In the fascinating world of medical AI, applications that utilize machine learning to solve problems tend to amaze in unexpected ways. True to form, DeepMind, the reigning world Go champion, chess champion, and eternal victor in popular eSports such as Warcraft, may have cut its teeth in the gaming world to get where it is today, but its latest forays into medical research have now dominated the single most daunting problem-at-large in the life sciences of recent decades: the protein folding problem, first introduced in 1972.
The protein folding problem, or the question of how amino acid sequences can be predicted to fold up into their characteristic atomic structures, is at the heart of drug and vaccine development, diagnostics, and countless facets of the biomedical field that deal with the nature of protein chains. It is also astronomically complex, with the sheer range of possible permutations limiting even the most advanced methodologies a matter of trial-and-error. The premise of the problem rests on the assertion by Nobel laureate Christian Anfinsen that a protein’s folded structure should theoretically be predictable with a knowledge of its specific amino acid sequence – and, implicitly, the power to analyze some 10^300 possible variations for a given sequence. Modeling these structures is essential to understanding biological processes in which proteins function as the mechanical doers of metabolic tasks, but even the most effective methods from an accuracy perspective have been overwhelmed by the sheer scale of the problem. With the power of predictive medical AI that iteratively narrows down the possibilities, though, scientists have an entirely new means of exploration at their disposal that can deliver in hours what might cost a lifetime to uncover in a laboratory.
AlphaFold is groundbreaking not just for medical AI research, but for biomedical research in general. The challenges it can tackle are exactly the type we are encountering so acutely in the 2020 pandemic, and it is difficult even to fathom just how much research it could empower and accelerate if applied widely. When we look at AI trends from a localization perspective, we tend to ask how it can empower existing solutions, and also how it might generate new products and services that require translation. Sometimes, however, it simply changes the entire game, and this appears to be one such instance.
In seeking analogies to AI’s growing role in communications – likewise crucial to the success of industries – AlphaFold seems to suggest that machine learning has far more to offer us than we might expect. By exceeding 90% consistency with previously discovered data on protein structures, the algorithm’s results actually call those benchmarks into question, as the discrepancies could just as likely be attributed to human error. Seen through a localization lens, it is fascinating to imagine a future in which it is not human linguists who evaluate machine translations, but rather machines that evaluate linguists and human communications, identifying patterns or common misunderstandings that tend to go unnoticed. Considering the sheer number of possible language pairs, a truly comprehensive language AI could be developed to assess, say, the consistency of content delivered into multiple languages from a source text. While all of this is speculative, it is worth noting that the traditional paradigm of costly, time-intensive research on protein structures may now be a fixture of the past – a sign of just how great the leaps and bounds of technological advancement are becoming.
CSOFT works closely with technology providers to deliver their products and services across borders and languages through rapid, accurate, and culturally-appropriate multilingual localization solutions. With our worldwide network of linguists and subject matter experts, CSOFT can help leaders in this emerging field translate their products for a broader range of users and enter new markets at the pace of technology today. Learn more at csoftintl.com.