degraders
Skill56 starsChanged 3 months agoMITView on GitHub
Work on bifunctional degraders and molecular glues, where potency comes from a ternary complex rather than occupancy. Use this skill to apply the property rules that govern this beyond-rule-of-five space, reason about linker length, attachment vector and E3 ligase choice, prepare inputs for ternary complex structure prediction, and interpret degradation readouts — DC50, Dmax, cooperativity, and the hook effect that makes a dose-response curve turn over. Also trigger on PROTAC, molecular glue, targeted protein degradation, E3 ligase, cereblon, VHL, ternary complex, DC50, Dmax, hook effect, cooperativity, or PROTAC-DB.
In this repository36 other files it gives its agents.
- adaptyvSkill
- admet-predictionSkill
- antibody-engineeringSkill
- autodock-vinaSkill
- binding-site-analysisSkill
- boltzSkill
- chemblSkill
- chemical-spaceSkill
- clinicaltrialsSkill
- datamolSkill
- deepchemSkill
- depmapSkill
- diffdockSkill
- esmSkill
- free-energy-perturbationSkill
- generative-designSkill
- glycoengineeringSkill
- immunogenicitySkill
- medchemSkill
- molecular-dynamicsSkill
- molfeatSkill
- ncats-araxSkill
- oligonucleotidesSkill
- openfdaSkill
- open-targetsSkill
- patent-landscapeSkill
- pkpd-translationSkill
- primekgSkill
- protein-binder-designSkill
- pytdcSkill
- rdkitSkill
- retrosynthesisSkill
- rowanSkill
- tamarindSkill
- target-safetySkill
- uniprot-rcsbSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 29
Read references/degrader-modalities.md for how this pharmacology differs, references/ternary-complex.md before modelling or designing a …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Targeted Protein Degradation
A degrader does not inhibit a protein — it recruits an E3 ligase to it, the protein is destroyed, and the degrader is released to do it again. That single change makes an "undruggable" target tractable, because a degrader needs only a binding site, not a functional pocket. It is the largest expansion of small-molecule target space in the last decade, and almost every developability heuristic you know is calibrated for something else.
No installation, no network, no key for the bundled scripts — they implement the property rules, linker arithmetic, and dose-response analysis. Ternary complex prediction needs an external tool (PRosettaC, DeepTernary, AlphaFold3) with its own licence and usually a GPU.
Read references/degrader-modalities.md for how this pharmacology differs, references/ternary-complex.md before modelling or designing a linker, and references/degrader-developability.md before judging a molecule — that one is judgement, not syntax.
The three scripts
| Script | Answers |
protac_properties.py | Is this molecule inside the bRo5 habitable band? |
ternary_setup.py | What do the prediction tools need, and what linker lengths do I make? |
degrader_triage.py | What do DC50, Dmax, and the curve shape actually say? |
Occupancy versus event
| Inhibitor | Degrader | |
| Requires | continuous occupancy | a transient encounter |
| Stoichiometry | 1:1 | catalytic |
| Removes | one function | the whole protein, scaffolding included |
| Duration | drug half-life | protein resynthesis rate |
| Needs | a functional pocket | any ligandable surface |
The last row is the point. No catalytic site, no allosteric mechanism, no functional consequence
of binding required — which is why binding-site-analysis names degradation first when a target
scores undruggable.
The duration row has a practical sting: a degrader can be fully cleared while its effect persists for days, so the usual exposure-response framing does not transfer.
Lipinski rejects every PROTAC
python skills/degraders/scripts/protac_properties.py windows| Property | Window | Failure below | Failure above |
| MW | 700–1100 | not a complete bifunctional | permeability collapses |
| cLogP | 3–7 | too polar at this size | aggregation, promiscuity |
| TPSA | 150–250 | check the molecule is complete | passive permeability lost |
| Rotatable bonds | 8–20 | a degrader has a linker | entropic cost of the ternary complex |
Every window is two-sided, so optimising any of them in one direction is wrong. Filtering a degrader series on Ro5 discards all of it, including the molecules that work.
Permeability is the binding constraint and TPSA cannot see it. Successful oral degraders are molecular chameleons — polar in water, and in a membrane they fold to form intramolecular hydrogen bonds that shield that polarity. Measure PAMPA or Caco-2; do not predict it.
Model the right construct
VHL does not fold without Elongin B/C. CRBN needs DDB1. Predicting against the isolated domain gives a complex that cannot exist:
python skills/degraders/scripts/ternary_setup.py manifest --target 6BOY --e3 vhl
python skills/degraders/scripts/ternary_setup.py toolsThis is not pedantry. Benchmarking against curated crystallographic ternaries found AlphaFold3's apparent performance inflated by exactly those accessory proteins contributing interface area that has nothing to do with the degrader — while PRosettaC outperforms it on the degrader-specific interface. Tool choice here disagrees with the obvious default.
The other setup trap: the attachment atom is not the exit vector. A linker must leave each ligand from a solvent-exposed atom pointing at the partner. Buried atoms cannot be linked from, whatever the docking says.
Scan the linker; do not predict it
Length has a pair-specific window of roughly 4–20 heavy atoms. Too short and the proteins clash; too long and the entropic cost of ordering the complex swamps the interface enthalpy.
python skills/degraders/scripts/ternary_setup.py linkers --min 4 --max 16 --chemistry pegFind the window with a flexible PEG or alkyl series, then rigidify at the optimum — that is where oral exposure comes from, and rotatable bonds are charged twice here, once against permeability and once against cooperativity. Rigidifying first locks in a conformation before you know which one you want.
The hook effect breaks the dose-response
python skills/degraders/scripts/degrader_triage.py curve \
--conc 0.1,1,10,100,1000,10000 --remaining 95,70,25,8,15,60# HOOK EFFECT: degradation falls by 52.0 points above 100. At high concentration the
# degrader saturates both proteins separately, forming binary complexes that cannot
# become ternary.
dmax_pct 92.0
dc50 2.7826A sigmoid fit through a hooked curve returns a confident, meaningless DC50 — the script fits the descending limb only. And "no hook observed" may just mean the range stopped two logs early.
Dmax usually matters more than DC50. DC50 1 nM at 40% leaves most of the protein; DC50 100 nM at 95% removes it. Below ~50% degradation a phenotype is unlikely whatever the potency.
Four things that surprise people
- Degradation potency need not track binary affinity. A weaker binder forming a better ternary complex often degrades better. Ranking a series by target Kd is an expensive mistake.
- A complex that forms may not ubiquitinate. Lysines must be presented to the E2 in the right geometry; forming the interface does not guarantee it.
- CRBN degrades IKZF1/3 whether you asked or not. The IMiD scaffold brings its own neosubstrates.
- Off-target degradation has no inhibitor analogue. Global proteomics is the standard selectivity experiment and there is no computational substitute. Run it early.
Composing with the rest of the bundle
binding-site-analysis→ here: when the pocket scores undruggable, this is the first alternative — a degrader needs a binding site, not an inhibitory pocket.rdkit/datamol→ before: compute the descriptorsprotac_properties.py checkreads.medchem→ carefully: its rule sets are calibrated for Ro5 chemistry and will reject degraders.uniprot-rcsb→ before: co-crystal structures for both ligands, to read the exit vectors.admet-prediction→ carefully: every public ADMET model is extrapolating on bRo5 chemistry.pkpd-translation→ after, with a protein-turnover compartment added.
Reporting results honestly
Give DC50 and Dmax, say whether a hook was seen and over what range, and give cooperativity if measured. Name the E3 and the linker. Report permeability as measured, never inferred from TPSA. Say whether global proteomics selectivity has been run. And when quoting property windows, call them bRo5 conventions — the molecule is supposed to violate Lipinski.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

