De Novo Design of Antimicrobial Peptides: From Sequence Templates to Artificial Intelligence

Bacterial resistance, eradicating the efficiency of many antibiotics, is globally increasing and poses a threat to public health. In parallel, an all-time low number of newly approved antibiotics can be witnessed as several pharmaceutical companies have withdrawn from antibiotic R&D. Together with the steady increase in cost and prolongation of the whole drug discovery process, the need for novel antibiotics has promoted the field of membranolytic antimicrobial peptides (AMPs) as alternative antimicrobial treatments. As AMPs directly act on microbial membranes and do not depend on molecular interactions with specific protein targets prone to mutations, it was hypothesized that these peptides would be less affected by known bacterial resistance mechanisms. Most known AMPs are naturally derived molecules with a role in host-defense of multi-cellular organisms. Yet, many of these peptides are either unspecific, metabolically instable or only exhibit moderate activity. Several attempts have been pursued in literature to either optimize natural AMPs or design new peptides with antimicrobial activity from scratch. The first, template-based approach, requires a known amino acid (AA) sequence with a desired activity to be improved. Enhancement strategies include step-wise deletion or exchange of residues with other AA or towards a different known sequence, employing evolutionary-inspired processes mimicking adaptation and selection, or randomly shuffling the whole peptide sequence. On the other hand, designing peptides de novo enables a broader exploration of chemical space, but demands for a priori knowledge of features important for activity. These de novo design strategies frequently employ statistical analysis tools and machine learning (ML) algorithms which learn from known AMPs. The trained models are then used for the activity prediction of computer-generated peptide sequences. Although both approaches have provided important insights on the structure activity relationship (SAR) of AMPs and helped to generate novel active peptides in the past, the exact features that render a peptide membranolytic on selected microbes are still a matter of debate. Therefore, we attempted to find novel features determining the activity and selectivity of cationic alpha-helical AMPs and use them for the de novo design of selective AMPs targeting resistant bacteria. In this thesis, we pursued this objective by four different approaches. In our first study, the known AMP Aurein 2.2d2 was gradually transformed into the synthetic amphipathic template AMP Klk14 through manual sequential morphing. Both peptides, as well as all intermediate designs that were synthesized, share the same overall amphipathic secondary structure. We obtained 18 novel peptides, of which a surprising high number of 11 were discovered to be growth-inhibiting against Staphylococcus aureus and six of them additionally active against selected resistant strains. One of the six was selective when compared to human erythrocytes and primary endothelial cells. Surprisingly, single AA exchanges between side chains of the same pharmacophore were able to completely eradicate activity. These activity-cliffs were unpredictable by current computational models. An ensuing physicochemical characterization and in-depth investigation of peptide features revealed C-terminal flexibility as a characteristic of our active and selective peptides and a potential new feature to be considered for AMP design. As current AMP activity prediction tools were shown to lack generalizability and specificity, we have prototyped active machine learning as a strategy to improve AMP prediction from AA sequences in a second study. In active learning, a ML model autonomously queries new compounds to be tested from a large available virtual library. A retrospective comparison of assorted picking strategies revealed that balanced, density-weighted selection intertwining greediness with curiosity simultaneously improves the prediction model and retrieves active AMPs. In four prospective active learning cycles, the model iteratively picked batches of 10 sequences that were synthesized and tested on bacteria. The experimental feedback was used to retrain and improve the model towards a final perfect classification. Out of 40 synthesized peptides, 22 were tested active against Staphylococcus aureus and / or Escherichia coli, which were all cationic peptides designed as amphipathic helices. Aforementioned ML models can implicitly reveal molecular features important for the observed activity. In our third study, we made use of this characteristic and designed sparsely connected artificial neural network models. Optimization of their architecture was achieved through evolutionary algorithms and eventually allowed cascaded feature extraction from the used molecular descriptor. We found that a peptide's charge density and aggregation propensity were both momentous to the model for distinguishing membrane-active AMPs from transmembrane regions of proteins. Comparably, a linear hydrophobicity gradient was considered crucial by our networks to tell apart helical AMPs and other helical inactive peptides. We validated our final models by synthesis and testing of the top and bottom five predicted actives from a large in silico sequence library and achieved 70% accuracy in our predictions. Finally, we put artificial intelligence in charge of exploring and designing sequences in novel areas of antimicrobial peptide space. By omitting molecular descriptors and directly training a recurrent neural network (RNN) model on the single-letter representation of helical AMP sequences, we were able to sample the chemical space around these peptides. The generated novel sequences closely resemble the training data and retained molecular features important for AMP activity. This first time application of RNNs in the field of AMPs propels the use of generative models for peptide de novo design. Yet, this theoretical foundation still needs to be validated by actual synthesis and activity tests in future studies. In summary, we have demonstrated the potential of connecting a priori knowledge obtained through manual design efforts to ML algorithms for prospective peptide de novo design. In the context of the ever increasing costs of drug development, such a machine-guided design approach might be a promising option to foster the discovery of novel innovative antibiotics in the future