AI-Powered Drug Discovery: The Software Behind It

The best-known AI-powered drug discovery software predicts how molecules fit together, so chemists can rule out weak candidates before anyone synthesises them. Whether that produces better medicines is a separate question, and the evidence so far is more mixed than the headlines suggest.

This guide covers what the software does, what the clinical data shows, and where the engineering gets difficult.

What AI-powered drug discovery software actually does

At the centre sit models that predict the three-dimensional structure of molecular complexes. AlphaFold 3, described in a May 2024 Nature paper, predicts joint structures of proteins, nucleic acids, small molecules, ions and modified residues in one framework.

On the PoseBusters benchmark of 428 protein-ligand structures released in 2021 or later, the authors report that it beats classical docking tools such as Vina. It does this without being handed the solved structure that docking tools use as privileged information.

A model alone isn't a product, though. Chemists need somewhere to submit jobs, a place to store results, and an interface they'll open every day. That surrounding layer is the part a software team builds, and it decides whether the model gets used at all.

Does AI-discovered mean more drugs get approved?

Not yet, at least not beyond the earliest trial phase. A Boston Consulting Group analysis in Drug Discovery Today, reported by HT World in May 2024, put Phase I success for AI-discovered molecules at 80 to 90 percent, against an average of 40 to 65 percent for human-discovered drugs.

Later stages look different.

By Phase II the gap closes. The same analysis found success of about 40 percent, which the coverage describes as in line with industry averages, and the authors flag a limited sample size. They call it "a first view of the potential", which is the right level of confidence to carry into a business case.

For anyone commissioning software, the practical reading is to treat these tools as a way to narrow the field of candidates earlier. Don't budget on the assumption that they remove late-stage risk, because the data doesn't show that yet.

Where structure prediction falls short

A predicted structure is a hypothesis for the lab to test. The AlphaFold 3 authors are open about the limits, and each one has a direct consequence for how the surrounding software should behave.

• As a generative model, it can invent plausible-looking structure in disordered regions. Predicted regions should never sit in the interface looking identical to measured ones.
• It sometimes violates chirality, at a rate of 4.4 percent on the PoseBusters set. Automated checks belong in the pipeline, before any output reaches a chemist.
• It predicts static structures, not the range of shapes a molecule adopts in solution. For cereblon, the paper says it predicted only the closed state, even without a bound ligand.
• Accuracy on antibody-antigen complexes keeps improving as the number of model seeds rises, up to 1,000. That means a job queue and visible compute cost, not a button that runs instantly.

None of these is a reason to avoid the tools. They're a reason to build the checks and the labelling that make the output safe to act on.

Can you build a product on AlphaFold 3?

Photo by Kechno Studio on Pexels

Check the licence before the architecture. When the paper was published, the AlphaFold Server was limited to non-commercial use, with restrictions on which ligands and covalent modifications it accepted, and the paper states "Code is not provided."

A commercial product can't assume access to that model on those terms. Terms may have changed since publication, so confirm the current position for every model you plan to depend on. If you can't swap a model out later without rewriting the product, the dependency is your biggest risk, and you should design the model layer as a replaceable component from the first sprint.

What it takes to build drug discovery software around a model

Most of the work is unglamorous. Data has to flow in from instruments and files in whatever formats labs use, results need provenance so a scientist can see which model version produced them, and the interface has to make uncertainty visible instead of burying it.

We've built healthcare products end to end, including OptimalMD's website, members portal and mobile app, from design through development to launch. That was patient-facing work rather than research tooling, so it's a different problem, but the discipline of owning a product from first design to launch carries over. Our apps and SaaS development and AI and automation teams build this kind of platform.

Regulation is the other thing to watch. The FDA's January 2025 draft guidance on AI in regulatory decision-making covers AI used to produce information supporting decisions about safety, effectiveness or quality. It's a draft with non-binding recommendations, but anyone whose model output may end up in a submission should read it before deciding what the software has to record.

Teams extending into clinical tools can also compare notes with our pieces on AI diagnostics software and clinical decision support systems. And once research tooling connects to patient-facing systems, our QA checklist for HIPAA apps applies.

Start with one narrow workflow, such as ranking candidate molecules for a single target, and ship that to a small group of chemists. Their corrections will tell you more about the next feature than any roadmap will.

Frequently asked questions

Can I use AlphaFold 3 in a commercial product?

Not through the AlphaFold Server as described in the paper, which was limited to non-commercial use. Terms may have changed since, so check the current licence before committing.

Do AI-discovered drugs succeed more often in trials?

In Phase I, the BCG analysis found a higher success rate than for human-discovered drugs. In Phase II, success was about 40 percent, in line with industry averages, on a limited sample.

Is the FDA guidance on AI final?

The version I read is a January 2025 draft marked as non-binding recommendations. Check the FDA's guidance page for its current status before relying on it.

Cover photo by Google DeepMind on Pexels