Neural interaction explainable AI predicts drug response across cancers

Abstract

Personalized treatment selection is crucial for cancer patients due to the high variability in drug response. While actionable mutations can increasingly inform treatment decisions, most therapies still rely on population-based approaches. Here, we introduce neural interaction explainable AI (NeurixAI), an explainable and highly scalable deep learning framework that models drug–gene interactions and identifies transcriptomic patterns linked with drug response. Trained on data from 546 646 drug perturbation experiments involving 1135 drugs and molecular profiles from 476 tumors, NeurixAI accurately predicted treatment responses for 272 targeted and 30 chemotherapeutic drugs in unseen tumor samples (Spearman’s rho >0.2), maintaining high performance on an external validation set. Additionally, NeurixAI identified the anticancer potential of 160 repurposed non-cancer drugs. Using explainable artificial intelligence (xAI), our framework uncovered key genes influencing drug response at the individual tumor level and revealed both known and novel mechanisms of drug resistance. These findings demonstrate the potential of integrating transcriptomics with xAI to optimize cancer treatment, enable drug repurposing, and identify new therapeutic targets.

Graphical Abstract

Graphical Abstract.

Introduction

The identification of oncogenic driver mutations has led to the development of a large number of targeted cancer drugs in the last two decades [1, 2]. These precision therapies have led to a paradigm shift in oncology by directly addressing the molecular mechanisms of underlying tumor pathology, often leading to significant survival benefits. However, therapeutic response remains highly variable, even among patients harboring the same oncogenic mutations.

Single-gene markers alone are often insufficient for accurately predicting treatment response. Apart from a few well-characterized resistance mutations, the intricate interplay of molecular mechanisms that drive therapy resistance cannot be fully understood by analyzing the mutational landscape alone. This highlights the urgent need for prediction models that integrate broader molecular profiles to anticipate tumor responses across diverse drugs. While mutational profiling is increasingly used in clinical diagnostics, gene expression profiling has shown promising results in predicting the functional tumor behaviors, including cancer cell vulnerabilities [3] and susceptibility to drug treatment [4, 5]. However, many existing approaches rely on simple, often linear models trained on small datasets [4, 6], limiting their ability to capture complex, nonlinear drug–tumor interactions.

Here, we introduce neural interaction explainable AI (NeurixAI), a deep learning framework designed to model drug responses across tumors and drug classes by leveraging large-scale molecular profiling data. NeurixAI efficiently integrates extensive drug screening datasets to uncover nonlinear relationships between drug activity and tumor transcriptomic profiles. Its architecture enables high-throughput analysis of >19 000 genes, while maintaining high predictive accuracy and low computational cost, mitigating the risk of data overfitting. We applied NeurixAI to a dataset from the DepMap database [7], which includes large-scale drug screening results and detailed molecular characterizations of cancer cell lines. By incorporating a variant of the explainable artificial intelligence (xAI) algorithm layer-wise relevance propagation (LRP) [8–13], NeurixAI enables the identification of key genes associated with drug response in individual tumors. This interpretability provides mechanistic insights into how gene expression influences drug response across diverse cancer types.

By providing a scalable and interpretable solution for predicting tumor behavior in response to drug treatment, NeurixAI holds the potential to improve personalized treatment strategies, guide drug repurposing, and uncover therapeutic targets in oncology.

Materials and methods

Model

NeurixAI combines two multilayer perceptrons, and which embed gene vectors (representing tumors) and drug vectors, respectively, in a shared 1000-dimensional latent space. We call these latent vectors the tumor latent vector (TLV) and drug latent vector (DLV), respectively. is a fully-connected neural network with one hidden layer of dimension 10 000 and an output layer of dimension 1000. has the same structure but a smaller hidden layer with 5000 neurons. NeurixAI uses both models’ outputs to compute the inner product as the final model’s output.

NeurixAI is highly efficient for predicting all interactions between a set of n drugs and m tumor profiles. While conventional models have to predict each of the n × m interactions, separately, NeurixAI merely has to model n predictions for the DLV and m predictions for the TLV. This process returns two matrices and of size n × 1000 and m × 1000, respectively. The interactions are then computed via a single large matrix multiplication where the resulting matrix I is of size n × m and contains all interactions. Because the entire architecture, including both neural networks and the inner product, is differentiable, the neural network weights can be trained end-to-end using gradient descent to model complex nonlinear interactions that predict drug responses for specific drug–tumor pairs.

Data

The data were downloaded from the DepMap database [7]. Drug screening results for PRISM repurposing data were obtained from the 19Q4 published file “secondary-screen-dose-response-curve-parameters.csv”. Sanger screening data were obtained from Drug_sensitivity_AUC_(Sanger_GDSC2).csv.

PRISM data consisted of 476 cancer cell lines treated with 1135 different drugs, resulting in 546 646 drug response measurements (including duplicate and missing experiments). RNA expression data were available for all cell lines for 19 193 protein-coding genes.

The 1135 drugs comprised compounds from three categories. Fifty drugs were defined as chemotherapies, 453 drugs as targeted drugs, and 538 were repurposed non-cancer drugs and 93 drugs were not further specified.

Targets

The log area under the drug-response curve was used as a measure of the response of a cell line to a specific drug. Since we were interested in predicting differences in a drug’s treatment effect between cell lines and not in predicting the overall cell toxicity of each drug, the log-transformed area under the drug response curve (log AUC) was standardized for each drug to zero mean and unit variance. To model the interaction between a drug and a cell line, NeurixAI requires an informative vector representation of both described in the following.

Cell line representation

The provided transcriptome data comprising expression data of 19 193 genes were used as a description for each cell line (“OmicsExpressionProteinCodingGenesTPMLogp1.csv”). No genes or cell lines were removed. Log transformation from the original publication was kept.

For each fold, RNA expression data were standardized to zero mean and unit variance on the training sets. The test sets were standardized using the normalization parameters calculated on the training sets.

Drug representation

The drug representation was composed of three vectors that were concatenated into one 2659-dimensional input vector. First, a one-hot encoding of dimension 1135. Second, for all drugs for which DepMap provided SMILES codes, an extended connectivity fingerprint with a distance of 6 was constructed using RDKIT [14]. If no SMILES code was provided, the representation was set to a baseline of zeros.

Third, NeurixAI was provided with prior knowledge about drug target similarity. The knowledge about drugs having a similar target spectrum may allow NeurixAI to share trained insights between drugs. To this end, a network was constructed in which each drug was represented as a node. A second class of nodes included all known targets and target pathways. Each drug node was connected to its specific target nodes and target pathways (Supplementary Fig. S1). In this network, drugs with similar target spectra are in close proximity to each other, while drugs that are functionally different show high distance along the network.

To interpret these network distances and give them as input to the neural network, a Python implementation (https://github.com/eliorc/node2vec) of Node2Vec was then used to create an embedding vector for each drug node [15]. We generated embedding vectors of dimension 500, with a walk length of 20 using 200 walks. Embeddings of drug nodes were then used as a drug descriptor. We note that these embeddings are similar between drugs if these drugs have similar reported targets, but that the embeddings contain no information about the specific target of a drug, and Node2Vec generates the identical embeddings if target names are permuted or deleted.

Model training

Five-fold cross-validation was conducted. For each fold, cell lines were partitioned into “training” (80%) and “test” (20%) cell lines and a training set was built containing the drug response experiments of all “training” cell lines with each of the 1135 drugs. The test set contained all drug response experiments of “test” cell lines with each of the 1135 drugs. This ensured that test samples exclusively represented data from unseen tumor cells. The process was repeated five times, each time changing the train-test split, to assess the model’s generalization performance.

Within each fold, each gene expression feature was standardized in the training dataset. The test set was standardized using the parameters from the training dataset.

During each iteration, training occurred with an equal probability for either the drug or the gene network. Stochastic gradient descent was used with a minibatch size of 128, a momentum of 0.9, and weight decay parameter of 10⁻⁵. Training started at a learning rate of 0.05. PyTorch’s exponential learning rate decay was applied with gamma set to 0.9. Dropout of 0.05 was applied during training on the hidden layer of each NN. To account for potentially very large products at the output, we treat them as effective statistical outliers and use for that purpose the Huber loss instead of the common square loss. We further applied gradient clipping (clip_grad_norm in PyTorch) with maximum norm set to 1.0 and updated the weights of either NN^T or NN^D at each iteration to achieve more stable optimization. NeurixAI was trained for 50 epochs.

Support vector regression from sklearn with default hyperparameters was applied as a baseline model. Ridge, Lasso, and Elastic Net regression implementations from sklearn were used with alpha set to 0.1. For Lasso regression, the max_iter parameter was set to 4000 to improve convergence. Supplementary Fig. S2 shows loss and performance changes during training. Different hyperparameters were tested for training (learning schedule, dropout rate) and model architecture (hidden layer width, model depth) but did not lead to improved performance (Supplementary Fig. S3).

Layer-wise relevance propagation

NeurixAI’s predictions were explained using a simple extension of LRP, which we contribute and that deals with the product-type interaction layer, and that we explain below. Let us slightly expand the equations of the NeurixAI model to explicitly reveal the feature representations associated with the two data modalities.

We consider the explanation of y in terms of the input , i.e. the attribution of the predicted response to the tumor’s genes. In particular, we derive our explanation as

where reads “explains a in terms of b” or “attribute a onto b”. (Note that when a and b are vectors, is a matrix). The equation linking the second and third expressions enforced the explanation to be linear with respect to the intermediate representation . The rightmost expression is the multiplication of a vector and a matrix, where the latter contains the attribution of a tumor’s neural network features in terms of the actual tumor features. Because neural networks mapping input to representation have a higher level of complexity and nonlinearity, we computed such a matrix of explanations using LRP, in particular, the th row of is given by the LRP computation , and we use for LRP similar parameters as in previous studies [16, 17], in particular, propagating the scores from layer to layer using the generalized LRP-gamma rule. Note that our formulation is general enough so that other attribution methods such as Shapley value sampling [18] and integrated gradients[19] can be used in case they lend themselves better to the task. We further note that the expression in the equation above separates what depends on the drug () and what depends on the tumor (). This means that rather than having to explain every individual tumor–drug matching, one only needs to compute the explanation matrix associated with each tumor gene profile once and then multiply that matrix with the different drug representations in order to get the desired explanations. This is especially advantageous when the space of different drugs is large.

Statistical analysis

Statistical analyses were conducted in R statistical packages. All tests were two-sided, and differences were regarded as significant if P< .05.

The predictions of NeurixAI and baseline models were compared to the experimental ground truth using Spearman’s rho. If the model could perfectly rank cell lines of the test fold according to the ground truth drug response, Spearman’s rho would assume the value 1. If experimental results were available for fewer than 10 cell lines in a fold, no results were calculated for this fold.

The k-means algorithm with k = 10 was used to find clusters of important genes based on the LRP scores (Supplementary Figs S9–S20).

Supplementary Figs S23–S25 show genes that had differential importance among drugs of a drug class. These genes were identified by computing the difference between the drugs of a drug class for which a given gene has the highest and the lowest importance rank, respectively. Additionally, the highest percentile rank among drugs had to be above 0.9, as we regarded differences between low percentile ranks as less interesting.

Figure 1 was generated using biorender.com. All other figures were generated using ggplot2 [20].

Figure 1.

Workflow for predicting drug sensitivity using latent interactions. (A) The dataset comprises experimental data measuring the response of 476 cell lines to 1135 different compounds. This experimental drug response serves as a label for the machine learning workflow. (B) NeurixAI predicts the outcome of a tumor–drug interaction (i.e. the response of a cell line to a specific drug) based on the latent interaction between the tumor RNA profile and the drug representation. Both RNA profile and a drug representation are mapped into a shared latent space. The experimental drug response is then modeled as the inner product between both latent vectors. (C) The NeurixAI explanation via LRP identifies the contribution of each gene to the tumor’s drug response for a specific drug. Created in BioRender. Mochmann, L. (2025) https://BioRender.com/quoc7qw.

Results

NeurixAI architecture and DepMap data

We developed NeurixAI to predict response to precision therapies from transcriptomic profiles by harnessing large-scale experimental data to model tumor-specific drug responses (Fig. 1). NeurixAI combines two data types, the molecular tumor profile (i.e. the transcriptome) and a drug description vector, and maps them to abstract representations, a TLV and a DLV using two neural networks (NN_Tumor and NN_Drug). The prediction of the drug response is then modeled as the inner product of DLV and TLV. This provides tumor- and drug-specific drug response predictions depending on the learned DLV and TLV representations in latent space (Fig. 1B). Importantly, to predict the outcome of a large number of drugs interacting with many different tumor cell lines, the respective neural network needs to map each drug and each cell line only once, making this approach highly efficient. The number of neural network calculations scales linearly with the number of drugs and tumors. In combination with the xAI approach LRP [8, 9], NeurixAI enables the prediction and understanding of relationships among drugs and molecular tumor properties.

Our study leveraged a large set of drug screening experiments from the PRISM project [6, 21] consisting of 476 cancer cell lines treated with 1135 different drugs. NeurixAI was applied to model the outcome of tumor cell viability in 546 646 experiments (including duplicate and missing experiments; Fig. 1A) [6, 22]. More specifically, we used RNA expression data of 19 193 protein-coding genes for each tumor, as the transcriptome is highly informative of phenotypic differences associated with drug response [4–6].

Drugs were characterized by three aspects that convey different types of information: First, an 1135-dimensional one-hot encoding was used. Second, an extended connectivity fingerprint [23] that is similar between drugs whose structural formulas are similar (applied, e.g. in [24]), and third, a representation created using the node embedding algorithm Node2Vec [15], which conveys prior information about whether drugs have been reported to have a similar target landscape (see the “Materials and methods” section). These three vectors were concatenated into a 2659-dimensional input vector. Notably, this representation remains agnostic to specific target genes.

Prediction of tumor-specific drug response by neural interactions

We trained NeurixAI to predict how a tumor will respond to a given drug, using information about the tumor’s gene expression and the drug’s properties. To measure drug response, we used the log AUC, which captures how effectively a drug inhibits tumor growth. Since some drugs are generally more toxic than others across all tumors, we standardized the log AUC for each drug—setting its mean to zero and variance to one. This removed baseline differences in overall drug potency and forced the model to focus on predicting how individual tumors respond differently to the same drug.

To benchmark NeurixAI’s performance, we performed a five-fold cross-validation. In each fold, NeurixAI was trained to predict the interactions of 80% of cell lines with all 1135 drugs, while the interactions of the remaining 20% of cells lines were held out as a test set.

For comparison, several regularized regression models, including support vector regression, were selected as baseline models and fitted for each drug separately. This is in contrast to NeurixAI, which incorporates the information of all tumor–drug interactions within one comprehensive model.

To assess the performance of each model, we compared the experimental (ground truth) drug response with the predicted drug response for each drug, separately. NeurixAI showed the lowest mean squared error (0.92) compared to all other models (0.93–1.2). As a more interpretable measure of performance, we used Spearman’s rho (Fig. 2B) that assumes the value 1, if the model perfectly ranks the test cell lines of a drug according to their ground truth response. NeurixAI outperformed all baseline models across all cross-validation folds (Fig. 2 and Supplementary Table S2). Stratified by drug classes, it achieved the highest performance in 54 out of 59 drug classes (91.5%), compared to Lasso (1.7%), Ridge (3.3%), SVR (3.3%), and Elastic Net (0%) (Supplementary Fig. S4). NeurixAI showed best performance for the MEK inhibitors trametinib, AZD8330, cobimetinib, and AS-703026 (Spearman’s rho ≥0.58; Supplementary Table S1) and the tyrosine kinase inhibitor poziotinib (Spearman’s rho = 0.64; Supplementary Table S1).

Figure 2.

Results of tumor–drug interaction prediction. (A) Median prediction performance (Spearman’s rho) for different drug classes (1135 compounds in total). Whiskers indicate interquartile range. Point size indicates the number of drugs per class. Classes containing <5 drugs and drugs belonging to multiple classes are not shown. The color indicates whether the majority of drugs belong to chemotherapy, targeted therapy, or non-cancer medication. A threshold of 0.2 is highlighted. HLM inhibitor, histone lysine methyltransferase inhibitor. (B) Comparison of NeurixAI model performance against baseline models. Spearman’s rho between predictions and ground truth across cell lines is used as a performance measure. (C) Proportion of drugs predicted by models with Spearman’s rho >0.2 (“captured” drugs). Across all cell lines, NeurixAI captured more drugs than baseline models. Numbers indicate different folds of the cross-validation.

The performance of drug response predictions varied widely between drug classes (Fig. 2A). The most accurate prediction was achieved for MEK inhibitors (median rho across drugs: 0.48, IQR: 0.35), but there was a wide performance variability among drugs within this class. Other drugs with good prediction performance were MDM2 inhibitors (median rho across drugs: 0.43, IQR: 0.04), HMG-CoA reductase inhibitors (median rho across drugs: 0.41, IQR: 0.17), tubulin polymerization inhibitors (median rho across drugs: 0.40, IQR: 0.1), EGFR inhibitors (median rho across drugs: 0.40, IQR: 0.17), and protein synthesis inhibitors (median rho: 0.38, IQR: 0.12; Fig. 5A).

We defined that the models “capture” a drug if the Spearman’s correlation between predicted and experimental response was >0.2, a threshold used in previous studies [25]. On average, NeurixAI captured 526 out of 1135 drugs (46%). Overall, 60% of targeted therapies (272 out of 453), 60% of chemotherapies (30 out of 50), and 29.7% of non-cancer therapies (160 out of 538) were predicted with Spearman’s correlation between prediction and experimental ground truth above 0.2. Ninety-three drugs were not specified as cancer or non-cancer drugs (64 captured, 68.3%). NeurixAI consistently performed better than the baselines also for higher thresholds than 0.2 (Supplementary Fig. S5).

As an independent validation, we tested the trained models on drug screening data from the Sanger Institute [26], selecting only cell lines not used for model training. NeurixAI captured 51.1% of drugs on the Sanger dataset, outperforming all baseline models (Supplementary Fig. S6 and Supplementary Table S3).

xAI identifies relevant molecular predictors of drug response

By using the xAI method LRP, NeurixAI reveals the underlying relationships between drug response and gene expression profiles for each tumor–drug pair. While these associations do not imply causation, they offer valuable insights into the molecular features most predictive of drug sensitivity. Such associations can serve as biomarkers for treatment stratification and may guide further experimental validation to uncover potential causal mechanisms underlying drug resistance or efficacy.

Due to the extent of results (influence of 19 193 genes for interactions between 476 tumors and 1135 drugs), we focused on a selection of 12 drugs from different drug classes for which NeurixAI showed high prediction performance (Fig. 3 and Supplementary Table S1). To quantify the importance of a gene for the prediction of drug response across cell lines, we computed the mean of the absolute LRP scores (i.e. the mean influence of the gene for and against a high drug response) over all tumors in concordance with previous xAI applications in other fields [16, 17, 27–29].

Figure 3.

Top-ranked genes as predictors for drug response. Network plots show genes identified as most relevant for predicting each drug’s response by NeurixAI. Colored nodes indicate the top 10 most important genes for drug response prediction in 12 drugs. Node size indicates importance. Blue nodes indicate higher gene expression associated with lower LRP scores (stronger drug response). Red nodes indicate higher gene expression associated with higher LRP scores (weaker drug response). The network structure was derived from the STRING database [30] and shows the functional relationships between genes. Gray nodes indicate linker genes. The top 30 most important genes per drug are shown in Supplementary Fig. S8.

Genes that code for known target proteins ranked among the most important genes for respective drugs (Supplementary Fig. S7). For EGFR inhibitors, the genes ERBB2 (percentile ranks >99.7% for poziotinib, dacomitinib, canertinib, pelitinib, and lapatinib) and EGFR (percentile ranks >98.2% for poziotinib, dacomitinib, canertinib, and pelitinib; lapatinib percentile rank: 95.16) were among the most important genes. ERBB2 and EGFR were not important for drug–tumor interaction with vincristine, vinblastine, or idasanutlin (ERBB2: <10%, EGFR: <60%). For idasanutlin, the gene coding for the known target protein, MDM2, was identified by NeurixAI as important for drug–tumor interactions (percentile rank: 99.99%). Interestingly, NeurixAI assigned less importance to MAP2K1 for MEK inhibitors (percentile ranks <20%). Instead, MAP2K6 (percentile ranks >99.6%) and MAP3K21 (percentile ranks >98.6%) were predicted to be highly relevant.

NeurixAI identifies off-target genes as key modulators of drug response

The effect of drugs on tumor tissue is influenced not only by the interaction between the drug and its target gene but also by gene products that modify the drug’s response, along with factors that dictate the overall viability of tumor cells. Knowledge of genes that modulate drug response may reveal potential resistance mechanisms or represent potential targets for secondary therapy. We quantified gene importance for each drug by computing the mean absolute LRP values across cell lines, providing a robust measure of overall contribution to response prediction (Fig. 3). All genes shown in Fig. 3 showed significantly higher importance when compared to the background distribution of absolute LRP values of all genes using the Mann–Whitney U (MWU) test (P< .001, FDR correction).

NeurixAI relied on similar gene expression patterns for the response prediction of drugs from the same drug category.

For dasatinib, NeurixAI relied strongly on the expression of LPAR1, ZNF853, MAL2, TUSC3, and IGFBP5. k-means clustering revealed a cluster of four genes (BCO1, HNF1B, HAVCR1, and PDZK1IP1; Supplementary Fig. S9) that contributed to strong drug response, in particular, in kidney tumors (average LRP score = −1.75) compared to all other organs (average LRP score = −0.005, P< .001, MWU test). Previous studies have identified an influence of LPAR1 and members of the IGFBP family on tumor response to dasatinib [31, 32].

For the prediction of EGFR inhibitor response as well as the response to ibrutinib, higher expression of ARL17A (and ARL17B), EMP3, H3-2, and LRRC37A2 contributed to increased drug resistance. High expression of ZNF880, SYT16, and FAM25A contributed alongside high ERBB2 expression the most to improved drug response. A k-means clustering analysis revealed that MIEN1, PGAP3, and ERBB2, often alongside STARD3, showed similar behavior across cell lines (Supplementary Figs S10–S15). The effect of EMP3 on EGFR inhibitor response is well documented and has been linked to the trafficking of EGFR into endosomes [33]. Evidence supporting an interaction between ZNF880 and EGFR inhibitor response has been reported [34].

For MEK inhibitors, NeurixAI assigned the most importance to ANXA2R, NPL, ODF3L1, CCDC188, SH3TC2, and TBC1D12. NPPC, TMEM272, and PFKFB4 contributed to increased drug response in skin cancer for cobimetinib (P< .001, MWU test) (Supplementary Figs S16 and S17). Higher expression of NPL, CCDC188, SH3TC2, and ODF3L1 also contributed to higher drug response.

For several of these genes, there has already been strong evidence for an influence on drug response.

High ANXA2R contributed to reduced drug response—a relationship that has been previously reported for MEK inhibitors [35]. ANXA2R and SH3TC2 have also been identified as one of the top influences on cancer response to MEK inhibitors before in drug perturbation experiments; however, the precise mechanism is not well established. NPPC has been shown to directly influence MEK signalling [36], possibly through cGMP [37]. Crucially, PFKFB4 has been shown to sustain the MAPK/ERK pathway and c-Myc activity: knocking down PFKFB4 in cancer cells leads to cell-cycle arrest and inhibition of glucose metabolism, concurrently inactivating the MEK/ERK/c-Myc signaling axis [38]. Targeting PFKFB4 alongside MEK inhibitors has been suggested as a strategy to prevent this escape mechanism [38].

High PITPNM3 expression has previously been linked to chemotherapy resistance, but no association with MEK inhibitor response has yet been reported. [39]. TBC1D12, a member of the Rab GTPase-activating protein family, may indirectly modulate inhibitor response through autophagy and receptor signaling dynamics [40].

CCDC188 and ODF3L1 have not been described to be involved in mechanisms related to drug response before and have been associated usually to spermatogenesis [41]. However, there have been recent reports of a strong prognostic value of ODF3L1 [42], while CCDC188 has been associated with MET pathway activity [43] and chemotherapy response in head and neck cancers [44]. However, it must be assumed that the relationship between CCDC188 and ODF3L1, while correlational, do not play a mechanistic role in response to MEK inhibitors.

To predict idasanutlin response, NeurixAI relied in particular on EDA2R, MDM2, CDKN1A, SESN1, and ZMAT3. (Supplementary Figs S8, S18). According to interactions found in the STRING database [30], these genes showed close functional relationships (Fig. 3) and there has been evidence for interactions with idasanutlin via pathways associated with p53 [45–47]. For vincristine and vinblastine, the genes ABCB1, TUSC3, RAB34, and RPRD1A were of high importance. According to NeurixAI, ABCB1, known for its role in multidrug resistance [48, 49], particularly contributed to a reduced drug sensitivity for vincristine and vinblastine in cancers from the liver, large intestine, and the kidney (P< .001, MWU test) (Supplementary Figs S19–S21). This association is supported by several studies for these cancer types [50–53]. While the role of TUSC3 in drug resistance is less established, previous studies suggest that high TUSC3 expression enhances glioblastoma sensitivity to temozolomide [54].

Class-specific variability of gene influence on drug response

Overall, drugs belonging to the same class exhibited similar LRP profiles, suggesting a comparable contribution to NeurixAI’s drug response predictions (Supplementary Fig. S22). However, many genes also contributed differentially across drugs, even within the same drug class (Supplementary Figs S23–S25).

For EGFR inhibitors, several genes differed in percentile ranks with >60 percentage points (Supplementary Fig. S23). TMCC2 was important in lapatinib (percentile rank 92%) but not in pelitinib (percentile rank 27%). In contrast, NeurixAI relied on the genes ATG4C (percentile rank 96%, others: 31%–86%) and MUC13 (94%, others: 23%–84%) for the prediction of response to pelitinib. PLPBP was important for poziotinib (percentile rank 95%) and dacomitinib (percentile rank 95%) but not for lapatinib (percentile rank 35%). KIAA1549L was important for the response to canertinib (percentile rank 94%). For poziotinib, NeurixAI put the least importance on this gene (percentile rank 20%).

For MEK inhibitors, several genes had differential importance between drugs (Supplementary Fig. S24). For drug response to trametinib, NeurixAI predicted a dependence on the genes NUDT7(93%) and CKAP4 (91%), while these genes had percentile ranks below 80% for cobimetinib. For cobimetinib, WWC1 (93%), SLC35F2 (91%), and ALDOC (92%) were assigned higher importance than for trametinib (<80%).

Supplementary Fig. S25 shows genes with differential importance for the response to the tubulin polymerization inhibitors vinblastine and vincristine. For vincristine, NeurixAI relied strongly on STK19 (94%), NID2 (93%), and TAF11 (92%). These genes had percentile ranks below 80% for vinblastine. The genes NTF3 (95%), CAMKK2 (92%), and TRAPPC11 (92%) were important for vinblastine response prediction but had percentile ranks below 80% for vincristine.

Since NeurixAI obtained information about the proximity of drugs based on prior knowledge of their common targets, a bias toward the high similarity of LRP profiles between similar drugs may be assumed. However, while training NeurixAI without prior knowledge about drug target similarities makes it rely on other gene expression patterns, these are still similar between drugs of the same drug class (Pearson’s r of LRP profiles between drugs ranging between 0.82 and 0.95; Supplementary Fig. S22B). Differences between LRP scores derived from NeurixAI with and without prior knowledge were stronger for low relevance attributions (overall Pearson’s r = 0.56) but decreased when selecting only the top 0.1% most important genes (Pearson’s r = 0.8). This suggests that these models have higher agreement when genes contribute strongly to the drug response but agree less with each other when the gene effects are minor.

Here, we reported the LRP profiles of NeurixAI with prior knowledge, as the model performed better in predicting drug response overall (Supplementary Fig. S26). In practice, choosing the appropriate approach might be best decided on a case-by-case basis and supported by comparing the model predictions with the observed drug response if this information is available.

Quantitative relationships between gene expression and drug response

The previous description of “gene importance” summarized the contribution of each gene for and against drug response across cell lines. Therefore, it did not enable quantitative conclusions about specific tumor–drug interactions. To address this limitation, we utilized the sample-wise LRP explanations to quantify the impact of gene expression on drug response (Fig. 4).

Figure 4.

Relationship between gene expression and contribution to drug sensitivity. Each point indicates the expression of a gene and its associated contribution to drug response. Arrows indicate that positive LRP scores contribute to increased drug resistance, while negative LRP scores contribute to increased drug sensitivity. The influence of gene expression on drug sensitivity varied strongly among drugs.

For EGFR inhibitors, as well as dasatinib and ibrutinib, EGFR gene expression had a strong influence on cell viability (Fig. 4). High gene expression led to a strongly negative LRP score (contribution to a strong drug response), an effect that surpassed a linear relationship (Supplementary Table S4). ERBB2 exhibited a comparable pattern. Additionally, high ERBB2 expression correlated with a reduced predicted drug response for the MEK inhibitors cobimetinib and trametinib (Fig. 4).

MDM2 expression strongly affected the predicted drug response for the MDM2 inhibitor idasanutlin. Low MDM2 expression led to a high LRP score (contribution to weak drug response), while high MDM2 expression was assigned a low LRP score (contribution to strong drug response). For other drugs, changes in MDM2 expression were not associated with a decisive shift in drug response.

We showed that NeurixAI relied on the gene expression of MAP2K6 and MAP3K21 to predict the response of MEK inhibitors (Supplementary Fig. S7). Figure 4 shows that these genes contributed differently. While high MAP2K6 levels suggested a contribution to a strong drug response, MAP3K21 contributed to a strong drug response only when expression levels were low.

ABCB1 (MDR-1) is a gene known to be associated with multidrug resistance [48, 49]. NeurixAI predicted a clear association between higher ABCB1 expression and lower drug response (higher LRP score, i.e. contribution to higher log AUC) only for vincristine and vinblastine. In contrast, ABCB1 expression did not influence the response to the remaining targeted therapies.

To further investigate this relationship, we analyzed the relationship between ABCB1 expression and LRP scores for all drugs of the PRISM drug screen for which NeurixAI made relevant predictions (Spearman’s rho between prediction and experimental ground truth >0.2). The slope between ABCB1 expression and predicted contribution to drug resistance was significantly higher in chemotherapies (median: 0.39, IQR: 0.5) compared to targeted therapies (median: 0.11, IQR: 0.34, P< .001, MWU test) and non-cancer therapies (median: 0.14, IQR: 0.37, P < .001, MWU test) (Supplementary Fig. S27).

Discussion

Diagnostic mutational profiling by next-generation sequencing is the cornerstone of precision oncology and has led to substantial advances in the clinical management of cancer patients [55–58]. However, even within the same tumor type and mutational profile, treatment response varies widely among patients, likely due to the mutational complexity and tumor heterogeneity. In this context, gene expression profiling provides complementary functional insights that can enhance the prediction of treatment response [4, 6, 21]. In this study, we demonstrated the potential of integrating large-scale drug perturbation data and transcriptomic profiling with xAI to better understand how molecular profiles modulate drug response and predict potential resistance mechanisms. Our deep learning framework, NeurixAI, efficiently models the interactions between transcriptomic profiles and diverse drugs, leveraging over 500 000 drug perturbation experiments. This approach enabled reliable drug response predictions for the majority of targeted and chemotherapeutic agents. Furthermore, NeurixAI is designed for computational efficiency, allowing the rapid evaluation of millions of tumor–drug interactions by representing both tumor and drug characteristics in a joint latent space. These abstract molecular representations (tumor–latent and drug–latent vectors) capture drug–tumor interactions comprehensively, supporting large-scale predictions with high accuracy and low computational cost.

Across all drugs, our results reveal notable differences in the predictability of various drug classes. This suggests that, in future clinical applications, drug response prediction may be valuable for guiding treatment with certain drug classes, while offering limited utility for others.

NeurixAI showed the highest predictive performance for targeted therapies, such as MEK and EGFR inhibitors, whereas responses to several chemotherapeutic and non-cancer drugs were more difficult to predict. This is expected, as the mechanisms of these compounds are often nonspecific and their efficacy may not be strongly associated with distinct transcriptomic patterns, making them harder to infer from RNA profiles [59, 60]. However, NeurixAI successfully predicted the response to tubulin polymerization inhibitors, a class of chemotherapies, highlighting its ability to capture transcriptomic signals even for certain non-targeted drugs.

NeurixAI’s performance likely benefits from its ability to learn across the full spectrum of tumor–drug interactions, enabling it to generalize patterns observed in one drug class to others with shared pharmacological or mechanistic features. To facilitate such transfer learning, we provided the model with structured drug representations that capture relevant properties, thereby allowing it to extrapolate learned relationships across related compounds.

Using LRP [8–10, 61], we identified key gene expression patterns associated with drug resistance. As expected, high ABCB1 expression was strongly linked to drug resistance [48], but NeurixAI also identified low TUSC3 and RAB34 expression as modulators of chemotherapy response. For targeted therapies, NeurixAI placed a strong emphasis on drug target genes and closely related proteins but also identified off-target genes influencing drug response. NeurixAI identified numerous genes previously reported to influence drug response within specific drug classes. While our xAI approach highlights genes with strong associations to drug response, several of the identified genes have also been described to exert direct, mechanistic effects. These findings suggest that NeurixAI may aid in elucidating drug action, identifying candidate compounds, and advancing drug development. Our findings emphasize that drug response is not solely dictated by a target gene but is also influenced by a broader network of gene interactions. By providing instance-wise explanations through LRP, NeurixAI not only identifies the overall importance of a gene but also quantifies nonlinear relationships between gene expression and drug response. This ability to dynamically assess gene–drug interactions at the individual tumor level could be valuable for identifying candidate drug targets or mechanisms responsible for drug resistance that could be targeted by secondary drugs.

Our results show that gene importance profiles vary substantially across drug classes, suggesting class-specific determinants of drug response. These findings could guide future experimental validation by enabling the construction of low-dimensional predictive scores based on key genes, and may inform drug development by highlighting molecular factors that modulate class-specific mechanisms of action.

Despite its robust predictive performance, translating prediction models into clinical applications remains a significant challenge [62]. We showed that NeurixAI is capable of making relevant predictions for more than half of the tested drugs on an external dataset. However, the generalization of models such as ours to the clinical real-world setting is particularly challenging. For clinical translation, it is essential that predictive models demonstrate the ability to generalize to patient-derived tumor samples. This remains a substantial challenge due to fundamental differences between in vitro cell lines and primary tumor specimens [63]. Patient samples typically exhibit significant cellular heterogeneity, comprising malignant cells alongside stromal, immune, and normal epithelial components—features largely absent in cell line models [63]. Moreover, drug response in vivo is shaped by complex physiological factors, including pharmacokinetics, immune interactions, and the tumor microenvironment, all of which can modulate therapeutic effectiveness [64]. Bridging this gap will require model adaptation and validation using patient-derived data [65]. Future studies integrating single-cell analyses of patient-derived tumor models could help bridge this gap.

Our study demonstrates that explicitly modeling drug response interactions using large-scale perturbation experiments, transcriptomics, and xAI provides critical insights into drug response and resistance. NeurixAI not only achieved high predictive accuracy but also revealed complex tumor- and drug-specific gene expression patterns associated with drug response. This integrative approach holds potential for precision oncology, offering a scalable framework to identify novel drug targets and optimize treatment strategies for cancer patients.

Supplementary data

Supplementary data is available at NAR Cancer online.

Conflict of interest

K.-R.M., F.K., and G.M. hold patents related to this work (9558550; 20180018553). K.-R.M. and F.K. are co-founders of the computational pathology start-up Aignostics.

Funding

J.K. was supported by a German Research Foundation (DFG)-funded clinician scientist program (FU 356/12-2). This work was partly funded by the Bundesministerium für Bildung und Forschung (BMBF) (under grants BIFOLD24B, BIFOLD25B, 01IS18025A, and 01IS18037A). Furthermore, K.-R.M. was partly supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grants funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University and No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation).

Data availability

All data are available at https://depmap.org/portal/. The code written in support of this publication is available at https://github.com/PhGK/NeurixAI/ (https://doi.org/10.5281/zenodo.15536558).