Qualitative analysis

What the models learned

The enrichment factor says which method found more hits. It does not say how these methods did that. We dig into two distinct screens to understand the difference between Gemini and AssayFormer, as well as how they combine.

How Gemini and AssayFormer differ across two screens

The figure follows an NF-κB / TNF signaling screen on the left and an AAV transgene silencing screen on the right. Each screen is run with AssayFormer, Gemini-3.1-Pro, and the AssayLoop handoff from Gemini to AssayFormer. Within each panel, the upper histogram shows the Reactome composition of the genes proposed in each round; the lower histogram shows the pathways represented among the hits. For example, DNA repair is the largest category in AssayFormer's first proposed batch for the AAV screen.

Gemini and AssayFormer begin with substantially different pathway compositions.
Gemini finds more hits in the first three rounds, but its yield drops afterward.
When AssayFormer takes over after round three, it proposes a different set of genes that contains more hits.
AssayFormer proposes more genes in the Other category, consistent with Gemini favoring better-known genes and pathways.
Gemini does not always use the full 100-gene acquisition budget.

How feedback changes the next batch

In a representative NF-κB / TNF signaling screen, AssayFormer recovers 60 of 169 true hits in ten rounds. At round four, the 24 hit labels already observed rerank the next batch: the label-aware policy recovers 11 hits spanning nine pathway categories, compared with seven hits spanning four categories when the labels are withheld. Leave-one-label-out counterfactuals identify which observations drive individual rank changes. For example, CYFIP1 rises from rank 1,078 without labels to rank 18 with labels; withholding just the observed KCTD10 hit lowers CYFIP1 from rank 18 to rank 62.

AssayFormer trajectory on an NF-kappa-B screen, a comparison of the next batch with and without hit labels, and rank changes caused by experimental feedback. — **Figure 3E–G: a worked AssayFormer trajectory.** The complete trajectory (E), composition of the fourth batch with and without labels (F), and changes in candidate rank induced by the observed outcomes (G).

LLMs converge on similar biological programs

The heatmap aggregates every requested gene across the 20 test screens and assigns it to a top-level Reactome pathway. We calculate a random baseline as the composition expected from uniform sampling over the full candidate universe. Each cell compares a model's observed pathway share with that random expectation: orange is over-represented, blue is under-represented, and the color scale is capped from 0.5× to 2×.

The LLMs have strikingly similar acquisition signatures. They favor RNA and protein metabolism while under-representing bulk metabolism and developmental biology. Differences among LLMs are small compared with their differences from methods such as Screen-kNN, and their effective pathway counts fall in a narrow range. Kimi-K2.6 and Claude Opus-4.8 also request fewer unique genes because they often return fewer than the requested 100 genes, possibly reflecting constraints from biological safety training. Hover any cell to see its raw pathway share, EP-D, total unique genes, and percentage of Reactome-annotated picks. The diversity explorer explains effective pathways in more detail.

What one hit tells AssayFormer about the rest

For a probe gene and a target gene, influence is the change in AssayFormer's predicted score for the target when the probe is added to the context as a hit. The interactive matrix uses 12 canonical cancer drivers as probes and 19 representative functional-module genes, plus ATM, as targets. Orange means the observation raises the target score; blue means it lowers it. Figure 5B deliberately uses a fixed, generic screen description and averages over ten randomized 50-gene background histories, isolating context-independent relationships learned by the model.

In the full analysis, influences among cancer drivers are generally weak, consistent with these genes acting through distinct pathways such as RTK/RAS, PI3K, p53, Wnt, TGF-β, and homologous-recombination repair. The strongest effects occur within functional modules whose members co-vary, including ribosome biogenesis, mitochondrial complex I and Fe-S assembly, and mRNA decapping. Influence is also directional rather than symmetric: the correlation between the matrix and its transpose is only 0.16, and 44% of reciprocal pairs have opposite signs. For example, observing MDM2 raises PFDN4 by 0.33, whereas observing PFDN4 lowers MDM2 by 0.46.

After removing pairs already represented in STRING, CORUM, SIGNOR, MSigDB, or Reactome, the strongest remaining positive and negative influences provide candidate hypotheses rather than established interactions.

From generic to screen-specific context

The generic context in Figure 5B motivates a stricter question: because AssayFormer conditions jointly on the screen description and observed history, do learned influences change in real, specific assay contexts? Figure 5C searches for relationships that are substantially stronger in one held-out screen than across the other 19. Both genes must be true hits in the focal screen, and pairs already represented in STRING, CORUM, SIGNOR, MSigDB, or Reactome are excluded.

Three screen-specific gene links with model influence in the matching screen versus other screens and a literature-supported biological rationale. — **Figure 5C: literature-supported, screen-specific gene links.** DCAF1 and ZNF638 converge on the HUSH complex; IGHMBP2 and GCN1 connect through the response to stalled ribosomes; and GTF3C1 and the ACTR5-containing INO80 complex converge at tRNA genes. The darker bar is influence in the matching screen and the lighter bar is the mean across the other 19 test screens.