How AI is Solving Science’s Biggest Unsolved Problems
Google DeepMind just solved what scientists called ‘impossible’ — here’s how AI is rewriting the rules of discovery.
The Problem That Stumped Science for Half a Century
For 50 years, the protein folding problem stood as one of biology’s most frustrating mysteries. Scientists knew that a protein’s function depends entirely on its three-dimensional shape, but predicting that shape from its amino acid sequence required years of painstaking laboratory work. Each protein structure could take a team of researchers months or even years to determine experimentally.
Then, in 2020, DeepMind’s AlphaFold changed everything.
The artificial intelligence system didn’t just make incremental progress on the protein folding problem — it essentially solved it. AlphaFold can now predict protein structures with accuracy comparable to experimental methods, but in minutes rather than months. The breakthrough was so significant that Science magazine called it one of the most important scientific advances of the decade.
Understanding the Magnitude of AlphaFold’s Achievement
To appreciate what DeepMind accomplished, you need to understand why protein folding was considered so intractable. A typical protein consists of a chain of hundreds of amino acids that can fold into an astronomical number of possible shapes. The mathematics are dizzying: even a small protein with 100 amino acids could theoretically adopt more configurations than there are atoms in the universe.
Traditional approaches to solving protein structures involved either X-ray crystallography — growing protein crystals and bombarding them with X-rays — or cryo-electron microscopy, which requires freezing proteins and imaging them with powerful electron microscopes. Both methods are expensive, time-consuming, and don’t work for all proteins.
Computational approaches existed before AlphaFold, but they were limited. Earlier algorithms might take weeks to produce a rough approximation of a protein’s structure, and the results were often unreliable. Scientists attempted to crowdsource the problem through initiatives like Folding@home, distributing calculations across millions of home computers, but progress remained frustratingly slow.
AlphaFold took a different approach. Rather than trying to simulate the physical folding process, DeepMind’s team trained a deep learning system on the Protein Data Bank — a database containing structures of more than 170,000 proteins painstakingly determined through decades of experimental work. The AI learned to recognize patterns in how amino acid sequences relate to final structures, discovering subtle relationships that human researchers had missed.
The results were stunning. In the 2020 CASP14 competition — the Olympics of protein structure prediction — AlphaFold achieved median accuracy scores above 90, far exceeding any previous computational method. For many proteins, its predictions were indistinguishable from experimentally determined structures.
The Ripple Effects Across Scientific Research
AlphaFold’s impact extends far beyond academic prestige. Understanding protein structures is fundamental to drug discovery, disease research, and biotechnology. Pharmaceutical companies can now rapidly predict how potential drug molecules might interact with disease-causing proteins. Researchers studying genetic diseases can understand how specific mutations affect protein function. Engineers designing new enzymes for industrial applications can iterate through designs in silico before moving to the lab.
DeepMind made AlphaFold’s predictions freely available to researchers worldwide, releasing structural predictions for over 200 million proteins — essentially all proteins known to science. This represents hundreds of thousands of years’ worth of traditional experimental work, delivered instantly and free of charge.
The scientific community has responded enthusiastically. Publications citing AlphaFold now number in the thousands, spanning fields from medicine to environmental science. Researchers have used AlphaFold predictions to understand antibiotic resistance mechanisms, design new vaccines, study plastic-degrading enzymes, and investigate the molecular basis of cancer.
Decoding the Language of Whales
While DeepMind was revolutionizing molecular biology, researchers at MIT were applying AI to an entirely different “unsolved problem” — understanding animal communication.
Sperm whales communicate through sequences of clicks called codas, producing distinctive patterns that vary between different whale populations. For decades, marine biologists have recorded these vocalizations, suspecting they might represent a sophisticated communication system. But human researchers couldn’t crack the code. The patterns seemed complex, contextual, and unlike human language in fundamental ways.
Enter Project CETI (Cetacean Translation Initiative), an ambitious collaboration between marine biologists, linguists, and AI researchers. The project deployed advanced underwater recording equipment to capture sperm whale communications in unprecedented detail, amassing a database of thousands of coda sequences with contextual information about the whales producing them.
What happened next mirrors the AlphaFold breakthrough: AI found patterns that humans couldn’t see.
By analyzing approximately 9,000 sperm whale codas using machine learning algorithms originally developed for natural language processing, the MIT team discovered what they call a “phonetic alphabet” in whale communication. The AI identified systematic variations in the timing, rhythm, and structure of clicks that function similarly to phonemes in human language — basic units of sound that combine to create meaning.
The system revealed that sperm whale communication is more sophisticated than previously recognized. Whales don’t simply repeat fixed patterns; they appear to combine click elements in rule-governed ways, suggesting a form of combinatorial communication. The AI detected contextual patterns indicating that certain coda types correlate with specific behaviors or social situations.
This discovery was only possible because machine learning algorithms could process vast amounts of acoustic data and identify subtle statistical patterns across thousands of examples — a task that would overwhelm human researchers. The AI didn’t bring preconceptions about what whale language “should” look like; it simply found structures in the data.
Why AI Succeeds Where Humans Struggle
The common thread connecting AlphaFold and Project CETI reveals something profound about scientific discovery in the 21st century: the biggest remaining unsolved problems often involve pattern recognition at scales beyond human cognitive capacity.
Humans are remarkably good at finding patterns in small datasets. We excel at intuitive reasoning, forming hypotheses, and designing experiments. But we struggle with high-dimensional pattern matching across massive datasets. Our working memory is limited, our attention is selective, and we’re prone to confirmation bias — seeing patterns we expect rather than patterns that actually exist.
Modern AI systems, particularly deep learning models, have complementary strengths. They can simultaneously consider millions of data points, identify subtle correlations across hundreds of variables, and discover patterns without preconceptions about what they’re looking for. They don’t get tired, distracted, or biased by existing theories.
This doesn’t mean AI replaces human scientists. Rather, it creates a powerful partnership. Humans still formulate the questions, design the studies, interpret the results, and understand what discoveries mean in broader contexts. But AI can handle the computational heavy lifting — the exhaustive pattern searching that was previously impossible.
Consider why the protein folding problem resisted solution for so long. Human researchers understood the basic physics of protein folding and could predict simple structures. But the relationship between sequence and structure involves complex, non-linear interactions between hundreds of amino acids. No human could simultaneously consider all these relationships. Traditional computers could calculate them but couldn’t learn from examples to improve predictions.
Deep learning bridges this gap. AlphaFold’s neural networks encode millions of parameters learned from existing protein structures, capturing subtle patterns about how amino acids interact in three-dimensional space. This encoded knowledge allows rapid, accurate predictions for new proteins.
Similarly, Project CETI succeeded because machine learning could identify acoustic patterns across thousands of whale vocalizations while accounting for contextual variables like whale identity, social grouping, and behavioral state — a multidimensional pattern matching problem that exceeds human cognitive capacity.
The Future of AI-Accelerated Discovery
These breakthroughs suggest we’re entering a new era of scientific discovery where AI doesn’t just assist research but enables entirely new categories of investigation.
Consider other long-standing unsolved problems where AI might make similar contributions:
Climate modeling and weather prediction: Current models struggle with long-term accuracy because atmospheric systems involve countless interacting variables. AI systems trained on historical weather data and satellite observations are already improving forecast accuracy, potentially enabling better climate predictions and extreme weather warnings.
Drug discovery and development: Beyond protein folding, AI is being applied to predict drug efficacy, identify side effects, optimize molecular structures, and even discover entirely new classes of therapeutic compounds by exploring chemical space that human chemists might never consider.
Materials science: Researchers are using AI to predict properties of novel materials before they’re synthesized, potentially discovering new superconductors, battery materials, or structural compounds. DeepMind has already applied similar approaches to discover new crystal structures.
Genomics and personalized medicine: AI systems are identifying subtle genetic patterns associated with disease risk, drug responses, and treatment outcomes by analyzing genomic data from millions of individuals — population-scale pattern recognition impossible through traditional analysis.
Fundamental physics: Even in theoretical physics, AI is being explored as a tool for discovering underlying equations from experimental data or identifying patterns in phenomena like turbulence that have resisted theoretical understanding.
The key insight is that many remaining scientific mysteries involve systems too complex for reductionist approaches but potentially amenable to pattern recognition from sufficient data. As we generate ever-larger scientific datasets — from genomic sequences to astronomical surveys to particle physics experiments — AI becomes increasingly valuable for extracting meaning from this information deluge.
The Limitations and Challenges
This optimistic picture requires important caveats. AI isn’t a magic solution to all scientific problems, and these technologies face significant limitations.
First, AI is only as good as its training data. AlphaFold succeeds because decades of experimental work created a large database of known protein structures. For problems where high-quality training data doesn’t exist, current AI approaches struggle. This limits AI’s applicability to truly novel problems where we lack examples to learn from.
Second, AI systems typically function as “black boxes” — they produce predictions without explaining their reasoning. When AlphaFold predicts a protein structure, it can’t explain why specific amino acids adopt particular orientations. This lack of interpretability means AI discoveries often require follow-up work to understand the underlying mechanisms.
Third, AI can identify correlations but doesn’t understand causation. In the whale language research, AI detected patterns in vocalizations but can’t tell us what the whales are actually communicating about. That requires additional research and interpretation by human scientists.
Fourth, there’s a risk of over-relying on AI predictions without experimental validation. AlphaFold’s predictions are highly accurate but not perfect. Scientists must still verify computational predictions through experiments, especially when the stakes are high, as in drug development.
Finally, the computational resources required for cutting-edge AI research create equity concerns. Training models like AlphaFold requires massive computing power and expertise concentrated in well-funded institutions and tech companies. There’s a risk that AI-accelerated discovery becomes available primarily to wealthy institutions, potentially exacerbating existing disparities in scientific research.
Conclusion: A Partnership, Not a Replacement
The breakthroughs from DeepMind, MIT, and other research groups herald an exciting new chapter in scientific discovery. But the most important lesson isn’t that AI replaces human researchers — it’s that AI enables a powerful partnership between human creativity and machine pattern recognition.
Humans remain essential for asking the right questions, designing experiments, interpreting results in context, and understanding the broader implications of discoveries. AI excels at the computational pattern matching that often constitutes the bottleneck in solving complex problems.
The protein folding problem wasn’t solved by AI alone — it required decades of experimental work to create training data, human insight to frame the problem appropriately for machine learning, and ongoing human interpretation of AI predictions. Project CETI’s whale language discoveries depend on marine biologists’ field expertise, linguists’ theoretical frameworks, and continued human analysis to understand what the patterns mean.
As we face increasingly complex scientific challenges — from climate change to disease to fundamental questions about the universe — this human-AI partnership will become ever more valuable. The problems that stumped researchers for decades often did so because they exceeded human cognitive limitations, not because they were fundamentally unsolvable. AI provides new cognitive tools to tackle these challenges.
The next decade will likely bring more examples of AI cracking problems once deemed impossible. But these breakthroughs will succeed precisely because they combine machine capabilities with human insight, creating a whole greater than the sum of its parts. That’s the real revolution: not AI replacing human discovery, but AI amplifying human potential to understand our world.
Frequently Asked Questions
Q: What is the protein folding problem that DeepMind solved?
A: The protein folding problem is the challenge of predicting how a chain of amino acids will fold into a three-dimensional protein structure. For 50 years, determining protein shapes required expensive, time-consuming laboratory experiments. DeepMind’s AlphaFold AI system can now predict these structures in minutes with accuracy comparable to experimental methods, essentially solving a problem that stumped scientists for decades.
Q: How did AI help decode whale language?
A: MIT’s Project CETI used machine learning algorithms to analyze approximately 9,000 sperm whale vocalizations, discovering a ‘phonetic alphabet’ in their communication. The AI identified systematic variations in click timing, rhythm, and structure that function like phonemes in human language. This revealed that whale communication is more sophisticated than previously thought, with rule-governed combinations of sound elements rather than just fixed patterns.
Q: Why can AI solve problems that human scientists couldn’t?
A: AI excels at pattern recognition across massive datasets with many variables — tasks that exceed human cognitive capacity. While humans are excellent at forming hypotheses and intuitive reasoning, we struggle with high-dimensional pattern matching across millions of data points. AI systems can simultaneously consider countless variables and identify subtle correlations without preconceptions, making them ideal for problems involving complex pattern recognition in large scientific datasets.
Q: Does AI replace human scientists?
A: No, AI doesn’t replace human scientists but rather creates a powerful partnership. Humans remain essential for formulating questions, designing studies, interpreting results, and understanding broader implications. AI handles computational pattern recognition that was previously impossible. The biggest breakthroughs, like AlphaFold, succeed because they combine human insight and creativity with machine pattern-matching capabilities.
Q: What are the limitations of AI in scientific discovery?
A: AI faces several important limitations: it requires large amounts of high-quality training data; it operates as a ‘black box’ without explaining its reasoning; it identifies correlations but doesn’t understand causation; predictions still require experimental validation; and the massive computational resources needed create equity concerns. These limitations mean AI is a powerful tool but not a magic solution to all scientific problems.
