“You always get what you screen for”
At a recent Bits ‘n Bio meetup, I met Akhila Kosaraju, M.D. Akhila is CEO and co-founder of Phare Bio, a nonprofit working with the Collins Lab at MIT on AI-driven antibiotic discovery, including predictive screening and generative design.
AI can dramatically change the scale and speed at which chemical space can be explored. Instead of experimentally testing every molecule, computational approaches can prioritize candidates before committing resources to physical screening.
But as I listened, I kept coming back to a lesson I’ve learned repeatedly in R&D:
You always get what you screen for.
I wrote recently about a high-throughput screen at Dow AgroSciences that was very good at identifying improved strains – except the improvement disappeared at production-relevant scale. The screen wasn’t failing. It was answering the question we’d asked it. It was just the wrong question.
An AI screen is also answering a question. What patterns, structures and interactions has it learned to associate with activity?
More importantly, what hasn’t it learned?
Every screen has biases built into it. In an experimental screen, those biases come from things like assay conditions, scale, readouts and thresholds. In an AI screen, they also come from the data the model was trained on and the patterns it has learned to recognize. The mechanisms are different, but the underlying problem is the same: the screen helps determine what you can see.
But what happens to molecules the model doesn’t recognize as promising? If they’re deprioritized before experimental testing, you may never know the hit was there.
Akhila is well aware of that risk. She told me that their plan includes rescreening previously evaluated molecules as the model evolves, since a molecule the model deprioritized yesterday may look very different as the model learns more.
That’s not a limitation unique to AI. Every screening strategy makes choices about what deserves a closer look and what doesn’t. The difference is that AI can make those choices across a vastly larger chemical space, potentially making both the opportunity – and the blind spots – much larger.
Whether you’re screening microbial strains in a microtiter plate or molecules in silico, the same rule applies: You always get what you screen for.
The harder question is what you might be screening out.
Leave a Reply