Research · Computational Biology
Can a biomedical AI pick the best drug to fight Parkinson's?
We set out to screen approved drugs against a Parkinson's target using IBM's open MAMMAL foundation model, hoping to rank existing medicines by how well they bind it. This is an honest write-up of what we found — including a clear negative result, one genuinely useful capability, and a hard look at where AI helps and where it does not.
Important — please read
This page documents a computational experiment. It is not medical advice, and nothing here is a treatment recommendation. There is currently no approved drug that cures or halts the progression of Parkinson's disease — every approved medicine manages symptoms. The "standard of care" ranking below is summarized from published clinical consensus, not produced by our model. Treatment decisions belong to a qualified neurologist.
Findings at a glance
AUROC 0.29
Drug-target binding screen
Worse than a coin flip. The model scored by protein and ignored the drug, so it could not rank candidates at all.
AUROC 0.87
Blood-brain-barrier model
Benchmark-level and independently validated. A genuinely useful early filter for whether a molecule can reach the brain.
≈1.00*
Toxicity / FDA models
*Looks perfect, but that is training-set memorization. The FDA model failed on real approved drugs it had not seen.
Abstract
We asked a simple question: given an open biomedical foundation model and the full list of approved drugs, can we rank which one best binds a Parkinson's-relevant target? We chose MAO-B — the validated target of real Parkinson's drugs — precisely because we could check the answer. The binding model failed this test decisively (AUROC 0.29; known MAO-B inhibitors ranked in the bottom half). Diagnostics showed why: the model scores almost entirely from the protein and is nearly blind to the drug. We then pivoted to tasks the same model is actually good at. Its blood-brain-barrier predictor reached benchmark-level performance (AUROC 0.87) on independent controls and is a legitimate early screen. Its toxicity and FDA-approval models posted near-perfect benchmark scores that turned out to be memorization — they failed on approved drugs they had not been trained on. The headline is methodological: validate a model on a question whose answer you already know before you trust it on one you do not.
Method — validate before you trust
Every screen here follows the same rule: before drawing any conclusion, run the model on compounds whose status is already known and measure whether it separates them (reported as AUROC, where 1.0 is perfect and 0.5 is random). Target sequences came from UniProt, drug structures from PubChem, and every molecule was parsed and canonicalized with RDKit so a typo could never masquerade as a result. All models are the public IBM MAMMAL checkpoints (a 458-million-parameter encoder-decoder) run locally on a GPU.
Result 1 — The binding screen failed, and that is the headline
We screened 60 drugs against MAO-B and seeded in 10 known MAO-B inhibitors as hidden positive controls. A working model would float them to the top. Instead they sank to the bottom half — AUROC 0.286, worse than random — and the apparent "top hits" (digoxin, simvastatin) have no MAO-B rationale whatsoever.
To find out why, we fed the model textbook nanomolar binders and obvious non-binders. The pattern was unmistakable: the score is driven by the protein and barely moves with the drug.
| Protein | Drug | Predicted pKd |
|---|---|---|
| ABL1 kinase | imatinib — a real nanomolar binder | 7.22 |
| ABL1 kinase | glucose — a non-binder | 6.85 |
| Carbonic anhydrase II | acetazolamide — a real nanomolar binder | 6.04 |
| Carbonic anhydrase II | water | 6.07 |
| MAO-B | safinamide — an approved MAO-B inhibitor | 5.41 |
Glucose scores almost the same as imatinib; water scores the same as acetazolamide. This is the well-documented "target-prior" failure of binding models — they learn an average affinity per protein and ignore the molecule. For the goal of ranking drugs against one target, the model is the wrong instrument, and we did not pretend otherwise.
Result 2 — Blood-brain-barrier prediction works
The same model has single-molecule property heads, where the target-prior trap cannot occur. We tested its blood-brain-barrier (BBB) predictor on 20 compounds with textbook-known status. It reached AUROC 0.87 — in line with published state of the art — and is directly relevant to Parkinson's: a drug cannot treat a brain disease it cannot reach.
Correctly identified as NON-penetrant
vancomycin, amoxicillin, sulfasalazine — probability ≈ 0.00
Correctly identified as penetrant
caffeine, diazepam, haloperidol, donepezil — probability ≈ 1.00
Over-called (the model's positive bias)
dopamine, domperidone, atenolol — flagged as crossing, but they do not
The honest caveat: the model over-predicts penetration (it labels most molecules as "crosses"), partly because the underlying dataset is skewed that way. So trust the ranking and the confident "no" calls; be skeptical of borderline "yes" calls. Every approved Parkinson's drug we tested was predicted to cross — correct, since they are all central-nervous-system drugs — with the telling exception that carbidopa and entacapone (designed to stay OUT of the brain) were over-called, exactly as the bias predicts.
Result 3 — When a perfect score is a warning sign
We also ran the model's clinical-toxicity and FDA-approval heads. On their own benchmark they scored AUROC 0.998 and 1.000 — suspiciously perfect. The reason is that our validation sample overlapped their training data, so they were graded on what they had memorized. The proof came when we applied the FDA-approval model to real approved Parkinson's drugs it had to generalize to: it rated pramipexole at 0.12 and rasagiline at 0.22 — i.e. "probably not approved" — even though both are approved. Impressive paper metrics, no real-world reliability.
The practical question
So what actually fights Parkinson's today?
Our model could not rank this for us — so the table below is not an AI result. It summarizes the established clinical standard of care (per movement-disorder treatment guidelines) ordered by symptomatic role. Two honest points first: these drugs treat symptoms, none of them stop or reverse the disease; and the "best" choice is individual, depending on age, symptoms, and stage. The MAO-B inhibitors at rank 3 are the class whose mechanism this study targeted.
| # | Drug / class | How it works | Role |
|---|---|---|---|
| 1 | Levodopa + carbidopa | Dopamine precursor; carbidopa keeps it intact until it reaches the brain | Most effective symptom control — the long-standing gold standard |
| 2 | Dopamine agonists — pramipexole, ropinirole, rotigotine | Directly stimulate dopamine receptors | Often first-line in younger patients or levodopa-sparing strategies |
| 3 | MAO-B inhibitors — rasagiline, safinamide, selegiline | Block the MAO-B enzyme, slowing the breakdown of dopamine | Early monotherapy or add-on — and the exact mechanism this study targeted |
| 4 | COMT inhibitors — entacapone, opicapone | Extend the duration of each levodopa dose | Add-on for end-of-dose wearing-off |
| 5 | Amantadine | NMDA receptor antagonism (among other effects) | Reduces levodopa-induced dyskinesia |
| 6 | Istradefylline | Adenosine A2A receptor antagonist | Add-on to reduce off-time |
| 7 | Anticholinergics — trihexyphenidyl, benztropine | Rebalance acetylcholine against dopamine | Mainly tremor, typically in younger patients |
On "blocking" the disease
There is no approved therapy that blocks or cures Parkinson's. Disease-modifying approaches — targeting alpha-synuclein aggregation, LRRK2, GBA, or repurposing GLP-1 drugs — are active areas of clinical research, but all remain investigational. Anyone who tells you a marketed drug "blocks Parkinson's" is overstating the evidence.
Honest conclusion
We did not find a new drug, and we will not claim we did. What we found is more useful than a fake headline: a clear map of where this AI helps and where it misleads. Binding-affinity ranking against a single target — the thing we most wanted — is exactly where it failed. Blood-brain-barrier screening is where it earned its keep. And a model can post a perfect benchmark score while being unreliable in the real world, which is why we validated every step against answers we already knew. That discipline is the actual finding.