Ten to twenty seconds instead of months of laboratory work
En los últimos días de julio, DeepMind reveló que su sistema de inteligencia artificial AlphaFold había cartografiado las estructuras tridimensionales de 200 millones de proteínas de un millón de especies, completando en segundos lo que antes exigía meses de laboratorio. Este hito no es solo un triunfo computacional: es una reconfiguración del ritmo al que la humanidad puede comprender la vida misma. La misma herramienta que podría acelerar la cura de enfermedades olvidadas también obliga a la ciencia a preguntarse hasta dónde debe llegar el conocimiento abierto.
- AlphaFold resuelve en 10 a 20 segundos un problema que durante décadas consumió meses de trabajo experimental costoso y lento.
- La base de datos, gratuita y accesible para cualquier científico del mundo, cubre prácticamente todas las proteínas humanas conocidas y las de un millón de especies.
- Investigadores que estudian enfermedades desatendidas como la leishmaniasis y el Chagas ven por primera vez una herramienta que podría nivelar el campo de juego científico global.
- La misma precisión que promete salvar vidas abre un debate urgente: el conocimiento sobre estructuras proteicas podría ser usado para diseñar armas biológicas o toxinas.
- DeepMind consultó a más de treinta expertos en bioseguridad antes de publicar la base de datos, concluyendo que los beneficios superan los riesgos, aunque la inquietud persiste.
A finales de julio, DeepMind anunció que su sistema AlphaFold había cartografiado las estructuras tridimensionales de aproximadamente 200 millones de proteínas provenientes de un millón de especies distintas, abarcando en la práctica el universo completo de proteínas humanas conocidas. Lo que antes requería meses de trabajo en laboratorio ahora ocurre en entre diez y veinte segundos.
Para entender la magnitud del logro, conviene recordar qué son las proteínas: moléculas complejas que ejecutan casi todas las funciones críticas de las células vivas. Se construyen a partir de cadenas de aminoácidos y se pliegan en formas tridimensionales únicas que determinan su función, su interacción con otras moléculas y la manera en que las enfermedades se desarrollan o los fármacos pueden intervenir. Predecir esa forma a partir de la secuencia de aminoácidos —el llamado problema del plegamiento de proteínas— había resistido solución durante décadas debido al número astronómico de configuraciones posibles.
AlphaFold, desarrollado por DeepMind —la empresa de inteligencia artificial con sede en Londres adquirida por Google en 2014—, fue entrenado con miles de estructuras proteicas conocidas hasta aprender los patrones subyacentes con suficiente profundidad como para predecir formas que nunca había visto. En 2020 ganó una competencia internacional de predicción de estructuras con una precisión comparable a la de los métodos experimentales, y la revista Science lo nombró el avance científico de 2021.
La base de datos resultante es abierta y gratuita, y funciona como un motor de búsqueda: se introduce una proteína y en segundos se obtiene su estructura predicha. Para quienes investigan enfermedades desatendidas que afectan principalmente a poblaciones pobres, como la leishmaniasis o el Chagas, esta aceleración podría ser transformadora: los costos bajan, los tiempos se comprimen y los fármacos podrían avanzar más rápido.
Demis Hassabis, director ejecutivo de DeepMind, describió el momento como el inicio de una nueva era de biología digital. Sin embargo, la misma capacidad que permite diseñar medicamentos que salvan vidas podría, en teoría, facilitar la creación de armas biológicas. DeepMind consultó a más de treinta expertos en bioseguridad y ética antes de publicar la base de datos, y concluyó que los beneficios superan los riesgos. Ewen Birney, director del Instituto Europeo de Bioinformática, argumentó que quienes tendrían tanto la intención como la capacidad de weaponizar esa información son una minoría vanishingly pequeña. La base de datos ya está en línea. Lo que ocurra a continuación depende de cómo la comunidad científica decida usarla.
In late July, DeepMind announced that researchers worldwide had used its AlphaFold system to map the three-dimensional structures of roughly 200 million proteins drawn from a million different species—essentially capturing the entire known universe of human proteins. What might sound like an abstract computational milestone carries profound weight for medicine and biology: the same work that once consumed months of painstaking laboratory effort now takes between ten and twenty seconds.
To understand why this matters, you need to know what proteins are and what they do. They are large, complex molecules that perform nearly every critical function in living cells. Built from chains of smaller units called amino acids—of which there are twenty different types—proteins fold into unique three-dimensional shapes after they are synthesized. That shape determines everything: what the protein can do, how it interacts with other molecules, how disease takes hold, how drugs might intervene. For decades, biologists have grappled with what they call the protein-folding problem: given the sequence of amino acids, predict the final three-dimensional form. It is a puzzle that has resisted easy solution because the number of possible configurations is astronomical, and experimental methods to determine actual structures are slow and expensive.
AlphaFold, a product of DeepMind—the London-based artificial intelligence company acquired by Google in 2014—was built to solve exactly this problem. The system uses deep neural networks trained on thousands of known protein structures. It learned the underlying patterns well enough to predict the shapes of proteins it had never seen before. When DeepMind entered a biennial protein-structure prediction competition in late 2020, AlphaFold's performance was nearly as accurate as experimental work but vastly faster. Science magazine named it the advance of 2021, ahead of COVID antivirals and CRISPR gene editing.
The recent announcement represents a quantum leap in scale. Rather than predicting structures one at a time, researchers used AlphaFold to map 200 million proteins across a million species. The resulting database is free and open to any scientist on Earth. It works like a search engine: type in a protein, retrieve its predicted structure in seconds. For researchers working on neglected diseases—leishmaniasis, Chagas disease, conditions that primarily affect poor populations and thus attract minimal funding—this acceleration could be transformative. The cost of research drops. The speed of discovery accelerates. Drugs that might have taken years to develop could move faster through the pipeline.
Demis Hassabis, DeepMind's chief executive, called it the beginning of a new era of digital biology. Eric Topol, a prominent American scientist, declared it a great day for artificial intelligence in the life sciences. Yet the announcement also surfaced a harder question: what happens when you give the world the ability to predict protein structures with such speed and precision? The same knowledge that enables the design of life-saving medicines could theoretically enable the creation of biological weapons or toxins.
DeepMind consulted more than thirty biosecurity and ethics experts before releasing the database publicly. The company concluded that the benefits—accelerated drug discovery, faster development of defenses against biological threats, the simple accumulation of human knowledge—outweighed the risks. Ewen Birney, director of the European Bioinformatics Institute, argued that entities with the intent and capability to weaponize such information would be vanishingly small, and that AlphaFold would ultimately make human biologists more effective at their work, not less. The database is now live, accessible to anyone with an internet connection. What comes next depends on how the scientific community uses it.
Citações Notáveis
Essentially, this covers the entire universe of proteins. We are at the beginning of a new era of digital biology.— Demis Hassabis, CEO of DeepMind
A great day for artificial intelligence in the life sciences.— Eric Topol, American scientist
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that a computer can predict protein shapes faster than a lab can?
Because the shape is everything. You can't design a drug to fight a disease if you don't know what the target protein looks like. Labs used to spend months or years figuring that out. Now it takes seconds. That's not just faster—it changes what's possible.
But couldn't this knowledge be dangerous? Couldn't someone use it to build something harmful?
Yes, that's the real tension. DeepMind asked that question directly. They talked to thirty experts in biosecurity. The conclusion was that the people with the intent to cause harm are already small in number, and they'd face enormous practical barriers even with this tool. The upside—curing diseases that have been neglected because they're poor—seemed to outweigh the downside.
Who benefits most from this?
Researchers in the poorest parts of the world, studying diseases that rich countries ignore. Leishmaniasis, Chagas disease. These kill thousands of people a year but get almost no funding because the victims are poor. Now a scientist in a lab with limited resources can do work that would have been impossible before.
Is this the end of the protein-folding problem?
It's the end of one chapter. We can now predict structures with remarkable speed. But understanding what those structures mean, how to actually use that knowledge to build better drugs—that's still hard. The tool is revolutionary. The work is just beginning.