1
00:00:00,000 --> 00:00:07,331
A genome-wide CRISPR screen puts a question to every gene: is this gene a hit? There are twenty thousand candidate genes.

2
00:00:08,331 --> 00:00:14,508
But what is a CRISPR screen? A screen tests which genes control a specific cellular behaviour or trait.

3
00:00:14,858 --> 00:00:26,288
You set up an experiment. Say, neuroblastoma cells, and whether they resist iron-mediated cell death. Then you switch one gene off, or on, and see whether the outcome changed.

4
00:00:26,638 --> 00:00:31,838
Do that across the whole genome, and you learn which genes control the behaviour you care about.

5
00:00:32,338 --> 00:00:42,900
A few of those genes are hits. A hit is a gene where switching it off, or on, actually produces the effect you are looking for. In a typical screen, only one in a hundred are hits.

6
00:00:43,250 --> 00:00:47,827
You do not know which ones they are. That is the whole point of running the screen.

7
00:00:48,327 --> 00:00:59,432
But testing all of them is often out of reach. The readout is too slow, or too expensive. So you get a budget: one thousand genes, five percent of the library.

8
00:00:59,782 --> 00:01:01,402
Which thousand should you pick?

9
00:01:02,052 --> 00:01:08,253
You don't have to choose all at once. Spend the budget in rounds. One hundred genes, ten times.

10
00:01:08,603 --> 00:01:16,010
After each round you learn which of those hundred were hits, and that answer can change what you pick next. That is the loop.

11
00:01:16,360 --> 00:01:26,718
This is AssayBench-Loop: one thousand three hundred and eighty-nine real CRISPR screens, split by time. Train on the past, test on twenty screens from the future.

12
00:01:27,218 --> 00:01:35,462
We can picture any strategy with a recovery curve: the fraction of all hits you have found, against the fraction of the library you have sampled.

13
00:01:35,812 --> 00:01:43,377
Pick genes at random, and you find hits at exactly the rate you sample them. A straight line. That is the thing to beat.

14
00:01:43,877 --> 00:01:54,540
One way to fill that first round is to ask a language model. It has read the literature, so it already has an opinion about which genes are plausible, before you measure anything.

15
00:01:55,490 --> 00:02:01,350
And it works. Gemini three point one Pro finds hits at nearly five times the random rate.

16
00:02:01,700 --> 00:02:10,156
But watch what happens as results come back. The curve flattens. The prior is strong. The learning from feedback is weak.

17
00:02:10,656 --> 00:02:22,615
AssayFormer is the opposite animal. A transformer trained across all thirteen hundred historical screens. It takes the history as input: every gene tested so far, and whether it was a hit.

18
00:02:22,965 --> 00:02:28,213
It starts colder. Before the first batch, it knows little about this particular screen.

19
00:02:28,563 --> 00:02:32,263
But every round of feedback sharpens it, and it keeps climbing.

20
00:02:32,763 --> 00:02:41,560
LLMs have a strong prior but weak adaptation. AssayFormer has a weak prior but strong adaptation. So use each one where it is best.

21
00:02:41,910 --> 00:02:49,264
Let the language model drive the first three rounds. Then hand the accumulated history to AssayFormer, and let it run.

22
00:02:49,614 --> 00:02:51,033
That is AssayLoop.

23
00:02:51,683 --> 00:03:00,545
Five point seven times better than random. Twenty-seven point seven percent of hits in the screen are found by sampling only five percent of the library.

24
00:03:00,895 --> 00:03:08,632
And because the policy is learned from past experiments, it improves as those accumulate, with no sign yet of saturating.

25
00:03:08,982 --> 00:03:12,589
Biology in the loop. Try it on our website today!