A protein’s function is dictated by its structure, and the structure is set by the amino‑acid sequence. Conventional protein design first fixes a target fold and then uses a machine‑learning model to generate sequences that could adopt that fold. In nature many different sequences can produce the same structure, and a single sequence can adopt multiple conformations depending on flexibility or functional triggers. Hence AI‑driven design must recognise that there are many viable answers, not a single “correct” sequence.
Amy E. Keating and colleagues, in a recent PNAS paper, argue that measuring success by whether a model reproduces the evolution‑selected sequence is suboptimal. To address this they built PottsMPNN in the Department of Biology – a framework that embeds the physical principles governing protein structure and stability. The model captures the sequence‑energy landscape, i.e., the relationship between each amino‑acid identity and overall protein stability, improving both sequence generation and mutation‑stability prediction. Adding PottsMPNN to a design pipeline enables the creation of structurally feasible proteins whose sequences bear no resemblance to any native protein.
Birnbaum’s team highlights the role of noise – deliberately perturbing structures during training – which reduces the tendency to over‑mimic native sequences and expands the diversity of generated sequences. PottsMPNN uses a pairwise distribution to model interactions between amino acids, allowing physical modeling of all 20 possible residues at any pair of positions, a key advantage over other methods in representing the sequence‑energy landscape. They also introduced evolutionarily related sequences during training so the model learns that different sequences can fold into the same structure. Although this still leans on evolutionary information, experiments show that as reliance on native sequences wanes, the model’s structural compatibility and energy predictions improve, even for completely novel proteins.
In the age of AI, the ability to design any protein on demand could dramatically broaden the scope of biological engineering. Birnbaum is optimistic, expecting that task‑specific fine‑tuning will further sharpen predictions of mutation outcomes. Keating concludes that these methods move the field toward designing useful, non‑natural proteins for diverse applications and lay a stronger foundation for future breakthroughs.
Review