Multi-signal computational pipeline for identifying small-molecule inhibitors of TBXT (Brachyury) - chordoma's master transcription factor and lineage-defining oncogenic driver.
A reproducible end-to-end pipeline that integrates docking (Vina, GNINA), generative co-folding (Boltz-2), free-energy refinement (MMGBSA), target-specific QSAR (RF + XGBoost on 653 measured Naar SPR Kd), and T-box paralog selectivity into a strict 7-criterion filter chain - yielding 137 organizer-compliant hit candidates and a top-4 ranked submission.
Built for the TBXT Hit Identification Hackathon 2026 (tbxtchallenge.org), hosted by muni.bio in Boston at Pillar VC, with platform support from Rowan and onepot.ai.
Note on history. This repository is the cleaned, post-hackathon version of the pipeline. The original development repository - including the full commit history from the pre-event build phase and the hackathon weekend - is preserved at
anandsahuofficial/tbxt-hit-id-dev.
If you use this pipeline or adapt it for your own targets, please cite:
Sahu, A. and contributors (2026). tbxt-hit-id: Multi-signal computational pipeline for TBXT (Brachyury) hit identification. https://github.com/anandsahuofficial/tbxt-hit-id
A machine-readable CITATION.cff is included for
automated citation tools (GitHub renders this as a "Cite this
repository" button in the right sidebar).
Chordoma is a rare, locally aggressive cancer of the notochord remnant (skull base, mobile spine, sacrum). It has no FDA-approved targeted therapy and a 5-year survival around 70%. The transcription factor TBXT (Brachyury) is its master regulator and is required for tumor cell survival - silencing TBXT collapses chordoma cell viability. This makes TBXT the most-validated drug target in chordoma, but it is also a long-considered "undruggable" transcription factor with a shallow DNA-binding domain pocket and high T-box paralog homology.
The G177D variant (rs2305089, allele frequency ~0.42) is enriched
in > 90% of Western chordoma cases and creates a unique pocket
(site F: Y88 / D177 / L42) that engages the variant residue directly -
a structural feature absent from the other 16 T-box paralogs and
therefore intrinsically selective.
This repo is the open computational pipeline that selected hit candidates against that pocket from a 570-compound novelty-filtered pool.
| Pool screened | 570 novelty-filtered compounds |
| Pass all 7 strict criteria | 137 organizer-compliant candidates |
| Submitted to experimental program | 24 compounds (4 to judges + 20 first batch) |
| Best predicted Boltz-2 Kd | 3.2 µM (dual-engine 1.02× agreement) |
| Cost to source all 4 picks | $875 via onepot.ai 100% catalog match |
| Site coverage on top 4 picks | 4 / 4 at site F (variant residue D177) |
The 4 picks (results/top4.csv):
| # | ID | Boltz Kd (run A / B) | gnina Vina/pKd | Cost | Risks |
|---|---|---|---|---|---|
| 1 | FM002150_analog_0083 | 3.2 / 3.26 µM | -5.01 / 3.94 | $125 | low/low |
| 2 | FM001452_analog_0104 | 3.7 / 4.97 µM | -5.77 / 4.03 | $250 | med/med |
| 3 | FM001452_analog_0201 | 8.16 / 8.76 µM | -6.07 / 4.69 | $375 | high/med |
| 4 | FM001452_analog_0171 | 8.32 / 8.17 µM | -6.19 / 4.44 | $250 | med/med |
The full 137-candidate pool with every per-criterion pass/fail flag
is in results/all_candidates_tiered.csv.
| Signal | What it catches |
|---|---|
| Vina ensemble (6 receptor confs) | Geometric fit, receptor flexibility |
| GNINA CNN pose + pKd | Vina-trap detection, ML affinity |
| TBXT-specific QSAR (RF + XGBoost on 653 measured Naar SPR Kd) | Target-specific affinity |
| Boltz-2 generative co-folding (two independent backends) | Independent affinity + binder/non-binder classifier |
| MMGBSA implicit-solvent refinement (top 30) | Free-energy refinement |
| T-box paralog selectivity (16 paralogs) | Off-target risk |
Each signal has a known failure mode that another signal in the stack catches. No pick depends on a single score.
C1 onepot.ai 100% catalog match (similarity = 1.000)
C2 strictly non-covalent
C3 Chordoma rule: MW ≤ 600, LogP ≤ 6, HBD ≤ 6, HBA ≤ 12
C4 lead-like ideal: 10–30 HA, HBD+HBA ≤ 11, < 5 rings, ≤ 2 fused
C5 PAINS-clean + no acid halides / aldehydes / diazo / imines /
polycyclic > 2 fused / long alkyl
C6 Tanimoto < 0.85 to Naar / TEP / prior_art_canonical
C7 ESOL log S > -5 (DMSO @ 10 mM + aqueous @ 50 µM)
570 → 137 strict-pass → 4 picks. See docs/filter_chain.md.
| Tier | Definition | Count |
|---|---|---|
| T1 GOLD | All criteria + Kd ≤ 5 µM + low/low risk | 0 (empty by design - honest finding, not overclaimed) |
| T2 SILVER | All criteria + Kd ≤ 10 µM + soluble | 16 |
| T3 BRONZE | All criteria + Kd ≤ 50 µM, borderline solubility | 89 |
| T4 RELAXED | All criteria + Kd ≤ 100 µM | 32 |
- Two independent Boltz-2 runs on separate compute backends: 4 / 4 picks agree within 1.34×
- 10-seed GNINA pose-stability at site F across all 570 compounds: identifies pose-stable picks (σ < 0.05)
- Rowan ADMET (49 properties × 4 picks): all 4 ADMET-profiled
- Rowan pose-analysis MD (explicit-solvent, 5 ns × 1 traj + 1 ns equil): protein-ligand RMSD trajectories captured per pick
- muni.bio
onepottool: all 4 picks at similarity = 1.000 with price + chemistry_risk + supplier_risk attached
Every pick is supported by multiple independent lines of evidence.
The recommended path is container-first - one CUDA-enabled image
shipped from GHCR carries every binary and Python dep, including
GNINA (which is otherwise glibc-fragile on HPC). A native conda
path is also available - see setup/README.md.
# 1. Clone
git clone https://github.com/anandsahuofficial/tbxt-hit-id
cd tbxt-hit-id
# 2. Pull the all-batteries-included container (~6-12 GB; one-time)
bash setup/pull_container.sh
# → ./tbxt-hit-id.sif
# 3. Fetch the receptor + bulk data assets
bash setup/fetch_receptor.sh # PDB 6F59:A from RCSB
bash setup/fetch_data.sh # candidate pool + Naar SPR Kd + receptor ensemble
bash setup/fetch_data.sh --include-poses # OPTIONAL: pre-computed scores (~600 MB, enables --demo)
# 4. Confirm setup is ready (recommended before a multi-hour pipeline run)
apptainer exec --bind $PWD tbxt-hit-id.sif bash setup/smoke_test.sh
# 5. Run the pipeline end-to-end (the container has every binary + Python dep baked in)
apptainer exec --nv --bind $PWD tbxt-hit-id.sif bash examples/reproduce_top4.sh
# or:
apptainer exec --bind $PWD tbxt-hit-id.sif bash examples/reproduce_top4.sh --demo
# 6. Inspect results
cat results/top4.csvTotal runtime - full mode: ~6 hours on a single RTX 4090 (24 GB)
for the 570-compound pool; ~12 hours for the full HPC variant matrix.
Demo mode: < 2 minutes, no GPU. See
setup/README.md for HPC notes, the native conda
path, and the container-internals reference.
Data availability. The bulk data bundle (570-compound pool, Naar SPR Kd training set, 6-conformer crystallographic ensemble, and optional pre-computed scoring outputs) is hosted as a separate Hugging Face dataset and downloaded by
setup/fetch_data.sh. If the dataset is not yet published,fetch_data.shwill print a clear error message; the curated post-pipeline outputs inresults/are committed directly to this repo and can be inspected without any data fetch.
This pipeline is target-agnostic. To screen against a different protein:
- Replace the receptor and pocket centroid in
src/pipeline/define_pockets.py - Drop your candidate compound pool into
data/pool.csv(SMILES + ID columns) - (Optional) Re-train the QSAR model on your target's measured-affinity data
- Re-run
examples/reproduce_top4.sh
The 6-signal architecture, 7-criterion filter chain, tier ranking, and cross-validation harness all generalize.
tbxt-hit-id/
├── README.md ← you are here
├── LICENSE NOTICE ← MIT + provenance
├── AUTHORS.md CITATION.cff
├── environment.yml ← one-command conda env
│
├── docs/ ← deep-dive methodology
│ ├── methodology.md ← 6-signal pipeline
│ ├── filter_chain.md ← 7-criterion strict gate
│ ├── tier_definitions.md
│ └── architecture.png
│
├── slides/ ← judges-facing deck
│ ├── slides.md slides.pdf
│ ├── architecture.png ← pipeline graphic (in deck)
│ └── renders/ ← 2D + 3D pose PNGs (4 picks)
│
├── results/ ← curated post-pipeline outputs (committed)
│ ├── top4.csv top5to24.csv all_candidates_tiered.csv
│ └── selected/ ← cross-val summary, MD attempt log, variants
│
├── src/ ← pipeline source (33 Python modules)
│ ├── pipeline/ ← vina, gnina, boltz, mmgbsa, qsar, paralog, consensus
│ ├── filters/ ← 7-criterion strict gate (+ PAINS + onepot membership)
│ ├── ranking/ ← 4-tier classifier
│ ├── enumeration/ ← one-pot reaction enumeration
│ └── viz/ ← render helpers (2D + 3D pose)
│
├── setup/ ← container + env + cold-start data fetchers
│ ├── Containerfile ← CUDA + GNINA + conda env + src baked in (CI builds → GHCR)
│ ├── pull_container.sh ← apptainer/singularity pull from ghcr.io
│ ├── fetch_receptor.sh ← PDB 6F59:A from RCSB
│ ├── fetch_data.sh ← candidate pool + Naar Kd + receptor ensemble (HF)
│ ├── HPC.md ← Singularity GNINA recipe + Boltz cache + SLURM
│ └── README.md ← quick-start + troubleshooting
│
├── .github/workflows/
│ └── container.yml ← builds + pushes container to ghcr.io on every push to main
│
├── tools/
│ └── render_slides.py ← Markdown → HTML → Chromium PDF renderer
│
└── examples/
└── reproduce_top4.sh ← one-command end-to-end (--full or --demo)
If you use this work, please cite:
@software{sahu2026tbxt,
author = {Sahu, Anand and contributors},
title = {tbxt-hit-id: Multi-signal computational pipeline
for TBXT (Brachyury) hit identification},
year = {2026},
url = {https://github.com/anandsahuofficial/tbxt-hit-id},
note = {TBXT Hit Identification Hackathon 2026, Boston}
}A machine-readable CITATION.cff is included.
- Public computational methods over-predict Kd by 6–25× at the µM regime. Realistic SPR for these 4 picks: 18–200 µM range.
- The pipeline is designed for hit identification, not lead optimization - none of the picks are expected to be sub-µM at SPR without follow-up SAR.
- The T1 GOLD tier is empty by design - no compound in the pool simultaneously hits Kd ≤ 5 µM AND low/low chemistry/supplier risk. This is surfaced honestly rather than gamed by relaxing the gate.
- Team Lead: Anand Sahu - pipeline architecture, methodology, multi-signal consensus integration, final pick selection, live demo
- Team contributors: see
AUTHORS.mdfor the full list of simulation, data-generation, and analysis contributors
- muni.bio -
onepotcatalog membership tool, CLI, and platform credits - Rowan - ADMET, docking, and pose-analysis MD platform
- onepot.ai - virtual catalog and one-pot synthesis library
- Pillar VC - Boston event venue
- TBXT Hackathon - organizers, mentors, and TEPs
- Naar SPR Kd dataset (653 measured TBXT-binding affinities) - TBXT Hackathon 2026
- PDB 6F59 chain A - TBXT G177D + DNA construct (matches the hackathon CF Labs SPR assay)
- muni.bio onepot.ai catalog - ~1.4M one-pot-accessible virtual compounds
Substantial portions of code, scripts, and prose were drafted with AI assistance (Claude) under the team lead's direction. All scientific decisions, parameter choices, and final outputs were reviewed and accepted by the team lead, who is responsible for all outcomes of this work.
MIT - see LICENSE. Use it, adapt it, share it.
Attribution required; see NOTICE for the authorship +
provenance record and CITATION.cff for the
preferred citation.
chordoma · TBXT · Brachyury · T-box transcription factor ·
G177D · rs2305089 · drug discovery · virtual screening ·
hit identification · multi-signal consensus · Vina · GNINA ·
Boltz-2 · co-folding · MMGBSA · QSAR · ADMET ·
MD pose analysis · paralog selectivity · onepot.ai ·
muni.bio · Rowan · Pillar VC · TBXT Hackathon 2026