Why Hasn't AlphaFold Truly Solved the Protein Folding Problem?

AI breakthroughs in biology are often reduced to "AlphaFold solved protein folding," but in a deep conversation, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido offered a strikingly different verdict: the protein folding problem is far from "solved." AlphaFold merely opened a door, and behind it lies a vast uncharted territory.

What AlphaFold Actually Achieved — and What It Missed

AlphaFold's core contribution lies in protein structure prediction: given an amino acid sequence, it can infer the protein's three-dimensional structure with extremely high accuracy. This was once regarded as one of the hardest prediction problems in biology, and the arrival of AlphaFold2 caused a sensation in the CASP assessment.

But Kohli and Candido emphasized that structure prediction is not the same as understanding folding. AlphaFold outputs a static final conformation, whereas protein folding is a dynamic process—how the peptide chain folds in real time within the cell, what the intermediate states are, how misfolding leads to disease. These questions AlphaFold does not answer.

More critically, AlphaFold relies on evolutionary information from known structure databases (multiple sequence alignments). What it learns is a statistical mapping of "what kind of sequence corresponds to what kind of structure," not the physical mechanism of folding. Candido pointed out that this means AlphaFold's capabilities are significantly limited when facing proteins lacking homologous sequences, or when one needs to understand folding dynamics.

CASP (Critical Assessment of Structure Prediction) is the international authoritative blind test competition in the field of protein structure prediction, held every two years. Participants must predict the structures of protein sequences whose structures have not yet been publicly released, and the true structures subsequently solved experimentally serve as the evaluation standard. CASP uses metrics such as GDT-TS to measure the similarity between predictions and true structures. AlphaFold2 achieved a median score of about 90 GDT-TS in CASP14 in 2020, far exceeding the previous level of about 60, and is widely regarded as a watershed event in the field. What made this result sensational is that it meant that for a large number of single-domain proteins, computational prediction accuracy had approached that of experimental methods. But the CASP assessment itself focuses on static structure similarity and does not examine folding pathways, dynamics, or intermediate states—which is precisely the structural reason why the two researchers emphasize what AlphaFold "misses."

Multiple Sequence Alignment (MSA) is a core component of AlphaFold's input. The logic is: homologous proteins across different species have undergone hundreds of millions of years of evolution, and residues that are jointly conserved in sequence often are close to each other in three-dimensional space or serve key functions. By aligning a large number of homologous sequences, co-variation signals between residues can be extracted, and spatial contact relationships can then be inferred. This evolutionary information is essentially the source of the "statistical mapping"—what AlphaFold learns is not physical folding rules, but the correlation between "which sequence patterns correspond to which structural features." Therefore, when the target protein lacks sufficient homologous sequences (such as orphan proteins, artificially designed proteins, or certain viral proteins), MSA information is sparse, and prediction accuracy drops markedly. This dependence on evolutionary databases also explains why AlphaFold struggles to answer the dynamical question of "how the folding process occurs."

The Applicable Boundary of the Bitter Lesson in Biology

The conversation repeatedly returned to the "Bitter Lesson" proposed by Rich Sutton—that reliance on large-scale computation and general learning methods will ultimately prevail over elaborate designs dependent on human domain knowledge. AlphaFold is often seen as a victory of this idea in biology.

Kohli takes a more cautious stance. He believes that the Bitter Lesson has a clear applicable boundary in biology. Proteins are not language, and biological data is far less abundant and cheap than internet text. Obtaining a protein structure requires expensive experiments (such as cryo-electron microscopy or X-ray crystallography), and the cost and speed of data generation constitute a fundamental bottleneck.

Candido added a practical perspective: Biohub's work involves a large amount of wet-lab validation, and AI predictions must form a closed loop with experimental data. Pure scaling cannot bypass the question of "where does the data come from." Both researchers believe that AI breakthroughs in biology are more likely to come from the synergy between computational methods and experimental methods, rather than simply piling on compute.

Cryo-electron microscopy (Cryo-EM) and X-ray crystallography are currently the two mainstream experimental methods for resolving protein three-dimensional structures. X-ray crystallography requires crystallizing the protein, then reconstructing electron density from X-ray diffraction patterns to model atomic coordinates, but many proteins are difficult to crystallize. Cryo-EM rapidly freezes samples into amorphous ice, images them with an electron beam, and performs three-dimensional reconstruction; in recent years it has become mainstream due to a resolution revolution, but it still requires expensive equipment, specialized operation, and large amounts of purified sample. Both types of methods typically require weeks to months to resolve a single structure, and not all proteins are amenable. This high-cost, low-throughput data production model is precisely the "fundamental bottleneck" Kohli and Candido refer to—in sharp contrast to the nearly zero-cost, near-infinite supply of internet text corpora, and it determines that the biological field cannot simply copy the scaling path of large language models.

What Is Still Missing to Build AI That Truly Understands Biology

When asked "what else is needed for AI to truly understand biology," the two gave complementary answers.

Kohli emphasized the importance of multi-scale modeling: biological systems span multiple scales—atoms, molecules, cells, tissues—while current AI models are mostly confined to a single scale. Protein folding is only a problem at one of these scales; to understand cellular behavior and disease mechanisms requires unified cross-scale modeling capability.

Candido pointed out from the perspective of data infrastructure that standardization and scaling of experimental data is an urgent priority. Current biological experimental data is scattered, varies in format, and is difficult to reuse, which severely constrains the training and validation of AI models. Biohub is trying to establish a more systematic mechanism for data generation and sharing.

Both mentioned that interpretability is a key obstacle to the deployment of AI in biomedicine. If a model predicts that a mutation is pathogenic but cannot explain why, it is difficult to adopt at both the clinical and regulatory levels. This is fundamentally different from AlphaFold's positioning as a research tool—a research tool can simply give an answer, but medical decision-making cannot.

Multi-scale modeling in biology refers to establishing a unified computational description across different spatial and temporal scales: from chemical bonds and conformational changes at the atomic scale (picoseconds to nanoseconds), to molecular complexes and pathways (microseconds to seconds), to cellular behavior, tissue morphology, and even organ function (minutes to years). Current AI models are mostly trained for a single scale—AlphaFold handles atomic coordinates, single-cell models handle gene expression, imaging models handle tissue morphology—but disease often originates at one scale and manifests at another. For example, a point mutation changes protein stability at the atomic level, triggers stress pathways at the cellular level, and ultimately manifests as degenerative pathology at the tissue level. If a model cannot pass information between scales, it is difficult to form a holistic understanding of biological systems. This is also why Kohli lists it as the next frontier.

Implications for Practitioners in AI + Biology

The value of this conversation lies in breaking a prevailing narrative inertia: equating AlphaFold's success with "AI has solved biological problems." Kohli and Candido's position is clear—AlphaFold proved the power of deep learning methods in biological structure prediction, but it neither solved the physical understanding of folding nor established a true AI understanding of biological systems.

For practitioners focused on AI for Science, this implies several practical judgments: data acquisition and experimental closed-loop capabilities may determine success or failure more than model architecture; cross-scale modeling is the next frontier; and interpretability is not a nice-to-have but a prerequisite for deployment. Treating AlphaFold as a starting point rather than an endpoint is a more accurate positioning of this technology.

One-line Summary

AlphaFold only opened the door to protein folding; AI's true understanding of biology still faces three major gaps: data, scale, and interpretability.

Paragraph Summary

In a conversation, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido pointed out that AlphaFold's success is often misread as "the protein folding problem has been solved," but what it actually does is structure prediction rather than understanding folding mechanisms—it learns a statistical mapping between sequence and structure, relies on evolutionary information in multiple sequence alignments, and does not involve folding dynamics. Kohli believes the Bitter Lesson has a clear boundary in biology: structure data is costly and slow to obtain, and pure compute expansion cannot bypass the data bottleneck. From a practical perspective, Candido emphasizes that AI predictions must form a closed loop with wet experiments, and that standardization and scaling of experimental data is an urgent priority. Both also point out that multi-scale modeling (from atoms to cells) and interpretability are key gaps for AI to truly understand biology and be deployed in medical scenarios. AlphaFold should be seen as a starting point, not an endpoint.