Self-assembly is a process capitalized by nature for converting chemically simple building blocks into hierarchically ordered structures and materials to function in living systems cooperatively.¹ Among all the building blocks, proteins are the most commonly utilized and form diverse structures ranging from oligomers and nanospheres to tubes and hierarchical architectures.² These structures with unique physical properties play crucial roles in vast biological functions, for instance, structural support, cargo transport, and microbial defence.³ The formation of these structures is predominantly driven and governed by non-covalent interactions, conferring the dynamism and flexibility of these structures. By studying the fundamental principles that control the self-assembly processes of native proteins, functional materials can be replicated in a bottom-up approach, whereby synthetic building blocks are designed for assembling into desired architectures with specific properties.⁴ In this context, biomimetic peptides with comparatively shorter sequences than proteins emerge as ideal building blocks.⁵ They provide a versatile approach for mimicking complex structures and materials with structural and functional diversity in a more tractable but less expensive approach.

The self-assembly behaviours of peptides depend strongly on their amino acid sequences.⁶ To date, the design of self-assembling peptides still relies on the examination of natural amino acid sequences, professional expertise in the peptide field, or laboratory discoveries by serendipity. However, these inefficient approaches fail to meet the growing demands for functional peptide materials since the design range of self-assembling peptides can be extremely broad. The amount of possible amino acid combinations for a peptide can reach up to 20ⁿ, where 20 is the number of commonly available amino acids and n is the amino acid number in the peptide.⁷ For instance, tripeptides (n = 3) containing 8000 combinations are intractable for any rigorous experimental study. Even though a brute-force computational search based on coarse-gained molecular dynamic (MD) simulations has been proposed by Ulijn’s team to overcome the search bias and has successfully identified several self-assembling tripeptides.⁸ Due to the high computational costs, it is impossible to extend it to longer sequence lengths (n > 3). Accurate prediction of the assembling processes and discovery of the assembling peptides remain challenging.

Artificial intelligence (AI) is an emerging inter-discipline that integrates computer science, mathematics, and psychology among others to emulate human cognitive functions for tasks requiring intelligence, such as language understanding and decision making.⁹ It holds great potential to navigate through the vast search space among amino acid combinations and present a subset displaying the most promising possibilities. Recently, writing in Nature Chemistry, Sankaranarayanan’s team introduced an “AI expert” that combined the Monte Carlo tree search (MCTS) algorithm with coarse-gained MD simulations for identifying self-assembling peptides that exhibit high aggregation propensities in aqueous solution.¹⁰

Performing in an autonomous way, the AI expert first utilized MCTS to generate peptide sequences before utilising MD simulations to estimate the aggregation propensity of the generated sequences and provide feedback to improve the quality of MCTS and guide future searches (Figure 1). Notably, an innovative concept of uniqueness function within the MCTS objective function, together with a random forest based surrogate model were included to boost the performance of the MCST algorithm. Compared to MCTS without the random forest model, random search, and brute-force search, the AI expert enabled by MCTS and random forest scheme was demonstrated to be the most efficient in identifying tripeptide with the highest-scoring.

Having validated the efficiency of AI experts, Sankaranarayanan and colleagues went on using it to discover self-assembling pentapeptides. Among the 3.2 million (20⁵) pentapeptide permutations, the AI expert first evaluated approximately 6600 cases based on MD simulations and then selected the top 100 candidates for further MD simulations using more rigorous parameters and longer timescales to improve the estimates of aggregation propensity. Nine candidates with top scores were filtered out and six of them were observed to aggregate according to experimental dynamic light scattering and atomic force microscope measurements. Comparatively, six out of eleven pentapeptides proposed by human experts were found to aggregate. By analysing the sequence similarity of identified peptides, the AI expert was found could not only recover known sequences similar to the human experts but also discover unknown sequences that deviate notably from the existing ones, indicating its potential to overcome the bias from humans and accelerate the discovery of peptides.

Integrated with coarse-grained MD, machine learning, and experiments, the AI expert was demonstrated to be an efficient “human-in-the-loop” framework for the discovery of self-assembling peptides. Future efforts were anticipated to be made to connect the developed AI expert to a robotic platform that can synthesise and characterise innovative peptides. As such experimental feedback could be digested by AI expert directly, supporting the search to progress in an iterative manner. Additional information obtained from the simulations, for instance, the aspect ratio and number of the peptides, their morphologies, and their moments of inertia were expected to be included to improve the MCTS scoring function. Similar AI strategies could also be developed for screening and discovering more functional peptides and peptide assemblies, providing great potential for further innovations in peptide-based novel therapeutics and functional materials.

Author contributions

YJS conceived and wrote the draft, HGH reviewed and edited the draft. Both authors approved the final version of the manuscript.

Financial support

The work was supported by the Shanghai Pujiang Program (No. 21PJ1404100).

Acknowledgement

None.

Conflicts of interest statement

The authors declare no conflicts of interest.

Open access statement

This is an open access journal, and articles are distributed under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License, which allows others to remix, tweak, and build upon the work noncommercially, as long as appropriate credit is given and the new creations are licensed under the identical terms.