Wei-Tse Hsu
@weitse-hsu.bsky.social
- Postdoc in Drug Design at Oxford Biochemistry (Biggin Lab).
- Ph.D. from the Shirts Group at CU Boulder.
- Keen on compchem, deep learning & education.
- Rookie runner.
- Originally from Taiwan.
- Check my MD tutorials: https://weitsehsu.com/
- Ph.D. from the Shirts Group at CU Boulder.
- Keen on compchem, deep learning & education.
- Rookie runner.
- Originally from Taiwan.
- Check my MD tutorials: https://weitsehsu.com/
🚀 Bottom line:
With careful filtering, co-folding predictions can indeed teach ML about binding affinity.
👉 Read the full JCIM paper: pubs.acs.org/doi/full/10....
Work with Aniket Magarkar
@boehringerglobal.bsky.social and @philbiggin.bsky.social @ox.ac.uk
(6/6)
With careful filtering, co-folding predictions can indeed teach ML about binding affinity.
👉 Read the full JCIM paper: pubs.acs.org/doi/full/10....
Work with Aniket Magarkar
@boehringerglobal.bsky.social and @philbiggin.bsky.social @ox.ac.uk
(6/6)
Can AI-Predicted Complexes Teach Machine Learning to Compute Drug Binding Affinity?
We evaluate the feasibility of using co-folding models for synthetic data augmentation in training machine learning-based scoring functions (MLSFs) for binding affinity prediction. Our results show th...
pubs.acs.org
January 20, 2026 at 7:27 PM
🚀 Bottom line:
With careful filtering, co-folding predictions can indeed teach ML about binding affinity.
👉 Read the full JCIM paper: pubs.acs.org/doi/full/10....
Work with Aniket Magarkar
@boehringerglobal.bsky.social and @philbiggin.bsky.social @ox.ac.uk
(6/6)
With careful filtering, co-folding predictions can indeed teach ML about binding affinity.
👉 Read the full JCIM paper: pubs.acs.org/doi/full/10....
Work with Aniket Magarkar
@boehringerglobal.bsky.social and @philbiggin.bsky.social @ox.ac.uk
(6/6)
🔎 SI highlights:
- AEV-PLIG beats Boltz-2 in 4 target classes in the FEP benchmark (loses 1, ties 6); both are competitive with FEP+ in some cases.
- ipLDDT & ligand pLDDT are also effective filters; pTM, PAE, PDE are not
- Boltz confidence seems to generalize better than its structure module
(5/6)
- AEV-PLIG beats Boltz-2 in 4 target classes in the FEP benchmark (loses 1, ties 6); both are competitive with FEP+ in some cases.
- ipLDDT & ligand pLDDT are also effective filters; pTM, PAE, PDE are not
- Boltz confidence seems to generalize better than its structure module
(5/6)
January 20, 2026 at 7:27 PM
🔎 SI highlights:
- AEV-PLIG beats Boltz-2 in 4 target classes in the FEP benchmark (loses 1, ties 6); both are competitive with FEP+ in some cases.
- ipLDDT & ligand pLDDT are also effective filters; pTM, PAE, PDE are not
- Boltz confidence seems to generalize better than its structure module
(5/6)
- AEV-PLIG beats Boltz-2 in 4 target classes in the FEP benchmark (loses 1, ties 6); both are competitive with FEP+ in some cases.
- ipLDDT & ligand pLDDT are also effective filters; pTM, PAE, PDE are not
- Boltz confidence seems to generalize better than its structure module
(5/6)
❓ Are co-folding predictions good enough to train scoring functions?
👉 Yes — with careful filtering. We see no performance difference b/w models trained on:
- experimental structures
- corresponding co-folding predictions
This holds across AEV-PLIG, EHIGN, and RF-Score.
(4/6)
👉 Yes — with careful filtering. We see no performance difference b/w models trained on:
- experimental structures
- corresponding co-folding predictions
This holds across AEV-PLIG, EHIGN, and RF-Score.
(4/6)
January 20, 2026 at 7:27 PM
❓ Are co-folding predictions good enough to train scoring functions?
👉 Yes — with careful filtering. We see no performance difference b/w models trained on:
- experimental structures
- corresponding co-folding predictions
This holds across AEV-PLIG, EHIGN, and RF-Score.
(4/6)
👉 Yes — with careful filtering. We see no performance difference b/w models trained on:
- experimental structures
- corresponding co-folding predictions
This holds across AEV-PLIG, EHIGN, and RF-Score.
(4/6)
❓ When can we trust a co-folding prediction?
👉 From reproducing HiQBind with Boltz-1x, a few simple heuristics are recommended high-quality cofolding augmentation:
1️⃣ single-chain systems
2️⃣ Boltz confidence > 0.9
3️⃣ train–test similarity > 60%
(3/6)
👉 From reproducing HiQBind with Boltz-1x, a few simple heuristics are recommended high-quality cofolding augmentation:
1️⃣ single-chain systems
2️⃣ Boltz confidence > 0.9
3️⃣ train–test similarity > 60%
(3/6)
January 20, 2026 at 7:27 PM
❓ When can we trust a co-folding prediction?
👉 From reproducing HiQBind with Boltz-1x, a few simple heuristics are recommended high-quality cofolding augmentation:
1️⃣ single-chain systems
2️⃣ Boltz confidence > 0.9
3️⃣ train–test similarity > 60%
(3/6)
👉 From reproducing HiQBind with Boltz-1x, a few simple heuristics are recommended high-quality cofolding augmentation:
1️⃣ single-chain systems
2️⃣ Boltz confidence > 0.9
3️⃣ train–test similarity > 60%
(3/6)
❓ How much can data augmentation actually improve scoring?
👉 Short answer: only if the added data are high-quality. Adding BindingNet v1 clearly improved performance, but v2 did not—despite being 10x larger—due to its substantially lower quality.
Quality beats quantity.
(2/6)
👉 Short answer: only if the added data are high-quality. Adding BindingNet v1 clearly improved performance, but v2 did not—despite being 10x larger—due to its substantially lower quality.
Quality beats quantity.
(2/6)
January 20, 2026 at 7:27 PM
❓ How much can data augmentation actually improve scoring?
👉 Short answer: only if the added data are high-quality. Adding BindingNet v1 clearly improved performance, but v2 did not—despite being 10x larger—due to its substantially lower quality.
Quality beats quantity.
(2/6)
👉 Short answer: only if the added data are high-quality. Adding BindingNet v1 clearly improved performance, but v2 did not—despite being 10x larger—due to its substantially lower quality.
Quality beats quantity.
(2/6)