Deep learning approaches

Early neural approaches predominantly utilized convolutional neural networks (CNNs) to capture local syntactic triggers, as demonstrated by models like DCNN [23] and SCNN [24]. While highly efficient at extracting explicit morphological cues within localized windows, CNNs inherently struggle to capture long-range dependencies across complex sentence structures. To bridge this gap, recurrent neural networks (RNNs) [25–27] and graph neural networks (GNNs) [28,29] were introduced to model sequential continuity and dependency trees. More recently, Transformer-based models (e.g., BioBERT [30], MEAT-BioBERT [31]) have established state-of-the-art results by leveraging global self-attention mechanisms for deep semantic modeling. Advanced variants have further enhanced performance by incorporating external knowledge graphs or multi-task learning strategies [32–35].

However, single-paradigm architectures often face an inherent trade-off: CNNs may miss the global context required for logical inference, while Transformers, despite their global receptive field, typically suffer from a quadratic computational complexity () and can overlook fine-grained local syntactic triggers due to dispersed attention weights.

State-space models and hybrid frameworks

State-space models (SSMs), particularly the recently proposed Mamba architecture [36], have emerged as a powerful alternative for long-sequence modeling. Unlike Transformers, which calculate pairwise attention scores, Mamba compresses historical context into a continuous latent state, enabling highly efficient linear-time propagation () [37,38]. This mathematical property makes Mamba exceptionally suitable for modeling the implicit logical dynamics that span long distances in biomedical narratives.

Despite its success in general domains, the application of selective SSMs in biomedical relation extraction remains underexplored. While Mamba excels at modeling sequential continuity (implicit logic), it lacks the explicit token-to-token alignment capability of attention mechanisms or the sharp focus of CNNs on localized triggers. Consequently, current approaches often suffer from a single-granularity bias, focusing excessively on either local syntax or global semantics.

To address these limitations, we propose DuSSM as a parallel reasoning framework. Unlike previous hybrid models that rely on sequential layer stacking, DuSSM utilizes a parallel architecture to structurally disentangle feature spaces. It assigns the morphological pattern recognition task to an explicit stream (powered by a local 1D-CNN) and delegates the long-range logical reasoning task to an implicit stream (driven by Mamba). This structural synergy allows the reasoning module to simultaneously capture high-frequency triggers and subtle semantic evolution, ensuring robust determinism and high precision even on severely imbalanced datasets.

Dual-stream manifold decomposition

As illustrated in Fig 2, DuSSM operates on the hypothesis that biomedical relations exhibit a signal duality: high-frequency explicit triggers (e.g., enzymatic verbs) and low-frequency implicit logic (e.g., causal trajectories across long distances). To prevent the mutual interference of these features, we formalize the overall feature space as the direct sum of two orthogonal subspaces:

(1)

where acts as a local pattern recognizer to capture specific syntactic triggers, and serves as a global semantic integrator. We implement this via a four-stage pipeline: (1) Contextual encoding: Mapping tokens to a high-dimensional manifold using a PubMedBERT backbone; (2) Dual-stream feature separation: Projecting embeddings into two specialized subspaces; (3) Entity-aware fusion: Calibrating semantic features at precise entity coordinates; (4) Imbalance-aware optimization: Refining the decision boundary via an adaptive focal manifold.

Crucially, the necessity of a parallel dual-stream design, as opposed to conventional sequential hybrid models, stems from the problem of feature interference. In a standard sequential architecture, if an explicit local trigger (captured by the CNN) is fed serially into a state-space module, its sharp, high-frequency signal is inevitably smoothed out or diluted by the continuous recurrent state transitions of the global sequence. Conversely, if Mamba precedes the CNN, the global semantic logic is fragmented by the local pooling operations.

To mathematically prevent this interference, DuSSM introduces a strictly parallel topological design. By processing the projected space through two independent, orthogonal pathways, the Explicit Stream acts as an uncorrupted high-pass filter for local triggers, while the Implicit Stream independently models the global semantic trajectory without being biased by localized noise. This structural disentanglement ensures that both morphological patterns and long-range logic are preserved in their native feature manifolds before being robustly integrated at the final entity-aware fusion stage.

Input representation and feature projection

Let be an input sequence containing target entities e₁ and e₂. We leverage PubMedBERT [39], pre-trained on 14 million PubMed abstracts, to ensure the latent representation is natively aligned with the clinical domain. In our implementation, we freeze the lower layers of the encoder to retain general syntactic knowledge while fine-tuning the upper layers to adapt to the relational extraction task.

It is important to acknowledge that the contextual encoding phase relies on a Transformer backbone, which theoretically possesses a quadratic computational complexity and a hard maximum sequence limit of 512 tokens. However, clinical sentences in standard DDI corpora are typically short (e.g., average length ≈ 42 tokens, truncated at a maximum of 300 tokens in our preprocessing). Within this practical length distribution, the quadratic overhead of the encoder is mathematically and empirically negligible. The true computational bottleneck and logical entanglement in modern biomedical extraction typically occur in the subsequent relational reasoning and feature fusion phases. Therefore, our core architectural contribution lies in replacing traditional computationally heavy cross-attention mechanisms with the proposed linear-time DuSSM reasoning module.

To bridge the contextual encoder and the reasoning module, the contextual embeddings are subsequently projected into a shared lower-dimensional latent space via a linear transformation, yielding :

(2)

Dual-stream disentanglement module

To resolve the entanglement inherent in monolithic serial models, we design two parallel pathways that process independently.

Explicit stream: Local pattern recognizer.

The explicit stream is engineered as a local 1D-CNN to capture lexical spikes with high sensitivity. From a signal processing perspective, the local convolution functions as a high-pass filter that isolates sparse, high-energy morphological cues. For a kernel window K, the response at position t is formulated as:

(3)

This mechanism ensures that precise triggers are preserved while background semantic noise is effectively attenuated.

Implicit stream: Selective state-space dynamics.

In parallel, the implicit stream models the global semantic trajectory using a selective state-space model (Mamba). Mamba maps the projected input token to a latent state via discretized dynamics:

(4)

where . Crucially, the model achieves logical selectivity through data-dependent parameters:

(5)

This formulation allows for dynamic semantic filtering: the step size adaptively contracts to skip parenthetical clinical background and expands to record critical logical evolution. This ensures a strict complexity while maintaining a stable semantic memory orthogonal to the explicit stream’s local responses.

Feature fusion and orthogonal calibration.

To establish a robust decision boundary, we integrate the disentangled views at entity coordinates and . The fused representation ensures that localized triggers do not overshadow long-range logical constraints. We achieve this by concatenating the independent features along the channel dimension:

(6)

The final classification vector integrates the global representation with the target-specific disentangled semantics via concatenation:

(7)

The output probabilities are derived via .

While more complex fusion mechanisms were considered during the architectural design, we intentionally adopted straightforward concatenation to preserve the strict orthogonality of the disentangled features. The core philosophy of DuSSM is to structurally prevent local morphological noise from interfering with global semantic logic. Advanced fusion strategies that dynamically mix and project these spaces risk re-entangling the features prior to classification, inherently negating the benefits of the dual-stream separation. Furthermore, concatenation introduces zero additional parameters and strictly maintains the computational efficiency of the framework. In our preliminary testing, introducing a gated fusion module yielded negligible F1 fluctuations (within ±0.2%) but noticeably increased inference latency. Therefore, concatenation serves as the most robust, interpretable, and computationally optimal strategy for this disentangled architecture.

Imbalance-aware optimization

To mitigate the majority-class dominance of negative samples (e.g., DDI-false or none classes), we utilize an adaptive focal loss:

(8)

By setting , the model prioritizes hard-to-classify positive interactions, refining the decision manifold to ensure high clinical reliability.

Datasets and preprocessing

To rigorously evaluate the generalizability and robustness of DuSSM, we conduct experiments across four heterogeneous biomedical benchmarks. These datasets represent various relational scales and semantic complexities in the biomedical domain:

• DDI Extraction 2013 [40]: A standard benchmark for drug-drug interactions, containing texts from DrugBank and MedLine with five categories (Mechanism, Effect, Advise, Int, False).
• ChemProt [41]: A complex chemical-protein interaction dataset from BioCreative VI, known for its high-density of none relations (>90%), making it ideal for testing clinical reliability.
• GAD [42] & EU-ADR [43]: Two diverse corpora focused on gene-disease associations. We follow the standard practice of 10-fold cross-validation on these datasets to ensure statistical significance.

Table 1 summarizes the data distribution across all four benchmarks. For preprocessing, all target entities were blinded with generic tokens (e.g., @DRUG$, @GENE$) to force the model to rely on structural logic rather than entity-specific memorization.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Statistical distribution of the four biomedical benchmarks used in evaluation.

https://doi.org/10.1371/journal.pcbi.1014480.t001

Implementation details

DuSSM was implemented using PyTorch 2.4.0 on a workstation with an NVIDIA RTX-4090 GPU (24GB). We initialized the contextual backbone with PubMedBERT, which is pre-trained on 14M PubMed abstracts, ensuring a specialized vocabulary for biomedical entities.

The initial hidden representations were mapped to a latent space of . For the explicit stream, we employed a local 1D-CNN with a kernel size of K = 3 and a dropout rate of 0.3. For the implicit stream, the Mamba module was configured with a state dimension , a local convolution width , and an expansion factor of 2. We adopted a differential learning rate strategy: for the fine-tuned upper layers of the encoder and for the dual-stream reasoning modules, optimized via AdamW with a 10% linear warmup. To mitigate the severe class imbalance observed in ChemProt and DDI 2013, we employed focal loss () with empirically derived class-specific weighting factors (e.g., for the DDI dataset). The maximum sequence length was fixed at 300 tokens across all experiments to maintain a consistent comparison of computational overhead.

Main results

Overall performance comparison.

Table 2 presents a comprehensive comparison of DuSSM against standard methods categorized in the recent survey by Dou et al. [44], alongside several contemporary models published in 2024 and 2025. We selected representative baselines across different architectural paradigms, including early neural models (SCNN [24], MCCNN [45], ATT-LSTM [28]), standard Transformer-based approaches (R-BERT [29], MEAT-BioBERT [31]), and a recent advanced architecture (CA-SQBG [46]).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Performance comparison with baseline models on DDI Extraction 2013.

https://doi.org/10.1371/journal.pcbi.1014480.t002

As shown in Table 2, DuSSM achieves an F1-score of 82.27% and an Accuracy of 94.56%. Our model significantly outperforms early deep learning baselines, surpassing the dependency-based MCCNN by +12.06% and the attention-based ATT-LSTM by +4.97% in F1-score.

To ensure a rigorous evaluation against recent advancements, we compared DuSSM with the contemporary 2025 year model CA-SQBG as well as recently emerging structured reasoning and hybrid sequence models [13,47–49]. Compared to the CA-SQBG architecture (73.10%) and other recent graph-neural-network (GNN) paradigms, our dual-stream approach yields a substantial +9.17% improvement in F1-score. While recent GNN-based and structured models often heavily rely on complex external knowledge graphs, multi-stage syntactic parsing, or hybrid topological mapping to achieve competitive scores [48,49], DuSSM demonstrates that highly robust performance can be achieved purely from unstructured text. By structurally separating morphological patterns from semantic evolution, DuSSM effectively matches or exceeds the relational sensitivity of these complex structured models, while simultaneously providing a strictly linear-time computational efficiency. This marks a significant step toward scalable clinical deployment.

Furthermore, while we highlighted CA-SQBG to demonstrate superiority over recent graph-based paradigms, it is crucial to position DuSSM against the strongest overall baseline in our study, MEAT-BioBERT. MEAT-BioBERT achieves a highly competitive F1-score of 80.90% by employing multiple entity-aware self-attention layers. However, this relies on a computationally heavy architecture. DuSSM not only surpasses MEAT-BioBERT by +1.37% in F1-score and +3.53% in recall, but it achieves this relational sensitivity using a structurally decoupled, linear-time reasoning module.

Regarding the precision metric, DuSSM (82.15%) successfully outperforms MEAT-BioBERT (81.00%). However, it is worth noting that the DREAM model achieves a marginally higher precision of 82.30%. This observation reveals an important precision-recall trade-off: DREAM’s slightly higher precision comes at a severe cost to its recall, which plummets to 74.70% (compared to DuSSM’s robust 84.43%). An over-conservative decision boundary that minimizes false positives by sacrificing recall leads to excessive false negatives—meaning actual, potentially dangerous drug interactions would be missed. By contrast, DuSSM provides a highly optimized precision-recall balance, ensuring that high specificity does not compromise the comprehensive detection of critical clinical alerts.

Class-wise analysis and reliability.

To assess clinical reliability, we analyze the performance across specific DDI types in Table 3. A critical challenge in pharmacovigilance is minimizing false alarms (False Positives). DuSSM exhibits exceptional performance in this regard, achieving a precision of 99.00% on the negative class (DDI-false). This high specificity ensures that the system rarely flags non-interacting pairs as dangers, significantly reducing the verification burden on clinicians.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Detailed performance of DuSSM per relation type.

https://doi.org/10.1371/journal.pcbi.1014480.t003

Crucially, our model demonstrates a substantial improvement in the Advise category. Sentences describing Advise are structurally complex and heavily reliant on implicit logical conditions. As shown in Table 3, DuSSM achieves a remarkable recall of 95.02% for DDI-advise. This result empirically supports our hypothesis that the implicit stream (Mamba) successfully captures long-range logical dependencies that traditional CNNs or shallow Transformers often miss.

In contrast, the DDI-int class shows lower performance (F1: 63.29%). This is primarily attributed to the extreme data scarcity (only 96 test instances), suggesting a potential avenue for future improvement via few-shot learning strategies. However, the high precision (80.65%) on DDI-int suggests that when the model does predict an interaction, it is highly trustworthy.

Fig 3 visualizes the confusion matrix on the test set. The model exhibits a strong diagonal dominance, particularly for the majority class (DDI-false), where 4552 instances were correctly identified with minimal leakage into positive classes. Notably, for the DDI-advise category, which typically involves complex conditional sentences, the model correctly classified 210 out of 221 instances, confirming the effectiveness of the implicit stream in capturing logical dependencies.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Confusion Matrix of DuSSM on the DDI Extraction 2013 test set.

https://doi.org/10.1371/journal.pcbi.1014480.g003

To further evaluate the clinical reliability, we present the Receiver Operating Characteristic (ROC) and Precision-Recall (PR) curves in Fig 4.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Discriminative performance analysis.

https://doi.org/10.1371/journal.pcbi.1014480.g004

As shown in Fig 4A, DuSSM achieves an Area Under the Curve (AUC) of 0.99 for the DDI-advise class and 0.98 for DDI-mechanism. The steep ascent of the ROC curves towards the top-left corner indicates a high true positive rate at low false positive rates, which is critical for minimizing alert fatigue in clinical decision support systems.

Fig 4B illustrates the Precision-Recall trade-off. Despite the extreme class imbalance, the DDI-advise category maintains a high Average Precision (AP) of 0.91. This empirically validates that our Mamba-based architecture successfully retains long-range context without being overwhelmed by local noise.

Robustness evaluation against baselines.

While Fig 4 illustrates the per-class ROC and Precision-Recall curves for our DuSSM framework, demonstrating strong discriminative ability across distinct interaction types, threshold-independent metrics such as AUC and AUPRC are crucial for comparing overall robustness against baseline methods under extreme class imbalance.

Since generating exact comparative curves requires raw prediction probabilities that are unavailable for externally reported baselines, we compare the macro-averaged metrics of DuSSM against our locally reproduced core baselines (the standard PubMedBERT encoder and the strong MEAT-BioBERT) in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Comparison of threshold-independent robustness metrics.

https://doi.org/10.1371/journal.pcbi.1014480.t004

As shown in Table 4, DuSSM achieves a AUC of 0.926 and a AUPRC of 0.766. Compared to the strongest baseline, MEAT-BioBERT, our model yields a substantial +0.045 improvement in AUPRC. This significant gain in AUPRC—a metric highly sensitive to false positives in imbalanced datasets—confirms that our dual-stream disentanglement effectively filters morphological noise, maintaining a highly robust precision boundary even at elevated recall thresholds compared to conventional architectures.

Mechanism analysis: Dynamic semantic gating & UMLS alignment

A persistent barrier in biomedical linguistics is the signal-to-noise bottleneck, where redundant clinical descriptors obscure distal causal logic. To substantiate that DuSSM transcends mere statistical co-occurrence, we analyze the input-dependent selection parameter (the discretization step size) within the implicit stream. To clarify its mechanistic calculation, is dynamically computed for each input token via a learned linear projection followed by a Softplus activation, formulated as . Conceptually, because Mamba treats sequence modeling as a continuous-time dynamic system, this discretization step size represents the “time interval” or “information bandwidth” allocated to the current token. Consequently, it functions as a dynamic memory gate. When is small, the gate narrows, restricting new information and forcing the model to persist its historical latent state (i.e., effectively ignoring the current token). Conversely, when is large, the gate opens widely, triggering a rapid update of the latent state with the new contextual information.

As visualized in the selection manifold (Fig 5), this mechanism effectively learns a concrete, domain-aware reading strategy. Rather than processing all words equally, the model automatically divides the input sequence into two distinct behavioral patterns:

• Semantic suppression: Upon encountering non-relational segments—such as irrelevant clinical filler words or background descriptions—the magnitude of undergoes a pronounced contraction (↓ 62% relative to the global mean). Mathematically, this forces the state-space into a preservation mode, effectively skipping tokens that reside outside the relational manifold.
• Logic capture: Conversely, when reading critical relational anchors (e.g., target entities like “@DRUG-A,” or explicit verbs like “inhibition,” “toxicity”), demonstrates sharp, synchronized surges. This dynamic gating ensures the hidden state undergoes high-fidelity transitions to record critical interaction logic, minimizing the risk of information decay.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Selection intensity () across biomedical sequences.

https://doi.org/10.1371/journal.pcbi.1014480.g005

Crucially, our knowledge-alignment analysis reveals that 78.4% of high- peaks precisely align with clinical entities or functional verbs indexed in the UMLS metathesaurus. Specifically, this alignment was calculated as the percentage of tokens within the top-10% activation range that exhibit an exact string match with verified UMLS semantic types (e.g., Pharmacologic Substance [T121], Disease or Syndrome [T047]) or a curated lexicon of biomedical interaction verbs. Furthermore, the synergy between explicit stream spikes () and implicit stream peaks () confirms a dual-verification mechanism: the explicit stream localizes the where (morphological triggers), while the implicit stream deciphers the how (causal evolution). This structural determinism provides a robust safeguard against the attention dispersion and logical hallucinations prevalent in monolithic Transformer-based architectures and large language models.

To further elucidate how DuSSM coordinates both local triggers and long-range logic as requested, we visualize the internal activation weights of both streams on a representative complex clinical sentence (Fig 6). For the Explicit Stream, we plot the normalized multi-scale CNN feature maps (prior to max-pooling). For the Implicit Stream, we visualize the temporal evolution of the Mamba hidden state magnitude ().

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Visualization of dual-stream activation weights on a sample DDI sequence.

https://doi.org/10.1371/journal.pcbi.1014480.g006

As illustrated in Fig 6, the Explicit Stream functions as a highly localized syntactic detector. Its activation weights form sharp, isolated spikes precisely at morphological triggers (e.g., explicit verbs such as “inhibits,” “metabolism”), while remaining entirely dormant across unrelated context words. In stark contrast, the Implicit Stream exhibits a continuous, long-range evolutionary pattern. It successfully maintains high activation from the subject entity (“@DRUG-A”) across the intervening clauses, integrates the semantic shift at the trigger verbs, and carries this contextual momentum to the object entity (“@DRUG-B”).

This dual-visualization explicitly confirms our architectural hypothesis: the CNN isolates the syntactic where (local triggers), while the Mamba block models the logical how (long-range entity relationships). Their orthogonal fusion forms a comprehensive and highly interpretable relational understanding.

Statistical significance test

To ensure that the performance improvements of DuSSM are statistically significant and not merely due to random chance, we conducted a paired t-test comparison against the strongest baselines, including the generative Med-Llama 3 (8B). We performed 10 independent runs for the non-generative models using different random seeds and evaluated the variance across Precision, Recall, and F1-score.

As reported in Table 5, DuSSM achieves a mean F1-score of 82.27% (±0.45), consistently outperforming the strong baseline R-BERT (, p < 0.01). Furthermore, when compared to the strongest non-generative baseline, MEAT-BioBERT, the performance gains achieved by DuSSM are statistically significant across all three metrics: Precision (p = 0.021), Recall (p = 0.015), and F1-score (p = 0.034).

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Statistical significance test across 10 independent runs.

https://doi.org/10.1371/journal.pcbi.1014480.t005

Crucially, when compared to the 8-billion parameter Med-Llama 3 (mean F1 83.25%), the statistical test yields a p-value of 0.086 (p > 0.05). This confirms that the slight numerical lead of the LLM in F1-score is not statistically significant. However, an analysis of the underlying metrics reveals a compelling trade-off: while the generative LLM achieves a higher recall (), DuSSM achieves a statistically significant superiority in Precision (82.15 vs. 80.15, p < 0.05). Therefore, our 160M-parameter DuSSM achieves statistically equivalent relational sensitivity to a massive LLM, while providing significantly higher precision on non-interactions (minimizing false alarms) and enabling strictly linear edge deployment.

Training dynamics and stability

To strictly assess the optimization efficiency and generalization capability of DuSSM, we monitored the training loss and F1 scores throughout the 22-epoch training process.

As illustrated in Fig 7A, the training dynamics exhibit two distinct phases: a rapid adaptation phase (Epochs 1–5) where the loss drops precipitously, and a stabilization phase (Epochs 6–22) where the curve flattens. This smooth convergence trajectory indicates that the proposed hybrid architecture maintains healthy gradient flow, effectively avoiding the vanishing gradient issues often associated with recurrent mechanisms.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Training dynamics of DuSSM.

https://doi.org/10.1371/journal.pcbi.1014480.g007

Fig 7B tracks the evolution of the F1 score. Crucially, the Dev F1 score closely follows the Training F1 score without exhibiting significant divergence or oscillation. The narrow gap between the training and validation curves confirms that the model generalizes well to unseen data rather than merely memorizing the training set. This stability validates the effectiveness of our regularization strategies in mitigating overfitting, despite the complexity of the parallel dual-stream design.

Ablation study and architectural topology analysis

To verify the contribution of each component and the rationale behind the parallel design of DuSSM, we conducted extensive ablation studies and topological comparisons. The results are summarized in Table 6.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Ablation study on component contributions and topological design.

https://doi.org/10.1371/journal.pcbi.1014480.t006

Visualization of feature representations

To intuitively validate the effectiveness of the proposed disentanglement strategy, we visualized the high-dimensional feature representations of the test set using t-SNE. We extracted the final classification vectors (before the Softmax layer) from both the BioBERT baseline and our DuSSM framework.

As illustrated in Fig 8A, the feature space of the baseline model exhibits severe entanglement. The minority classes (e.g., Advise and Mechanism) are heavily intertwined with the dominant False class, resulting in blurred decision boundaries and high false-negative rates. This observation supports our hypothesis that serial architectures struggle to differentiate subtle logical triggers from background noise.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. t-SNE visualization of learned feature spaces.

https://doi.org/10.1371/journal.pcbi.1014480.g008

In stark contrast, Fig 8B demonstrates that DuSSM learns a highly discriminative manifold. The interaction categories are projected into compact, well-separated clusters, with clear margins between the False class and positive interactions. Notably, the Advise category, which involves complex conditional logic, forms a distinct island in the DuSSM space, whereas it is scattered in the baseline. This confirms that the parallel explicit and implicit streams successfully disentangle morphological surface patterns from deep logical semantics, enabling the classifier to capture fine-grained relational nuances.

Computational efficiency analysis

In clinical deployment scenarios, such as real-time EHR monitoring, inference latency is as critical as accuracy. We compared the inference throughput (samples per second) of DuSSM against representative Transformer-based baselines (BioBERT-Base and R-BERT) on the same hardware environment (NVIDIA RTX-4090, Batch Size = 32).

As visualized in Fig 9, DuSSM achieves a throughput of 590 samples/s, substantially outperforming BioBERT (365 samples/s) and R-BERT (340 samples/s). Despite having a dual-stream architecture, DuSSM achieves a 1.6× speedup over the backbone model. This efficiency gain is primarily attributed to the linear complexity of the Mamba module, which eliminates the quadratic computational bottleneck of the self-attention mechanism found in pure Transformer architectures.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Inference throughput comparison.

https://doi.org/10.1371/journal.pcbi.1014480.g009

Extended evaluation on multiple biomedical corpora

To directly address the concerns regarding the generalizability of DuSSM, we extended our evaluation to three additional representative biomedical datasets: ChemProt, GAD, and EU-ADR. These experiments transition the framework from drug-drug interactions to chemical-protein and gene-disease association tasks.

Performance on ChemProt (complex multi-class relations).

The ChemProt dataset is significantly more challenging than DDI-2013 due to its technical terminology and fine-grained interaction types. As shown in Table 7, DuSSM achieves a weighted F1-score of 87.64%.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Detailed performance of DuSSM on ChemProt dataset.

https://doi.org/10.1371/journal.pcbi.1014480.t007

A pivotal finding in this experiment is the model’s performance on the None class. In biomedical relation extraction, the ability to correctly identify non-interactions is vital to prevent false alarms in clinical decision support. DuSSM achieved a Precision of 94.32% for the None class. This demonstrates that the dual-stream disentanglement successfully filters out linguistic noise in long sequences, providing the clinical reliability required for practical application.

Robustness on GAD and EU-ADR (gene-disease associations).

We further evaluated DuSSM on GAD and EU-ADR using 10-fold cross-validation to ensure statistical stability. These datasets represent a different semantic domain.

As summarized in Table 8, DuSSM maintains high performance across both corpora. On EU-ADR, the model achieved an exceptional recall of 98.86%, suggesting that the implicit stream is highly sensitive to the logical cues indicating associations between genes and diseases, even when the sentence structure is sparse. The consistency of results across 10 folds () confirms that DuSSM is not overfitted to specific dataset biases.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. 10-fold cross-validation results on GAD and EU-ADR.

https://doi.org/10.1371/journal.pcbi.1014480.t008

Benchmarking against established paradigms.

To rigorously validate DuSSM’s generalizability, we extended our evaluation to three diverse biomedical relation extraction benchmarks: ChemProt (chemical-protein), GAD, and EU-ADR (gene-disease).

It is worth noting that many of the baseline models evaluated in Table 2 are highly specialized architectures customized specifically for the DDI extraction task, and thus lack officially reported results across these diverse ontologies. Therefore, to provide a standardized and fair cross-domain evaluation, we benchmarked DuSSM’s overall F1-scores against the officially reported results of domain-agnostic foundational language models (BioBERT [30] and PubMedBERT [39]). Furthermore, since PubMedBERT serves as the contextual encoder for DuSSM, this comparison directly highlights the architectural gain provided by our reasoning module.

As presented in Table 9, DuSSM achieves highly competitive, state-of-the-art performance across all domains. Notably, on the EU-ADR dataset, DuSSM reaches an F1-score of 86.04%, outperforming the strong PubMedBERT baseline. Similarly, on the complex ChemProt corpus, DuSSM’s weighted F1-score (87.64%) significantly exceeds the literature-reported baselines. This sustained superiority empirically validates that the dual-stream disentanglement of local triggers and global logic is a universally effective framework for biomedical relation extraction, rather than a DDI-specific architectural artifact.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Generalizability benchmark: F1-score comparison on diverse biomedical datasets.

https://doi.org/10.1371/journal.pcbi.1014480.t009

Cross-dataset efficiency comparison.

To empirically validate the efficiency of the Mamba-based architecture across varying sequence complexities, we report the inference overhead in Table 10.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Empirical computational overhead across multiple datasets.

https://doi.org/10.1371/journal.pcbi.1014480.t010

The results show that the average inference latency remains below 2.2 ms per sample across all datasets, regardless of the number of relation types or sequence density. Furthermore, the peak GPU memory footprint is remarkably stable (approximately 1.8 GB), which is significantly lower than the memory demands of standard Transformer-based models when processing long abstracts. This provides strong empirical evidence for the linear-time complexity of DuSSM, making it suitable for high-throughput clinical processing.

Qualitative and scalability analysis

To substantiate the architectural advantages of DuSSM, we provide both a qualitative visualization of its internal mechanisms and an empirical validation of its scaling limits.

Visualizing feature disentanglement.

We visualize the saliency maps of the parallel streams in Fig 10 to demonstrate how DuSSM disentangles distinct feature spaces.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Saliency map comparison between the parallel streams in DuSSM.

https://doi.org/10.1371/journal.pcbi.1014480.g010

As shown in the visualization, the explicit stream (CNN) acts as a high-pass filter, generating sharp spikes on high-frequency syntactic triggers such as “inhibition” and “increased”. Conversely, the implicit stream (Mamba) exhibits a smoother activation profile, capturing the underlying semantic evolution and state transitions. This functional separation allows the model to remain robust when interaction triggers are embedded in noisy, long-range contexts, effectively addressing the feature entanglement issue common in serial architectures.

Isolated stress testing and scaling limits.

A core theoretical claim of the DuSSM reasoning module is its linear-time complexity . To rigorously evaluate the scalability limits of the proposed dual-stream architecture independently from the inherent 512-token limit and bottleneck of the static PubMedBERT encoder, we conducted an isolated computational stress test focusing strictly on sequence length resilience.

As illustrated in Figs 11 and 12, we bypassed the contextual encoder and fed simulated high-dimensional latent embeddings directly into the isolated reasoning modules. The results confirm a stark divergence in mathematical scalability. While standard Transformer-based classification heads exhibit a quadratic explosion in latency and memory usage beyond 1,024 tokens (eventually leading to out-of-memory constraints at 2,048 tokens), the DuSSM reasoning module maintains a stable, strictly linear trajectory up to 4,096 tokens.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Inference latency across varying sequence lengths.

https://doi.org/10.1371/journal.pcbi.1014480.g011

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Scalability analysis and sequence length stress test (Memory footprint).

https://doi.org/10.1371/journal.pcbi.1014480.g012

This empirical evidence demonstrates that the dual-stream disentanglement strategy provides a highly scalable and memory-efficient solution. By successfully maintaining linear complexity over extremely long sequences, DuSSM paves the way for future document-level extraction architectures—such as the direct processing of comprehensive electronic health records (EHRs)—once the underlying contextual encoder bottleneck is fully resolved by emerging linear foundational models.

Comparative analysis: Determinism vs. LLM hallucinations.

To evaluate the clinical readiness of DuSSM, we conducted a comparative benchmark against state-of-the-art large language models (LLMs), specifically GPT-4o and the domain-tuned Med-Llama 3 (8B). While LLMs demonstrate remarkable zero-shot reasoning capabilities, their deployment in pharmacovigilance is fundamentally constrained by probabilistic hallucinations—a tendency to infer drug interactions based on high-frequency statistical co-occurrence rather than grounded syntactic evidence.

The results in Table 11 reveal a critical reliability boundary. Although Med-Llama 3 marginally leads in overall F1-score due to its massive knowledge prior, DuSSM establishes a superior safety margin with a precision of 94.32% on non-interactive (None) samples. In high-stakes clinical environments, where false positives directly contribute to alert fatigue, DuSSM’s structural determinism offers a more robust safeguard than the generative heuristics of LLMs.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Performance and reliability comparison with SOTA LLMs.

https://doi.org/10.1371/journal.pcbi.1014480.t011

By enforcing a dual-stream feature separation of morphological triggers and long-range logic, DuSSM ensures that a biochemical relation is only flagged when explicit evidence aligns with logical validity. Furthermore, the computational efficiency and 160M parameter footprint facilitate privacy-preserving deployment on consumer-grade hospital hardware (Edge/PC), bypassing the data-leakage risks and latency overhead associated with closed-source cloud-based LLMs.

Synergy of dual-stream disentanglement

The primary motivation for DuSSM is the disentanglement of linguistic duality in biomedical texts. Standard serial architectures (e.g., CNN-Transformer or monolithic BERT-based models) often suffer from feature interference: local syntactic triggers (e.g., “inhibits,” “interacts”) are frequently confounded by global semantic noise in long, complex abstracts.

By employing a parallel design, the explicit stream (local 1D-CNN) acts as a morphological pattern recognizer to lock onto explicit triggers, while the implicit stream (Mamba) serves as a logical backbone to track long-range dependencies. As shown in our ablation study (Table 6), removing either stream leads to a significant performance drop, particularly in the “Advise” and “Mechanism” categories where both local triggers and conditional logic are required for correct classification. This orthogonal synergy ensures a more discriminative feature manifold, as visualized in the t-SNE analysis (Fig 8).

Architectural determinism vs. scaling heuristics

As large language models (LLMs) such as GPT-4o and Med-Llama 3 redefine the state-of-the-art in general NLP, our findings prompt a critical reflection on a fundamental question: Does massive parameter scaling inherently converge toward clinical reliability? We argue that DuSSM and generative LLMs represent fundamentally different methodologies in how medical knowledge is retrieved and validated. While LLMs exhibit remarkable zero-shot heuristic capabilities, they remain tethered to a probabilistic next-token prediction mechanism. This mechanism can occasionally lead to over-generation—a phenomenon where interactions might be inferred based on the statistical frequency of drug co-occurrence in training corpora rather than grounded syntactic evidence.

Crucially, we acknowledge that comparing a fine-tuned, task-specific architecture (DuSSM, 160M parameters) against general-purpose LLMs (e.g., Med-Llama 3, 8B parameters) is not a strict apples-to-apples comparison. These models serve fundamentally different paradigms. LLMs are versatile foundational engines capable of broad reasoning and generation across unseen tasks. In contrast, DuSSM is designed as a specialized diagnostic filter. The marginal, and sometimes statistically insignificant, performance gap between DuSSM and Med-Llama 3 does not diminish the value of LLMs. Rather, it highlights that for the specific, narrow task of clinical relation extraction—where edge-device deployment, low latency, and high specificity (rejecting false positives) are paramount—a lightweight, structurally decoupled module can match the deterministic performance of massive generative models at a fraction of the computational cost.

DuSSM establishes what we define as a highly reliable clinical baseline. By achieving a superior precision of 94.32% on non-interactive samples, DuSSM addresses the socio-technical crisis of alert fatigue. In high-throughput clinical workflows, excessive false positives act as a form of logical noise that erodes practitioner trust and contributes to diagnostic oversight. DuSSM mitigates this by employing a targeted dual-verification logic: a relation is only flagged when localized morphological anchors (explicit stream) and distal state-space trajectories (implicit stream) exhibit structural congruence.

This architectural transparency provides the determinism-by-design required for medical safety. Furthermore, while LLMs require massive, often cloud-dependent computational resources (), the complexity of the DuSSM reasoning module allows for decentralized, privacy-preserving deployment on edge devices within hospital intranets. This democratization of high-precision AI ensures that clinical decision support remains both scalable and ethically aligned with patient data sovereignty.

Structural robustness and scaling limits

A critical challenge for current relation extraction models is their performance degradation in long-form narratives. Our isolated stress tests (Fig 12) reveal that while standard Transformer-based classification heads suffer from attention dispersion and quadratic memory explosion () beyond 1,024 tokens, the decoupled DuSSM reasoning module maintains strictly linear memory scaling () and near-constant performance stability.

This resilience suggests that the disentanglement of transient surface signals from persistent logical trajectories represents a highly scalable computational paradigm. By demonstrating robust adaptability across diverse scientific domains—ranging from pharmacology (DDI) and chemistry (ChemProt) to genomics (GAD, EU-ADR)—our findings indicate that this explicit-implicit duality is an essential primitive for complex information processing. Thus, DuSSM is not merely a specialized tool for a single corpus, but a prototype for resource-efficient architectures that prioritize structural transparency and logical determinism over brute-force parameter scaling.

Clinical reliability and generalizability

Beyond DDI, our evaluation on ChemProt, GAD, and EU-ADR demonstrates that DuSSM effectively models the fundamental grammar of associations in scientific literature. The consistent performance across varying biomedical ontologies confirms that the dual-stream framework mitigates data scarcity issues common in genomic studies. In clinical environments, this stability translates into a trustworthy tool for real-time EHR monitoring on consumer-grade hardware, making advanced AI more accessible to resource-constrained hospital environments.

Limitations and future work

Despite its robust performance, DuSSM faces challenges in extremely data-scarce classes, such as the DDI-int category. As noted in our error analysis, a critical vulnerability of the dual-stream architecture emerges when both streams are deprived of strong signals: the explicit stream lacks morphological triggers, and the implicit stream lacks sufficient training instances to map the subtle topological logic.

To mitigate this limitation and enhance the practical utility of the architecture in real-world clinical deployments, we recommend implementing a confidence-aware gating mechanism. When the reasoning module outputs a low-confidence probability distribution across all relation types for a sentence containing recognized biomedical entities, the system should default to a flag for review state. This human-in-the-loop approach ensures that highly ambiguous, trigger-less interactions are not dismissed as false negatives, thereby maintaining clinical safety.

Future work will explore several targeted architectural directions to overcome this dual-stream bottleneck. First, to address the data scarcity of underrepresented classes, we will investigate using general-purpose LLMs as offline synthetic data generators to augment the training manifold of categories like DDI-int. Second, to provide the implicit stream with stronger signals when lexical cues are absent, we plan to inject external topological priors—such as relational subgraphs from the UMLS or DrugBank knowledge bases—directly into the Mamba module’s latent state. This knowledge-guided state-space modeling would allow the system to infer relations based on biochemical priors even when the text sequence lacks explicit structural hints. Additionally, once the contextual encoding bottleneck is fully bypassed by emerging linear foundational models, we aim to extend the DuSSM framework to document-level relation extraction, enabling the capture of complex interactions spanning across multiple paragraphs and complete electronic health records.