BINMAN
Blind-spot INventory of Molecular Adhesives and Neosubstrates
Selection state Nothing pinned yet. GLUE E3 TARGET

About BINMAN

generated 2026-10-16T05:25:18+00:00

Workflow

BINMAN pipeline schematic Described in the text below the figure. Inputs SourcesRCSB, AFDB,UniProt Ground truthPublisheddatasets 1 · Acquisitionhttpx, tenacity 130,391 entries tier 1 3,871 tier 2 34,222 tier 3 17,375 Fetch every PDB entry that could hold a glue. 2 · Geometrygemmi, FreeSASA 55,449 entries analysed 285,996 bridges found 19 failed Measure which ligands touch two proteins at once. 3 · Scan and triageDSSP, fpocket 21,717 degrons scanned 650 ligases triaged 1,650 lysines scored 582 failed Score what a degrader needs: tags, ligases, lysines. 4 · Modelmlx-lm, LoRA Qwen2.5-3B, 4-bit rank 8 over 32 layers one adapter, three heads Train one small model on the atlas's own records. Modules Glue Atlas Degron Scan E3 Triage Degradability BINMAN-LM · one adapter, three heads, no arithmeticbinman-qwen-2.5-3b-4bit-round15 A · query your question becomes a query object when you ask, then the parser checks it B · triage evidence class on 27,590 bridge rows at build time, the one model-derived column C · abstain answerable or not, and what is missing when the parser rejects a proposal
Described in words

The pipeline runs left to right in four stages. Two kinds of input enter on the left: the primary data sources (RCSB, AlphaFold DB, UniProt, InterPro), and separately the validation datasets that serve as ground truth. Stage 1, acquisition, catalogued 130,391 entries into priority tiers using httpx and tenacity. Stage 2, geometry, analysed 55,449 entries with gemmi and FreeSASA and found 285,996 bridges, with 19 entries failing. Stage 3, scan and triage, used DSSP and fpocket to produce 21,717 degron candidates, 650 triaged ligases and 1,650 scored lysines. Stage 4 trains BINMAN-LM with mlx-lm, which does text work only and never emits a number. Four modules leave on the right: Glue Atlas, Degron Scan, E3 Triage and Degradability. In plain terms: stage 1 fetches every PDB entry that could hold a glue, stage 2 measures which ligands touch two proteins at once, stage 3 scores what a degrader needs (tags, ligases and lysines), and stage 4 trains one small model on the atlas's own records. Below the four stages sits the trained layer: binman-qwen-2.5-3b-4bit-round15, one LoRA adapter with three heads. Head A, query, turns a question into a query object when somebody asks, and the parser then checks it. Head B, triage, assigns an evidence class to 27,590 bridge rows at build time, and is the only model-derived column in the atlas. Head C, abstain, says whether a question is answerable and what is missing when it is not, and is called when the parser rejects a proposal. None of the three produces a number.

Stage counts

StageRecords FailedWhy
bridges 55,468 19
catalogue 130,391 0
degrons 20,431 570
interface_persistence 40 0
ligases 650 12
lm_eval 14 0
lysines 12 0
triage_predict 6,510 0
trimmed_structures 3,839 0
validation_acquire 13 2

Ground truth

Step 1 · found in the atlas

9OHL · A1CBE

TRMT112-METTL5 bound to SAM and FWG-33B

LigandN-(4-{[(3aS,4R,6S,6aR)-6-(3-chlorophenyl)-2-propanoyloctahydrocyclopenta[c]pyrrol-4-yl]oxy}phenyl)acetamide
MethodX-RAY DIFFRACTION
Resolution1.29 Å
ΔSASA chain C/A442.3 Ų
ΔSASA chain D/B581.2 Ų
Bridging balance0.761
Buried fraction0.795
Heavy-atom contacts27 / 44
Ligand heavy atoms30

This passes the bridging filter because it buries more than the minimum area against both chains at once and makes enough heavy-atom contacts to each. Anything burying surface against only one chain, or classified from its structure as crystallisation furniture, is counted and kept but not called a bridge. 9OHL records 1 bridge(s) in total, 1 of them for A1CBE across different chain pairs; the one above is the widest of them.

nothing pinned
Loading the ternary complex.

Step 2 · does its substrate carry a degron?

TRMT112 (Q9UI30) carries a hairpin at residues 91 to 110, tip G104.

Tip residue G104
Hairpin 91–110
Mean pLDDT 88.37
Tip relative SASA 0.308
Geometry score 0.632

Accessions in this entry: Q9NRN9 Q9UI30

nothing pinned
The AlphaFold model loads here once the Degron Scan has run.

Step 3 · which ligase would you recruit?

Neither protein here is an E3 ligase, which is the usual case for a bridge found by geometry rather than by annotation. The question this module answers is which ligase you would recruit, and it ranks the human repertoire on ligandability, structural coverage, expression selectivity and family. The top of that ranking:

LigaseRank PocketStatus
DCAF15 1 0.988
DCAF8L1 DCAF 2 0.86
CCNF F-box 3 0.956
ASB8 SOCS-box 4 0.982

This ranking is of the repertoire, not of its fit to this particular target. Pairing a ligase with a substrate is what the Lens graph is for.

BINMAN-LM model card

base modelQwen/Qwen2.5-3B-Instruct
fine tune typelora
adapter rank8
adapter alpha160.0
adapter layers32
fusedFalse
adapter Dellboy/binman-lm-adapter
demo space https://huggingface.co/spaces/Dellboy/binman-lm
source https://github.com/bellcheddar/BINMAN

References

How to cite BINMAN

Deller, M. C. BINMAN: Blind-spot INventory of Molecular Adhesives and Neosubstrates.
https://github.com/bellcheddar/BINMAN

Cite BINMAN for the atlas itself, and cite every underlying resource separately: the atlas is a derivative of the databases listed in the reference table, and they deserve their own citation.