Source-linked AI summary
Design of metalloproteins and novel protein folds using variational autoencoders
Joe G Greener, Lewis Moffat, David T Jones
TL;DR
Finding sequences that fold into desired protein structures or functions is difficult because the sequence space is vast and sequence suitability is hard to assess. This paper uses conditional variational autoencoders to design metal-binding sequences and sequences for a novel topology. Generated sequences showed more copper-binding motifs than the HMM in reported comparisons, and one topology candidate showed little deviation during MD simulations.
Problem
Finding protein sequences that fold into desired structures or perform desired functions is difficult because there are vast numbers of possible sequences and sequence suitability is challenging to assess.
Source: Introduction · p. 1Method
Conditional variational autoencoders generate protein sequences conditioned on properties including metal binding and specified topologies.
Source: Introduction · p. 2, Model · p. 4Results
10% of VAE sequences had close histidines (8% for the HMM), and one 86-residue topology candidate showed little structural deviation over three 200 ns MD runs.
Source: Generation of metal binding sites · p. 6, Generation of proteins with a given… · p. 7Takeaways & Limitations
The authors conclude that CVAEs can carry out protein-design tasks conditioned on desired properties such as metal-binding sites and topologies.
Source: Discussion · p. 7The model does not reproduce residue covariation across sequence families; the dataset was limited to 50 homologues per PDB sequence and the model input was a single sequence.
Source: Discussion · p. 7
Abstract
from arXiv · showhide
The design of novel proteins has many applications but remains an attritional process with success in isolated cases. Meanwhile, deep learning technologies have exploded in popularity in recent years and are increasingly applicable to biology due to the rise in available data. We attempt to link protein design and deep learning by using variational autoencoders to generate protein sequences conditioned on desired properties. Potential copper and calcium binding sites are added to non-metal binding proteins without human intervention and compared to a hidden Markov model. In another use case, a grammar of protein structures is developed and used to produce sequences for a novel protein topology. One candidate structure is found to be stable by molecular dynamics simulation. The ability of our model to confine the vast search space of protein sequences and to scale easily has the potential to assist in a variety of protein design tasks.
Introduction
Protein design is difficult because sequence space is vast and assessing sequence suitability is challenging, while computational selection can rely on costly molecular dynamics. This work uses conditional variational autoencoders to explore metal-binding design and novel protein topologies.
Introduction
- Protein design faces a vast sequence space and difficult sequence suitability assessment, while computational selection often relies on costly force-field molecular dynamics.Protein structures can also be marginally stable, complicating the inverse folding problem.Source: Introduction · p. 1
- Metal-binding site design has mature methods, but machine learning is rarely used for design beyond computational validation, despite its use in site prediction.Novel protein topology design also had not seen pervasive machine-learning use.Source: Introduction · p. 1
- Earlier deep generative protein work generated antimicrobial peptides but did not explicitly account for structure or learn a compact latent representation.Other cited deep-learning sequence-generation methods used deterministic models and engineered features.Source: Introduction · p. 2
- Conditional variational autoencoders are proposed to narrow the protein-sequence search space and generate sequences conditioned on desired properties in two design tasks.The tasks are adding metal-binding sites to existing proteins and generating sequences for novel topologies.Source: Introduction · p. 2
Methods
The methods assemble sequence and structural data, train conditional variational autoencoders, and use distinct iterative or classifier-based sampling procedures for fold and metal-binding design.
Context free grammar
- A context-free grammar based on a periodic table of protein folds represents topology through structural elements, orientations, layers, and positions.Its topology strings describe the three basic forms and their secondary-structure arrangements.Source: Context free grammar · pp. 2–3
Data collection
- The fold model used 105,000 sequences assigned to 325 topology strings, while the metal-binding dataset contained 148,000 entries including homologues and negative examples.The initial PDB set contained 3,785 single-domain protein chains; metal-binding data covered eight metals.Source: Data collection · p. 2, Context free grammar · p. 3
Model
- The conditional VAE jointly trains inference and generative models by maximizing the evidence lower bound, balancing sequence reconstruction against divergence from a prior.The final sequence-generation model used a 16-dimensional latent space and batch normalization improved reconstruction accuracy.Source: Model · pp. 3–4
- Sequences were encoded as 140 × 22 tensors and concatenated with an eight-value metal code or a 1,265-value grammar representation; KL loss was introduced gradually during training.The model used one-hot amino-acid representations, with a padding symbol and a symbol for non-standard amino acids.Source: Model · p. 4
Sampling procedure - structure task
- For topology design, 1,000 sequences were sampled, structurally evaluated, and the best candidate was iteratively used to search its local latent space.Each cycle threaded sequences onto a template, relaxed structures, selected five candidates, and generated 10,000 ab initio structures; each iteration took around 500 CPU hours and was parallelizable.Source: Sampling procedure - structure task · p. 4
Sampling procedure - metalloprotein task
- For metalloprotein design, 1,000 sequences were drawn from the generator; a classifier ranked them for predicted metal-binding potential, using a 90%/10% training/validation split with no shared ECOD family and early stopping on validation loss.Source: Sampling procedure - metalloprotein task · pp. 4–5
Results
The CVAE generated sequences with requested metal-binding features and protein topologies, including candidates for a novel topology that remained structurally stable during the reported simulations.
Model behavior
- The model broadly reproduced family-level residue conservation, with a 0.72 Pearson correlation between generated-sequence and blastp-query Jensen-Shannon divergence profiles.The comparison used an MSA of 1,000 sequences generated from one lysozyme sequence and an MSA from a blastp query.Source: Results · p. 5
Generation of metal binding sites
- Median Rosetta energy scores were -182 for VAE sequences, -115 for HMM sequences, and -184 for the native sequence threaded through the same structure.The authors interpret the VAE scores as indicating likely adoption of the input structure, whereas HMM sequences may adopt another or no stable structure.Source: Generation of metal binding sites · p. 5
- 10% of VAE sequences had spatially close histidines and 4% had His-x3-His on an α-helix, versus 8% and 2% for HMM sequences.Among sequences with more histidines than the native sequence, the respective VAE/HMM proportions were 42%/17% for close histidines and 16%/4% for His-x3-His on an α-helix.Source: Generation of metal binding sites · p. 6
- The discriminator-selected MDM2 sequence contained two potential new copper-binding sites, while a selected TSG-6 sequence contained a potential new calcium-binding site.The copper sites were a linear His-His-Cys motif and a non-linear pair of histidines; the calcium site joined a model-added glutamate with a native glutamate three residues away.Source: Generation of metal binding sites · p. 6
- 87% of 1,000 generated plastocyanin sequences recovered His37, and 11% recovered the full His37-Cys84-His87 copper-binding site.The native plastocyanin sequences were excluded from training, while versions with mutated metal-binding residues remained.Source: Generation of metal binding sites · p. 6
Generation of proteins with a given topology
- For a known two-layer sandwich topology, 55% of generated sequences had at least 50% identity to a training sequence, and 18% scored within 10% of a native structure's energy.The energy comparison used four PDB structures with that topology; the authors describe the sequences as viable homologues in this case.Source: Generation of proteins with a given… · p. 6
- After three refinement iterations, the CVAE produced an 86-residue sequence with predicted structures similar to the modified 2CU6 backbone, though its central region differed in secondary structure.The novel topology was created by changing loops while retaining the secondary structural elements' 3D arrangement; refinement used ab initio structure generation and localized sampling.Source: Generation of proteins with a given… · pp. 6–7
- The candidate structure showed little RMSD or radius-of-gyration deviation across three 200 ns MD runs, which the authors describe as an indication of stability.They note that longer simulations and experimental testing are needed to verify the prediction.Source: Generation of proteins with a given… · p. 7
Discussion
The discussion presents CVAEs as a way to condition protein sequences on design properties, while noting limits in reproducing residue covariation and pointing to other possible design targets.
Discussion
- Even with 2 latent dimensions, homologues of each input sequence clustered together, sequences showed some separation by CATH class, and sequences of the same architecture tended to be close.The first latent dimension correlated with molecular weight as the model learned input sequence length.Source: Discussion · p. 7
- The model reproduced conservation patterns but not sequence-family covariation, with the authors citing a 50-homologue limit per PDB sequence and single-sequence rather than MSA input.Proposed directions include training on MSAs, using proteins with more homologues, and regularization to promote covarying outputs.Source: Discussion · p. 7
- Other proposed design targets include protein-protein interactions, allosteric sites, multiple conformations, linked PDB fragments, and sequences with selected motifs.The passage also describes activation maximization as a possible way to sample sequences with desired measurable attributes.Source: Discussion · p. 7
- The discussion frames vast sequence-space navigation and scaling design from specific proteins to all protein space as continuing challenges for protein design.It presents deep learning's ability to learn complex patterns without supervision and low computational cost after training as potential ways to address these challenges.Source: Discussion · p. 7
Author contributions statement
JGG, LM, and DTJ conceived and designed the study and reviewed the manuscript; JGG and LM performed the computational work, analysis, and drafting.
Author contributions statement
- JGG and LM carried out the computational work and analysis and drafted the manuscript; JGG, LM, and DTJ conceived and designed the study and reviewed it.Source: Author contributions statement · p. 10