“Progress in science depends on new techniques, new discoveries and new ideas, probably in that order.” -Sydney Brenner
When people talk about AI in drug discovery, the conversation almost always leaps to the dramatic possibilities: algorithms that will design drugs from scratch, or superintelligent systems replacing entire labs. That’s the story that captures the imagination and the press releases. But it also distracts us. What’s actually happening right now is quieter, less flashy, but in many ways more transformative. It’s not AI discovering drugs. It’s AI writing the code that makes computational chemistry work.
I don’t think most people outside the field realize just how big a deal this is. The bottleneck in computational chemistry has never really been in the theoretical methods. We have plenty of tools - RDKit, AutoDock, OpenMM, Biopython, Open Babel, and many more. Most of these do a decent job with workhorses of the field: docking, molecular dynamics, similarity searching, scaffold hopping. The problem has been getting them to work, and work together in particular. Until now, installing and debugging these packages was enough to make even experts tear their hair out. Dependencies broke, APIs shifted, compilers and file formats clashed, operating systems got in the way. Hours could vanish trying to make one program talk to another, and if you were new to the field the odds were you would simply give up.
With large language models like ChatGPT and Claude, that’s no longer true. Installation, which could be enough of a challenge to bring tears to your eyes, is now something you can get through in half an hour. The models diagnose dependency problems, suggest fixes, try alternate versions, and patch the quirks that would otherwise derail the process. What used to feel like trench warfare now feels like powered flight above the battlefield.
The same shift has happened with file formats. Every tool either needs it own formats or converts existing files into its own formats. There is an old joke about a conference where scientists decided to get rid of the ten existing file formats and replace them with the One Format to Rule Them All. The result? They now had eleven file formats. Most of the real frustrations in computational chemistry aren’t deep theoretical puzzles - they’re things like mismatched bond orders, missing charges, or PDBs that have extra lines some parser can’t read (parametrization used to be another, although modern force fields have mostly gotten rid of that one). These problems are small but relentless, the kind of thing that burns out motivation because fixing them feels like watching paint dry. Now you just paste the offending file into ChatGPT or Claude, and they tell you exactly what’s wrong and how to fix it. You can almost hear the collective sigh of relief from people who’s spent days chasing down these tiny, invisible errors.
Where this really starts to feel magical is in building pipelines. In the past, setting up even a modest workflow - say, generating analogs, docking them, running a round of MD minimization, filtering on properties, re-docking, and reprioritizing - was a serious project that would be tantamount to a character test. You’d have to hand off files between programs, write little scripts to convert formats, check each input and output, make sure versions matched, and intervene constantly to keep the thing from falling apart. It was tedious enough that most people avoided it unless they had to. But with these tools, you can just describe the workflow in plain English. “Take this molecule, enumerate a set of analogs with these property constraints, dock into this protein, minimize with MD, filter on these criteria, re-dock under stricter conditions, then generate focused analogs from this block library and reprioritize.” And out comes a working script, sometimes hundreds of lines long, that will do all of that. Often the only input it needs is a ligand file or a PDB. The script handles the rest: calling RDKit or Open Babel or whatever other dependency, retrying with a different tool if one fails, downloading a new version if there’s a compatibility issue, running the analysis, and consolidating the results. I have created dozens of such pipelines in the last few months, and sometimes it felt like I had just strapped on a jet pack.
And the scope of what’s possible is widening. Imagine starting with a library of thousands of ligands: an LLM can set up a pipeline that converts them into 3D conformers, docks them into a protein, filters out the top 1–2% by docking score, and then pushes those through short MD simulations in OpenMM to calculate free energies. At the end you get a neat CSV ranking compounds by docking and ΔG. Or say you want to do scaffold hopping: begin with a known lead, generate matched molecular pairs with RDKit, filter the results for drug-likeness, run them through an ADMET predictor, keep the promising ones, and dock them and pick the top ones. A process that would once have meant juggling five different formats and endless debugging now runs from one script. Even more specialized pipelines like covalent and antibody docking are (thanks to tools like AlphaFold) within reach. These were workflows that only seasoned computational chemists with lots of time and patience could manage. Now they’re accessible to almost anyone.
The tools aren’t perfect. ChatGPT sometimes gets stuck in loops; Claude tends to be better at debugging, though slower, and its scripts are longer. Neither gets everything right on the first try; you usually need to iterate, maybe delete a line or tweak a parameter. Occasionally a particular tool can test your patience for a day or two. But the key is that iteration is trivial. Most of the time you just paste the error message back in and out comes the fix. What used to be a serious obstacle is now something you almost don’t think about. The qualitative change from what went on before cannot be overstated.
I’m not someone who gets excited about every new piece of tech. I’ve been around long enough to see plenty of “revolutions” plateau into useful ordinary tools; docking, bound water analysis, molecular dynamics, protein folding and yes, lots and lots of AI hype. But here, I don’t hesitate to say it: when it comes to the process of doing computational chemistry, these models are revolutionary. They collapse time like nobody’s business. They turn hours, sometimes weeks, into minutes. They eliminate the motivation barrier, which is often the real obstacle to using open-source tools. The problem was never that these tools were inaccurate, but that they were brittle, finicky, and exhausting to set up. Now, overnight, that problem is gone. Computational chemists are supposed to focus on the chemistry, not the obstacles posed by the computation. These tools finally make that possible.
The consequences of this are hard to predict, but they will be real. Tools like Knime and Pipeline Pilot, built around making pipelining easier, will have to adapt. They aren’t obsolete ye because their GUIs still make them attractive, but the ground beneath them has shifted. More importantly, the roles of computational chemists themselves will change. Much of the work that once absorbed junior cheminformaticians or internal developers - the endless scripting, file conversion, pipeline construction - can now be done by LLMs. Those roles won’t disappear, but they will evolve. Specialists will need to move up the stack, finding the gaps these tools leave and filling them with unique expertise. There will be layoffs, new hires and unpredictable swings of fortune. And while the average scientist will greatly benefit from these tools, experts would also be foolish not to use them if it saves them hours and days of coding.
And here’s another impact that people miss: this is going to make open-source tools far more attractive. In a world where cost-conscious companies - especially startups - are already reluctant to buy expensive commercial packages, suddenly the main barrier to open source isn’t accuracy or features, but usability. If LLMs make it seamless to install, configure, and pipeline these tools, adoption will skyrocket. And not just adoption: development too. When barriers to using open-source tools fall, more people use them, more people contribute, and new projects emerge. We may see a wave of new open-source efforts like OpenADMET that would never have taken off before. In this sense, LLMs don’t just make existing tools more usable. They re-energize the entire open-source ecosystem in computer-aided drug design. And that’s a huge overall win.
And that’s why I keep coming back to this point when talking about new technologies in drug discovery: don’t miss what’s right in front of you. The hype about AI discovering drugs may or may not pan out the way people hope. But the quiet revolution has already arrived. In computational chemistry, LLMs aren’t just “useful.” They’ve changed the landscape. In the time savings that cut hours or weeks to minutes, in the way they handle dependencies, in how they remove the motivation barrier for both experts and average users, these tools are astounding. They let us get to the actual business of discovering drugs. And we would be wise to adopt them.