Disclosure: this article contains affiliate links. If you buy through them we may earn a commission at no additional cost to you. Our recommendations are based on the evidence recorded alongside this article, and prices were verified on the date shown.

What changed and why it matters

Science.org — the publishing arm of AAAS, which runs the Science family of journals — has published a piece that puts the question in its own headline: "So How Is AI Drug Discovery Doing, Really?" The post, bylined AnodicElegy and retrieved 2026-08-15, exists specifically to check real-world progress in AI-driven drug discovery against the claims made about it.

That a science-publishing outlet felt the need to ask the question directly, rather than assume the answer, is itself worth noting. When a field has generated enough noise that a serious outlet devotes a dedicated post to auditing it, that's a signal to read before you take any single vendor's claim about "AI drug discovery" at face value.

The specific findings live in the piece itself, and this page won't paraphrase them into something that sounds more definitive than the source. What it offers instead is a framework for reading that piece, and any claim like it, with the right amount of skepticism — plus a practical next step if reading it leaves you wanting to test part of the pipeline yourself.

The list

No summary can substitute for reading the source, so this list isn't a recap of what Science.org found — it's the checklist worth applying to that piece, and to every other AI-drug-discovery claim you'll run into after it. Use it before you repeat a headline as fact.

1. Which pipeline stage the AI actually touches. "AI drug discovery" gets used for target identification, molecule generation, and clinical trial design interchangeably, and those are three different processes with three different failure rates. A claim about one stage doesn't validate the others, and a headline rarely tells you which one it means.

2. Whether a compound reached human trials, or just a screen. A computational hit is cheap to generate; a molecule that clears Phase I is the actual signal that something worked. Read closely for that distinction before crediting any candidate as "AI-discovered" in the way headlines imply.

3. Who is making the claim, and what they gain from it. A company announcing its own AI-discovered candidate has different incentives than a peer-reviewed journal or an outlet like Science.org auditing the field from the outside. Neither source is automatically right, but they're not interchangeable either, and conflating them is how hype survives fact-checking.

4. What "success" is being measured against. The useful comparison is AI-assisted timelines against the traditional discovery baseline, not against zero. The relevant question is whether a method is faster or cheaper than what already exists, not whether it produced a result at all.

5. Whether the reporting is dated. A retrieval date attached to a claim — like the 2026-08-15 check on the Science.org piece cited here — tells you the information hasn't gone stale. Treat any claim that carries no retrieval date with more caution, since you have no way to know how current it is.

6. What it takes to run this kind of analysis yourself. If your next step is screening compound libraries, training models on molecular data, or standing up a research-adjacent app to sit on top of either, the compute has to hold up under real workloads before any of the reading above matters in practice. A small always-on instance is the usual starting point for testing a pipeline before scaling it, and you can Check current price for DigitalOcean Basic Droplet (2 vCPU / 4 GB) to see what a 2 vCPU / 4 GB instance costs today.

The table below collapses the first five items into a quick reference — keep it open in a tab next to the source piece.

What to do next

None of the checklist above settles whether AI drug discovery is "working" in the sense of shipping approved drugs faster — that's the exact question Science.org's piece exists to interrogate, and the honest move is to read it directly rather than take a secondhand summary of it as the final word.

What the checklist does settle is more practical: if your next step after reading is building or testing any part of an AI-assisted research pipeline, the infrastructure decision is separable from the open science question above, and it's one you can price out now rather than later.

• For storing screening outputs and datasets — structure files, molecular data dumps, model checkpoints — Check current price for DigitalOcean Spaces object storage before you commit to a bucket size. Best for teams already generating enough output data to outgrow local disk. Skip it if you're still prototyping on a single machine with nothing to archive yet.
• For a relational store behind compound or trial metadata — Check current price for DigitalOcean Managed PostgreSQL rather than guessing at what a managed instance costs. Best for a pipeline that needs queryable structured records, not flat files. Skip it if your data never leaves a notebook.
• For comparing DigitalOcean against other providers before you commit to one — See current pricing at the merchant. Best for teams that haven't picked a provider yet and want the current numbers side by side. Skip it if you're already locked into infrastructure elsewhere and switching costs outweigh the comparison.

If you're also evaluating AI agent tooling to help triage or summarize research literature like the Science.org piece, the comparison in "Show HN Agentic Ship compared: which one is worth buying" is a reasonable next stop before you commit to a stack on top of whichever infrastructure you land on.

The decision in front of you isn't "is AI drug discovery real" — that's Science.org's question, not this page's, and it deserves a direct read rather than a paraphrase that flattens the nuance. The decision this page can actually help with is narrower: whether the infrastructure to test any part of this yourself is worth provisioning now, at the prices linked above, or worth waiting on until the science question above is clearer to you.