from most recent Harvard lawsuit. sums it up pretty succinctly I think

Lastly, Kaia checked to see if EPIC and its higher affinity mutant are able to protect exatecan from hydrolysis, which is not something serum albumin can do. For a drug that normally hydrolyzes in a few hours, EPIC was able to stabilize the lactone form for days! ✅

Since EPIC and exatecan aren't in the PDB, we wanted to see how co-structure predictors do on it. They each get the backbone right but differ at the ligand. The pose is correct but the modeling of the conformer is wonky. AF3 does the best. AF3 is also able to rank affinities via pLDDT of ligand! 😱

Kaia was able to crystalize EPIC and determine its structure to 2.0 Å resolution. It agreed pretty well with the LASEr design! RFAA had a hard time modeling the lactone ring of the drug, so there is some disagreement there. The lactone is buried as intended, and the goal was to hide it from water 👍

Ben didn't stop there. He wanted to improve affinity of EPIC for exatecan using computation alone. He used LASErMPNN to "proofread" EPIC's sequence using a predicted co-structure as input. LASEr suggested two mutations. Kaia verified that each improved binding 10x. 100x when combined (1 nM Kd)!

Kaia Slaw (no bluesky) experimentally tested 4 designs from NISE and 16 from COMBS. All 4 NISE designs bound! The highest affinity binder- which Ben and Kaia call "EPIC" - was pretty tight (0.1 uM Kd). Compared to COMBS (3 of 16 bound, tightest was 10 uM), NISE and LASErMPNN did a much better job!

Ben used NISE and LASErMPNN to design binders to exatecan, an anticancer drug prone to inactivation by hydrolysis. We also used a more "traditional" approach using COMBS and Rosetta to design binders. We could compare the methods head to head.

With the new co-structure predictors like RFAA, Boltz-1, and AF3, we can now extend self-consistency into the ligand dimension. And Ben's NISE algorithm maximizes this. Code repo here: github.com/polizzilab/N...

We all know in protein design about the goal of self consistency. That is, we want the predicted structure to look like the structure for which we designed the sequence.

Ben used LASErMPNN in combination with a protein-ligand co-structure predictor, RFAA, in an iterative algorithm called NISE that refines designs. NISE optimizes the sequence, structure, and ligand conformer together to improve the confidence of both models. It's a neural-network-only algorithm

Ben Fry (@benf549.bsky.social) was excited when proteinMPNN came out, which motivated him to train a new gNN called LASErMPNN to design sequences given protein-ligand co-structure. LASErMPNN does pretty well at this! The repo is available and even has the training code! github.com/polizzilab/L...