Note: It is about competition and is not affiliated with the company
I participated in both rounds of the competition.
Round 1: No Luck
Round 2: https://foundry.adaptyvbio.com/competition?design=97854d19-791b-4f22-8d60-18412ff193ff
Approach: Having participated in PDC round 1, I noticed a few things. First, although smaller sequences tended to get low IPAE scores (the primary metric used in PDC1), they did not produce better experimental results. Second, I only experimented with RFDiffusion workflow. At the same time, I obtained low IPAE structures and had two of these structures tested (based on the above workflow), neither bound effectively despite showing high expression. This could be due to two reasons: 1. Steric hindrance (Protein has many conformations) and 2. The binder doesn't fold into the predicted structure (a predicted structure is unstable). Keeping things in mind for round 2, I participated with three main objectives: 1. To experiment with longer sequences (greater than 100 amino acids), as they are more likely to form stable structures. 2. To test a workflow other than RFDiffusion. While I am not implying that it is not a good tool, it is computationally intensive and requires filtering thousands of structures, which is challenging to achieve on Google Colab. (I think most participants, including me, rely on Google Colab, which is unsuitable for this task.) 3. I chose to use a BindCraft-based workflow because it had high experimental success rates, and I wanted to understand why this might be the case, so I am still figuring it out. After experimenting with it for a week, I observed a few things. Pros: 1. It generated high-quality structures. (Comparing IPAE and ipTM scores, it rejects many structures in the background.) Cons: 1. It is much more compute-intensive than RFDiffusion. 2. It would be better to pass the whole structure, as it would likely lead to a better binder. However, it is so compute-intensive that passing the whole structure for a longer target is impossible. 3. Out of all the workflows I ran, there were two instances where, even after running for 24 hours, it didn’t converge (no structures generated). Also, even though Domain 1 has more epitopes than Domain 3, it was harder for the algorithm to generate a binder for Domain 1 than for Domain 3. Domain 3 has a higher amino acid count than Domain 1. Overall, I think BindCraft is also an interesting approach to binder generation. Like last time, I won’t keep high hopes. However, I hope some of the binders from this round are worthy of further research.
Problem Statement: https://design.adaptyvbio.com/
Protein Design Tools: https://design.adaptyvbio.com/tools
About EGFR: Epidermal Growth Factor Receptor (EGFR) is a transmembrane protein that plays a critical role in cell growth, differentiation, and survival. It is frequently overexpressed or mutated in various cancers, including non-small cell lung cancer, colorectal cancer, and head and neck cancer. This makes EGFR a crucial target for cancer therapies such as Cetuximab, an antibody with more than 1B USD in annual revenue. (Taken from Adaptyv Bio page)
Every protein designer can submit up to 10 designs.
Designs must adhere to the following criteria:
- Only natural amino acids
- Max 200 amino acids in length
- Single chain
- At least 10 amino acid edit distances from known binders
Designs will be ranked based on PAE_interaction with AF2 as described here. The top 100 designs will be selected for experimental testing. Additional designs will be chosen for their novelty, creativity in the design process, or other interesting factors.
Selected designs will be tested in our Affinity Characterization workflow, which includes protein expression and a high-quality kinetics assay to determine the protein's binding affinity to the target.
You can find more information in our RFdiffusion case study: https://www.adaptyvbio.com/blog/rfdiff_il7ra
PDB Protein link: https://www.rcsb.org/sequence/6aru
RFDiffusion: https://www.bakerlab.org/2023/07/11/diffusion-model-for-protein-design/
PyRosetta: https://www.pyrosetta.org/home
Paper link: https://www.nature.com/articles/s41586-023-06415-8
ColabDesign: https://github.com/sokrypton/ColabDesign
ColabFold: https://github.com/sokrypton/ColabFold
Binder design verification: https://github.com/nrbennet/dl_binder_design
paper link: https://www.nature.com/articles/s41467-023-38328-5
Peptide Binders With RFdiffusion: https://www.youtube.com/watch?v=nqzr9rNHlxA
Faster AlphaFold protein structure predictions using ColabFold: https://www.youtube.com/watch?v=gIbCAcMDM7E&t=2s
Convert PDB to FASTA: https://zhanggroup.org/pdb2fasta/pdb2fasta.cgi
How to design a binder: https://github.com/RosettaCommons/RFdiffusion#binder-design