Am I Hallucinating or Can AI Now Design Cancer-Curing Antibodies?

Colby T. Ford, PhD
4 min readJun 8, 2022


Your feed may have recently been flooded with cool, AI-generated images rendered from tools like OpenAI’s DALL-E 2 or other models. The concept of an AI system creating new information (rather than just interpreting it) is called “hallucination”. So, does this mean that we can cause other models like, say, AlphaFold2 to hallucinate new proteins for us? Yes!

Some DALL-E Hallucinations for “a robot with an antibody”

Making a Faux Pembro

Pembrolizumab is an immune checkpoint inhibitor that targets PD-1. This monoclonal antibody, sold under the brand name Keytruda by Merck, allows the body’s T cells to clear tumor cells by inhibiting the immune system’s checkpoint process that would otherwise tell the T cells to skip over them. This popular drug has been use in combination therapies for cancers ranging from non-small cell lung cancer to melanoma to breast cancer. Pretty cool.

Pembrolizumab Fab (in green) bound to PD-1 (in purple). PDB: 5JXE. Made with Protein Imager.

I asked the AlphaFold2 system to hallucinate a new pembro-like Fab. Luckily, the amazing Sergey Ovchinnikov has a new Google Colab notebook for just this sort of task. I simply provided the PDB ID: 5JXE and specified the chains I wanted to use as a reference. Then, after a bit, the system came up with a new Fab molecule. I give you…


“Pemfauxlizumab” made with ColabDesign and AlphaFold2

Big Dock Energy?

So, how does the binding affinity compare between pembrolizumab and my made up pemfauxlizumab? Using HADDOCK 2.4, I docked each Fab structure with PD-1 to compare the metrics.

As you can see from the metrics above, pembrolizumab is certainly superior at binding to PD-1 compared to pemfauxlizumab. The hallucinated Fab has a similar electrostatic energy to pembrolizumab, but the interaction is more distant and has a lower Van der Waals interaction with PD-1.

Comparison of the binding angles of Pembrolizumab (in green) and Pemfauxlizumab (in blue) bound to PD-1 (in purple).

Well, I guess that means I’m not going to make a billion-dollar blockbuster drug out pemfauxlizumab just yet…

Thoughts About Protein Hallucination

Let’s talk about what this hallucination got right…and wrong.

What’s really powerful is the ability for the protein folding system to create another random amino acid sequence that conforms into a very similar tertiary shape as a given reference structure.

Interestingly, these systems have little knowledge about the intricacies of antibody design. So, while pemfauxlizumab (structurally) may look like a normal Fab section of an antibody, its sequence is quite different.

The hallucinated structure’s sequences differ from what I would’ve expected of a real antibody’s Fab chains. While they may fold into similar structures, their sequence similarity is quite low.

If we were to just make this protein in a lab and put it in a human, these vast differences in the sequences could be quite immunogenic. This would likely lessen the antibody’s efficacy and safety, even if it still binds to the target well. While this is cool to generate a novel protein from AI, it’s not likely that we could use this in vivo in its current state.

Even in this current form, I could see hallucinations of protein structures being useful for generating sets of potential sequences that may bind to a desired target. Then, with docking tools, we could evaluate these candidates quickly and experimentally test ones that pass our simulations. This has the potential to revolutionize antibody design.

Going forward, we’ll likely see more and more AI-aided drug design, especially in the development of monoclonal antibody therapeutics. I’m excited for this technology as it may allow us to explore more broadly in silico and think further outside of the box!

Stay curious…




Colby T. Ford, PhD

Cloud genomics and AI guy and aspiring polymath. I am a recovering academic from machine learning and bioinformatics and I sometimes write things here.