11/ Discovery mode: a ~10^8 affibody library against Mdm2–p53 and Myc–Max returned de novo inhibitors of both. The Myc–Max hit is the head-to-head — same proteins, same library design, and binder selections had given us only non-inhibitors.

10/ Near-complete coverage lets us test a live argument in the field. Singles predicting higher-order fitness: r = 0.54. Add doubles (median): r = 0.61. Take the best double-based prediction: r = 0.90. The information is in the doubles — you just can't tell which pairing dominates.

8/ Then we pushed it as a measurement tool. Deep mutational scan (DMS) of pen-Raf: four positions fully randomized (NNK), 160,000 protein variants. We recovered 99% of full-length protein variants - perhaps one of the largest DMS experiments ever performed.

7/ Mock selections: 10 inhibitor phage spiked into 10^10 empty phage. After 3–8 overnight passages the inhibitor takes over — at least 10^14-fold relative enrichment, in days. Works for several important human therapeutic targets: KRas–Raf, Mdm2–p53, and Myc–Max.

5/ On infection RNAP-N is recruited to both targets. The control PPI drives gIII (phage need it). The target PPI drives a dominant-negative gIII (poison). Disrupt the target PPI → poison off → phage replicates. Disrupt anything else → no gIII → dead.

4/ PANCS-Inhibitors makes phage replication depend on PPI disruption - not just binding. Phage carry RNAP-N fused to a zipper (ZP1) plus a candidate inhibitor. The E. coli host holds two trimolecular complexes: your target PPI, and a control PPI (eg ZA–ZB).

2/ The problem: There are 10s of thousands of human PPI targets. Binder discovery is not becoming more routine, but binders interact with hot-spots, which may or may not target a specific interface, and allosteric sites are hard to predict a priori.

I'm becoming this cranky old man....

The results were striking: Female SCN1a+/− mice showed 50% mortality by P50. With CIRTS-4GT3 treatment? Only 13% mortality. We also saw significantly higher seizure thresholds in treated mice—key functional improvements.

We tested this in Dravet syndrome—a severe epilepsy caused by SCN1a haploinsufficiency affecting 1:15,000 people. AAV9 delivery of CIRTS-4GT3 targeting SCN1a to neonatal mice increased NaV1.1 protein ~25% in cortex and hippocampus.

Key advantages of CIRTS-4GT3: (1) Flexible gRNA design targeting 5' or 3' UTRs, (2) fits in single AAV vectors, (3) made entirely from human proteins (reduced immunogenicity), (4) protein boost scales with endogenous mRNA levels—no overexpression toxicity.

We screened 11 translational effector domains and optimized eIF4GI truncations to create CIRTS-4GT3—a compact 601 amino acid activator that doubles target protein expression. It works by recruiting eIF3 and the translation machinery to guide RNA-targeted mRNAs.

9/ Then we went hunting. We profiled NSD3 degradation across ovarian cancer models and found something unexpected: some lines (ES-2) were exquisitely sensitive while others (CAOV-3) were completely resistant—independent of NSD3 expression levels. New biology to explore.

8/ We swapped RNF8's substrate-recognition domain for our NSD3 binder → a mini-protein degrader that potently depleted endogenous NSD3 in colorectal cancer cells and completely blocked proliferation in NSD3-dependent lines.

7/ But binders are just the beginning. We next asked: can we turn these into degraders? We screened 9 E3 ligases and found RNF8—previously unexplored for TPD—was the most potent, driving 90% target depletion.

6/ Key outcome: The binders are all selective and worked in mammalian cells, not just E. coli. We could use them to relocalize proteins in live mammalian cells.

5/ The timeline: Day 1: design constructs. Day 8: genes arrive. Day 17: start selections. Day 20: all 6 selections showed high titers (!). Day 26: sequence-verified, function-validated binders for ALL THREE targets. Affinities ranged from 58 nM to 1.8 µM.

4/ The targets: NSD3 (histone methyltransferase), NMNAT2 (NAD+ biosynthesis), and CSF1R (macrophage receptor)—structurally diverse, clinically relevant, and with few existing targeting tools. A real test.

3/ The setup: We asked oncologist and head of @uchicagocancer.bsky.social Kunle Odunsi to pick 3 cancer targets without telling us in advance. At 5pm on a Tuesday, he emailed us three gene names. The clock started. No cherry-picking. No optimization. Just: can we get binders?

1/ Check out our newest paper where we ask: How fast can we experimentally discover binders from scratch? And we mean scratch: a blinded study. TLDR: 26 days. And the binders work…and led to new cancer biology. We’re coming for you AI…. chemrxiv.org/engage/chemr...

And it worked! Now we have binders that use a secondary hot spot, but that is LC3B selective. This demonstrates the system can target specific epitopes/regions within a single protein - precision at the molecular level! 15/n

So… we used PANCS-Spec-binders to force the selections to find binders that bind the next hottest hot spot on LC3B that IS NOT the LIR motif. This is an “Epitope-specific” selection. 14/n

We figured out this is because the LC3B binders decided to bind at the “LIR motif”, which is shared between GABARAP and LC3B (side note, the LC3B binders were inherently selective – so interesting!). 13/n

Proof of Concept #2 - LC3B Region Targeting: As a second example, we developed binders specific to the LIR (LC3 Interacting Region) of LC3B. This stemmed from our initial PANCS-binder paper, where all our binders were incredibly selective…except for one. Our LC3B binder also bound GABARAP. =( 12/n

Swapping this since amino acid between HRas and KRas completely inverts the specificity! Not what we would have guessed, but that is the power of unbiased discovery. 11/n

How do out binders bind HRas selectively? Mapping the binding modes of our initial non-isoform selective was relatively easy. Alphafold predicted the site of binding, which we confirmed biochemically and by a really nice X-ray structure…9/n

Proof of Concept #1 - RAS Specificity: In our initial PANCS-Binders paper, the binders we got that bound HRas also bind KRas (like N-LHY). This makes sense, since they are almost identical and share a key hotspot. Now, we can create binders like N-WYN that specifically bind HRAS! 8/n

How It Works: The system uses iterative rounds of simultaneous positive selection (bind your target) AND negative selection (avoid similar proteins). 6/n

Why This Matters: Many proteins share "hot spots" - common binding surfaces that make specific targeting nearly impossible. This is especially problematic for protein families like RAS (crucial in cancer), where isoforms are >90% identical but have distinct biological roles. 5/n

The Problem: Protein binders are crucial research tools, but they often suffer from promiscuous binding - hitting multiple similar proteins when you only want ONE specific target. Or, binding your protein, but at the wrong site. Enter: PANCS-spec-Binders! 2/n