FSSM-DDI: Fusion State Space Model for predicting drug-drug interaction using social-media and drug descriptions

Abstract

Background

Predicting drug-drug interactions (DDIs) from social-media and drug descriptions is crucial for healthcare, drug regulation, and pharmaceutical research, yet remains a challenging task. Existing deep learning and Transformer-based approaches often struggle with modeling long-range dependencies within sentences or entail high computational costs, limiting their practical applicability in large-scale drug safety screening.

Results

To overcome these limitations, we propose a Fusion State Space Model (FSSM) for DDI prediction (DDIP). FSSM leverages a selective State Space Model (SSM) to efficiently capture long-range syntactic and semantic dependencies within sentences, while an Interaction-based Selective Filtering (ISF) module mitigates information redundancy from multimodal inputs. Experiments on the DDIExtraction-2013 corpus demonstrate that FSSM achieves an F1-score of 75.23%, matching state-of-the-art Transformer models while requiring significantly less training time. Ablation studies confirm that each architectural component contributes meaningfully, with social-media embedding providing the largest performance gain (8.93% drop upon removal) and the ISF module contributing a 4.80% improvement.

Conclusions

FSSM offers a promising balance of predictive performance and computational efficiency for drug-drug interaction prediction. The model’s linear complexity enables large-scale screening, real-time applications, and deployment in resource-constrained settings, making it a valuable tool for drug discovery and drug safety assessment.

Background

Multidrug combination therapy is a common and effective strategy for treating complex or concurrent diseases, particularly in elderly patients with multiple conditions. Drug combination can enhance treating effect, but it may lead to unexpected DDIs and adverse side effects, increasing the risk of patient injury or even death [1]. Before combination therapy, it is necessary to understand the potential DDIs from the existing database to the side effects of the drug combination. There are some valuable DDI databases, such as PharmGKB, Drug2ank 4.0 and Therapeutic Target Database [2, 3], but not all DDIs are recorded in the current datasets. The possible combination therapy of many types of drugs makes it difficult to test all DDIs through experiments and determine previously unknown DDIs.

There are a large number of potential DDI in the extensive medical social-media corpus. Due to the complexity of the corpus, accurately predicting these DDIs remains an important and challenging task. Simple deep learning (DL) models often fail to fully capture the comprehensive information of drug pairs, multi-feature fusion may introduce redundancy, and the imbalance between positive samples and negative samples can significantly impair the performance of the classifier, leading to biased predictions. Most existing methods mainly focus on the prediction of whether drugs interact with each other, while ignoring drug description knowledge (DDK), which is useful for DDI prediction (DDIP). On the Drugs.com website (https://www.drugs.com/amoxicillin.html), as shown in Fig. 1, the DDK contains valuable information, such as usage, dosage, side effects, warnings, and the relationship with other drugs. All of these are related to DDIP.

Fig. 1.

DDK in Drugs.com

Drug-related websites and datasets generally provide detailed DDK, including disease types, drug composition, drug effects, and potential relationships between different drugs, all of which contribute to judging whether DDIs exist. To overcome the limitations of DL and Transformer-based approaches while leveraging DDK, and inspired by Mamba-DTA [4], we propose a Mamba-based DDIP method named FSSM (Fusion State Space Model). Unlike existing approaches, FSSM explicitly integrates external DDK to enhance DDIP with low computational complexity.

While state space models (SSMs) like Mamba have shown promise in general sequence modeling, directly applying them to DDIP faces two challenges: (1) standard SSMs lack mechanisms to filter cross-modal noise when fusing social-media text with structured drug descriptions; and (2) existing fusion strategies often treat drug features independently, ignoring the interactive nature of DDI. To address these limitations, FSSM introduces a task-specific Interaction-based Selective Filtering (ISF) module that dynamically weights pairwise drug features—a design tailored specifically for DDI rather than a generic combination of off-the-shelf components.

A number of DDIP methods have been proposed in the literature, which can be divided into four major categories: traditional machine learning (ML), deep learning (DL), Transformer, and SSM-based methods.

Traditional ML-based methods rely on handcrafted features. Mei et al. [5] proposed a drug-target spectrum representation for DDI characterization. Zhao et al. [6] categorized DDIP methods and public datasets while addressing methodological limitations. Zhang et al. [7] examined AI applications in DDIP across datasets and ML techniques. Wang et al. [8] conducted a multidimensional comparison of DDIP studies, covering datasets, drug properties, and computational methods, ultimately outlining current challenges and future directions for ML-based DDIP research. The performance of these ML-based methods depends heavily on extracted handcrafted features, which limits their generalization capability.

In recent years, many DL models, including convolutional neural networks (CNNs), long short-term memory (LSTM) networks, gated recurrent units (GRUs), and their variants, have been applied to DDIP. Karim et al. [9] introduced a knowledge graph (KG) embedding and convolutional-LSTM approach, constructing a large-scale dataset of 28.9 million DDI pairs and building a comprehensive knowledge graph using multiple embedding techniques. Wu et al. [10] developed a hybrid framework combining stacked CNN and bidirectional GRU (BiGRU) to extract entity location and lexical information, enhanced by attention pooling. Deng et al. [11] proposed DDIMDL, a multimodal model that integrates chemical targets, substructures, enzymes, and pathways to predict 65 types of DDI events. Graph convolutional networks (GCNs) have also been applied to DDIP. Feng et al. [12–14] proposed three modified GCNs for DDIP, including a directed graph attention model, a five-layer graph attention network, and an end-to-end architecture with a GCN encoder. Wang et al. [15] developed a multi-kernel GCN that identified top potential DDIs for specific cancer drugs.

Multi-feature integration strategies have shown particular promise. Qiu et al. [16] and Wang et al. [17] provided systematic reviews and frameworks contextualizing these methodological advancements. Chen et al. [18] introduced multi-scale feature fusion evaluated across multiple DDI tasks. Huang et al. [19] combined different fusion strategies with adaptive weighted scoring. Zhang et al. [20] integrated dual BiGRU with squeeze-and-excitation attention, using molecular descriptors and Group Lasso for feature selection. Wang et al. [21] proposed multi-view fusion with graph isomorphic networks and knowledge graph embeddings. Zhang et al. [22] demonstrated the superiority of multi-feature approaches using CNN-extracted features from multiple drug characteristics. Asfand et al. [23] constructed a multimodal CNN for DDIP using enzymes, pathways, chemical structures, and targets, showing improved accuracy and overfitting resistance compared to simpler models.

Despite these advances, DL models incorporating attention mechanisms remain challenged by (1) the enormous combinatorial space of potential drug regimens; (2) inherent noise, heterogeneity, and class imbalance in multimodal biomedical datasets; and (3) limitations of conventional CNN architectures in effectively capturing long-range dependencies—where “long-range” refers to dependencies spanning distant positions within a single sentence (e.g., linking drug mentions across 15–30 tokens). For example: “Although Drug A is metabolized by CYP3A4 and often prescribed with caution, concurrent use with Drug B may lead to severe interactions.” Here, the interaction signal depends on linking “Drug A” and “Drug B” across more than 15 words.

While deep neural networks can capture nonlinear relationships through their layered architectures, they are insufficient for extracting long-range dependencies within drug characteristics, resulting in inaccurate and inefficient feature representations. To address these limitations, several Transformer-based approaches have advanced the DDIP field. Transformer models are effective for DDIP by capturing both long-range dependencies and short-range contextual patterns, achieving remarkable performance. Ioannis et al. [24] presented an end-to-end model, TP-DDI, which utilizes Transformer architecture with biomedical domain pre-trained weights (BioBERT). Wu et al. [25] proposed a DDIP method based on M-Transformer and KG, using multi-head attention to capture complex relationships among drug features while discarding redundant information. Zhang et al. [26] constructed a Semantic Cross-Attention Transformer (SCAT) model to integrate extracted features and long-range dependencies, implementing DDIP via a feature-joint classifier. Zhao et al. [27] constructed a dual-source Transformer framework, BIOFUSE-DDI, integrating a dual-source attention mechanism for feature extraction, a bimodal interaction enhancement module for feature fusion, and a sequential Transformer optimization module for comprehensive interaction pattern learning. However, despite their success, Transformer-based methods have limited ability to model remote dependencies due to their high computational complexity.

SSMs are a novel class of sequential modeling architectures based on classical control system theory [28], capable of capturing long-range dependencies with linear computational complexity. Mamba is a selective structured SSM that alleviates the modeling constraints of CNNs through global receptive fields and dynamic weighting, serving as an effective alternative to traditional Transformers. Compared with the self-attention mechanism in Transformers, Mamba can effectively capture long-range dependencies in long sequence data through state space equations and can handle longer sequences without significantly increasing computational complexity [28, 29].

In the DDIP task, numerous groups or fragments are responsible for functions other than drug binding, while binding-relevant fragments constitute only a very small proportion—meaning there is substantial noise information. When processing noisy data, attention mechanisms may be affected by unstable attention weights. Inspired by Mamba, we constructed a hybrid model, FSSM, which utilizes DDK to significantly improve DDIP results.

Results

Experimental setup

To assess the generalization ability and reliability of the models, a lot of experiments are conducted on the DDIExtraction2013 corpus (SemEval-2013 Task 9.2), and compared it against six state-of-the-art DDIP methods: Hybrid Neural Networks on Biomedical Literature (HNNBL) [10], DBGRU-SE [20], CNN-DDI [22], TP-DDI [24], MTrans [25], and SCATrans [26]. All experiments are conducted under a unified hardware and software environment (Intel Core I7-6300, 3.4 GHz, GeForce 1080, 16 GB memory, Ubuntu 16.04.2 LTS) to ensure reproducibility.

Overall performance comparison

Table 1 presents the DDIP results of seven methods on the DDIExtraction2013 corpus (mean ± std over fivefold cross-validation). FSSM achieves the best balance between performance and efficiency: its F1 reaches 75.23%, on par with the top-performing SCATrans, while the training time is only 115 min, less than half of SCATrans (230 min) and MTrans (254 min). Traditional DL-based methods (HNNBL, CNN-DDI) showed higher training efficiency but significantly lower F1 scores (≤ 71.61%). Transformer-based models (MTrans, SCATrans) achieves excellent performance but suffers from high computational complexity. The standard deviations show that FSSM achieves the most consistent performance across different data splits (± 0.46% F1), compared to Transformer-based methods (± 0.49–0.60% F1) and DL-based methods (± 0.82–1.12% F1).

Table 1.

DDIP results of seven methods on DDIExtraction2013

Method	Result
	Precision (%)	Recall (%)	F1 (%)	Training time (min)
HNNBL	72.71 ± 0.89	71.38 ± 0.94	72.04 ± 0.91	123 ± 5
DBGRU-SE	71.29 ± 0.76	71.94 ± 0.88	71.61 ± 0.82	111 ± 4
CNN-DDI	71.24 ± 1.02	67.83 ± 1.21	69.49 ± 1.12	104 ± 3
TP-DDI	74.79 ± 0.58	73.06 ± 0.62	73.91 ± 0.60	233 ± 8
MTrans	75.80 ± 0.51	74.12 ± 0.55	74.95 ± 0.53	254 ± 9
SCATrans	76.72 ± 0.47	73.80 ± 0.52	75.23 ± 0.49	230 ± 7
FSSM	76.04 ± 0.45	74.45 ± 0.48	75.23 ± 0.46	115 ± 4

Performance across sentence lengths

To further assess Mamba’s suitability for sentence-level biomedical texts, we analyze model performance across different sentence lengths in the test set. Table 2 shows that while all models perform similarly on very short sentences (< 10 tokens), performance disparities become more pronounced as sentence length increases. Mamba exhibits the smallest performance degradation as sentence length grew, with a drop of only 1.1% from the shortest to the longest sequences, compares to 7.9% for BiLSTM and 2.0% for Transformer.

Table 2.

F1-scores (%) of different models across sentence length intervals

	F1
Sentence length	BiLSTM	Transformer	Mamba
< 10 tokens	74.2	75.1	75.3
10–20 tokens	72.8	74.9	75.1
20–30 tokens	70.1	74.2	74.8
> 30 tokens	66.3	73.1	74.2

Type-wise performance analysis

Table 3 shows the DDIP results with and without negative samples. After removing negative samples, F1 scores for all four types of positive DDI increases significantly—especially the Int type, which improves by approximately 17%. This directly demonstrates that the absolute dominance of negative samples in the original data seriously interferes with the model’s learning of positive interactions, particularly for rare types.

Table 3.

DDIP results with and without negative samples (%)

Dataset		Negative	Mechanism	Effect	Advice	Int
With negative samples	Precision	96	80	74	76	52
	Recall	94	75	71	73	51
	F1	95.0	77.4	72.5	74.5	51.5
Without negative samples	Precision	–-	84	78	79	70
	Recall	–-	82	76	78	67
	F1	–-	83	77.0	78.5	68.4
Improvement (%)			+ 4.0	+ 4.5	+ 5.6	+ 16.9

Ablation study results

To rigorously evaluate the contribution of each component of FSSM, ablation experiments are conducted under a consistent fixes train-test split using 5FCV. Table 4 summarizes the results for twelve model configurations.

Table 4.

Average precisions of FSSM variants (%)

No	Variants	Experimental operation description	Avg. precision (%)	Drop vs. full model
1	GloVe Embedding	Using pre-trained GloVe vectors for word initialization	75.24	− 0.80
2	Word2vec Embedding	Using pre-trained Word2vec vectors for word initialization	75.18	− 0.86
3	w/o DDK Embedding	Removing the drug description knowledge (DDK) input channel and its corresponding embedding layer, using only social-media text	73.25	− 2.79
4	w/o Social Media Embedding	Removing the social-media text input channel and its corresponding embedding layer, using only DDK	67.11	− 8.93
5	w/o ISF Module	Removing the Interaction-based Selective Filtering (ISF) module. The features from the two drugs extracted by Mamba are directly concatenated for classification	71.24	− 4.80
6	Standard SSM Backbone	Replacing the selective SSM (Mamba) with a standard (non-selective) State Space Model as the backbone	75.71	− 0.33
7	w/o PCA Reduction	Removing the PCA dimensionality reduction step. The concatenated high-dimensional unified vector is fed directly into the model	75.86	− 0.18
8	w/o POS Feature	Removing the Part-of-Speech (POS) tag feature vectors from the embedding layer	75.93	− 0.11
9	BiLSTM Backbone	Replacing the model backbone with a BiLSTM network	72.28	− 3.76
10	Dot-Product Attention	Replace ISF with attention-weighted sum of VF1​ and VF2​	73.25	− 2.79
11	Standard dot-product attention	Replace ISF with standard dot-product attention	72.89	− 3.15%
12	FSSM (Full)	The complete proposed model incorporating all designed components	76.04	–

The results in Table 4 reveal that social-media text embedding is the most critical contributor; its removal causes the largest performance drop (8.93%), underscoring the indispensable value of user-generated content for DDI signal mining. The ISF module ranks second, with a 4.80% drop upon ablation, validating its efficacy in filtering noise and focusing on interaction-relevant features. Incorporating DDK provides a solid performance gain of 2.79%, confirming the necessity of multimodal fusion.

Discussion

Interpretation of key findings

FSSM achieves strong overall performance in DDIP, attaining an F1-score of 75.23% while requiring substantially less training time (115 min) than Transformer-based methods (230 + min). This efficiency-performance trade-off is particularly valuable for large-scale drug safety screening, where computational resources may be limited.

The ablation studies quantify each component’s contribution. Removing social-media embedding causes the largest drop (8.93% F1), confirming that user-generated content serves as the primary information source for detecting DDI signals. The ISF module contributes a 4.80% improvement, validating its effectiveness in filtering irrelevant pharmacological information shared by drug pairs. DDK incorporation yields a 2.79% gain, confirming the value of multimodal fusion. The selective SSM (Mamba) offers a modest 0.33% advantage over a standard SSM, indicating it is retained primarily for content-aware filtering and computational efficiency rather than dramatic performance gains.

Comparison with existing approaches

Unlike Transformer-based methods that rely on quadratic-complexity self-attention, FSSM achieves linear computational complexity through the selective SSM architecture. This fundamental difference explains the substantial reduction in training time (115 min vs. 230 + min for MTrans and SCATrans) while maintaining competitive F1 scores.

Compared to traditional DL methods (HNNBL, CNN-DDI), FSSM demonstrates superior capability in capturing long-range dependencies within sentences. As shown in Table 2, Mamba’s performance degradation across increasing sentence lengths (only 1.1% drop) is substantially smaller than BiLSTM (7.9% drop), confirming that SSM-based architectures better preserve semantic cues even when key drug mentions are separated by multiple clauses—a common pattern in clinical narratives.

Type-wise performance variation

The type-wise analysis (Table 3) reveals that benefits vary across DDI categories. The rare Int type shows the most substantial improvement after negative filtering (+ 16.9% F1), compared to modest gains (+ 4.0–5.6%) for frequent types. This indicates that ISF is particularly effective at amplifying weak signals for rare interactions, which are often the most clinically significant but difficult to detect.

Limitations and future directions

Several limitations of this study should be acknowledged. First, while ablating ISF yields a 4.80% drop, this single result does not reveal its performance across varying noise levels. Future work will evaluate ISF on datasets with controlled noise ratios, as its benefit should theoretically increase with noisier data.

Second, the interpretability analysis presented here is preliminary. For a test instance predicting interaction between Ciprofloxacin and Warfarin, the model highlighted social-media tokens (“bleeding” 0.32, “bruising” 0.28) and DDK features (“CYP1A2 inhibition” 0.45, “vitamin K antagonist” 0.38). While these weights provide decision clues, FSSM currently lacks end-to-end visualization tools—a direction for future work.

Third, the current study focuses on sentence-level long-range dependencies within the DDIExtraction2013 corpus. Future work should assess generalization on additional datasets (e.g., Twosides, DrugBank) and evaluate performance on cross-sentence and cross-document contexts.

Fourth, the selective SSM’s modest advantage over standard SSM (0.33%) suggests room for architectural improvements. Future work should isolate long-range modeling contributions via controlled experiments and explore hybrid architectures that combine the strengths of SSMs and attention mechanisms.

Conclusions

This paper proposed FSSM, a multimodal hybrid framework integrating social-media text with drug description knowledge (DDK) for DDIP. By adopting the Mamba module with linear complexity, FSSM efficiently captures intra-sentence long-range dependencies while significantly reducing computational overhead. The Interaction-based Selective Filtering (ISF) module dynamically integrates drug pair features, filtering redundant information to focus on interaction-relevant signals.

Experiments on DDIExtraction-2013 showed that FSSM achieves comparable performance to state-of-the-art Transformer models (75.23% F1) with less than half the training time (115 min vs. > 230 min). This efficiency-performance trade-off enables large-scale screening, real-time applications, and deployment in resource-constrained settings.

Future work will (1) isolate long-range modeling contributions via controlled experiments; (2) assess generalization on additional datasets (e.g., Twosides, DrugBank); (3) evaluate ISF across varying noise levels; and (4) develop end-to-end visualization tools for improved interpretability.

Methods and dataset

Dataset

The DDIExtraction2013 dataset is available from the SemEval-2013 Task 9.2 repository. It is a benchmark dataset for DDIP, and is documented in the original task description [30]. It comprises 792 DrugBank documents and 233 MEDLINE abstracts, with all drug entities and drug pairs annotated at the sentence level. Following standard practice, we conduct DDIP experiments without requiring named entity recognition. The dataset contains 5002 positive DDI instances and 33,486 negative instances. Positive instances are further categorized into four subtypes: mechanism, effect, advice, and int (interaction).

Data preprocessing

All sentences from DDIExtraction2013 are preprocessed through tokenization, normalization, stopword removal, and case unification [5–7]. Specifically, all words are lowercased, numbers are replaced with a “drug” tag via regular expressions, target drugs are labeled as “Drug1” and “Drug2,” and other drug mentions are labeled as “Drug0.” The task is formulated as predicting triples of the form (Drug1, DDI, Drug2). Each sentence is encoded into a 300-dimensional vector using Word2Vec, with padding or truncation applied as needed to achieve uniform length.

We extract 11,930 drugs and their corresponding DDK entries from DDIExtraction2013. Drug descriptions are embedded using the Stanford GloVe model, pretrained on a DDI-oriented corpus to leverage existing interaction knowledge. Missing DDK entries are represented as zero vectors.

In DDIP, preprocessing of social-media texts and DDK is an important foundation for constructing efficient models. The process mainly includes four key steps:

• Text normalization processing: First, unified processing of unstructured text on social-media and standardized drug descriptions is carried out, including: converting continuous text into word element sequences through word segmentation tools, implementing lowercase letters to eliminate word form variations, and using regular expression techniques to handle numeric strings and special symbols.

• Standardization of drug entities: For drug names in the text, a unified standardized labeling scheme is adopted: specific labels (such as Drug1, Drug2) are used to replace the target drug names, and uniform identifiers (such as Drug0) are used for non-target drug entities to retain syntactic relationships between drugs while eliminating noise caused by name complexity.

• Sequence structuring processing: To meet the input requirements of the model, the text sequence is standardized: the sequence length is unified through truncation or padding operations, the attention mask mechanism is adopted to distinguish real content from padding positions, and position encoding is constructed to retain sequential relationships between words.

• Negative sample optimization strategy: To address sample imbalance, the following measures are adopted: a negative instance filtering mechanism based on syntactic structure, difficult sample mining to enhance model discriminative power, and data augmentation techniques to augment training sample diversity.

FSSM for DDIP

The framework of FSSM is shown in Fig. 2, consisting of five modules, i.e., embedding, feature extraction by Mamba, Interaction-based Selective Filtering (ISF), feature fusion, and DDIP.

Fig. 2.

Overview architecture of FSSM

Input social-media corpus, drug description, and each candidate drug in the input sentence: (1) the embedding layer extracts word embedding of corpus text, part of speech (POS), position embedding (relative distances between a word), and word embedding of drug description; (2) feature extraction is conducted; (3) the extracted vectors are input into ISF to filter the irrelevant information; (4) the multimodal features are fused; (5) Softmax classifier is adopted to perform DDIP. Mamba and ISF are two main modules, as shown in Fig. 3A and B.

Fig. 3.

The structures of Mamba and ISF. A Mamba. B ISF

Vector embedding

In DDIP, a multimodal embedding framework is developed to integrate social-media texts with DDK (such as Pharmacological characteristics). For social-media text, a triple-encoding strategy is applied: (1) BERT-Base generates contextualized representations of user-reported experiences, (2) dependency parsing extracts semantic relations, and (3) domain-adaptive training on clinical corpora enhances understanding of specialized terms. For DDK, drug indications and mechanisms are encoded via GloVe embedding. Cross-modal alignment is obtained through attention-based semantic weighting between symptom descriptions and side effect texts, feature fusion via gated recurrent units, and residual connections that preserve original modality integrity. For social-media text, we adopt GloVe embeddings (300-dim) pretrained on biomedical corpora to balance representation quality and computational efficiency. Preliminary experiments with BERT-based encoding showed marginal F1 improvement (< 0.5%) at threefold computational cost; thus, GloVe is selected for the final model to prioritize efficiency.

The embedded vectors of drug1 and drug2 are concatenated and reduced by PCA with 128-dimensional vector, noted as Vector₁ and Vector₂, respectively.

Feature extraction

The embedded vectors Vector1 and Vector2 of Drug1 and Drug2 are input into the feature extraction module to extract discriminative spatial structural features, respectively. The processing is given as follows:

The embedded vectors Vector₁ and Vector₂ of Drug1 and Drug2 are input into feature extraction module to extract the discriminative spatial structural features, respectively. The processing is given as follows,

$$V{M}_i=lin(SSM(SiLU(conv(lin(pool(Vecto{r}_i)))))\otimes SiLU(lin(pool(Vecto{r}_i))))$$ 1

$$V{F}_i=pool(ReLU(conv(ReLU(conv(V{M}_i)))))$$ 2

where i = 1,2, Vector_i is 128-dimensional vector of the ith drug obtained after data preprocessing and embedding, VM_i is the features acquired by the Mamba module, VF_i is the discriminative features of drug, $\otimes$ is element-wise multiplication, pool(.) is max-pooling operation, lin(.) is fully connection operation, conv(.) is 1-D convolution operation, and SSM(.) is SSM operation, SiLU(.) and ReLU(.) are the SiLU and ReLU activation functions.

Interaction-based Selective Filtering (ISF)

ISF module is used to model DDI and filter out irrelevant information. Its processing is as follows,

$$VFW=Sig(lin(SiLU(lin(V{F}_1+V{F}_2))))$$ 3

$$VISF=V{F}_1\otimes (1-VFW)+V{F}_2\otimes VFW$$ 4

where VF₁ and VF₂ are the discriminative features of drug1 and drug2, respectively, Sig(.) is the Sigmoid activation function and VFW is the fusion feature weight that controls the selectivity of two drug features, VISF is the key feature of DDI obtained by ISF when focusing on a particular DDI.

ISF targets redundant pharmacological information shared by drug pairs (e.g., common metabolic pathways) that is irrelevant to interaction prediction. Unlike standard gating (single-vector filtering) or self-attention (token-wise weighting), ISF operates on paired features (VF1​ and VF2​) to compute a joint weight, performing soft-selection over the interaction space—amplifying interaction-relevant differences while suppressing shared, non-informative components.

Feature fusion, DDIP, and model training

VFW and VISF are concatenated to obtain the DDIP feature vector, as follows,

$$VDDI=lin(caten(VFW,VISF)$$ 5

where VFW and VISF are the discriminative features of drug1 and drug2 obtained by ISF, caten(.) is concatenation operation.

VDDI is input into Softmax classifier to predict DDI. Cross-entropy function as the loss function is utilized to avoid the gradient vanishing problem during model training. The loss of FSSM is defined by minimizing the distance between the prediction and the actual label, as follows:

$$Loss=-\frac{1}{m}\sum _{i=1}^m\left(l_{\mathit{AB}},,log,,{\widehat{l}}_{\mathit{AB}},+,\left(1,-,l_{\mathit{AB}}\right),log,,\left(1,-,{\widehat{l}}_{\mathit{AB}}\right)\right)$$ 6

where $l_{\mathit{AB}}$ and ${\stackrel{⌢}{l}}_{\mathit{AB}}$ are the real DDI-label for the drug pair (Drug1, Drug2) and predicting DDI probability of the ith DDI type belonging to {Advice, Mechanism, Effect, Int, Negative}.

The super-parameters of FSSM are optimized by minimizing its loss between the predicted DDI class of the corpus and the true DDI label. Softmax classifier is adopted to implement DDIP by performing a normalized probability score for each DDI category.

Evaluation Metrics

The model performance is evaluated by average precision (P), average recall (R), and average F1-score (F1), calculated as follows:

$$P=\frac{1}{5}\sum _{l\in S_{\mathit{DDI}}}{P}_l,R=\frac{1}{5}\sum _{l\in S_{\mathit{DDI}}}{R}_l,F1=\frac{1}{5}\frac{2PR}{P+R}$$ 7

where S_DDI = {Mechanism, Advice, Int, Effect, Negative} is the DDI label set, P_l and R_l of the lth instance are calculated by,

$$\begin{array}{c}{P}_l=\frac{\#\mathrm{DDI}\mathrm{is}\mathrm{the}l\mathrm{th}\mathrm{class}\mathrm{and}\mathrm{is}\mathrm{classified}\mathrm{as}\mathrm{the}l\mathrm{th}\mathrm{class}}{\#\mathrm{Classified}\mathrm{as}\mathrm{the}l\mathrm{th}\mathrm{class}}\\ R_l=\frac{\#\mathrm{DDI}\mathrm{is}\mathrm{the}l\mathrm{th}\mathrm{class}\mathrm{and}\mathrm{is}\mathrm{classified}\mathrm{as}\mathrm{the}l\mathrm{th}\mathrm{class}}{\#\mathrm{Drugpair}\mathrm{the}l\mathrm{th}\mathrm{class}}\end{array}$$ 8

Abbreviations

CNN

Convolutional neural network

DDIP

Drug-drug interaction prediction

DDI

Drug-drug interaction

DDK

Drug description knowledge

DL

Deep learning

FSSM

Fusion State Space Model

GCN

Graph convolutional network

GRU

Gated recurrent unit

HNNBL

Hybrid Neural Networks on Biomedical Literature

ISF

Interaction-based Selective Filtering

KG

Knowledge graph

LSTM

Long short-term memory

ML

Machine learning

PCA

Principal Component Analysis

POS

Part-of-Speech

ReLU

Rectified Linear Unit

SCATrans

Semantic Cross-Attention Transformer

SiLU

Sigmoid Linear Unit

SSM

State Space Model

5FCV

Fivefold cross-validation

Authors' contributions

SWZ carried out manuscript preparation and writing, TZ provided programming and statistical analysis, DWW conducted data acquisition. All authors performed a manuscript review and approved the content of the manuscript.

Funding

This paper was supported by the National Natural Science Foundation of China (No. 62172338).

Data availability

The DDIExtraction 2013 dataset supporting the conclusions of this article is publicly available from the SemEval-2013 Task 9 repository (http://www.cs.york.ac.uk/semeval-2013/task9/). The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation. The source code and processed datasets supporting this study are publicly available at https://github.com/zsw-group/FSSM.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.