Researchers from Russia’s Artificial Intelligence Research Institute (AIRI) and Germany’s Constructor University have introduced DiMA, an innovative AI model for generating artificial proteins essential for new drug development. According to a press release from AIRI, the new model is 100 times more compact than existing analogues while outperforming them in efficiency.
“DiMA is capable of directly generating protein sequences that have never been found in nature but precisely match researcher-specified criteria, such as a particular 3D structure or functional properties,” said Pavel Strashnov, Lead Researcher at the Protein Design Group of the AI-based Drug Discovery Center at AIRI.
Various types of language models have previously been used to create artificial proteins. Autoregressive models function similarly to ChatGPT, generating protein sequences letter-by-letter. Discrete models generate the entire sequence, or “word,” at once. However, both approaches require large model sizes and vast amounts of training data.
DiMA employs a fundamentally new approach based on continuous Gaussian diffusion, a technique actively used in image-generating neural networks that gradually removes “noise” from a created picture. The model was first trained to generate a diverse range of proteins that are always biologically viable and do not literally replicate natural ones. The task was then “narrowed down” by fine-tuning the model to generate proteins with specific criteria. “This approach not only expands the understanding of theoretically possible protein configurations but also solves specific applied problems in biotechnology and medicine,” the press release notes.
By 2030, the global AI in healthcare market is projected to grow 6.5-fold – from $22 billion to $130-160 billion. A similar surge is expected in Russia: from 12 billion rubles in 2024 to 78 billion by 2030.