Biomateriomics is even more interdisciplinary now

Engineers and scientists, as well as artists, have long been inspired by the beauty and functionality of nature’s designs. Japan designed high-speed trains to cut through the air as smoothly as the kingfisher cuts through water, for example, but useful designs can also be found at a microscopic level. The study of biology in combination with materials science is called biomateriomics. An Italian research team sees great potential in the application of generative artificial intelligence to this already interdisciplinary field. They have described this potential, and the associated limitations and challenges, in an open access review article titled “Generative Artificial Intelligence for Advancing Discovery and Design in Biomateriomics,” published May 1 in Intelligent Computing, a Science Partner Journal.

According to the authors, “The importance of biomateriomics in materials science cannot be overstated. It offers a unique lens through which scientists can observe and learn from nature’s ingenious solutions to complex material challenges.” Biomateriomics has biomedical applications such as tissue engineering, regenerative and personalized medicine, drug delivery, and drug discovery. Insights from biomateriomics can also be applied in the development of new types of sustainable energy and sustainable materials. However, exploring the huge design space can be labor-intensive and expensive. Generative artificial intelligence methods are seen as a way to conduct biomateriomics research that would otherwise be impracticable, as these methods are designed for complex multiscale, multimodal, interconnected scenarios.

Seeking to guide engineers, policymakers and other researchers, Raffaele Pugliese and his team cover three main topics in their review, as described below.

The history of generative artificial intelligence in biomateriomics

Prior to 2014, artificial intelligence applications in biomateriomics primarily consisted of traditional machine learning methods used in conjunction with existing datasets to classify materials and predict their properties. From this background, AI tools and methods gradually took on a more creative role:

• Generative adversarial networks and variational autoencoders began to be used to design materials with specific targeted properties.
• Reinforcement learning began to be used for optimizing the search for new and useful biomaterials.
• Artificial intelligence methods began to be combined with physics-based simulations to ensure that generated material designs were constrained within realistic bounds.
• Large language models began to be used to link textual information from multiple disciplines, thus increasing the creativity and efficiency of idea generation.
• Denoising diffusion techniques for image generation began to be used to transform textual input into visual output.

The next wave of innovation could come from quantum computation and simulation.

Bioinspired material design via generative artificial intelligence models

Models of several types are surveyed in the review: generative adversarial networks, diffusion models and large language models, including multi-agent systems.

Generative adversarial networks attempt to create outputs that have the same characteristics as the input dataset. One part of the network creates a candidate output, and another acts as a discriminator, judging whether the candidate is a good enough match. In biomateriomics, generative adversarial networks, and transformer-enhanced generative adversarial networks used with natural language processing have been successfully used to suggest new structures based on leaves, gyroids, and elk antlers, for example.

Diffusion models are trained to remove visual noise from blurred images of a specific target type. After training, a model is given an image consisting entirely of noise, which it uses to create a new image using the learned de-noising process. In biomateriomics, RFdiffusion, Chroma, FoldingDiff and RoseTTAFold have been successfully used to suggest new three-dimensional protein structures.

Large language models are natural language processing tools that read and write text. They can be fine-tuned for specific tasks or contexts, and may include search functionality (i.e., retrieval-augmented generation). In biomateriomics, BioinspiredLLM and LifeGPT have been used to generate scientific hypotheses for possible future research and to make predictions about future states of complex systems, and SciAgents collects information from multiple large language models and an interdisciplinary knowledge graph to help analyze and systematize various stages of the research process.

Challenges and ethical considerations

According to the authors of the review, integrating artificial intelligence tools into the biomateriomics research process requires careful consideration of a number of issues and challenges that call for an ethical-by-design approach to ensure responsible development:

• Obtaining sufficient amounts of high-quality labeled data is often difficult and expensive.
• Explainable artificial intelligence techniques must be developed to improve transparency and interpretability.
• Standardized validation protocols are needed for AI-driven biomaterial designs.
• Frameworks to assess and mitigate environmental impacts associated with AI-driven materials must be established.
• The computing power required to operate artificial intelligence models requires large amounts of electrical power.
• Accessible tools and interfaces to democratize AI-driven biomateriomics need to be developed.

The authors recommend proactively integrating ethical principles and practices directly into the development process so that biomateriomics can develop sustainably.