# AI use cases for Boltz (jwohlwend/boltz) Repo: [jwohlwend/boltz](https://github.com/jwohlwend/boltz) | Use cases: 3 | Page: https://dedupedai.com/use-cases/jwohlwend-boltz/ Boltz is a family of open models for biomolecular interaction prediction. Boltz-1 was the first fully open source model to approach AlphaFold3 accuracy, and Boltz-2 jointly models complex structures and binding affinities, approaching the accuracy of physics-based free-energy perturbation (FEP) methods while running about 1000x faster. Code and weights are MIT licensed for academic and commercial use, and the boltz command runs predictions from YAML input files. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Predict a protein complex structure with Boltz: https://dedupedai.com/use-cases/jwohlwend-boltz/predict-a-protein-complex-structure-with-boltz/ - Screen small molecules for likely binders: https://dedupedai.com/use-cases/jwohlwend-boltz/screen-ligands-for-binders-with-boltz-affinity/ - Compare small changes to a lead molecule: https://dedupedai.com/use-cases/jwohlwend-boltz/compare-ligand-modifications-with-boltz/ ## How to predict a protein complex structure with Boltz with Boltz Repo: [jwohlwend/boltz](https://github.com/jwohlwend/boltz) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Describe your biomolecules in a YAML file and let Boltz predict the structure of the complex. **Who it is for:** A structural biologist who wants a quick open-source structure model of a complex before setting up experiments. **The problem:** Getting a structure prediction for a complex usually means a hosted service, a queue or a closed model. **The outcome:** Predicted complex structures written by the latest Boltz model on your own machine. ### Steps 1. Create a fresh Python environment, as the README recommends 2. Install with `pip install boltz[cuda] -U` (drop [cuda] on CPU-only or non-CUDA GPU hardware) 3. Write an input YAML file describing the biomolecules you want to model, following docs/prediction.md 4. Run `boltz predict input_path --use_msa_server` with input_path pointing at the YAML file 5. Check every option with `boltz predict --help` ### Prompt to paste into your agent ```text Using jwohlwend/boltz, install it in a fresh environment, write an input YAML for following docs/prediction.md, and run `boltz predict input_path --use_msa_server`. Tell me where the predicted structures were written. ``` ### Caveat Predictions are model outputs, not experimental structures; validate anything you act on. Risk, platform: The CPU version is significantly slower than the GPU version, so large complexes realistically need an NVIDIA GPU. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/jwohlwend-boltz/predict-a-protein-complex-structure-with-boltz/ ## How to screen small molecules for likely binders with Boltz Repo: [jwohlwend/boltz](https://github.com/jwohlwend/boltz) | Difficulty: Advanced | Time: half a day | Basis: From the README Use Boltz-2's affinity output to separate likely binders from decoys in a hit-discovery screen. **Who it is for:** A drug discovery scientist triaging a list of candidate ligands against one target. **The problem:** Physics-based free-energy methods are accurate but far too slow to run on every candidate in a screen. **The outcome:** A per-ligand affinity_probability_binary score from 0 to 1 that ranks which molecules look like binders. ### Steps 1. Install with `pip install boltz[cuda] -U` 2. Write one YAML per protein and ligand pair that asks for the affinity property, as described in docs/prediction.md 3. Put the YAML files in one directory and run `boltz predict input_path --use_msa_server` on that directory for batched processing 4. Read affinity_probability_binary for each ligand to detect binders from decoys 5. Send only the top candidates on to slower methods or the lab ### Prompt to paste into your agent ```text With jwohlwend/boltz, build a directory of YAML inputs pairing with each ligand in , requesting affinity, and run `boltz predict input_path --use_msa_server` on it. Rank the ligands by affinity_probability_binary. ``` ### Caveat The README says affinity_probability_binary is for hit discovery and affinity_pred_value for lead optimization; do not mix them up. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/jwohlwend-boltz/screen-ligands-for-binders-with-boltz-affinity/ ## How to compare small changes to a lead molecule with Boltz Repo: [jwohlwend/boltz](https://github.com/jwohlwend/boltz) | Difficulty: Advanced | Time: half a day | Basis: From the README Use affinity_pred_value to see how small modifications change predicted binding during hit-to-lead work. **Who it is for:** A medicinal chemist choosing which analogs of a lead compound to synthesise next. **The problem:** Synthesising and assaying every analog is slow and expensive. **The outcome:** Predicted log10(IC50) values for each analog, so you can prioritise which ones to make. ### Steps 1. Install from GitHub for daily updates if you want the newest code: `git clone https://github.com/jwohlwend/boltz.git` then `cd boltz; pip install -e .[cuda]` 2. Write YAML inputs for the lead and each modified analog against the same target, requesting affinity 3. Run `boltz predict input_path --use_msa_server` on the directory of inputs 4. Compare affinity_pred_value, reported as log10(IC50) derived from an IC50 in micromolar, across analogs 5. Use the results to rank which analogs to synthesise first ### Prompt to paste into your agent ```text Using jwohlwend/boltz, predict affinity for my lead compound and these analogs against , then make a table of affinity_pred_value per analog with a note on which modifications helped. ``` ### Caveat Updated Boltz-2 evaluation and training code are marked as coming soon in the README. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/jwohlwend-boltz/compare-ligand-modifications-with-boltz/