In this project, we will explore the application of deep learning techniques to optimize the performance of PETase, an enzyme that can degrade polyethylene terephthalate (PET), a common plastic. The goal is to enhance the thermal stability and catalytic efficiency of PETase through computational methods.
In this aim, we will create a deep learning model that can predict the thermal stability of different PETase variants. We will use a dataset of known PETase variants and their corresponding stability data to train the model. The model will be designed to take into account various features of the enzyme, such as amino acid sequence and structural properties.
In this aim, we will synthesize the wild-type PETase and a selection of variants that are predicted to have improved stability based on the deep learning model. We will use recombinant DNA technology to express and purify these enzymes for further testing.
In this aim, we will assess the catalytic efficiency of the synthesized PETase variants. We will perform enzymatic assays to measure the rate of PET degradation by each variant under various conditions. This will allow us to determine if the predicted improvements in stability also translate to enhanced catalytic performance.
In this visionary aim, we will implement a closed-loop system that integrates computational predictions with experimental validation. We will use laboratory automation to rapidly test the predicted PETase variants and feed the results back into the deep learning model for further refinement. This iterative process will enable us to continuously improve the model’s predictive accuracy and optimize the performance of PETase variants in a more efficient manner.
ProteinMPNN is a pre-trained graph neural network that generates protein sequences for given protein backbone structures. It learns the conditional probability of protein sequence autoregressively conditioned on the backbone structure:
$$\log p(\mathbf{s} | \mathbf{X}) = \sum_{i=1}^L \log p( s_{\sigma(i)} | s_{\sigma(1)}, \ldots, s_{\sigma(i-1)}, \mathbf{X} )$$
where:
ProteinMPNN can be used to score amino acid types on a specific position $k$ by the following marginal probability:
$$\log p(s_k | \mathbf{s}_{\backslash k}, \mathbf{X})$$
The difference in likelihood, $\log p(s_k = \text{M} | \mathbf{s}{\backslash k}, \mathbf{X}) - \log p(s_k = \text{W} | \mathbf{s}{\backslash k}, \mathbf{X})$, is used to estimate the mutational effect of substituting the amino acid W on position $k$ to amino acid M.
We benchmarked the original pre-trained ProteinMPNN on the SKEMPI dataset. The likelihood difference only shows weak correlation with experimental $\Delta \Delta G$, which elucidates the gap between the structural compatibility information learned by ProteinMPNN and the experimental catalytic performance.
The visionary aim of this project is to close the loop between computational predictions and experimental validation via laboratory automation. To achieve this, we implement an assaywise alignment strategy that integrates the deep learning model’s predictions with experimental data from enzymatic assays.
We repurposed the Direct Preference Optimization (DPO) method to align ProteinMPNN with experimental experimental values.
Each mutation entry is treated as an ordered pair of preference:
Intuitively, the DPO loss function increases the likelihood of the favorable amino acid type and penalizes the unfavorable one:
$$\mathcal{L} = \begin{cases} -\log \sigma\left( \beta \log \frac{p_\theta(\text{W}|\mathbf{s}, \mathbf{X})}{p_{\text{ref}}(\text{W}|\mathbf{s}, \mathbf{X})} - \beta\log \frac{p_\theta(\text{M}|\mathbf{s}, \mathbf{X})}{p_{\text{ref}}(\text{M}|\mathbf{s}, \mathbf{X})} \right) & \Delta \Delta G > 0 \ -\log \sigma\left( \beta \log \frac{p_\theta(\text{M}|\mathbf{s}, \mathbf{X})}{p_{\text{ref}}(\text{M}|\mathbf{s}, \mathbf{X})} - \beta\log \frac{p_\theta(\text{W}|\mathbf{s}, \mathbf{X})}{p_{\text{ref}}(\text{W}|\mathbf{s}, \mathbf{X})} \right) & \Delta \Delta G < 0 \end{cases}$$
A limitation of preference optimization using only individual mutations from the dataset is that the model learns if a mutation is favorable or not, but it does not learn to compare different mutations directly. To address this, we augmented the training data by combining pairs of mutations.
Suppose there are two mutations at the same position in the dataset:
We can combine these by inverting the first mutation into D30A ($\Delta\Delta G = +1.0$) and then merging it with the second to create D30K ($\Delta\Delta G = 3.0$).
While data augmentation introduces necessary ordering between mutations, we cannot merge every possible pair. Doing so might introduce artificial ordering between mutations from different experimental settings. Therefore, we use an assay-wise approach: (1) Two mutations are only merged if they belong to the same assay; (2) “same assay” is defined by sharing the same protein structure, the same publication source, and the same experimental method.
By restricting the scope of merging, we guarantee that relative ordering is introduced only within the same assay and not across inconsistent experimental conditions.
The PETase sequence we used is A0A0K8P6T7:
There are multiple structures of PETase available in the Protein Data Bank (PDB). We used the structure with PDB ID 5XFY for ProteinMPNN inference. We scored all 19 possible amino acid substitutions on each position of the PETase sequence and ranked them by the likelihood difference compared to the wild-type amino acid.
According to the predictions, top 10 mutations are:
The codon-optimized for E. coli DNA sequence submitted for synthesis is as follows:
The vector used for cloning is pET-28a(+), which contains a T7 promoter.
I used the Ginkgo CFPS protocol for cell-free protein synthesis of PETase. The reaction mixture contains the following components:
Below is the photo of the pipetting process for creating the CFPS reaction mixture:
The reaction was mixed and centrifuged to remove bubbles. Then, it was incubated at 30°C for 24 hours.
Below is the incubated CFPS reaction mixture (in the leftmost PCR tube) after 24 hours:
To verify the expression of PETase, we performed SDS-PAGE analysis on the CFPS reaction mixture. The expected molecular weight of PETase is approximately 28 kDa.
Below is the protocol followed for SDS-PAGE analysis:
We will perform a two-step dilution to ensure the protein is concentrated enough to see clearly.
Step 1 (Intermediate 1:10 Dilution):
Mix 2 µL CFPS stock + 18 µL PBS.
Total volume: 20 µL.
Step 2 (Final 1:20 Loading Sample):
Mix 10 µL of the 1:10 dilution (from Step 1).
Add 5 µL 4X Bolt LDS Sample Buffer (No DTT added).
Add 5 µL DNase-free water.
Total volume: 20 µL.
Final Concentration: 1:20 CFPS in 1X Sample Buffer.
Load Ladder: Add 3 µL of PageRuler prestained ladder.
Load Samples: Load 15 µL of your 1:20 prepared sample per lane.
Tip: Using 15 µL instead of 10 µL will deliver even more protein to the lane to help with the “faint band” issue.
Run Settings:
Run at 180 V (constant).
Stop when the dye front is near the bottom.
We only have ~500 ng of DNA template from the initial synthesis, which is not enough for multiple rounds of CFPS. To create more DNA template, we first cloned the synthesized DNA into the pET-28a(+) vector and transformed it into E. coli for amplification. After growing the transformed bacteria, we used a mini-prep kit to extract the plasmid DNA.
Below is the protocol followed for the culturing process, adopted from the protocol of Week 6 Gibson Assembly lab:
The next day, I saw colonies on the agar plates, which indicates successful transformation. I picked 3 colonies and grew them in liquid culture for mini-prep.
The liquid culture was heated at 30°C for 12 hours to allow the bacteria to grow.
First, I transferred the liquid culture to centrifuge tubes (1.5mL each) and centrifuged them to pellet the bacteria. Then, I removed the supernatant.
After that, I followed the mini-prep kit protocol to extract the plasmid DNA from the bacterial pellets.
The final eluted DNA was quantified using a spectrophotometer, and I obtained a concentration of approximately 30~50 ng/µL, which is sufficient for more CFPS.
| 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|
| 
| 
| 
| 
|
To evaluate the catalytic capability of the synthesized PETase, I bought a PET film from Amazon. The PET film was cut into small pieces, and was preprocessed by heating at 95°C for 30 minutes to increase its surface area and make it more accessible to enzymatic degradation.
| 
| 
|
|---|---|---|
| PET film bought from Amazon | PET film cut into small pieces and heated at 95°C for 30 minutes | Small PET pieces soaked in the PETase solution |
I soaked the preprocessed PET pieces in the CFPS reaction mixture containing the synthesized PETase for 48 hours at 30°C.
