Huge thanks to Colin, Minkyu, and Aymen for all the help, and @alisiafadini.bsky.social, @moalquraishi.bsky.social for the guidance!!
These results suggest that AF3 models may learn (uncalibrated) implicit conformational landscapes, and that the right interface can unlock new capabilities without retraining. We're excited to explore interpretability, transferability, and applications of ConforNets.
🧵 7/8
2. Learn a desired state from one protein that can then be induced in many other proteins
🧵3/8
We introduce ConforNets, a mechanism for conformational control in AlphaFold3 models
- SoTA at producing diverse conformations on every multistate benchmark (N=104)
- Novel capability: transfer state from one protein to another
Outperforms BioEmu, ConforMix and AFsample3
🧵1/8
We train “conformational prompts” on 3 therapeutically important protein families and show strong induction of desired states:
- GPCR active: 24% → 79%
- Kinase DFG-out: 6% → 23%
- Transporter outward-facing: 16% → 57%
Built-in templates cause no observable shift.
🧵6/8
On multi-state benchmarks such as membrane transporters, ConforNets double recovery of alternate states from 24.3% → 51.1% over the OF3p baseline, at the cost of 3 GPU-minutes for a 400-residue protein.
🧵4/8
Yet, the above benchmarks only ask whether an alternate state is observed among hundreds of samples. Can we prompt the model to consistently predict states that are rarely sampled? A "conformation prompt”?
🧵 5/8
arXiv: arxiv.org/abs/2604.18559
Code: github.com/aqlaboratory...
All training, inference, and benchmarking data/code is open-sourced!
🧵 8/8
Our key idea is to operate globally across Pairformer channels, instead of residues, to broadly sculpt conformational preferences. This makes it possible to:
1. Maximize unsupervised objectives like conformational diversity by coordinately training multiple ConforNets
🧵2/8
