Rowan

Some problems (protein folding) are a good fit for simulation, while others (lit review) are good for LLMs—the toughest problems, however, span both categories. Our latest Rowan post examines one such case: NMR-based structure elucidation by AI agents. The forward structure-to-NMR-spectrum problem is well-studied and can be done by DFT or ML models. In contrast, the inverse spectrum-to-structure task is hard and typically done by humans, although recent Anthropic work ("Making Claude a Chemist") showed that Claude is decent: https://lnkd.in/gVG_fc48 We imagined that giving an LLM access to spectrum-prediction tools could give it the ability to check its own work and find the correct answer—to test this, we examined cases where NMR spectra had been misassigned in the literature and evaluated models on their ability to return the correct SMILES string from the experimental 13C NMR shieldings. Broadly, our hypothesis appears correct: GPT-5.6 Sol runs using tools have a higher success rate than baseline runs, with increasing returns to effort. On "xhigh" effort, Sol gets 52% of the cases correct with NMR tools but only 21% correct without tools. External tool calls provide a nice way to follow agents' reasoning. Agents often quickly determine the correct functional groups but require multiple rounds of simulation to figure out relative stereochemistry or connectivity, illustrating why simulation helps. Agents using tools also reason for longer and burn more tokens than regular agents, with correspondingly better performance. One interpretation of this is that simulation tools allow agents to test their hypotheses and reason coherently for longer (the "world model" view). We think that combining design and simulation tools with AI agents is the future of computational chemistry, and we're excited to continue benchmarking + building in this space! Read the full blog post here: https://lnkd.in/gwk5iQ4Q

Many scientists new to structure-based drug discovery intrinsically grasp how and when to run docking or co-folding, but aren't quite sure what to do once they've found a promising pose. In our latest blog post, "What to Do with a Pose," we provide a quick guide to the various next steps that scientists take here: visual inspection, short MD trajectories, interaction analysis, and more. https://lnkd.in/gmJittni

Rowan's new protein-preparation workflow makes it simple to model proteins even with substantial unresolved regions. Here's 5-HT2A from PDB 9AS8, where only the transmembrane helices are resolved; the workflow preserves the resolved structure while adding the unresolved loops:

Rowan's new protein-preparation workflow makes it simple to model proteins even with substantial unresolved regions. Here's 5-HT2A from PDB 9AS8, where only the transmembrane helices are resolved; the workflow preserves the resolved structure while adding the unresolved loops:

We're sharing a big set of structure-based drug design improvements today! - We've added a new protein preparation workflow that uses Boltz-2 or PDBFixer to fill in missing residues & atoms, caps chains, offers multiple protonation methods, and outputs MD-ready systems. This allows e.g. running MD/FEP on proteins with large unresolved crystallographic regions, which was very painstaking before. - We've added Gnina as a docking engine. Gnina scores poses with a neural network and outperforms Vina & other docking methods at pose retrieval and binder enrichment on a number of benchmarks. It can also be used for covalent docking, which is handy! - Our co-folding pipeline now supports folding CCD-specified non-canonical residues and nucleotides. These can be specified when using Boltz, Chai, or OpenFold. - We now run MM/GBSA single points on docking, co-folding, and analogue docking poses when requested. MM/GBSA interaction energies can be used filter out unphysical poses + better resolve ligand geometries. - We've added a number of new outputs to our MD workflows to help with analysis including SASA, polar SASA, radius of gyration, RMSF, new trajectory clustering options, MM/GBSA, and hydration-site analysis. Our MD is also twice as fast now thanks to hydrogen-mass repartitioning (>1 µs/day on dasatinib•ABL, e.g.). - Our FEP graph creation can now automatically identify and add “missing” compounds. These synthetic intermediates help break difficult alchemical transformations into multiple easy transformations. - Finally, our FEP workflow now supports iteratively adding and removing ligands without re-running already-computed legs. This enables tighter computational coupling with real-world iterative design cycles. These changes address problems and limitations that we learned about from real-world testing + benchmarking with customers, pilots, and hackathons. We're excited to continue making Rowan's CADD offerings more powerful, easier to use and gain insight from, and faster. Read the full release newsletter here: https://lnkd.in/gf7vNe_c

We've been working on making Rowan's FEP workflow easier to use and more intuitive since launch. Here's a few of the changes we've made: (1) Rowan can now show the atom mapping for each leg, allowing users to double-check if atoms in the two ligands are correctly being assigned. (2) You can now automatically view the overlay of the two ligands, with or without the protein pocket, to check that the poses are properly aligned (and gain insight into why the affinity might change). (3) Each leg shows the windows from the automatic lambda scheduler; for well-behaved legs like this one, the overlap between all adjacent windows is green.(Easy perturbations take fewer lambda windows to reach convergence, while difficult perturbations require more.) (4) The thermodynamic consistency of the output is automatically checked. While perfect RBFE calculations should sum to zero along every close cycle, in practice cycles often have a small non-zero sum. Rowan checks every cycle and automatically highlights it on the RBFE graph. We're continuing to iterate with customers to make Rowan's FEP more powerful and easier to use. If you're interested in fast, accurate, & intuitive FEP calculations to accelerate lead optimization, try Rowan!

Today Rowan's launching openconf, an open-source package that uses Monte Carlo simulations to generate conformer ensembles for drug-like molecules. openconf is built for speed. With default settings, openconf generates 50–100 conformers for maraviroc in <3 s on my laptop; with "rapid" settings, openconf runs in <100 ms for drug-like fragments (fast enough for large-scale virtual screening). Before openconf, we were continually struggling with accuracy in high-throughput contexts: ETKDG works well for a few conformers but often fails to recapitulate the whole ensemble, while CREST struggles to run quickly enough. This caused issues in pKa, strain, FEP posing, &c and was a constant source of decreased accuracy. openconf combines a lot of known tricks: clever Monte Carlo moves, ETDKG seeds, low-mode searching, adaptive move-probability reweighting, iterative pool deduplication, and early stopping. Switching Rowan workflows to openconf has led to dramatically improved runtimes (>4x for pKa) and increased accuracy in many cases. We put extra care into macrocycles, a notoriously tough case. openconf uses kinematic inverse closure, crankshaft moves, and ring amide flips to generate decent ensembles for cyclic peptides & other macrocycles. Pictured below is a ensemble (lowest 20 conformers shown) generated with default settings on the pan-KRAS inhibitor AMG 401; the run took 12 seconds and generated 177 distinct conformers. openconf is freely available on GitHub under the MIT license, and we want people to use it! Consider trying it out if you're working in this space, and feel free (or even encouraged) to leave an issue or open a PR if you find problems. Here's the link: https://lnkd.in/gddHSj3N You can find many more technical details + updates on our other open-source scientific projects in our latest Rowan newsletter: https://lnkd.in/gXxCat9q Update 6/18/26: a few comments pointed out that the MMFF94 energy rankings are quite bad. This is true! openconf returns a lot of conformers by default, with the intention that you can use your favorite method (NNP, xTB, DFT, etc) to reoptimize + rescore the ensemble. Depending on what you're hoping to do, it can also make sense to cluster the conformers to get a small-but-diverse ensemble; we've written about Rowan's approach to this problem on our newsletter (https://lnkd.in/gN7pH4zj).