Naslov
Soft computing for constructive peptide design and peptide activity prediction

Autori
Mauša, Goran ; Otović, Erik ; Njirjak, Marko ; Erjavac, Ivan ; Kalafatovic, Daniela

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Skup
Chemistry and Biology of Peptides Gordon Research Conference

Mjesto i datum
Oxnard, California, USA, 30.10.-04.11.2022

Vrsta sudjelovanja
Poster

Vrsta recenzije
Nije recenziran

Ključne riječi
peptide ; design ; prediction ; soft computing

Sažetak
Soft computing is a set of probabilistic algorithms, which are robust to imprecision and tolerant to uncertainty, that enable us to grapple with analytically intractable problems and make up for the lack of theoretical knowledge. In our project we apply a wide range of soft computing models to predict peptide activity, construct novel peptides and cover the chemical search space. More in depth, we tackle the problems of (1) sensivity of highly accurate predictive models, (2) building predictive models with low amount of available data, (3) interpretability of neural network-based classifiers, (4) ability to generate new peptide sequences, and (5) coverage- based parallel exploration of chemical space. Focused on the category of terapeutic peptides, we addressed the issue of sensivity of highly accurate predictive models (Erjavac et al. 2022) and proposed the sequential properties representation scheme to improve their predictive power (Otovic et al. 2022). This offered us the foundations to employ deep learning models improved by transfer learning for the prediction of poorly researched peptide activities, like catalitic peptides. To gain insight into the decision process of black-box neural network models we employ the Grad-Cam technique, which enables us to pinpoint the properties and residues important for the precition outputs and analyze their behaviour. The capability of these models to generalize knowledge is used to construct de novo peptides of desired property. For this purpose we experimented with generative adversarial networks, variational auto encoders and genetic algorithms guided by machine- learning and analyzed to which extent they are succesfull in producing varying terapeutic peptides. Finally, we developed a methodology based on multi-objective genetic algorithm for the design of maximally diversified random peptide libraries which is able to simultaneously search for multiple libraries and take into account their cross-library and intra-library diversity. We envision that these strategies will maximize the chance of successful identification of active peptides improving the environmetal impact of failed experimental attempts.

Izvorni jezik
Engleski

Znanstvena područja
Kemija, Računarstvo

POVEZANOST RADA