Artificial Intelligence's Transformative Role in Advancing Biology

Key Insight

Proteins, comprising 75 percent of the dry body weight, are life's building blocks, performing myriad vital tasks from forming tendons to acting as enzymes. Understanding how a protein folds into its specific shape is crucial for determining its function, yet this has historically been an immense challenge. Traditional brute-force computational methods for predicting protein folding were estimated to take longer than the age of the known universe, severely hindering developments in fields like drug discovery. Recognizing this bottleneck, scientists established the biannual Critical Assessment for Structure Prediction (CASP) competition in 1993 to foster progress in solving the protein folding problem, which remained arduous despite steady advancements over decades.

A significant breakthrough occurred at CASP13 in 2018 when DeepMind's AlphaFold, a new entrant with no prior track record, surpassed 98 established teams. AlphaFold successfully predicted 25 out of 43 of the most difficult protein structures in mere hours, a substantial improvement over the second-place team which predicted only 3 structures. This achievement was not based on traditional pharmaceutical expertise or conventional algorithmic methods, but on expertise in machine learning and AI, specifically using deep generative neural networks trained on known protein structures. The impact was so profound that in 2020, Scientific American declared, 'One of the Biggest Problems in Biology Has Finally Been Solved,' leading to the retirement of the CASP competition itself as AlphaFold had effectively cracked a half-century-old grand challenge.

The public release of AlphaFold2 in 2022 further accelerated biological research, with over 1 million researchers accessing the tool within 18 months, addressing questions ranging from antibiotic resistance to the origins of life. Previously, approximately 190000 protein structures were available in databases, representing only 0.1 percent of known proteins; DeepMind uploaded 200 million structures, covering almost all known proteins, transforming a process that once took weeks or months into seconds. This exponential advancement signals a profound convergence, or 'superwave,' where AI is indispensable for processing biology's sheer complexity. AI models, including large language models fine-tuned on biochemical data, are now learning the 'language' of biology and chemistry, generating plausible candidates for new molecules and predicting their properties in simulation. The convergence extends to brain-computer interfaces, exemplified by a paralyzed man spelling words with implanted electrodes in 2019, Neuralink's 2021 pig brain implants, and Cortical Labs teaching neurons in a vat to play Pong, all pointing towards an entangled future where AI and synthetic biology become almost interchangeable, leading to an age of biomachines and synthetic life.