Search

Abstract Poly ADP-ribose polymerase 1 (PARP1) is an attractive therapeutic target for cancer treatment. Machine-learning scoring functions constitute a promising approach to discovering novel PARP1 inhibitors. Cutting-edge PARP1-specific machine-learning scoring functions were investigated using semi-synthetic training data from docking activity-labelled molecules: known PARP1 inhibitors, hard-to-discriminate decoys property-matched to them with generative graph neural networks and confirmed inactives. We further made test sets harder by including only molecules dissimilar to those in the training set. Comprehensive analysis of these datasets using five supervised learning algorithms, and protein–ligand fingerprints extracted from docking poses and ligand only features revealed one highly predictive scoring function. This is the PARP1-specific support vector machine-based regressor, when employing PLEC fingerprints, which achieved a high Normalized Enrichment Factor at the top 1% on the hardest test set (NEF1% = 0.588, median of 10 repetitions), and was more predictive than any other investigated scoring function, especially the classical scoring function employed as baseline.

Background: Gemcitabine is a first-line chemotherapy for pancreatic adenocarcinoma (PAAD), but many PAAD patients do not respond to gemcitabine-containing treatments. Being able to predict such nonresponders would hence permit the undelayed administration of more promising treatments while sparing gemcitabine life-threatening side effects for those patients. Unfortunately, the few predictors of PAAD patient response to this drug are weak, none of them exploiting yet the power of machine learning (ML). Methods: Here, we applied ML to predict the response of PAAD patients to gemcitabine from the molecular profiles of their tumors. More concretely, we collected diverse molecular profiles of PAAD patient tumors along with the corresponding clinical data (gemcitabine responses and clinical features) from the Genomic Data Commons resource. From systematically combining 8 tumor profiles with 16 classification algorithms, each of the resulting 128 ML models was evaluated by multiple 10-fold cross-validations. Results: Only 7 of these 128 models were predictive, which underlines the importance of carrying out such a large-scale analysis to avoid missing the most predictive models. These were here random forest using 4 selected mRNAs [0.44 Matthews correlation coefficient (MCC), 0.785 receiver operating characteristic–area under the curve (ROC-AUC)] and XGBoost combining 12 DNA methylation probes (0.32 MCC, 0.697 ROC-AUC). By contrast, the hENT1 marker obtained much worse random-level performance (practically 0 MCC, 0.5 ROC-AUC). Despite not being trained to predict prognosis (overall and progression-free survival), these ML models were also able to anticipate this patient outcome. Conclusions: We release these promising ML models so that they can be evaluated prospectively on other gemcitabine-treated PAAD patients.

The interaction between PD1 and its ligand PDL1 has been shown to render tumor cells resistant to apoptosis and promote tumor progression. An innovative mechanism to inhibit the PD1/PDL1 interaction is PDL1 dimerization induced by small-molecule PDL1 binders. Structure-based virtual screening is a promising approach to discovering such small-molecule PD1/PDL1 inhibitors. Here we investigate which type of generic scoring functions is most suitable to tackle this problem. We consider CNN-Score, an ensemble of convolutional neural networks, as the representative of machine-learning scoring functions. We also evaluate Smina, a commonly used classical scoring function, and IFP, a top structural fingerprint similarity scoring function. These three types of scoring functions were evaluated on two test sets sharing the same set of small-molecule PD1/PDL1 inhibitors, but using different types of inactives: either true inactives (molecules with no in vitro PD1/PDL1 inhibition activity) or assumed inactives (property-matched decoy molecules generated from each active). On both test sets, CNN-Score performed much better than Smina, which in turn strongly outperformed IFP. The fact that the latter was the case, despite precluding any possibility of exploiting decoy bias, demonstrates the predictive value of CNN-Score for PDL1. These results suggest that re-scoring Smina-docked molecules with CNN-Score is a promising structure-based virtual screening method to discover new small-molecule inhibitors of this therapeutic target.

Abstract Doxorubicin is a common treatment for breast cancer. However, not all patients respond to this drug, which sometimes causes life‐threatening side effects. Accurately anticipating doxorubicin‐resistant patients would therefore permit to spare them this risk while considering alternative treatments without delay. Stratifying patients based on molecular markers in their pretreatment tumors is a promising approach to advance toward this ambitious goal, but single‐gene gene markers such as HER2 expression have not shown to be sufficiently predictive. The recent availability of matched doxorubicin‐response and diverse molecular profiles across breast cancer patients permits now analysis at a much larger scale. 16 machine learning algorithms and 8 molecular profiles are systematically evaluated on the same cohort of patients. Only 2 of the 128 resulting models are substantially predictive, showing that they can be easily missed by a standard‐scale analysis. The best model is classification and regression tree (CART) nonlinearly combining 4 selected miRNA isoforms to predict doxorubicin response (median Matthew correlation coefficient (MCC) and area under the curve (AUC) of 0.56 and 0.80, respectively). By contrast, HER2 expression is significantly less predictive (median MCC and AUC of 0.14 and 0.57, respectively). As the predictive accuracy of this CART model increases with larger training sets, its update with future data should result in even better accuracy.

Machine learning-based models have been widely used in the early drug-design pipeline. To validate these models, cross-validation strategies have been employed, including those using clustering of molecules in terms of their chemical structures. However, the poor clustering of compounds will compromise such validation, especially on test molecules dissimilar to those in the training set. This study aims at finding the best way to cluster the molecules screened by the National Cancer Institute (NCI)-60 project by comparing hierarchical, Taylor–Butina, and uniform manifold approximation and projection (UMAP) clustering methods. The best-performing algorithm can then be used to generate clusters for model validation strategies. This study also aims at measuring the impact of removing outlier molecules prior to the clustering step. Clustering results are evaluated using three well-known clustering quality metrics. In addition, we compute an average similarity matrix to assess the quality of each cluster. The results show variation in clustering quality from method to method. The clusters obtained by the hierarchical and Taylor–Butina methods are more computationally expensive to use in cross-validation strategies, and both cluster the molecules poorly. In contrast, the UMAP method provides the best quality, and therefore we recommend it to analyze this highly valuable dataset.

Abstract Background Oncology drugs are only effective in a small proportion of cancer patients. Our current ability to identify these responsive patients before treatment is still poor in most cases. Thus, there is a pressing need to discover response markers for marketed and research oncology drugs. Screening these drugs against a large panel of cancer cell lines has led to the discovery of new genomic markers of in vitro drug response. However, while the identification of such markers among thousands of candidate drug-gene associations in the data is error-prone, an appraisal of the effectiveness of such detection task is currently lacking. Methods Here we present a new non-parametric method to measuring the discriminative power of a drug-gene association. Unlike parametric statistical tests, the adopted non-parametric test has the advantage of not making strong assumptions about the data distorting the identification of genomic markers. Furthermore, we introduce a new benchmark to further validate these markers in vitro using more recent data not used to identify the markers. Results The application of this new methodology has led to the identification of 128 new genomic markers distributed across 61% of the analysed drugs, including 5 drugs without previously known markers, which were missed by the MANOVA test initially applied to analyse data from the Genomics of Drug Sensitivity in Cancer consortium. Conclusions Discovering markers using more than one statistical test and testing them on independent data is unusual. We found this helpful to discard statistically significant drug-gene associations that were actually spurious correlations. This approach also revealed new, independently validated, in vitro markers of drug response such as Temsirolimus-CDKN2A (resistance) and Gemcitabine-EWS_FLI1 (sensitivity).

(1) Background: Inter-tumour heterogeneity is one of cancer’s most fundamental features. Patient stratification based on drug response prediction is hence needed for effective anti-cancer therapy. However, single-gene markers of response are rare and/or may fail to achieve a significant impact in the clinic. Machine Learning (ML) is emerging as a particularly promising complementary approach to precision oncology. (2) Methods: Here we leverage comprehensive Patient-Derived Xenograft (PDX) pharmacogenomic data sets with dimensionality-reducing ML algorithms with this purpose. (3) Results: Combining multiple gene alterations via ML leads to better discrimination between sensitive and resistant PDXs in 19 of the 26 analysed cases. Highly predictive ML models employing concise gene lists were found for three cases: paclitaxel (breast cancer), binimetinib (breast cancer) and cetuximab (colorectal cancer). Interestingly, each of these multi-gene ML models identifies some treatment-responsive PDXs not harbouring the best actionable mutation for that case. Thus, ML multi-gene predictors generally have much fewer false negatives than the corresponding single-gene marker. (4) Conclusions: As PDXs often recapitulate clinical outcomes, these results suggest that many more patients could benefit from precision oncology if ML algorithms were also applied to existing clinical pharmacogenomics data, especially those algorithms generating classifiers combining data-selected gene alterations.

Abstract Many computational methods to predict the macromolecular targets of small organic molecules have been presented to date. Despite progress, target prediction methods still have important limitations. For example, the most accurate methods implicitly restrict their predictions to a relatively small number of targets, are not systematically validated on drugs (whose targets are harder to predict than those of non-drug molecules) and often lack a reliability score associated with each predicted target. Here we present a systematic validation of ligand-centric target prediction methods on a set of clinical drugs. These methods exploit a knowledge-base covering 887,435 known ligand-target associations between 504,755 molecules and 4,167 targets. Based on this dataset, we provide a new estimate of the polypharmacology of drugs, which on average have 11.5 targets below IC50 10 µM. The average performance achieved across clinical drugs is remarkable (0.348 precision and 0.423 recall, with large drug-dependent variability), especially given the unusually large coverage of the target space. Furthermore, we show how a sparse ligand-target bioactivity matrix to retrospectively validate target prediction methods could underestimate prospective performance. Lastly, we present and validate a first-in-kind score capable of accurately predicting the reliability of target predictions.

To address the problem of resistance to paclitaxel treatment, we have investigated to which extent is possible to predict Breast Cancer (BC) patient response to this drug. We carried out a large-scale tumor-based prediction analysis using data from the US National Cancer Institute’s Genomic Data Commons. These data sets comprise the responses of BC patients to paclitaxel along with six molecular profiles of their tumors. We assessed 10 Machine Learning (ML) algorithms on each of these profiles and evaluated the resulting 60 classifiers on the same BC patients. DNA methylation and miRNA profiles were the most informative overall. In combination with these two profiles, ML algorithms selecting the smallest subset of molecular features generated the most predictive classifiers: a complexity-optimized XGBoost classifier based on CpG island methylation extracted a subset of molecular factors relevant to predict paclitaxel response (AUC = 0.74). A CpG site methylation-based Decision Tree (DT) combining only 2 of the 22,941 considered CpG sites (AUC = 0.89) and a miRNA expression-based DT employing just 4 of the 337 analyzed mature miRNAs (AUC = 0.72) reveal the molecular types associated to paclitaxel-sensitive and resistant BC tumors. A literature review shows that features selected by these three classifiers have been individually linked to the cytotoxic-drug sensitivities and prognosis of BC patients. Our work leads to several molecular signatures, unearthed from methylome and miRNome, able to anticipate to some extent which BC tumors respond or not to paclitaxel. These results may provide insights to optimize paclitaxel-therapies in clinical practice.

Drug combinations are of great interest for cancer treatment. Unfortunately, the discovery of synergistic combinations by purely experimental means is only feasible on small sets of drugs. In silico modeling methods can substantially widen this search by providing tools able to predict which of all possible combinations in a large compound library are synergistic. Here we investigate to which extent drug combination synergy can be predicted by exploiting the largest available dataset to date (NCI-ALMANAC, with over 290,000 synergy determinations). Each cell line is modeled using primarily two machine learning techniques, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), on the datasets provided by NCI-ALMANAC. This large-scale predictive modeling study comprises more than 5,000 pair-wise drug combinations, 60 cell lines, 4 types of models, and 5 types of chemical features. The application of a powerful, yet uncommonly used, RF-specific technique for reliability prediction is also investigated. The evaluation of these models shows that it is possible to predict the synergy of unseen drug combinations with high accuracy (Pearson correlations between 0.43 and 0.86 depending on the considered cell line, with XGBoost providing slightly better predictions than RF). We have also found that restricting to the most reliable synergy predictions results in at least 2-fold error decrease with respect to employing the best learning algorithm without any reliability estimation. Alkylating agents, tyrosine kinase inhibitors and topoisomerase inhibitors are the drugs whose synergy with other partner drugs are better predicted by the models. Despite its leading size, NCI-ALMANAC comprises an extremely small part of all conceivable combinations. Given their accuracy and reliability estimation, the developed models should drastically reduce the number of required in vitro tests by predicting in silico which of the considered combinations are likely to be synergistic.

Background and purpose: Identifying the macromolecular targets of drug molecules is a fundamental aspect of drug discovery and pharmacology. Several drugs remain without known targets (orphan) despite large-scale in silico and in vitro target prediction efforts. Ligand-centric chemical-similarity-based methods for in silico target prediction have been found to be particularly powerful, but the question remains of whether they are able to discover targets for target-orphan drugs. Experimental Approach: We used one of these in silico methods to carry out a target prediction analysis for two orphan drugs: actarit and malotilate. The top target predicted for each drug was carbonic anhydrase II (CAII). Each drug was therefore quantitatively evaluated for CAII inhibition to validate these two prospective predictions. Key Results: Actarit showed in vitro concentration-dependent inhibition of CAII activity with submicromolar potency (IC50 = 422 nM) whilst no consistent inhibition was observed for malotilate. Among the other 25 targets predicted for actarit, RORγ (RAR-related orphan receptor-gamma) is promising in that it is strongly related to actarit’s indication, rheumatoid arthritis (RA). Conclusion and Implications: This study is a proof-of-concept of the utility of MolTarPred for the fast and cost-effective identification of targets of orphan drugs. Furthermore, the mechanism of action of actarit as an anti-RA agent can now be re-examined from a CAII-inhibitor perspective, given existing relationships between this target and RA. Moreover, the confirmed CAII-actarit association supports investigating the repositioning of actarit on other CAII-linked indications (e.g., hypertension, epilepsy, migraine, anemia and bone, eye and cardiac disorders).

In silico models to predict which tumors will respond to a given drug are necessary for Precision Oncology. However, predictive models are only available for a handful of cases (each case being a given drug acting on tumors of a specific cancer type). A way to generate predictive models for the remaining cases is with suitable machine learning algorithms that are yet to be applied to existing in vitro pharmacogenomics datasets. Here, we apply XGBoost integrated with a stringent feature selection approach, which is an algorithm that is advantageous for these high-dimensional problems. Thus, we identified and validated 118 predictive models for 62 drugs across five cancer types by exploiting four molecular profiles (sequence mutations, copy-number alterations, gene expression, and DNA methylation). Predictive models were found in each cancer type and with every molecular profile. On average, no omics profile or cancer type obtained models with higher predictive accuracy than the rest. However, within a given cancer type, some molecular profiles were overrepresented among predictive models. For instance, CNA profiles were predictive in breast invasive carcinoma (BRCA) cell lines, but not in small cell lung cancer (SCLC) cell lines where gene expression (GEX) and DNA methylation profiles were the most predictive. Lastly, we identified the best XGBoost model per cancer type and analyzed their selected features. For each model, some of the genes in the selected list had already been found to be individually linked to the response to that drug, providing additional evidence of the usefulness of these models and the merits of the feature selection scheme.

Docking scoring functions can be used to predict the strength of protein-ligand binding. It is widely believed that training a scoring function with low-quality data is detrimental for its predictive performance. Nevertheless, there is a surprising lack of systematic validation experiments in support of this hypothesis. In this study, we investigated to which extent training a scoring function with data containing low-quality structural and binding data is detrimental for predictive performance. We actually found that low-quality data is not only non-detrimental, but beneficial for the predictive performance of machine-learning scoring functions, though the improvement is less important than that coming from high-quality data. Furthermore, we observed that classical scoring functions are not able to effectively exploit data beyond an early threshold, regardless of its quality. This demonstrates that exploiting a larger data volume is more important for the performance of machine-learning scoring functions than restricting to a smaller set of higher data quality.

Background: Selected gene mutations are routinely used to guide the selection of cancer drugs for a given patient tumour. Large pharmacogenomic data sets, such as those by Genomics of Drug Sensitivity in Cancer (GDSC) consortium, were introduced to discover more of these single-gene markers of drug sensitivity. Very recently, machine learning regression has been used to investigate how well cancer cell line sensitivity to drugs is predicted depending on the type of molecular profile. The latter has revealed that gene expression data is the most predictive profile in the pan-cancer setting. However, no study to date has exploited GDSC data to systematically compare the performance of machine learning models based on multi-gene expression data against that of widely-used single-gene markers based on genomics data. Methods: Here we present this systematic comparison using Random Forest (RF) classifiers exploiting the expression levels of 13,321 genes and an average of 501 tested cell lines per drug. To account for time-dependent batch effects in IC50 measurements, we employ independent test sets generated with more recent GDSC data than that used to train the predictors and show that this is a more realistic validation than standard k-fold cross-validation. Results and Discussion: Across 127 GDSC drugs, our results show that the single-gene markers unveiled by the MANOVA analysis tend to achieve higher precision than these RF-based multi-gene models, at the cost of generally having a poor recall (i.e. correctly detecting only a small part of the cell lines sensitive to the drug). Regarding overall classification performance, about two thirds of the drugs are better predicted by the multi-gene RF classifiers. Among the drugs with the most predictive of these models, we found pyrimethamine, sunitinib and 17-AAG. Conclusions: Thanks to this unbiased validation, we now know that this type of models can predict in vitro tumour response to some of these drugs. These models can thus be further investigated on in vivo tumour models. R code to facilitate the construction of alternative machine learning models and their validation in the presented benchmark is available at http://ballester.marseille.inserm.fr/gdsc.transcriptomicDatav2.tar.gz.

Machine learning (ML) has become a crucial component of early drug discovery [...]

It has recently been claimed that the outstanding performance of machine-learning scoring functions (SFs) is exclusively due to the presence of training complexes with highly similar proteins to those in the test set. Here, we revisit this question using 24 similarity-based training sets, a widely used test set, and four SFs. Three of these SFs employ machine learning instead of the classical linear regression approach of the fourth SF (X-Score which has the best test set performance out of 16 classical SFs). We have found that random forest (RF)-based RF-Score-v3 outperforms X-Score even when 68% of the most similar proteins are removed from the training set. In addition, unlike X-Score, RF-Score-v3 is able to keep learning with an increasing training set size, becoming substantially more predictive than X-Score when the full 1105 complexes are used for training. These results show that machine-learning SFs owe a substantial part of their performance to training on complexes with dissimilar proteins to those in the test set, against what has been previously concluded using the same data. Given that a growing amount of structural and interaction data will be available from academic and industrial sources, this performance gap between machine-learning SFs and classical SFs is expected to enlarge in the future.

Background: Selected gene mutations are routinely used to guide the selection of cancer drugs for a given patient tumour. Large pharmacogenomic data sets were introduced to discover more of these single-gene markers of drug sensitivity. Very recently, machine learning regression has been used to investigate how well cancer cell line sensitivity to drugs is predicted depending on the type of molecular profile. The latter has revealed that gene expression data is the most predictive profile in the pan-cancer setting. However, no study to date has exploited GDSC data to systematically compare the performance of machine learning models based on multi-gene expression data against that of widely-used single-gene markers based on genomics data. Methods: Here we present this systematic comparison using Random Forest (RF) classifiers exploiting the expression levels of 13,321 genes and an average of 501 tested cell lines per drug. To account for time-dependent batch effects in IC 50 measurements, we employ independent test sets generated with more recent GDSC data than that used to train the predictors and show that this is a more realistic validation than K-fold cross-validation. Results and Discussion: Across 127 GDSC drugs, our results show that the single-gene markers unveiled by the MANOVA analysis tend to achieve higher precision than these RF-based multi-gene models, at the cost of generally having a poor recall (i.e. correctly detecting only a small part of the cell lines sensitive to the drug). Regarding overall classification performance, about two thirds of the drugs are better predicted by multi-gene RF classifiers. Among the drugs with the most predictive of these models, we found pyrimethamine, sunitinib and 17-AAG. Conclusions: We now know that this type of models can predict in vitro tumour response to these drugs. These models can thus be further investigated on in vivo tumour models.