Why It Took Fifty Years: Protein Structure Prediction and Its Long Trail of Failures

A review published in 1990 left behind the report card of that era in numbers. A protein is a chain of small parts linked in a long line, and this chain winds into coils like springs in places or becomes strands spread out like folding screens. In the problem of guessing for each spot where a single part is placed—that is, each position in the chain—whether it is a coil, a strand, or neither of the three, the best methods at the time got 63 to 65% right. Yet when a similar protein was already present in the database, the accuracy rose to about 90%. The conclusion of the review was this: predicting the three-dimensional shape of an entire protein is practical only when the structure of a similar protein is already known (PMID 2126456).

The author of this review is the person who made one of the two representative methods the review compared, and the 90% figure is also the score of the method he co-created. He wrote his own report card.

The parts that make up a protein are about twenty kinds of amino acids. The order in which amino acids are placed in the chain is called the amino acid sequence. The chain clumps together by itself inside the cell to become a distinct three-dimensional mass. The event where a single-line chain clumps into a determined three-dimensional shape like this is folding, and the shape that comes out after folding is called the structure of that protein. Structure prediction is the task of guessing this shape by looking only at the sequence.

The shape that comes out after folding determines what that protein does. A digestive enzyme grabs only a specific molecule because its shape fits that molecule precisely, and if the shape is distorted, it cannot grab it. This principle that shape determines function is called "form follows function."

Christian Anfinsen’s thermodynamic hypothesis answers where the shape comes from. This hypothesis has been considered the central pillar of protein science for half a century. To quote Anfinsen directly, the three-dimensional structure of a native protein in its normal physiological milieu is the structure with the lowest Gibbs free energy for the whole system. Therefore, the native structure is determined by the totality of interatomic interactions and, hence, by the amino acid sequence (PMID 40750093).

A native protein is a protein in its original shape that does its job inside the cell. Free energy is a physics scale that measures how "comfortable" a certain state is, and the lower it is, the more comfortable it is. This means the chain rolls down into the most comfortable shape among all the shapes it can go to, and which shape is most comfortable is determined by the sequence.

If so, it should be possible to calculate the shape by looking only at the sequence. This is why the conclusion of the earlier review is strange. The answer is inside the sequence, yet the calculation to pull out that answer did not work for nearly half a century.

A major leap occurred in 2020, and that story is covered by "From Sequence to Shape: How AlphaFold Cracked a 50-Year Problem". This article traces the path before it. We look at where attempts hit walls and what those failures left behind that made the final leap possible.

The Answer Is in the Sequence

Even knowing the principle, the calculation could not be done. The simplest method is to scan through every shape a protein can take one by one and pick the most comfortable one. However, the number of possible shapes is so enormous that doing this would take billions of years. Real proteins find their own shapes in a biologically reasonable amount of time (PMID 36659994). What kind of shortcut does nature use?

Physics answered this question relatively early on. When you assign an energy score to each shape and lay them out like the elevation of land, a protein is like a ball rolling across that topography. If the ground is flat, the ball wanders aimlessly, but if the ground slopes to one side, no matter how many places there are to go, it quickly reaches the lower side. A 1997 review summarizes this shift in thinking. The old picture of proteins searching shapes haphazardly was discarded. The view that replaced it is that much of the energy landscape is tilted toward the native state, and this slope can dominate the folding. Then, because there are multiple paths to the native state, there is no need to assume one specific route (PMID 9269572). Today, this theory explains the range of folding times by chain length and even predicts the maximum size of proteins that can fold (PMID 36659994).

However, this answer did not provide a prediction algorithm. Knowing that nature folds via a shortcut and reproducing that shortcut on a computer are completely different things. The same 1997 review also noted this point. The new optimism had to be complemented by more detailed research to uncover the structural and energetic factors creating that slope (PMID 9269572).

Anfinsen's dogma left two questions behind. Why do proteins fold, and how does the sequence determine its own folding? The former question had a plausible answer called the thermodynamic hypothesis, but the latter question remains an unsolved problem to this day (PMID 37165178).

Attempts to bypass this wall branched into three paths: copying structures that were already solved, calculating from the bottom up using physical laws, and reading the traces left by evolution. These three paths did not arrive in a neat sequence; they overlapped in time and borrowed parts from one another.

The First Path — Copying What Is Already Solved

The first era is summarized by the conclusion of the 1990 review mentioned earlier: if you have a similar answer sheet, you can solve it. Cousin proteins that split from the same ancestor resemble each other in both sequence and shape. This kind of family relationship is called homology. Homology modeling is a method where a cousin protein whose structure is already solved is used as a template, the new sequence is placed on top of it, and the misaligned parts are fixed. It is close to copying someone else's answer sheet and changing only the name, and because it relies on homology, it is also called homology-based modeling.

Copying was powerful. Even 20 years later, in 2011, homology modeling remained "the most accurate computational structure prediction method". About 65,000 structures were piled up in the public warehouse that collected experimentally solved structures, and automated prediction programs churned out about 1.9 million models for sequences whose structures were unknown. Drug design targeting proteins of unknown structure was also opened up thanks to this method (PMID 21521153).

However, there were two ceilings.

The first ceiling is accuracy. When a new sequence and a template are placed side by side, the ratio of positions where the same amino acid is placed is called sequence identity. As of 2006, homology modeling with a sequence identity of over 30% was already close to the natural limit obtainable from the template. To go beyond that, a technique to pull the model closer to the native structure and refine it (model refinement) was needed. The protein trying to be predicted is called the target, and the area where the biggest progress was observed in recent years was actually the difficult targets. There, choosing the right template and accurately aligning the sequences remained the main problems (PMID 16510277).

The second ceiling is more fundamental. If there is no similar answer sheet, it is difficult to produce a usable model. Anfinsen's dogma was that "the sequence is all you need," but the method that actually worked was "you must have a similar answer sheet." This gap was the remaining homework.

An Invention, Not an Algorithm — The Decision to Wear Blindfolds

The next turning point came outside calculation methods, in the testing system.

If you tweak your method after seeing the correct answer and say "our method fits well," you prove nothing. Yet papers are mostly written after the correct answers are made public.

The first CASP was held in 1994. CASP is a blind structure prediction test held once every two years. The name is an English acronym meaning "Critical Assessment of Techniques for Protein Structure Prediction." The organizers provide only the sequences of proteins whose structures have already been solved but not yet released, and participants submit predictions without knowing the answers (PMID 31116374). Afterward, when the structures are revealed, independent evaluators grade them. A blind test and independent evaluation—these two are the core components of CASP (PMID 34533838). The editions are numbered sequentially. CASP does not make predictions itself; it only grades them. That grading became the baseline distinguishing what counts as progress from what is an illusion ever since.

Once the system was established, tests targeting its blind spots followed. Among CASP participants, some teams submitted predictions with human intervention applied to program outputs, and these human-assisted predictions outperformed automated servers. An automated server is an internet program that runs predictions without human intervention once a sequence is entered. The first CAFASP, published in 1999, borrowed targets from CASP3 to evaluate fully automated servers separately. The goal was to reveal the performance of the programs themselves rather than human-assisted scores. Automated servers frequently found distant similarities that pairwise sequence comparison methods missed. However, this test was hosted by a separate group, not the CASP organizers, and was not blinded since it was conducted after the structures were revealed. The authors themselves wrote that the rules and procedures differed, making it incomparable to CASP3. Furthermore, this evaluation was for "users of our methods," and some of the evaluators had created the programs being evaluated (PMID 10526371).

The material forming the backbone of this timeline is the summary paper of each CASP edition. Those summary papers are written directly by the people who run CASP. The same names, including John Moult, continue as authors from his solo comprehensive evaluation in 1999 to the CASP14 summary. Therefore, this article traces both the places where the summaries recorded success and where they recorded failure, and separately confirms the 2020 leap using a paper by an independent evaluation team rather than the organizers.

The records of CASP document both success and stagnation.

What the Test Bench Recorded — Ten Rounds

In 1999, an evaluation combining the CASP results up to that point was published. There were several improvements on the homology modeling side. As comparative methods that reference multiple relative sequences at once (sequence profile methods) improved, they also had a major impact on figuring out which known folding shape a new sequence belongs to. Model refinement remained a serious challenge. The side of building from scratch without templates was at the level of getting fragments of 40 to 60 residues roughly correct, and it was said that 「still a long way to go」 for general problems to be solved. Still, the final sentence is an optimism that the field is maturing into a practical technology that produces useful models for many sequences (PMID 10600698).

The 2007 CASP7 summary reports four advances. They are better accuracy than when using only a single best-matching template, the pursuit of automated servers, the emergence of methods to gauge how correct a model might be, and a rapid increase in practical applications (PMID 17918729). The 2009 CASP8 summary also records achievements first. When model refinement was evaluated blindly for the first time, substantial improvement was possible depending on the conditions, and there were advances in methods for judging accuracy and in regions absent from the main template. Then it drives the point home. There is a fact that must be placed against these advances: viewed by existing CASP metrics, no progress in model quality compared to CASP7 is detected for either template-based or template-free prediction (PMID 19774620).

The 2011 CASP9 summary notes that the pace of progress is 「more gradual compared to the early CASPs」 while also lifting four advances. They are accuracy in some target classes, the ability to pick the most accurate model among several, short 「new fold」 models resembling no known structure, and performance higher than expected revealed by new analysis in regions not derived from the most obvious templates (PMID 21997831).

The CASP10 summary published in 2014 looks back that ten CASPs covered nearly 20 years, during which there was tremendous progress in methods and model accuracy. A notable event in this round was that models refined using a method that calculates how atoms move over time (molecular dynamics) consistently improved for the first time.

When a protein folds, two positions far apart in the chain often sit right next to each other in space. This close proximity of two positions within a folded protein is called a contact. The ability to use sparse contact information—obtained through experiments for just a few locations—in modeling was tested for the first time in this round, and the results were encouraging. And there is one decisive sentence. New methods for predicting contacts from sequences hold considerable promise but have not yet made an impact in the CASP test (PMID 24344053).

The summary notes that the nature of targets is also changing. Structures with irregular shapes have increased, and structures where multiple domains (units that fold independently within a protein) or subunits (individual chains in a multi-chain protein) gather have also increased. Typical cases of already known folds have decreased. Taking this into account, the overall accuracy of models is steadily improving, and the authors wrote that improvements in regions without templates are particularly driving that.

While surrounding technologies continued to improve while the test became harder, the pace slowed down and targets without any templates still failed to catch up with homology-based models. Contact prediction, the key to opening that wall, was already out there, but had not yet made any mark on the test bench.

The Second Path — From the Ground Up with Physics

Where copying hit a wall, a second attempt emerged. It is to not use an answer key at all. Physics-based calculation is a method where you write down the equations for the forces acting between atoms and use a computer to find the shape where that equation gives the lowest score. This equation that scores each shape is called an energy function, and the lower the score, the more comfortable the shape. It is an approach that translates Anfinsen's hypothesis literally into an algorithm.

The leading contender was a program called Rosetta. However, Rosetta was not a pure physical simulation, and a method published in 2000 illustrates that compromise. First, short fragments are picked from proteins whose structures are already known. The criteria for picking them are whether their sequences are similar and whether they match signals obtained from experiments. Next, the fragments are assembled to find the shape with the lowest energy function. This energy function gives good scores to shapes where parts that dislike water are buried inside, shapes where chain strands are paired up, and shapes that satisfy experimental constraints well. Experimental constraints are clues that experiments reveal, such as "these two atoms are close to each other."

The grades were good. NMR is an experimental method that measures proteins in solution using a magnetic field, and an X-ray crystal structure is a structure obtained by solidifying a protein into a crystal and exposing it to X-rays. A computational model with a small amount of experimental clues—about one experimental constraint per site—was in some cases closer to the X-ray crystal structure than the published NMR experimental structure of the same protein (PMID 11200525). Such cases already existed in 2000. However, that was when experimental data was included alongside.

The grades when given only the sequence were left in numbers by a benchmark in 2006. A benchmark is a comparative test that pits various methods against the same problems. Out of 38 test proteins, an energy function called SimFold got 12 right within 6.5 angstroms. That is about one in three. An angstrom (Å) is a unit of length roughly the size of a single atom. This score was equal to the score when the scratch-built version of Rosetta, publicly available at the time, was run with default settings. When the candidates produced by the two methods were combined and judged together, the number increased to 16, four more than when using either side alone (PMID 16294329). And the criterion of 6.5 angstroms itself is far from an accuracy that hits individual atoms.

Why did it get stuck? The path of exhaustively scanning all shapes remained closed (PMID 36659994), and why the force-equation-based approach failed at accurate prediction became a review topic in its own right (PMID 37165178). Nature rolls down a sloped floor, but the computer was not given the exact shape of that floor.

The Third Path — The Traces Left by Evolution

The third path came from an entirely different place, neither physics nor copying, but from the records of evolution.

Consider two positions touching each other within a protein. If one amino acid changes to a larger one through a mutation, the contact is disrupted and the protein is broken. However, if a change where the opposite side becomes correspondingly smaller happens together, the shape is maintained. Therefore, over long periods of evolution, pairs of positions that touch in three dimensions leave behind traces of having changed together. This phenomenon where such pairs of touching positions have changed together as they evolved is coevolution.

A group of related proteins split from the same ancestor is called a family. A table where a bunch of sequences from a family are collected, aligned position by position, and stood side by side from top to bottom is a multiple sequence alignment. By counting which positions changed together in this table, you can reconstruct which areas touch each other without looking at the 3D shape. This result, which indicates whether each pair of positions touches or not, is called a contact map.

The idea itself is old. The reason it didn't work for a long time was indirect correlation. If A touches B and B touches C, then A and C, which are not touching, appear to change together. Direct coupling analysis is a method that separates pairs that truly touched and changed together from pairs that appear to have changed together through an intermediate bridge. In 2011, this method was efficiently implemented and applied to many protein domains, correctly predicting a significant number of contacts using only sequence information. For most of the domains investigated, the overall outline of the contact map was also reproduced. The same paper specified that this method presupposes a sufficiently large number of homologous sequences (related sequences from the same family), and thanks to advances in genome sequencing, such data is being secured rapidly (PMID 22106262).

In this final condition lies the core of the era. To read coevolutionary signals, statistics are needed, and statistics need many sequences. Even after the method to separate indirect correlations emerged, for this method to work properly, the data had to follow.

The condition was soon fixed into numbers. In 2013, a method came out that used structural context information along with coevolutionary information to predict contacts more accurately with fewer homologous sequences. The conditions presented by this study are two. First, even after counting sequences that are more than 90% identical to each other as one, if the number of aligned sequences exceeds five times the protein length (number of positions), the prediction is likely to be accurate. Second, for that prediction to be useful for structure calculation, the aligned sequences must be closer to the target than the closest relative whose structure is known. At the time, there were 422 families satisfying this condition in a protein family database. The same paper also details a catch: the larger the family, the more robust the signal, but it also increases the likelihood that a plausible model can be made using only the relative's structure, rendering contact information unnecessary (PMID 24009338). The places where coevolution works well and the places where it is strictly necessary were misaligned.

Even so, proof began to emerge in blind tests. In CASP11, while Rosetta found and refined shapes, when contacts predicted by coevolution were set as conditions to be respected, model accuracy generally improved. The effect was particularly clear when the number of sequences in the family exceeded three times the protein length. The highlight was a large and complex target called T0806. It was a protein with no relatives of known structure at all, and an unprecedented accuracy emerged where the average deviation from the experimentally solved structure across 223 positions was less than 3.0 angstroms. However, humans lent a hand to this prediction, and a procedure that increased the amount of searching was used. The authors wrote that this result clearly shows that contacts obtained through coevolution in a blind prediction situation significantly increase the accuracy of template-free modeling (PMID 26677056). The official summary of the same round of CASP also acknowledged that thanks to the new method, several dazzling template-free models were produced. At the same time, right within that sentence, they also noted that models based on sequence homology to experimental structures remain the most accurate (PMID 27171127).

There were also limitations. Pairs of positions that change together are not necessarily close to each other. Analyzing 4,000 families, 25% of directly coevolving pairs were further than 5 angstroms apart, and 3% were further than 15 angstroms apart. However, the same paper largely explains these exceptions. 91% of the exceptions in the 5 to 15 angstrom range were actually touching in other structures of the same family. The authors' overall assessment is that, excluding repetitive structure proteins, directly coevolving pairs of positions are spatially close in at least one biologically meaningful conformation within the family (PMID 28784799).

A more painful critique came out in 2018. The limits of prediction were most prominent under strict criteria where two positions must be very close together to count as a contact, and geometrically tight contacts of this kind are the ones with the largest amount of information for determining structure. However, this paper also offered a prescription. When an index measuring whether those position pairs are likely to actually touch was used alongside the coevolution score, performance improved considerably, and that improvement was especially large under strict criteria (PMID 29953468).

Amplifiers — Machines That Grow Faint Signals

In the mid-2010s, two currents met. The noisy contact signals produced by co-evolution, and a learning machine skilled at extracting patterns from noise. Here, the learning machine refers to deep learning, that is, multi-layered computing networks that learn patterns on their own from many examples. That lineage is traced by "From Backpropagation to AlphaFold: What Deep Learning Stood On".

Deep learning had entered CASP a few years prior. According to a report by a team that participated in CASP13, deep learning made its debut at CASP10 in 2012. And after CASP11 in 2014, predicting the distance relationships between positions—initially guessing whether two positions were in contact—became the core direction that pushed the field forward (PMID 30985027). It was CASP13 in 2018 that changed the game by predicting not just whether they touched, but even how far apart they were (PMID 31589781). Although this is a paper where the team summarized their own results, this timeline does not conflict with the CASP summary papers.

An analysis by a team that participated in CASP12 breaks down with numbers how much each ingredient contributed. Scoring is done by choosing the most confident predictions for pairs of positions far apart in the chain, up to one-fifth of the protein's length, and looking at the proportion of those that actually make contact (precision). Measured across all targets in CASP12 this way, the baseline using only clues used for a long time had an average precision of 28.4%. Adding clues obtained from co-evolution raised it to 41.6%, and integrating all clues with machine learning pushed it to 56.3%. And the precision of contact prediction showed a correlation of 0.66 with the logarithm of the number of effective sequences included in the alignment. This means that as more sequences gathered in the alignment, the grades rose together (PMID 29047157). It was confirmed with numbers that the amount of data moves together with performance.

The CASP summary papers also confirmed this change. At CASP12 in 2016, new contact prediction methods doubled contact accuracy, and the model accuracy of proteins with no templates improved dramatically (PMID 29082672). Two years later at CASP13, deep learning predicting distances once again dramatically lifted the accuracy of three-dimensional structures. Regarding the problem of guessing how a single-chain protein (monomeric protein) coils up as a whole (folding topology), the summary paper writes this. With the caveat that there are enough known sequences in the protein family, the new methods effectively solve the long-standing problem of predicting the folding topologies of monomeric proteins. The number of sequences needed for alignment itself also decreased considerably (PMID 31589781).

By 2018, the ingredients for the answer had already gathered.

And 2020 — From Here, a Different Article

CASP13 featured a new entrant. The industry lab DeepMind entered for the first time, and their AlphaFold won first place in the novel fold prediction category (first place in the same round's template-based category went to the Zhang group). A short commentary summarizing that time reveals the lineage of this achievement. DeepMind's approach was built on two ideas developed by the academic world over the past decade. One is co-evolution analysis, which translates traces of correlated mutation in sequences into physical contacts in structure, and the other is reading those patterns with a deep learning neural network to turn them into contact maps (PMID 31116374). This commentary is a short piece written by one person right after CASP13 ended.

And in 2020, a leap occurred at CASP14 (PMID 37920879). The CASP14 summary noted that one research group's deep learning method consistently produced accuracy rivaling experimental structures, and in that sense, this result signified the solution to the classical protein folding problem, at least for single proteins. And immediately, it added that other research groups also significantly improved their performance (PMID 34533838).

What structure that method used to pull this off is covered by "From Sequence to Shape: How AlphaFold Cracked a 50-Year Problem". The story of moving from solved predictions to drawing and making desired proteins belongs to "From Prediction to Design — After AlphaFold, the Science of Drawing Proteins and Building Them". Here, just one thing is added.

The independent assessors who graded CASP14 wrote this. With the application of deep learning, the "high-accuracy" category expanded to all targets, and alongside that massive leap, CASP14's second-place method outperformed CASP13's first-place method. The assessors wrote that this demonstrates the role of community-based benchmarking where multiple teams compete on the same test (PMID 34218458). The leap belonged to one team, but the rise belonged to the entire community.

What Half a Century Left Behind

The phrase "the protein folding problem has been solved" is convenient, but inaccurate. That declaration of resolution came with the caveat "at least regarding single proteins" (PMID 34533838).

The biggest advancement in CASP15 in 2022 was that the computational accuracy of complexes rose significantly. A complex is a clump formed by multiple proteins sticking together. However, even that accuracy fell short of the single protein level (PMID 37920879). Even for single proteins, running the standard AlphaFold2 protocol with default settings yielded top-quality results for only about two-thirds of targets, while the remainder required generating many more candidates (PMID 37920879). The complex evaluations of CASP16, the most recent round, are more sobering. Protein complex structure prediction remains an unsolved challenge, and even the leading groups predicted just slightly over half of the targets with high accuracy. However, the same evaluation also notes that things generally improved compared to the previous round and that complex prediction is steadily making progress (PMID 41170922).

Neither of Anfinsen's two questions has been answered yet. How the sequence determines self-folding remains an unsolved problem (PMID 37165178). We have come to possess a machine that guesses the answer very well, but explaining the principle of why that answer is the answer is a different matter.

The ingredients for the answer had been scattered for a long time. Anfinsen's dogma, the physics answer that nature folds through shortcuts, the ceiling of copying, the place where going bottom-up with physics hit a wall, the statistical traces left by evolution, sequences that exploded through genome decoding, and learning machines grown in other fields. The leap of 2020 happened when these gathered in one place. The genealogy of those ingredients was already pointed out by a commentary reviewing CASP13 two years before (PMID 31116374).

What judged that the ingredients had gathered was a test. Anyone can say their method has gotten better. This field decided to take that judgment out of its own hands in 1994. Without that decision, the "no progress detected" of CASP8 (PMID 19774620), the double improvement of CASP12 (PMID 29082672), and the leap of 2020 (PMID 34533838) would all have simply remained individual claims. What will likely outlast any algorithm from this half-century is this evaluation method itself. Only the first chapter of the half-century problem has closed, and the test is still held every two years.

Source Papers

The nature of each paper is noted alongside. Figures that appear verbatim in the text are not repeated. Abstract comparison and grade tables are in review-notes.md.

Starting Point — Principles and the First Wall

• PMID 40750093 (2025) — Review. Concludes, even after examining amyloid fibrils, that in vivo structures are determined by sequence and physiological environment.
• PMID 36659994 (2022) — Review. The impossibility of exhaustive search, folding time scales, and predictions of the maximum size that can fold.
• PMID 9269572 (1997) — Review (Martin Karplus). The discarding of the random search picture, bias toward the native state, and a clue that optimism must be supplemented.
• PMID 37165178 (2023) — Single-author perspective paper. Anfinsen's two questions and an examination of the failure of force field-based approaches, suggestions on how to redefine those approaches, and an estimation of the causes of recent numerical method successes.

The First Path — Copying and Matching

• PMID 2126456 (1990) — Review (J. Garnier). Three-state prediction of helix, strand, and coil with a standard deviation of 6 to 8% per protein. ⚠️ The author is one of the authors of the comparison algorithms, and 90% of the performance belongs to the method he co-created.
• PMID 21521153 (2011) — Review. The status and scale of comparative modeling, and drug design for targets of unknown structure.
• PMID 16510277 (2006) — Review. Sequence identity above 30% approaches natural limits. The greatest progress is observed in difficult targets instead, where template selection and alignment accuracy are the main issues.

The Invention — Blind Tests

• PMID 31116374 (2019) — Letter-format commentary (single author, written right after CASP13). Points out the 2-year blindness cycle, DeepMind's first entry, and a lineage built upon two ideas from academia.
• PMID 34533838 (2021) — CASP organizers' summary. The core composition of blindness and independent evaluation, solved "at least as far as single proteins are concerned," with significant improvements from other groups as well.
• PMID 10526371 (1999) — Evaluation paper by a separate group, not CASP organizers. "Human + machine" is superior to automatic, and automated servers frequently find distant similarities that pairwise comparisons miss. ⚠️ The first round was not blind, and the organizers themselves wrote that goals, rules, and procedures make it incomparable with CASP3, while developers of evaluated servers were among the evaluators.

Records of the Testbed (All are summaries written by CASP organizers)

• PMID 10600698 (1999) — CASP founder (J. Moult) single-author review. Sequence profiles have a huge impact on fold recognition, and alignment quality limits model utility. The final sentence points to maturation into a practical technology.
• PMID 17918729 (2007, CASP7) — Four advances: accuracy exceeding single optimal templates, the catch-up of automated servers, model quality estimation, and the rapid increase in practical applications.
• PMID 19774620 (2009, CASP8) — Conditional improvements in the first blind evaluation of refinement, but no progress detected by existing metrics.
• PMID 21997831 (2011, CASP9) — Slowed pace, yet some target classes, model selection, short new folds, and higher-than-expected performance revealed by new analysis.
• PMID 24344053 (2014, CASP10) — Massive progress, the first sustained improvements in molecular dynamics refinement, encouraging first trials of utilizing sparse experimental contacts, predicted contacts not yet impactful, and steady progress given target changes.
• PMID 27171127 (2016, CASP11) — Homology-based remains the best. Refinement slightly improves almost all models, and accuracy estimation reaches a practical level. Also contains the first evaluation that sparse NMR restraints dramatically improve accuracy and that models can answer biological questions that prompted experiments.
• PMID 29082672 (2018, CASP12) — A 2-fold increase in contact accuracy and dramatic improvements in template-free targets, with template-based and accuracy estimation also improving.
• PMID 31589781 (2019, CASP13) — Dramatic improvements in distance prediction deep learning, the practical solution of the fold topology problem given sufficient sequences, and a reduction in required sequence count.
• PMID 37920879 (2023, CASP15) — Major progress from the 2020 experiment (CASP14), complex accuracy rose but fell short of single proteins, and limitations of basic parameters.

The Second Path — From the Ground Up with Physics

• PMID 11200525 (2000) — Original research (Baker lab). Experimental constraints enter both the fragment selection criteria and the energy function. With about 1 experimental restraint per position, models are in some cases closer to X-ray structures than to published NMR solution structures.
• PMID 16294329 (2006) — Original research benchmark. Ties with Rosetta ab initio default settings. Hydrophobic interactions are most decisive, and sequence-specific terms are also weak but positive.

The Third Path — Co-evolution

• PMID 22106262 (2011) — Original research. Separation of direct and indirect correlations, assuming multiple homologous sequences. Signals beyond domain interiors are captured, including alternative conformations, ligand-mediated binding, and inter-chain interaction signals.
• PMID 24009338 (2013) — Original research (Baker lab). More accurate predictions with fewer sequences through the integration of structural context. Two conditions and the number of families at the time; larger families may render contact information unnecessary.
• PMID 26677056 (2016) — Participating team's self-report (Baker lab). The effects and conditions of using co-evolutionary contacts as constraints. Utilized a procedure involving human intervention and increased search volume.
• PMID 28784799 (2017) — Original research (Baker lab). The ratio of distant outliers and their resolution. Outliers exceeding 15 Å are mostly explained by oligomer interfaces (35%), conformational changes (19%), and repeat protein alignment errors (27%).
• PMID 29953468 (2018) — Original research. Generous contact definitions raise hit rates while lowering information content. Using contact potential metrics together significantly improves performance, especially with strict definitions.

The Amplifier — Learning Machines (Both are participating teams' self-reports)

• PMID 30985027 (2019) — Deep learning's debut and shift to distance prediction. Noted that difficulties when homologous sequences are scarce remain an ongoing challenge.
• PMID 29047157 (2018) — All CASP12 targets, 3-tier breakdown of top L/5 long-range criteria.

Arrival and Beyond

• PMID 34218458 (2021) — Independent evaluation team's paper. High-accuracy category expanded to all targets, 2nd place outperformed the previous 1st place — the role of community benchmarking.
• PMID 41170922 (2026) — Evaluation paper. Complex prediction remains unsolved. However, it also notes overall improvements compared to CASP15 and steady progress in multimer prediction. Model ranking and selection are major bottlenecks.