The Drug Discovery Checklist

Some things work. Yours probably doesn't.

Drug discovery is a hard business, and like any other hard business, there are a lot of people who’d not only like to improve it, but think they have a clever idea to completely change how it’s done and catapult us into a new world of productivity. It reminds me of the proposals to build better spam filters (before GMail and some sender authentication mechanisms mostly solved it for most of us). Therefore, in the spirit of the (in)famous Slashdot Spam Proposal Checklist, I humbly offer the Drug Discovery Checklist, a tool to rapidly screen proposals on new ways to Make Drugs Fast.

(Yes, this is obviously meant to be tongue-in-cheek. Trying to advance therapeutic development is a noble cause. It’s OK, my ideas probably won’t work either.)

Your paper/post/pitch-deck advocates a

( ) machine learning ( ) genetics based ( ) patient stratification ( ) massively parallel experimental

approach to improving NDA success rates. Your idea will not work. Here is why it won’t work. (One or more of the following may apply to your particular idea, and it may have other flaws which will vary from jurisdiction to jurisdiction depending on regulatory mood, population stratification, and differences in IP protection.)

( ) Chemical space is really big. You won’t believe just how vastly, hugely, mind-bogglingly big it is. I mean, you may think there’s a lot of variation in a phage library, but that’s just peanuts to chemical space.
( ) Regulators will not buy into your proposed trial structure changes.
( ) Public markets will not tolerate such a high-risk proposal.
( ) VCs will not tolerate the likely low-multiple return.
( ) Considering targets in isolation fails to account for polypharmacologic effects.
( ) You tested on your training set.
( ) High ROC AUC does not imply high PPV for rare conditions.
( ) Frequentist analysis is insufficient to handle experimental uncertainty for high-dimensional data.
( ) Bayesian analysis is confusing to reviewer 3 and you’ll never get your Nature paper.
( ) Lots of things are polymers, but not everything is a polymer.
( ) The world is made up of more than the Europeans present in your favorite genetics dataset.
( ) Consumers/patients/hospitals will not give their data up for free.
( ) Everyone else has looked at those public data sets too.
( ) “Improving” on a bad baseline isn’t meaningful.

Specifically, your plan fails to account for

( ) Everyone else has sequencers too.
( ) No one does good sequencing. No, you don’t.
( ) Placebo is a pretty good standard of care.
( ) Most compounds in your screening library probably degraded or were mislabeled in the first place.
( ) Existing IP held by larger players who will sue you out of existence.
( ) Poor IP protection in the target market.
( ) Batch effects
( ) Genetic background
( ) Mice aren’t people.
( ) Organoids aren’t organs.
( ) Cell line heterogeneity
( ) hERG
( ) The liver and its amazing technicolor CYPs
( ) Adversarial examples
( ) Domain shift
( ) Congressional inquiries
( ) CRISPR off-target effects
( ) Alternative splicing
( ) HIPAA
( ) GDPR
( ) Insurers won’t pay for it.
( ) Formulation and manufacturing for this product will be impossible.
( ) Resistance evolves faster than you can get new drugs onto the market.

And the following philosophical objections may also apply:

( ) Ideas similar to yours are easy to come up with, yet none have ever been shown effective.
( ) Large proprietary genetic databases are on the wrong side of history.
( ) A business model built on raising prices of known compounds is on the wrong side of history.
( ) YOUR COMPOUNDS ARE AN AFFRONT TO OUR LORD AND MASTER LIPINSKI REEEEEEEEEE
( ) If you had that many GPUs, it would have better ROI to just mine cryptos.
( ) No one will accept your shitcoin as payment for their data.
( ) Making a drug that can only help a narrow slice of people is a dodge of our larger responsibility.
( ) Flow is a good for a rapper, less so for chemistry.
( ) Preventing future generic competition violates our contract with society.
( ) You know that ADMET matters too, right, not just Kd/IC50?
( ) Putting hard-working scientists out of a job is bad.
( ) If you put in that many nitrogens, you’re making real warheads, not chemical warheads.
( ) Have you even defined what you mean by epigenetics?
( ) Small molecules are dead, long live mAbs/RNAi/macrocycles/cell therapies.
( ) Your target market would rather use herbs and energy crystals than your drug.

Furthermore, this is what I think about you:

( ) Sorry dude, but I don’t think it would work.
( ) This is a stupid idea, and you’re a stupid person for suggesting it.
( ) Nice try, assh0le! I’m going to write STAT News and get them to burn your stock down!

(Thanks to the anonymous friends who reviewed early drafts of this checklist!)