Multimodal learning in clinical proteomics: enhancing antimicrobial resistance prediction models with chemical information

Abstract Motivation Large-scale clinical proteomics datasets of infectious pathogens, combined with antimicrobial resistance outcomes, have recently opened the door for machine learning models which aim to improve clinical treatment by predicting resistance early. However, existing prediction frameworks typically train a separate model for each antimicrobial and species in order to predict a pathogen’s resistance outcome, resulting in missed opportunities for chemical knowledge transfer and generalizability. Results We demonstrate the effectiveness of multimodal learning over proteomic and chemical features by exploring two clinically relevant tasks for our proposed deep learning models: drug recommendation and generalized resistance prediction. By adopting this multi-view representation of the pathogenic samples and leveraging the scale of the available datasets, our models outperformed the previous single-drug and single-species predictive models by statistically significant margins. We extensively validated the multi-drug setting, highlighting the challenges in generalizing beyond the training data distribution, and quantitatively demonstrate how suitable representations of antimicrobial drugs constitute a crucial tool in the development of clinically relevant predictive models. Availability and implementation The code used to produce the results presented in this article is available at https://github.com/BorgwardtLab/MultimodalAMR.

Table A2.Direct AMR prediction results with multi-drug models.The best average metric is highlighted for each dataset and split.Fig. A6: Fraction of the test samples with a positive label (i.e.displaying resistance to the drug).The random split allows us to sample with stratification to obtain consistent test splits.The species-drug zero-shot split relies on a heuristic to construct test sets that approximately constitute a specific percentage of the total available samples, leading to small fluctuations in the fraction of positive samples in each test split.The drug zero-shot setting, on the other hand, does not allow for any control on the imbalance of the test split.The resulting variability is a challenge for the machine learning models tested to overcome.For each DRIAMS collection site, 10 train/test splits are randomly selected, and a ResMLP model is trained using the same configuration presented for the results in the classification performance.The resulting metrics display a small level of variation.An analysis of variance test, however, confirmed that the choice of molecular fingerprint is not statistically significant.Supplementary figures for the drug recommendation task  Fig. A12: Truncated precision at cut-offs 1, 2, 3, 4, and 5 for the different recommendation set-ups.For the random baseline, baseline species, and spectrum similarity approaches, the number of top neighbours k is set to 30.

MACCS
Fig. A13: Chemical structures of representative drugs in four common antibiotic families.Penicillin: the beta-lactam ring (highlighted in red) is a key structural feature in beta-lactam antibiotics such as Penicillin, where resistance is frequently conferred by beta-lactamase enzymes, which hydrolyze the amide bond, rendering the antibiotic inactive.The beta-lactam ring ranks amongst the first in the SHAP feature importance analyses presented in this paper for most Penicillin antibiotics (such as Amoxicillin, Oxacillin, Ampicillin and Benzylpenicillin).Interestingly, antibiotics that are not susceptible to beta-lactamases (such as Cefepime and Aztreoman) do not follow this trend.Gentamicin: is an aminoglycoside antibiotic, as are Amikacin and Tobramycin.Although other resistance mechanisms involving these drugs have been reported, by far the most prevalent involve AG-modifying enzymes that target the glycoside rings and their aglycone components, which matches the top ranked features by the provided SHAP analysis (highlighted in red).Chloramphenicol encounters high-level bacterial resistance due to the enzyme chloramphenicol acetyltransferase.This enzyme mediates the transfer of an acetyl group from acetyl CoA to the primary hydroxyl group within the chloramphenicol molecule, which ranks as the top chemical feature in the provided SHAP analysis.Erythromycin is a macrolide (such as Azithromycin), a family of antibiotics where drug modification is also the prevalent resistance mechanism in place.Among others, hydrolyzation of the ester group by esterases in particular acts in the ester group next to the atoms highlighted in red (which rank first in the provided feature importance analysis for both mentioned macrolide antibiotics).

Fig. A2 :
Fig. A2: ROC and PRC curves for selected combinations of species and drugs using the ResMLP model.This figure is a reproduction of Figure 2 from Weis et al. (2022), which showcases how the use of a multi-drug model outperforms the more restricted models presented in their work.In the parentheses next to the AUC measures, the difference compared to the corresponding models in Figure 2 from Weis et al. (2022).

Fig. A4 :
Fig. A4: Full set of Precision-Recall curves for the comparison with the models presented in Figure 2 from Weis et al. (2022) Fig.A7: Comparison of the performance of a ResMLP model trained using the different types of molecular fingerprints.For each DRIAMS collection site, 10 train/test splits are randomly selected, and a ResMLP model is trained using the same configuration presented for the results in the classification performance.The resulting metrics display a small level of variation.An analysis of variance test, however, confirmed that the choice of molecular fingerprint is not statistically significant.
Fig.A8: AUPRC for the predictions in the drug zero-shot task on the four DRIAMS datasets.Drugs for which only one class of response was available have been excluded from the analysis.In orange, we represented as dots the highest Tanimoto index calculated by comparing the MACCS fingerprints of the target drug with the rest of the compounds in the dataset.This measure of similarity does not appear to share a pattern with the generalization performance measured by the AUPRC.
Fig. A11: (a) Comparison of the number of drugs recommended with different variations of the approach based on the similarity between spectra.The error bars represent the 95% confidence interval.(b) Comparison of performance of different variants of the spectrum similarity set-up based on the top k similarity.The y-axis shows the average precision across all individuals in the test set.The error bars represent the 95% confidence interval.

Table A3 .
Results of one-way Kruskal-Wallis analysis of variance for comparing the performance of the ResMLP model using the different categories of molecular fingerprints (MACCS, 1024-dimensional Morgan, Pubchem).Using the random split setting, 10 train/test splits were performed on each of the 4 DRIAMS collection sites.A ResMLP model was trained using the same configuration as the main section of the classification experiments.The resulting metrics were used to calculate the Kruskal-Wallis H statistic.The resulting p-values indicate that no fingerprint choice leads to significantly different results, using a significance threshold of 0.05.