].

The ensemble classifier achieved strong performance on the held-out test set. Using calibrated probabilities with a standard argmax decision rule, the model obtained an accuracy of 95.83% (95% CI 89.58-100.00) and a macro-F₁-score of 0.94 (95% CI 0.83-1.00; Figure 7). After applying class-specific threshold tuning, overall accuracy improved to 97.92% (95% CI 93.75-100.00) with a macro-F₁ of 0.97 (95% CI 0.89-1.00), indicating that threshold optimization enhanced minority-class recovery without loss of precision. The confusion matrix (Figure 7) shows that the model consistently identified steroid-binding sequences while only misclassifying a small number of cellular-class aptamers. These results confirm that short-range and mid-range sequence motifs, combined with minimal global descriptors (GC content and sequence length), are sufficient to distinguish between aptamers targeting cellular and steroidal molecules.

The gain-based feature importances, averaged across all ensemble models, are shown in Figure 8. Several of the short-string motifs are substrings of the longer (6-mer) sequences, as expected, because overlapping character n-gram windows capture the same underlying patterns. In addition to motif-level contributions, sequence length was identified as a frequently used feature based on weight-based importance. Given the sequence length in the feature importance list and its weight-based ranking (Table 1), we examined whether sequence length differed between the 2 target classes. Cell-binding aptamers were significantly longer (n=187; mean 83.4, SD 25 nt) than steroid-binding aptamers (n=49; mean 63.1, SD 19.5 nt; Mann-Whitney U=6917.0, P=3.9 × 10⁻⁸; Figure 9). In this example, while this finding indicates that sequence length holds a genuine class-splitting signal, it should not be correlated to the size of the target in general, that is, the same-sized library can be used not only to detect small molecules successfully but also for selection of large molecules, such as proteins and enzymes, that is, their active sites. In most cases, the sequence length is between 20 and 80 base pairs, and there is no specific protocol defining which library to select for SELEX. In this study, we used a random (N30) library: 5′-GGAGGCTC­TCGGGACGAC (N30) GTCGTCC CGCCTTTAG­GATTTACAG-3′ [14].

The 13 aptamers isolated in our laboratory are shown in Table 2. All sequences in this set were experimentally confirmed to bind to ALDO. The ALDO aptamer set shows strong sequence similarity but different ThT affinities, implying that relatively small sequence changes shift the 3D folding and binding affinities of aptamers for ALDO. Most sequences share conserved flanking motifs typical of a selected library scaffold, that is, frequent ACGACG/CGACG located at the 5′ and GC-rich termini (eg, AGTCGT/TGTCG), while the internal regions vary in length and composition. This pattern suggests that the main molecular recognition motifs for ALDO binding are not the specific base composition but rather a specific secondary-structure arrangement (eg, an exposed stacked stem, internal bulge, or groove) that can accommodate this steroidal target.

ThT K_d values range approximately 7 orders of magnitude, separating the selected aptamers into several subgroups. The tightest ALDO binders are ALDO 06 and ALDO 05 with K_d values of 0.3785 µM and 0.7306 µM, respectively, followed by ALDO 14 with a K_d value of 2.248 µM. Moderate-affinity ALDO binders include ALDO 15, ALDO 17, and ALDO 10, with K_d values in the ThT assay of 26.2 µM, 29.57 µM, and 36.68 µM, respectively. Weaker binders include ALDO 13 (K_d=225.7 µM) and ALDO 23 (K_d=281.8 µM). At the same time, several sequences show no binding under these conditions: ALDO 02, ALDO 11, ALDO 16, ALDO 22, and ALDO 07, with K_d values above 10,000 µM. We believe these nonbinding aptamers probably still bind to ALDO but likely fold into compact, thermodynamically stable complexes that are not suitable for ThT assay conditions. This is something that needs to be explored in the future—that is, determining K_d values for these sequences using other fluorescent assays and conditions. We tried to correlate the obtained in vitro K_d values to predicted folding free energies and observed that stability alone does not predict their binding to ALDO. Both MFold and NUPACK predict a wide range of stabilities, with MFold values ranging from −2.99 to −11.63 kcal/mol and NUPACK from −2.53 to −11.52 kcal/mol. The best aptamer binders, ALDO 06 and ALDO 05, are predicted to be fairly stable in both MFold and NUPACK, supporting the idea that a defined fold is necessary for high-affinity ALDO interactions. However, other examples demonstrate that high predicted stability is not sufficient, exemplified by ALDO 02, which is among the most stable by both models (MFold −11.63 kcal/mol and NUPACK −11.52 kcal/mol) but shows no binding to ALDO (>10,000 µM). On the other hand, ALDO 14 shows good binding affinity for ALDO with a K_d value of 2.248 µM, while relatively weak predicted stability (MFold −3.28 kcal/mol and NUPACK −4.44 kcal/mol), implying that the geometry of the fold (even if only modestly stable) can be more important than ΔG magnitude. Taken together, these comparisons indicate that ALDO K_d values are primarily sensitive to structural class/topology (which elements are exposed and how bases stack). In contrast, MFold/NUPACK ΔG values largely predict overall base-pairing stability without considering the formation of the target-binding complexes.

To evaluate generalization beyond the curated database, these aptamers were fed to the completely trained ensemble model and were not a part of the training or validation datasets. To assess the novelty of each of the 13 new laboratory-created aptamer sequences relative to the UTexas Aptamer Database sequences used to train the model, we calculated pairwise sequence similarity between the 2 groups using the normalized Levenshtein distance. The mean normalized distance between each of the 13 laboratory aptamers and its nearest neighbor in the training set was 0.40 (SD 0.07), with no sequence sharing more than 75% similarity with an aptamer in the training sequences. This confirmed that the 13 aptamers are genuinely unique, and the model’s performance on these sequences shows generalization rather than memorization. Detailed information on the Levenshtein distances for each of these sequences is provided in Table 3.

While both raw and calibrated argmax misclassified a subset of borderline cases, the final calibrated and thresholded ensemble successfully identified all 13 sequences, achieving 100% recall. To explore if the model confidence scores reflect binding quality, we calculated the Spearman rank correlation between the calibrated steroid-class probability and experimentally measured K_d values for the 13 laboratory-generated aptamers. An insignificant negative correlation (ρ=−0.05, P=.08), indicating no statistically significant or practically meaningful relationship between calibrated steroid-class probability and experimental K_d values, suggests that the model assigned higher confidence to sequences with lower K_d values (ie, better binding). Although this test did not reach statistical significance at the conventional P<.05 level, the results are still notable given the small sample size (n=13). Nevertheless, model confidence does not reliably predict binding affinity at this stage.

The raw and calibrated probability distributions for each sequence are presented in Figures 10-12. The apparent improvement in the reliability of probability estimates does not reduce confidence. Calibration rescales overconfident outputs, so that the predicted values more closely reflect the true likelihood of correct classification, thereby producing interpretable probabilities that generalize more accurately to new experimental sequences.

This preliminary study found that a sequence-only ensemble ML framework can accurately distinguish between cellular and steroidal DNA aptamers using TF-IDF motif representations. The calibrated XGBoost ensemble achieved strong classification performance on both the held-out test set and an independent set of experimentally derived steroid-binding aptamers, demonstrating the model’s ability to generalize beyond curated training datasets. These initial findings suggest that a purely sequence-based predictive framework can meaningfully complement wet-laboratory SELEX discovery by prescreening candidate sequences for binding potential. The ability to produce calibrated, interpretable probability estimates enables the selection of aptamer candidates for validation, potentially reducing experimental cycles and resource requirements. In addition, feature importance analysis revealed that the model potentially relied on recurring relevant sequence motifs and global sequence length for classification.

Prior approaches to aptamer discovery largely relied either on thermodynamic structure prediction (eg, NUPACK or MFold) or on training on aptamer-target pairs (eg, AptaBLE [12]). While these methods have significantly advanced the field, they are computationally expensive, scale poorly to large aptamer libraries, and are sensitive to inaccurate folding predictions. Moreover, models such as AptaBLE require retraining for each new target, limiting their applicability across the multiple categories commonly encountered in aptamer discovery. Hybrid methods, such as combining ML with genetic algorithms [18], have also been explored for aptamer design, but focus on generating aptamer candidates for a single predetermined target rather than classifying candidates across multiple target categories. By contrast, the sequence-only, TF-IDF framework presented here does not require structural modeling. It can be extended to multiple target classes, thereby providing a lightweight complement to the methods above.

In practice, a classifier of this kind could serve as an early triage step within an aptamer discovery pipeline: candidate sequences that emerge from initial SELEX rounds, in silico libraries, or public aptamer databases could be quickly scored for target-class compatibility before committing additional rounds of SELEX. In the evaluation case presented here, the model correctly classified all 13 laboratory-formulated SELEX-derived new ALDO aptamers as steroid-binding, achieving 100% recall on sequences confirmed to be genuinely novel to the training set (mean normalized Levenshtein distance 0.40, SD 0.07). As aptamer databases continue to grow, sequence-only classifiers of this kind could serve as a prescreening layer to quickly identify viable aptamer candidates for a target class, thereby reducing SELEX cycles and leading to faster, higher-throughput biosensor development.

Several limitations should be considered when interpreting the results of this study. The curated dataset was relatively small and imbalanced toward cell-binding aptamers, which may limit the diversity of sequence patterns learned by the model, despite the use of oversampling and class weighting. Additionally, although the model successfully generalized to 13 independently generated SELEX-derived steroid aptamers, further validation on larger, more diverse experimental datasets will be necessary to assess robustness across broader target classes. Finally, because the model selection was not benchmarked against alternative classifiers, the reported performance reflects the performance of the XGBoost ensemble specifically rather than a comparison with other algorithmic approaches, and it remains possible that other models could achieve comparable or improved performance on this task. Additionally, although feature importance analysis identified recurring motifs for classification, these only reflect statistical association within the dataset and model rather than actual structural roles in binding. Therefore, further investigation is needed to establish them as irrefutable determinants of target specificity.

Future work will extend this framework to multitarget classification. Regression on continuous K_d values, combined with structure-based features to capture higher-order sequence dependencies, will also be explored. Future work will also include systematic benchmarking against other classifiers to evaluate whether the observed performance is generalizable across various model architectures.