Natural Computing (2023) 22:659–670

[Link] (0123456789().,-volV)(0123456789().
,- volV)

Protein structure prediction with energy minimization and deep

learning approaches
Juan Luis Filgueiras1 • Daniel Varela1 • José Santos1

Accepted: 12 April 2023 / Published online: 8 May 2023

Ó The Author(s) 2023

Abstract
In this paper we discuss the advantages and problems of two alternatives for ab initio protein structure prediction. On one
hand, recent approaches based on deep learning, which have significantly improved prediction results for a wide variety of
proteins, are discussed. On the other hand, methods based on protein conformational energy minimization and with
different search strategies are analyzed. In this latter case, our methods based on a memetic combination between
differential evolution and the fragment replacement technique are included, incorporating also the possibility of niching in
the evolutionary search. Different proteins have been used to analyze the pros and cons in both approaches, proposing
possibilities of integration of both alternatives.

Keywords Protein structure prediction  Differential evolution  Evolutionary computing niching methods 
Crowding niching method  Deep learning

1 Introduction A first alternative in PSP is the use of templates with

structural information of resolved proteins (their structure
The structure of proteins largely determines their function, is known). For example, PSP methods of homology mod-
hence the great importance of determining their native eling are based on finding a PDB-resolved protein with a
three-dimensional structure. For this purpose, traditional homologous amino acid sequence, since with high
laboratory methods, such as X-ray crystallography, Nuclear homology the structures are the same. PSP threading is
Magnetic Resonance (NMR) and electron cryo-mi- another possibility if there are no resolved proteins with a
croscopy, are expensive and time-consuming. As an alter- homologous sequence. In the latter case, for a target protein
native, computational Protein Structure Prediction (PSP) and a library of possible templates (folds), threading
methods attempt to close the gap between the number of methods search for the fold in which the target sequence
proteins with known sequence (on the order of millions) best fits.
and the number of resolved proteins with known structure In the most difficult and challenging alternative of PSP,
(about 200,000 in the Protein Data Bank - PDB database called ab initio, only the primary sequence information of
([Link] the protein (its amino acid sequence) is used. This ab initio
prediction is based on Anfinsen’s dogma (Anfinsen 1973),
which states that the native structure of the protein is
& Juan Luis Filgueiras determined solely by the amino acid sequence, as well as
[Link]@[Link] that the native structure corresponds to the one with the
Daniel Varela lowest Gibbs free energy (thermodynamic hypothesis).
[Link]@[Link] Consequently, an alternative in ab initio PSP is the use of
José Santos search methods that attempt to discover the structure with
[Link]@[Link] minimum energy, once a protein representation and energy
1
Department of Computer Science and Information models have been established. The problem is that PSP
Technologies, CITIC (Centre for Information and energy landscapes are high-dimensional and full of local
Communications Technology Research), University of A minima. Thus, evolutionary computing search or
Coruña, A Coruña, Spain

123
660 J. L. Filgueiras et al.

optimization methods have been used intensively, given [Link]/). Moreover, RoseTTAFold allows the pre-
their global search capability in multidimensional and diction of accurate protein-protein complex models only
multimodal energy landscapes. from sequence information, as a new avenue to short-cir-
This possibility, based on the search for the minimum in cuit the traditional docking approach.
the energy landscape, has been one of the traditional PSP The objective of this paper is to analyze the advantages
approaches, with simplified lattice models for protein and problems of different ab initio PSP strategies (energy
representation (Santos and Diéguez 2011; Varela and minimization and deep learning-based approaches),
Santos 2017) and with atomic models (Garza-Fabre et al. extending the previous work presented at conference
2016), where metaheuristics, especially evolutionary and IWINAC 2022 (Filgueiras et al. 2022). The recent methods
bio-inspired approaches (Márquez-Chamorro et al. 2015), based on deep learning AlphaFold2 (Jumper et al. 2021)
play an essential role due to the complexity of the search and RoseTTAFold (Baek et al. 2021) were selected, ana-
landscape. With atomic models of protein representation, lyzing the problems of these approaches with proteins with
the Rosetta system ([Link] is few homologous sequences. On the other hand, the widely
one of the leading methods for PSP with these energy- used Rosetta PSP protocol for the discovery of energy-
based approaches. Rosetta ab initio protocol uses small minimized conformations is also considered, together with
protein fragments (from resolved proteins) and the classical our proposals based on memetic approaches between Dif-
Metropolis criterion (Metropolis et al. 1953) to decide ferential Evolution (Price et al. 2005) and the local search
whether a structural fragment replaces a part of the current provided by the Rosetta fragment replacement technique.
conformational structure of the target protein, with the goal The rest of the paper is organized as follows. Section 2
of finding the structure with the minimum energy. Working provides a brief summary of the main aspects of the PSP
with the fragment replacement technique as a local search approaches considered, while Sect. 3 discusses the com-
operator, our evolutionary/memetic computational solu- parison of PSP alternatives and their problems using
tions (HybridDE and CrowdingDE) (Varela and Santos selected proteins. Finally, a discussion based on the results
2019, 2020, 2022) outperform Rosetta ab initio protocol presented is provided in Sect. 4.
when searching for structures with minimal energy and
under the same number of conformational energy
evaluations. 2 Methods
Nevertheless, a more recent approach in PSP is to pre-
dict the contact map or the interdistance map between 2.1 Rosetta ab initio protocol
amino acids, which is a simpler representation of the pro-
tein three-dimensional structure. This prediction typically Two protein representations are used by the Rosetta sys-
uses the information from Multiple Sequence Alignment tem: coarse-grained and all-atom. The coarse-grained rep-
(MSA) of the target protein sequence as input to deep resentation only considers the main atoms of the protein
learning schemes, such as the initial approaches of trRo- backbone (with their dihedral angles), whereas the side
setta (Yang et al. 2020) and DeepMind’s AlphaFold chains are modeled with a pseudo-atom located at their
(Evans et al. 2018; Senior et al. 2020). DeepMind’s recent center of mass (Fig. 1).
deep learning-based method, called AlphaFold2 (Jumper The Rosetta ab initio PSP protocol (Rohl et al. 2004)
et al. 2021), has shown a very large improvement over ([Link] with the low-resolution
previous approaches, as demonstrated by the results in the protein representation, employs a search technique in
CASP (Critical Assessment of protein Structure Prediction) which a Monte Carlo procedure decides whether the
competition (CASP14 in 2020) ([Link] dihedral angles of small protein fragments can replace the
). In addition, in a partnership between DeepMind and the original ones (Kaufmann et al. 2010; Rohl et al. 2004). A
European Molecular Biology Laboratory (EMBL), Alpha- protein fragment is a group of consecutive amino acids of a
Fold2 was used to publicly provide structure predictions resolved protein. Fragments are selected by considering
(over 200 million protein structure predictions including their sequence similarity with respect to the window of
the  20,000 proteins expressed by the human genome) consecutive residues of the target protein into which the
([Link] fragments will be inserted. The decision regarding whether
Likewise, RoseTTAFold (Baek et al. 2021), also based the dihedral angles of a selected fragment replace those of
on a deep learning architecture, has shown a very consid- the target protein is based on the Metropolis criterion
erable improvement over the traditional counterpart based (Metropolis et al. 1953). This criterion always accepts the
on energy minimization, as well as over other deep learn- changes that improve the energy (lower values), while
ing approaches, as tested with proteins in the Continuous occasionally accepting dihedral angle changes that worsen
Automated Model EvaluatiOn (CAMEO) project (https:// the energy, with the probability of accepting the fragment

123
Protein structure prediction with energy minimization and deep learning approaches 661

Fig. 1 Rosetta’s coarse-grained

model for protein
representation. It considers only
the main atoms of the protein
backbone, while pseudo-atoms
represent the lateral residues. x,
/ and w dihedral angles encode
each protein conformation

depending on the increase in energy relative to the previous grained representation (encoding the dihedral angles for
state of the target protein. each amino acid). The hybrid combination integrates the
In Rosetta, the energy of a protein conformation is advantage of the global search of the evolutionary algo-
defined as a weighted linear combination of different rithm with the local search provided by the replacement of
energy terms that model the molecular forces acting fragments. The fragment insertion technique is used in DE
between the amino acid atoms. For example, steric overlap to refine the population solutions and to refine the DE
between backbone and side-chain atoms is penalized, while candidate or trial solutions. The memetic search follows a
other Rosetta’s energy terms correspond to van der Waals three-stage evolutionary process, as the fitness of the
interactions, electrostatics effects and solvation, hydrogen encoded conformations corresponds to different Rosetta
bonding, repulsion and scores related to protein secondary score functions in each stage, while the first stage of
structure (e.g., helix-strand packing and strand pairing). Rosetta is used to define the initial population (with par-
The detailed definition of the energy terms can be found, tially folded and different conformations). The memetic
for example, in Rohl et al. (2004), and the weight sets for version is detailed in Varela and Santos (2019, 2020).
the individual energy terms for the definition of every HybridDE outperforms the Rosetta ab initio protocol in
Rosetta score are detailed in ([Link] obtaining conformations with minimum energy and under
org). The Rosetta score function named score3 integrates the same number of conformational energy evaluations.
all of the energy components. Figure 3 corresponds to an example of the fitness evo-
In the search for protein conformations with minimum lution in the three-stage evolutionary process of HybridDE.
energy, the stochastic Metropolis Monte Carlo procedure is The example uses an evolution with a population size of
run thousands of times. For this, the Rosetta ab initio 100 individuals, 100 generations (in each of the three
protocol is divided into four stages, which use different evolutionary stages) and SARS-CoV-2 protein orf8 as
score functions (progressively incorporating new energy target. In each evolutionary phase, the fitness is associated
terms) and number of fragment insertion attempts. Detailed to the energy score that Rosetta uses in its corresponding
information about these four stages can be found, for phase and, consequently, it has different ranges in each
example in Rohl et al. (2004) and Varela and Santos evolutionary stage. Note that, using Rosetta nomenclature
(2022), while Fig. 2 illustrates schematically the search in Stage 3, score5 was the final energy (with energy terms
process of the Rosetta ab initio protocol with its four focused on compactness and secondary structure), although
stages. Rosetta uses the coarse-grained protein represen- score3 in the final Stage 4 is the energy score incorporating
tation and its fragment insertion technique (with the all energy components. This evolutionary process with
Metropolis criterion) throughout these four stages to gen- different stages, integrating in each of them more detailed
erate new structural conformations. The final conforma- energy functions, allows a progressive structural
tions (‘‘decoys’’) in this ab initio protocol, can be refined in refinement.
an ‘‘Ab initio Relax’’ procedure using the Rosetta’s full Nevertheless, obtaining energy-optimized conforma-
atomic model. tions is not the only goal when the energy landscape is
deceptive (such as the inaccurate Rosetta’s energy model).
2.2 HybridDE and CrowdingDE PSP approaches This occurs when the conformation with the minimum
energy does not correspond to the conformation closest to
We defined a memetic combination for PSP between Dif- the real native structure. Consequently, one strategy is to
ferential Evolution (DE) (Price et al. 2005) and the obtain a set of optimized conformations (with minimal
Rosetta’s fragment replacement technique (HybridDE energy) with structural diversity. For this purpose, we
memetic approach). The genetic population encodes pos- introduced niching methods such as crowding, fitness
sible protein conformations using the Rosetta’s coarse- sharing and speciation into the HybridDE version, thus

123
662 J. L. Filgueiras et al.

Fig. 2 Workflow of the Rosetta Ab initio protocol, working with the stage), different number of fragment insertion attempts, as well as
coarse-grained representation. Each of the four stages uses a fragments with length of 3 or 9 mers
particular score function (incorporating new energy terms in each

Fig. 3 Energy/fitness evolution of HybridDE in a run with SARS- line). There are three sequential evolutionary stages, corresponding to
CoV-2 protein orf8 as target: evolution of the average energy of the the same Rosetta stages (using the same fragment lengths and Rosetta
population (green line) and the energy of the best individual (red energy score functions)

forcing the search algorithm to obtain optimized confor- 2.3 AlphaFold and RoseTTAFold
mations in different niches (with different structural con-
formations) (Varela and Santos 2022). This increases the The recent deep learning-based methods of Alpha-
chances of obtaining candidate structures close to the Fold2Jumper et al. (2021) and RoseTTAFold Baek et al.
native structure. The crowding niching method was found (2021) for PSP were considered here. AlphaFold2 is the
to be the most useful niching technique in the application, latest version of DeepMind’s effort in PSP with deep NN
given its simple parameter decision process, defining the architectures, which is an improvement of the first version
CrowdingDE version, detailed in Varela and Santos AlphaFold (Evans et al. 2018; Senior et al. 2020). In the
(2020, 2022). case of AlphaFold2, the system receives the MSA of a

123
Protein structure prediction with energy minimization and deep learning approaches 663

target sequence as input information. MSA algorithms AlphaFold2 to progressively refine the final and predicted
provide the alignment of evolutionarily related protein model of the protein.
sequences, with a 2D matrix representation where the AlphaFold2 provides two confidence measures of the
horizontal axis represents the residues of the target protein predicted structures. The first is the predicted local-Dis-
and the vertical axis corresponds to homologous protein tance Difference Test (plDDT), a per-residue measure of
sequences with an optimized alignment with respect to the local confidence (on a scale from 0-100). The local-Dis-
other sequences (alignment that may be suboptimal since tance Difference Test (lDDT) is a superposition-free score
heuristics are used). that evaluates local distance differences of all atoms in a
The idea behind the use of MSA as input information is model (Mariani et al. 2013). The plDDT measure provided
that correlated mutations between residues indicate their by AlphaFold2 estimates how well the prediction would
spatial physical interaction. That is, if an amino acid agree with an experimental structure, since it predicts the
mutates at position i in a homologous sequence (with agreement of the lDDT (lDDT-Ca , considering only the
respect to the target sequence), and a correlated mutation atom Ca in each amino acid) between the predicted and real
appears at position j in the same homologous sequence, structures. The second metric is the Predicted Alignment
then it is likely that residues i and j are in contact in the Error (PAE). PAE (x, y) reports the expected position error
tertiary structure, since the correlated mutation tends to at residue x, when the predicted and real structures are
maintain the protein structure unchanged. aligned on residue y. Consequently, it provides a level of
AlphaFold2’s architecture is detailed in Jumper et al. confidence about the relative positions of the amino acids
(2021). Several self-attention operations are performed in (and different domains) of the protein.
the deep NN architecture. Attention enables the NN to RoseTTAFold (Baek et al. 2021) is also a recent method
guide the flow of information, by learning to select which based on deep learning and inspired by the DeepMind’s
aspects of the input information should interact with other framework, as the authors state. It is also an improvement
aspects of the same input. For instance, MSA representa- over the previous version of the same group at the
tions are processed with consecutive blocks of self-atten- University of Washington, called trRosetta (Yang et al.
tion in rows and columns. The first generates attention 2020). RoseTTAFold uses a three-track neural network to
weights for amino acid pairs, allowing identification of simultaneously process sequence, distance, and coordinate
which amino acid pairs are most closely related. The sec- information. The main new feature in RoseTTAFold Baek
ond attention process (in the vertical direction of MSA) et al. (2021), with respect to AlphaFold2, is the incorpo-
allows elements belonging to the same target amino acid ration of a third track in the deep NN design, which
position to exchange information, i.e., it determines which operates in the 3D coordinate space. As the authors state
protein sequences are most informative in the MSA input (Baek et al. 2021), this provides a tighter connection
information. between the protein sequence, residue-residue distances
One of the key modules of the internal architecture of and their orientations, as well as the coordinates of all
AlphaFold2 is the main network block called Evoformer, a atoms. Therefore, the neural architecture has 1D, 2D and
stack of several NN layers that performs feature embed- 3D tracks with attention mechanisms. There are connec-
ding. Evoformer works with an embedding of the MSA and tions between the three tracks to allow simultaneous
with an internal pair representation (a generalized version learning of relationships within and between sequences,
of a distogram, i.e., a map of interdistances between resi- distances and coordinates. RoseTTAFold (Baek et al. 2021)
dues). Both representations exchange information, as outperformed other PSP servers with recent structures
updates in the MSA embedding provide new information to submitted to PDB (CAMEO project ([Link]
change the structural hypothesis in the pair representation, )). RoseTTAFold also provides an estimate of the per-
and vice versa. residue precision, based on the estimated Ca RMS error,
Evoformer is followed by a NN module or Structure i.e., the predicted distance with respect to the native
Module that maps the embedding or abstract representation structure (using only the Ca atoms). This Ca RMS error is
of the Evoformer stack to concrete 3D coordinates of all estimated from the predicted lDDT-Ca , as detailed in Baek
atoms (as well as the per-residue confidence commented et al. (2021).
below). In this module, NN attention mechanisms (with The AlphaFold2 ([Link]
invariance to rotations and translations of the protein github/sokrypton/ColabFold/blob/main/[Link])
conformation in space) are used to progressively refine the and RoseTTAFold ([Link]
structure (which includes the side-chain atoms). Finally, github/sokrypton/ColabFold/blob/main/RoseTTAFold.
the predicted structure information is returned to the Evo- ipynb) ColabFold servers were used in the predictions.
former blocks. Consequently, these two steps (Evorformer Coupled with Google Colaboratory, ColabFold (Mirdita
and Structural Prediction) are repeated several times in et al. 2022) is a free and accessible platform for protein

123
664 J. L. Filgueiras et al.

folding that integrates the fast homology MMA search of the ColabFold servers was used, that is, using the MSA
MMseqs2 (Steinegger and Söding 2017) with AlphaFold2 information as input, while the servers provide five can-
and RoseTTAFold. The commented measures of both deep didate models, those with the highest prediction
learning-based methods will be used to analyze the results. confidence.

3.2 Examples with PDB proteins

3 Results
Proteins from PDB [20] were selected, i.e., proteins with
3.1 Setup of the PSP approaches known structure. Consequently, these proteins serve to test
whether the predictions are close to the folded and resolved
The PSP approaches discussed in the Methods section are structure deposited in PDB. Proteins with shallow MSA
used with different proteins. In the case of Rosetta ab initio information (few sequences homologous to the target
PSP protocol, taking into account its stochasticity, the protein or with low sequence identity) were selected in
protocol is run 1,000 times to generate 1,000 decoys order to test the behavior of deep learning-based
(candidate conformations). Rosetta parameter increase_- approaches.
cycles is set to 10 (which multiplies the default values of A first example is selected in which, even with the
fragment insertion cycles in the different Rosetta stages, scarce information provided by MSA, deep learning
details in Rohl et al. (2004) and Varela and Santos (2022)), methods show excellent prediction results. The example
as recommended on the Rosetta site ([Link] corresponds with protein 1r69 (phage 434 repressor, 69
[Link]). amino acids). Figure 4 shows the distribution of solutions
Regarding the approaches based on DE (HybridDE and generated by the different methods. This is a standard
CrowdingDE), the same setup used in Varela and Santos graph in PSP for evaluating the performance of energy
(2019, 2020, 2022) was employed. DE parameters were minimization methods, as it shows the distribution of the
experimentally adjusted to generate candidate conforma- optimized protein decoys (in terms of their distances from
tions in DE with slight variations with respect to their base the native structure), along with the optimization (in terms
individual (random conformation of the population in the of energy) obtained in the optimized solutions. The dis-
DE scheme used), in order to minimize conflicts between tances of the predicted/optimized conformations from the
atoms in the DE trial or candidate solutions. Therefore, a native structure are calculated with the RMSD (Root Mean
low weight factor (F ¼ 0:025) in the mutation operator is Squared Deviation), taking into account the Ca atoms of
needed, along with a high crossover probability each amino acid, superimposing each conformation with
(CR ¼ 0:99). Moreover, DE strategy DE/rand/1/bin was the native one. The energy of each conformation corre-
used, which provides low selective pressure (see Varela sponds to Rosetta’s coarse-grained representation (score3,
and Santos (2020, 2022) for details). which includes all individual energy terms).
With the same purpose of generating 1,000 decoy con- It should be noted that the RMSD distance can only be
formations, HybridDE and CrowdingDE were run 10 calculated with proteins whose native structure is known a
times, with a population of 100 solutions and over 100 priori, such as those considered here. That is, with these
generations in the 10 independent runs. Consequently, the proteins acting as a benchmark since their native structure
10 runs also generate 1,000 final solutions (joining the final is known, we can analyze the quality of the predictions in
populations of the runs). These DE-based energy mini- terms of their structural distance with respect to the native
mization approaches will use the same number of fitness/ reference structure.
energy evaluations as Rosetta ab initio for generating the Figure 4 shows that all AlphaFold2 and RoseTTAFold
1,000 decoys. It should be noted that energy evaluations predicted solutions are very accurate, with very low values
are synonymous with fragment insertion attempts, since an of RMSD with respect to the native structure deposited in
insertion attempt involves the energy evaluation of the PDB. On the contrary, the energy-based approaches
resulting conformation. The fair comparison is obtained (Rosetta ab initio, HybridDE and CrowdingDE) present
since, in HybridDE and CrowdingDE, parameter in- solutions with a large variety of RMSD values, with few
crease_cycles is set to 0.1 (100 times less than in the case decoys with values lower than 2 Ȧ (angstroms) and several
of Rosetta ab initio protocol), because the evolutionary solutions with large distances (in RMSD terms) to the
approaches refine 1,000 conformational solutions with the native structure.
same Rosetta approach based on fragment insertions, but Nevertheless, in terms of energy of the predicted solu-
over 100 generations. tions, the comparison is totally different, since the deep
For the deep learning-based approaches (AlphaFold2 learning-based approaches present solutions with higher
and RoseTTAFold), the default configuration provided by energies. This is due to the fact that, in the AlphaFold2 and

123
Protein structure prediction with energy minimization and deep learning approaches 665

minimization-based approaches, HybridDE and Crowd-

ingDE obtain solutions with better energy with respect to
Rosetta ab initio, showing the better ability of the evolu-
tionary approaches to sample the conformational space
under the same number of energy evaluations. However,
HybridDE obtains the best solution in energy terms, but it
corresponds to a solution that is farther away from the
native structure with respect to other solutions provided by
HybridDE and Rosetta. This shows the inaccuracies of the
Rosetta energy model, since the solutions with the best
energy do not have to correspond to those closest to the
native structure, defining a clear deceptive energy land-
scape for the search algorithms, in which the best
approaches in energy terms are those that provide the worst
solutions in RMSD terms.
To address the problem of deceptiveness in the energy
landscape, the inclusion of crowding in the memetic evo-
lutionary algorithm (CrowdingDE), allows us to obtain a
wider distribution of optimized solutions, as clearly shown
by the distributions in the violins in Fig. 4, but at the cost
(in this protein) of presenting solutions not as optimized (in
energy terms) as HybridDE. Nevertheless, the goal of
CrowdingDE is precisely this, to present a set of optimized
and structurally different proteins, as shown by the distri-
bution. However, even with the inclusion of crowding, in
this protein, CrowdingDE does not present solutions closer
to the native structure with respect to Rosetta ab initio.
Several comments can be made regarding the Alpha-
Fold2 and RoseTTAFold solutions. Figure 5 shows infor-
mation about the AlphaFold2 predictions. Figure 5
includes a representation of the MSA information, which is
the same input to AlphaFold2 and RoseTTAFold. The MSA
coverage graph shows that there is a large number of
homologous sequences, obtained from different genetic
databases, although without high sequence identity in most
cases. Nevertheless, the solutions provided by AlphaFold2
present quite high confidence in the predictions and in most
of the protein chain. This is shown by the high confidence
in the plDDT per residue (explained in Sect. 2.3), where
plDDT[90 corresponds to predictions modeled with high
accuracy, whereas areas where plDDT\50 should not be
considered reliable (Jumper et al. 2021). It is only at the
final amino acids that there is a drop in prediction confi-
dence, where there is also worse sequence coverage in the
Fig. 4 Energy (score3) vs. RMSD (from the native structure, in MSA information. Similarly, the PAE graphs for the five
angstroms - Ȧ) for proteins 1r69, 1ha8 and 2ksw. Gray: Rosetta best-rated prediction models (considering the average
ab initio. Blue: HybridDE. Red: CrowdingDE. Green: AlphaFold2 plDDT in the residues) show the high prediction confi-
solutions. Pink: RoseTTAFold solutions. The violin plots correspond
dence, where the bluer, the lower the estimated error. The
with the RMSD (upper) and energy (right) distributions. The quartiles
(black lines) and the median (light blue mark) are shown in the violins PAE graph again shows that only the 10 final amino acids
exhibit low confidence in the structural prediction.
RoseTTAFold solutions, some atoms have collisions in the Figure 6 includes the confidence with the RoseTTAFold
side chains, a problem that can be improved by further solutions. The graphs in Fig. 6 correspond to the estimated
refinement of the structure. Considering the energy Ca RMS error (commented in Sect. 2.3) in each amino acid

123
666 J. L. Filgueiras et al.

Fig. 5 Information about AlphaFold2 models with proteins 1r69, predicted AlphaFold2 models. Bottom figures: PAE (Predicted
1ha8 and 2ksw. For each protein, top left: MSA sequence coverage. Aligned Error) of the five highest-rated AlphaFold2 models
Top right: predicted Local Distance Difference Test (plDDT) of the

position, that is, the predicted distances (in Ȧ) between the That is, a lower value corresponds to a better prediction,
Ca positions (of each amino acid) of the native structure and a zero value corresponds to a perfect prediction of the
and the predicted RoseTTAFold conformational solution. native structure. For each protein in Fig. 6, its confidence

123
Protein structure prediction with energy minimization and deep learning approaches 667

Fig. 6 Estimated error (Ca RMS error, in Å) in each amino acid position for the RoseTTAFold solutions with the best average confidence and for
proteins 1r69, 1ha8 and 2ksw

graph corresponds to the RoseTTAFold solution with the (CrowdingDE), allows us to obtain a wider distribution of
best confidence (averaged over amino acid positions) of the optimized solutions. In fact, now CrowdingDE presents the
5 returned solutions (all five solutions exhibit fairly similar closest solution to the native structure.
confidence values). The confidence results of the Alpha- AlphaFold2 and RoseTTAFold present worse solutions
Fold2 solutions (using plDDT, Fig. 5) and the confidence in terms of RMSD distance from the native structure. As
results of the best RoseTTAFold solution (using Ca RMS, shown in Fig. 5, with protein 1ha8, again the MSA does
Fig. 6) are very similar, with only a drop at the end of the not present high sequence identity between the target
protein chain due to the poor quality of the MSA infor- protein and the homologous ones and, in this case, it does
mation at the end of the amino acid sequence. not include such a large number of homologous sequences
Regarding computing times, it should be noted that as in the previous protein. This information is not sufficient
Rosetta ab initio is not parallelized while HybridDE and for fairly high confidence in the predictions. This is shown
CrowdingDE are parallelized in MPI (Message Passage by the poor confidence in the plDDT per residue. Similarly,
Interface). Typical computing times are 45 min for each of the PAE graphs for the five best-rated prediction models
the parallelized 10 independent runs of HybridDE and show the low prediction confidence. The analysis with the
CrowdingDE (protein 1r69 as target). Rosetta Ab initio RoseTTAFold solutions is similar, as Fig. 6 shows that the
requires an average of 24.64 h to generate 1,000 solutions. confidence (estimated Ca RMS error) is worse with respect
The experiments were run in the Supercomputing Center of to the previous protein and with poor confidence (RMS
Galicia ([Link]), with Intel Xeon E5-2680 v3 pro- error greater than 1.5 Ȧ) at both ends of the protein, which
cessors at 2.50GHz and 1GB of RAM. The computing time is logical as the MSA information also has a drop at the
for the deep learning approaches is variable, since these protein ends.
were run in the ColabFold servers ([Link] The third example corresponds to protein 2ksw (a beetle
[Link]/github/sokrypton/ColabFold/blob/main/Alpha hemolymph protein, 46 amino acids). Figure 4 shows again
[Link], [Link] that HybridDE and CrowdingDE provide better average
rypton/ColabFold/blob/main/[Link]). It can energy of the optimized solutions, although Rosetta
vary from 4 min to 30 min (which includes MSA ab initio also discovers solutions with low energy. As in the
calculation). previous example, the AlphaFold2 and RoseTTAFold
The second example corresponds with protein 1ha8 solutions present high RMSD values from the native
(pheromone from protozoan E. Raikovi, 51 amino acids). structure and these are worse with respect to many solu-
As shown in Fig. 4, with protein 1ha8, now the energy tions of the energy minimization approaches. The reason is
minimization approaches return solutions closer to the that, in this case, the MSA map is simple, since there are no
native structure with respect to the deep learning approa- homologous sequences found in the genetic databases
ches. In the case of the former approaches, HybridDE and (except for the sequence itself which was found 3 times).
CrowdingDE obtain solutions with better energy with That is, it is an example with no information of proteins
respect to Rosetta ab initio, again showing the enhanced with similar sequence. Consequently, the AlphaFold2
ability of evolutionary approaches to sample the confor- models present very low confidence at all positions in the
mational space. Once again, HybridDE obtains the best protein chain, as can be seen in Fig. 5 with the PAE plots
optimized solution in energy terms, a solution that is far- and with the plDDT measure.
ther away from the native structure with respect to many Similarly, Fig. 6 shows that the confidence of the best
other solutions, which again shows the deceptiveness of the RoseTTAFold solution is also poor, with estimated Ca RMS
Rosetta energy landscape also with this protein. As with the errors between 2 and 5 Å. Figure 6 shows that, however,
previous example, the inclusion of crowding there is no match (considering the amino acid positions

123
668 J. L. Filgueiras et al.

with the worst confidence), between the AlphaFold2 solu- these latter approaches shows that the best RoseTTAFold
tions and the RoseTTAFold solution (Figs. 5 and 6). solution and the best AlphaFold2 solution exhibit low
The final example corresponds to protein orf8, a protein confidence at the ends and middle positions of the protein,
component of the SARS-CoV-2 virus with 104 amino acids especially the RoseTTAFold solution at the middle amino
(protein visualizations of predictions with SARS-CoV-2 acids.
proteins and the prediction methods can be seen in (https://
[Link]/ir/in845d-02/SARS-CoV-2_protein_pre
diction/[Link])). This protein has no homologous 4 Discussion and conclusions
proteins in the PDB database [20]. Even with the search for
homologous sequences in genetic databases, the MSA This study has performed a comparison between PSP
coverage is poor (Fig. 7, top right). Consequently, the approaches based on energy minimization and deep
prediction confidence of AlphaFold2 and RoseTTAFold learning. It is clear that energy minimization approaches
models is low. present better solutions in terms of minimized energy with
Neither approach presents accurate solutions in RMSD respect to deep learning approaches. In addition, in energy
terms, as shown by the distribution of solutions in the minimization-based alternatives, memetic approaches
energy vs. RMSA plot (Fig. 7, top left), showing the dif- show better sampling of the energy landscape with respect
ficulty of some proteins for the different PSP approaches. to the state-of-the-art Rosetta ab initio protocol. However,
Clearly, the solutions with deep learning approaches need a imperfections in the energy landscape do not allow the best
posterior refinement, especially in the case of the RoseT- optimized solutions with memetic approaches to corre-
TAFold solutions. Finally, the prediction confidence of spond to solutions closer to the native structure. The main

Fig. 7 MSA information input to AlphaFold2 and RoseTTAFold with highest rated AlphaFold2 model. Bottom right: Estimated error (Ca
SARS-CoV-2 protein orf8. Top left: Energy (score3) vs. RMSD with RMS error) per amino acid for the RoseTTAFold solution with the
different PSP approaches (same colors as in Fig. 4). Top right: MSA best average confidence
sequence coverage. Bottom left: PAE (Predicted Aligned Error) of the

123
Protein structure prediction with energy minimization and deep learning approaches 669

Table 1 Summary of the main conclusions that can be drawn from the comparison of the results with the different protein structure prediction
approaches considered
Approaches based on energy minimization obtain better energy-optimized conformations than those based on deep learning.
In the energy minimization-based approaches, the memetic approaches (HybridDE and CrowdingDE) better sample the energy landscape with
respect to Rosetta ab initio under the same number of energy evaluations/fragment insertion attempts.
The incorporation of the crowding niching method into the memetic algorithm (CrowdingDE) allows obtaining a set of optimized conformations
with a higher structural diversity with respect to HybridDE, which is useful in proteins with a deceptive energy landscape.
Deep learning-based approaches (AlphaFold2 and RoseTTAFold) provide better solutions, in terms of distance (RMSD) from the native structure,
for most proteins (with known structure). However, their predictions are not reliable when MSA information is poor.
If the input MSA provides sufficient information, deep learning-based approaches tend to provide high quality predictions of the protein
backbone, while the side chains tend to present collisions. A refinement process, based on energy minimization, can improve their initial
predictions.

conclusions on the comparison of the approaches consid- consensus refined model of the PSP approaches, thus
ered can be summarized in Table 1. integrating the different approaches.
Predictions with selected proteins show the dependence
Acknowledgements This study was funded by the Xunta de Galicia
of deep NN-based approaches on MSA input information.
and the European Union (European Regional Development Fund -
When the MSA information is not detailed enough, deep Galicia 2014-2020 Program), with grants CITIC (ED431G 2019/01),
NN-based approaches can present predictions with low GPC ED431B 2022/33 and IN845D-02 (funded by the ‘‘Agencia
confidence, as shown with selected proteins. As also stated Gallega de Innovación’’, co-financed by Feder funds, supported by the
‘‘Consellerı́a de Economı́a, Empleo e Industria’’ of Xunta de Galicia),
by Peng et al. Peng et al. (2022), there are proteins in
and by the Spanish Ministry of Science and Innovation (project
which deep learning based methods relying on MSA PID2020-116201GB-I00).
information present poor predictions, such as proteins from
viruses without homologous sequences in genetic data- Author contributions All authors contributed to the conception and
design of the study. Material preparation and data collection were
bases, which is the case of the last example with a protein
carried out mainly by J.L. Filgueiras. Analysis of the results was
of the SARS-CoV-2 virus. In this sense, a work to be done performed by all authors. All authors contributed equally to the
will be an analysis of the correlation between the quality of writing of the manuscript. All authors read and approved the final
the MSA information and the confidence of the prediction manuscript.
using a large number of proteins.
Funding Open Access funding provided thanks to the CRUE-CSIC
Nevertheless, it must be taken into account that proteins agreement with Springer Nature. This study was funded by the Xunta
with low sequence identity were chosen for the analysis, de Galicia and the European Union (European Regional Development
and the recent approaches based on deep learning present Fund - Galicia 2014-2020 Program), with grants CITIC (ED431G
2019/01), GPC ED431B 2022/33 and IN845D-02 (funded by the
predicted solutions with very low distances to the real
‘‘Agencia Gallega de Innovación’’, co-financed by Feder funds, sup-
native structure in the vast majority of proteins, showing a ported by the ‘‘Consellerı́a de Economı́a, Empleo e Industria’’ of
great leap forward in this problem of computational Xunta de Galicia), and by the Spanish Ministry of Science and
structural biology. For example, in the case of the human Innovation (project PID2020-116201GB-I00).
proteome, 58% of residues present a confident prediction
Data availability The data that support the findings of this study are
(pLDDT[ 70) with AlphaFold2 Tunyasuvunakool et al. available from the corresponding author upon reasonable request.
(2021).
However, deep learning-based methods cannot provide Declarations
different predicted structures for proteins mutated in a few
amino acids, as discussed in Buel and Walters (2022); Conflict of interest The authors have no relevant financial or non-
Callaway (2022); Peng et al. (2022). Moreover, predictions financial interests to disclose.
with deep learning schemes present solutions with high
Ethical approval Not applicable.
energy, in most cases due to conflicts between atoms in the
side chains. In all cases, a refinement process is needed to Open Access This article is licensed under a Creative Commons
resolve these conflicts. A refinement process that integrates Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as
the different approaches is a line of research to be explored,
long as you give appropriate credit to the original author(s) and the
using evolutionary algorithms to refine (in energy terms) source, provide a link to the Creative Commons licence, and indicate
the predicted conformations. The initial population may if changes were made. The images or other third party material in this
include different prediction models of different PSP alter- article are included in the article’s Creative Commons licence, unless
indicated otherwise in a credit line to the material. If material is not
natives, so that evolutionary optimization can obtain a

123
670 J. L. Filgueiras et al.

included in the article’s Creative Commons licence and your intended Márquez-Chamorro AE, Asencio-Cortés G, Santiesteban-Toca CE,
use is not permitted by statutory regulation or exceeds the permitted Aguilar-Ruiz JS (2015) Soft computing methods for the
use, you will need to obtain permission directly from the copyright prediction of protein tertiary structures: a survey. Appl Soft
holder. To view a copy of this licence, visit [Link] Comput 35:398–410
org/licenses/by/4.0/. Peng Z, Wang W, Han R, Zhang F, Yang J (2022) Protein structure
prediction in the deep learning era. Curr Opin Struct Biol
77:102495
References Prediction results of the SARS-CoV-2 unsolved proteins. https://
[Link]/ir/in845d-02/SARS-CoV-2_protein_predic
tion/[Link]
AlphaFold2 PSP database,. [Link] Price KV, Storn RM, Lampinen JA (2005) Differential evolution. a
AlphaFold2 server. [Link] practical approach to global optimization
ton/ColabFold/blob/main/[Link] Protein Data Bank. [Link]
Anfinsen CB (1973) Principles that govern the folding of proteins. Protein structure prediction center. [Link]
Science 181(96):223–230 Rohl CA, Strauss CEM, Misura KM, Baker D (2004) Protein structure
Baek Minkyung, DiMaio Frank, Anishchenko Ivan et al (2021) prediction using Rosetta. Methods Enzymol 383:66–93
Accurate prediction of protein structures and interactions using a RoseTTAFold server. [Link]
three-track neural network. Science 373(6557):871–876 rypton/ColabFold/blob/main/[Link]
Buel GR, Walters KJ (2022) Can AlphaFold2 predict the impact of Rosetta system. [Link]
missense mutations on structure? Nat Struct Mol Biol 29(1):1–2 Santos J, Diéguez M (2011) Differential evolution for protein
Callaway E (2022) What’s next for AlphaFold and the AI protein- structure prediction using the HP model. 6686:323–323
folding revolution. Nature 604:234–238 Senior AW, Evans R, Jumper J et al (2020) Improved protein
Continuous Automated Model EvaluatiOn (CAMEO). https:// structure prediction using potentials from deep-learning. Nature
[Link]/ 577:706–710
Evans R et al (2018) De novo structure prediction with deep-learning Steinegger M, Söding J (2017) Mmseqs2 enables sensitive protein
based scoring. In: Thirteenth critical assessment of techniques sequence searching for the analysis of massive data sets. Nat
for PSP, p 1–4 Biotechnol 35:1026–1028
Filgueiras JL, Varela D, Santos J (2022) Energy minimization vs. Tunyasuvunakool K, Adler J, Wu Z et al (2021) Highly accurate
deep learning approaches for protein structure prediction. In: protein structure prediction for the human proteome. Nature
Proceesings IWINAC 2022, lecture notes in computer science, 596:590–596
p 109–118 Varela D, Santos J (2020) Protein structure prediction in an atomic
Garza-Fabre M, Kandathil SM, Handl J, Knowles J, Lovell SC (2016) model with differential evolution integrated with the crowding
Generating, maintaining, and exploiting diversity in a memetic niching method. Nat Comput 21:1–15
algorithm for protein structure prediction. Evolut Comput Varela D, Santos J (2022) Niching methods integrated with a
24(4):577–607 differential evolution memetic algorithm for protein structure
Jumper J, Evans R, Pritzel A et al (2021) Highly accurate protein prediction. Swarm Evolut Computa 71:101062
structure prediction with AlphaFold. Nature 596:583–589 Varela D, Santos J (2017) A hybrid evolutionary algorithm for protein
Kaufmann KW, Lemmon GH, DeLuca SL, Sheehan JH, Meiler J structure prediction using the Face Centered Cubic lattice model.
(2010) Practically useful: what the Rosetta protein modeling In Proceedings ICONIP 2017, lecture notes in computer science,
suite can do for you. Biochemistry 49:2987–2998 p 628–638
Mariani V, Biasini M, Barbato A, Schwede T (2013) lDDT: a local Varela D, Santos J (2019) Crowding differential evolution for protein
superposition-free score for comparing protein structures and structure prediction. In: Proceedings IWINAC 2019, lecture
models using distance difference tests. Bioinformatics notes in computer science, p 193–203
29(21):2722–2728 Yang J, Anishchenko I, Park H, Peng Z, Ovchinnikov S, Baker D
Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E (2020) Improved protein structure prediction using predicted
(1953) Equation of state calculations by fast computing interresidue orientations. PNAS 117:1496–1503
machines. J Chem Phys 21(6):1087–1092
Mirdita M, Schütze K, Moriwaki Y, Heo L, Ovchinnikov S,
Steinegger M (2022) ColabFold: making protein folding acces- Publisher’s Note Springer Nature remains neutral with regard to
sible to all. Nat Methods 19:679–682 jurisdictional claims in published maps and institutional affiliations.

123