Machine learning overcomes human bias in the discovery of self-assembling peptides

Abstract

Peptide materials have a wide array of functions from tissue engineering, surface coatings to catalysis and sensing. This class of biopolymer is composed of a sequence, comprised of 20 naturally occurring amino acids whose arrangement dictate the peptide functionality. While it is highly desirable to tailor the amino acid sequence, a small increase in their sequence length leads to dramatic increase in the possible candidates (e.g., from tripeptide = 20 3 or 8,000 peptides to a pentapeptide = 20 5 or 3.2 M). Traditionally, peptide design is guided by the use of structural propensity tables, hydrophobicity scales, or other desired properties and typically yields <10 peptides per study, barely scraping the surface of the search space. These approaches, driven by human expertise and intuition, are not easily scalable and are riddled with human bias. Here, we introduce a machine learning workflow that combines Monte Carlo tree search and random forest, with molecular dynamics simulations to develop a fully autonomous computational search engine (named, AI-expert) to discover peptide sequences with high potential for self-assembly (as a representative target functionality). We demonstrate the efficacy of the AI-expert to efficiently search large spaces of tripeptides and pentapeptides. Subsequent experiments on the proposed peptide sequences are performed to compare the predictability of the AI-expert with those of human experts. The AI performs on-par or better than human experts and suggests several non-intuitive sequences with high self-assembly propensity, outlining its potential to overcome human bias and accelerate peptide discovery.

Molecular Modeling of Self-Assembling Peptides

“…The sequence space increases dramatically with the length of the peptides as 20 n , with n being the length of the sequence–an intractable problem to identify the best sequences for self-assembly. Thus, design strategies typically use heuristics on known substitution patterns (e.g., among hydrophobic residues) or by using charge complementarity to design new sequences, reducing the number of sequences to study into a manageable number . In recent years, advancements in machine learning for both protein structure prediction, , peptide binding, and sequence optimization have transformed computational biology.…”

Biomimetic mineralization based on self-assembling peptides

“…Machine learningbased techniques can hopefully be used to predict peptide sequences that can be assembled to form controlled morphologies and specific functions, thereby controlling their mineralization properties. [325][326][327] This will help in shortening the time of peptide design, reducing the work of experimental screening and optimization, improving the design success rate, and making it possible to design sequences with a better structure and function.…”

“…324 In addition, machine learning methods can be used to design peptide sequences with high self-assembly propensity, which can self-assemble into hydrogels. 325 Based on the above research studies, we believe that machine learning strategies for designing peptide sequences during mineralization should be developed, which is promising, although not yet directly relevant. Machine learning-based techniques can hopefully be used to predict peptide sequences that can be assembled to form controlled morphologies and specific functions, thereby controlling their mineralization properties.…”

Inverse design of viral infectivity-enhancing peptide fibrils from continuous protein-vector embeddings

“…Recently, an ML approach, in combination with Monte Carlo tree search and molecular dynamics simulations, was reported for predicting unexpected de novo -sheet rich self-assembling peptides. 22 While all these reports successfully identified either bioactive or self-assembling peptides separately, they did not predict bioactive self-assembling peptides.…”