Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens

Motivation

• CRISPR screens require sequential experiment selection because vast perturbation spaces make exhaustive testing infeasible under finite budgets.Experimental outcomes can guide subsequent actions in a lab-in-the-loop process, supporting adaptive discovery under constrained resources.Source: §1 · p. 2
• Existing benchmarks lack the scale and phenotype diversity needed to evaluate transferable acquisition policies or determine whether systems can learn across completed experiments.Prior methods generally fit predictive models or acquisition strategies independently for each screen, while LLM adaptation from feedback remains unresolved.Source: §1 · p. 2

Benchmark

• AssayBench-Loop provides 1,389 CRISPR screens across five phenotype categories, enabling systematic evaluation and learning acquisition strategies across historical experiments.The benchmark includes metrics accounting for hit-rate heterogeneity and incomplete ground truth.Source: §1 · p. 2

Method

• AssayLoop combines AssayFormer, a transformer policy trained across historical assays, with LLM-derived biological priors through an adaptive handoff.The LLM guides early acquisitions when assay-specific evidence is limited, after which AssayFormer uses accumulated observations to prioritize candidates.Source: §1 · pp. 2–3
• AssayLLM extends historical-screen learning directly to a 27B-parameter LLM through acquisition-trajectory fine-tuning and reinforcement learning over sequential campaigns.This serves as a proof of principle for learning adaptive acquisition behavior within an LLM.Source: §1 · p. 3

Results

• 5.67-fold enrichment over random selection and 27.7% hit recovery after assaying approximately 5% of candidates show AssayLoop outperforms adaptive-design methods, standalone LLMs, and AssayFormer alone.On temporally held-out screens, AssayLoop retains broad pathway coverage and transfers to unseen phenotype categories while improving with more historical training data.Source: §1 · p. 3

2.1 A large collection of historical CRISPR screens enables learning a transferable acquisition policy for adaptive hit discovery

• ASSAYBENCH-LOOP contains 1,389 CRISPR screens across five phenotype categories, split into 1,349 training, 20 validation, and 20 temporally held-out test screens.The benchmark emphasizes phenotype-diverse genome-wide evaluation and is substantially larger than existing narrow benchmarks.Source: §2.1 · pp. 3, 5, Figure 2 · p. 4
• The benchmark evaluates sequential acquisition at a fixed budget of 10 rounds of 100 requested genes using hit enrichment factor, nAUC, fraction of hits found, and pathway diversity metrics.Hit enrichment normalizes discovered hits against random selection while penalizing hallucinated genes and incomplete batches, but not valid genes absent from a screen library.Source: Figure 2 · p. 4, §2.1 · p. 5

2.2 ASSAYFORMER learns a feedback-conditioned acquisition policy across experiments, which ASSAYLOOP combines with LLM priors

• ASSAYFORMER learns one shared transformer policy that maps each screen description and accumulated gene–outcome history to acquisition scores over candidate genes.Training combines BPMF initialization, supervised prediction from randomly sampled histories, and GRPO fine-tuning with a context-delta reward.Source: §2.2 · p. 5
• ASSAYLOOP combines ASSAYFORMER’s feedback-conditioned strategy with LLM-derived biological priors through an adaptive handoff because early acquisitions have limited assay-specific information.The learned policy exploits cross-screen structure while adapting to perturbation outcomes observed during the current experiment.Source: §2.2 · p. 5

A. Active hit discovery paradigms

• ASSAYLOOP uses an LLM for the first k = 3 rounds, then hands off to ASSAYFORMER, combining biological priors with adaptation to accumulated assay feedback.ASSAYFORMER embeds the screen description and observed gene-hit history, scores untested genes, and is trained through BPMF initialization, BCE, and GRPO rollouts.Source: A. Active hit discovery paradigms · p. 7
• 60 of 169 true hits were recovered within 1,000 genes in a representative screen, while incorporating 24 observed hit labels increased the next-batch yield from 7 to 11 hits.The observed labels also expanded represented pathways from four to nine categories and altered candidate rankings through specific model dependencies.Source: A. Active hit discovery paradigms · p. 7

2.3 ASSAYLOOP improves adaptive hit discovery on held-out screens

• 5.67-fold enrichment, 21.7% nAUC, and approximately 27.7% of hits recovered at 5% effective library coverage make ASSAYLOOP outperform standalone LLMs and ASSAYFORMER alone.ASSAYLOOP also achieves higher mean EF than BioBO, Probability-of-Hit, LLMNN, and ICBR-EF, supporting the complementarity of LLM warm starts and feedback-conditioned adaptation.Source: Table 1 · p. 8, §2.3 · p. 9

2.4 Adaptive experimental capability scales with historical experience and transfers across phenotype categories

• ASSAYFORMER’s mean EF more than doubled across the evaluated training-set sizes without clear saturation, while increasing model capacity across nearly 40-fold yielded little additional benefit.These results indicate that performance in this regime is primarily limited by the number of historical screens rather than model size.Source: §2.4 · pp. 9–10
• Under leave-one-phenotype-out training, ASSAYFORMER reached EF 4.59 versus 4.83 with full training, with a smaller EF drop than Screen-kNN and performance above all-phenotype baselines.The results support robustness when training screens from the evaluated phenotype category are unavailable.Source: §2.4 · pp. 9–10
• BPMF gene-token initialization achieved EF 4.83 after reinforcement learning versus 4.40 for SVD, while removing screen conditioning or context-delta rewards reduced performance.Training-hit-matrix embeddings outperformed GenePT and K562-derived external biological embeddings.Source: §2.4 · pp. 10–11

2.5 ASSAYLOOP preserves biological breadth while enriching for hits

• ASSAYLOOP achieved EP-B 18.4, EP-S 46.6, and EP-D 70.4, whereas standalone LLMs showed lower within-batch and within-screen diversity but broad dataset coverage.Gemini-3.1-Pro reached EP-B 14.5, EP-S 34.6, and EP-D 66.2, while ASSAYFORMER reached EP-B 19.3, EP-S 46.7, and EP-D 65.0.Source: Figure 5 · p. 11, §2.5 · p. 12
• ASSAYLOOP increased EF from 4.83 to 5.67 while retaining comparable diversity to ASSAYFORMER across batches, screens, and the full dataset.Its diversity values were EP-B 18.4 vs. 19.3, EP-S 46.6 vs. 46.7, and EP-D 70.4 vs. 65.0.Source: Figure 5 · p. 11, §2.5 · p. 12

2.6 ASSAYFORMER encodes directional gene–gene influence patterns

• ASSAYFORMER encoded asymmetric directional updates, with 44% of reciprocal gene pairs showing opposite-sign influences, including MDM2→PFDN4 and PFDN4→MDM2.Adding MDM2 increased PFDN4’s score by 0.33, whereas adding PFDN4 decreased MDM2’s score by 0.46.Source: §2.6 · pp. 12–13
• Context-dependent influences identified DCAF1→ZNF638, IGHMBP2→GCN1, and GTF3C1→ACTR5 as hypotheses with stronger effects in focal screens than across other contexts.Both genes in each pair were hits in the focal screen, and database-annotated pairs were excluded; these examples do not establish direct genetic interactions.Source: A. Reactome pathway diversity · p. 12, §2.6 · p. 13

2.7 Post-training LLMs can improve sequential hit discovery

• Post-training improved standalone Qwen3.6-27B EF from 2.65 to 3.69, a 39% gain, narrowing the gap to ASSAYFORMER’s 4.83.SFT increased EF to 3.56 before GRPO reached 3.69, approaching the GLM-5.1 teacher’s EF of 4.0.Source: §2.7 · p. 13, Figure 6 · p. 14
• EF rose from 5.05 with ASSAYLLM warm-starting to 5.23 after handoff-aligned training, approaching the GLM-5.1 → ASSAYFORMER configuration at 5.27.The handoff objective rewards both hit discovery and construction of a useful history for ASSAYFORMER.Source: §2.7 · p. 14

• AssayBench-Loop spans 1,389 CRISPR screens across five phenotype categories and is two orders of magnitude larger than previous benchmarks.Its dedicated metrics address incomplete libraries and LLM-specific failure modes while enabling rigorous comparison and cross-experiment policy training.Source: §3 · p. 14
• 27.7% of hits were recovered while assaying 5% of the candidate library, with AssayLoop outperforming both LLM and AssayFormer components used alone.The result supports complementarity between biologically informed early prioritization and feedback-conditioned adaptation.Source: §3 · pp. 14–15
• Domain-specific post-training improved AssayLLM by 39% over the base model, while trajectory supervision, reinforcement learning, and handoff-aligned training improved adaptive acquisition.General-purpose LLMs were more effective as biological-prior sources than as feedback-adaptive policies, whose in-context adaptation remained limited.Source: §3 · p. 15
• AssayFormer performance improved consistently with the number of training screens, with no clear plateau, and showed strong leave-one-phenotype-out generalization.These results suggest that larger and more diverse screen collections may improve transferable acquisition policies.Source: §3 · p. 15
• The study is limited to retrospective genome-wide screens with sufficient hit signal, fixed-budget binary hit discovery, and labels that do not explicitly model replicate-level experimental noise.The policies also remain unvalidated prospectively in a live experimental loop, while richer readouts and heterogeneous costs remain future directions.Source: §3 · p. 15
• Natural-language task descriptions could extend the benchmark beyond current phenotypes and modalities to diverse perturbations and richer readouts.AssayLLM could incorporate new experimental data through post-training, supporting a coupled lab-in-the-loop cycle of biological discovery and model improvement.Source: §3 · p. 15

4.1 Problem formulation

• CRISPR hit discovery selects 10 batches of 100 genes, maximizing discovered hits under a 1,000-acquisition budget through a history-conditioned policy.The formulation treats screens as Bayesian adaptive experimental design with deterministic hit observations and utility equal to discovered hits.Source: §4.1 · p. 21

4.2 Dataset and benchmark construction

• AssayBench-Loop retains 1,349 pre-2021 training screens and uses 20 validation and 20 post-2021 test screens selected for genome-wide coverage, sufficient hits, low hit rates, and phenotype diversity.The benchmark follows a temporal split and limits evaluation sets because sequential-design evaluation is computationally intensive.Source: §4.2 · pp. 21–22

4.3 Evaluation metrics

• The primary enrichment factor normalizes discovered hits by acquired, hallucinated, and unfilled slots, while nAUC, hit recall, shortfall, essential-gene fraction, and pathway or gene diversity provide complementary evaluation.The enrichment denominator avoids penalizing valid genes absent from a screen library but penalizes hallucinated genes and unfilled slots.Source: §4.3 · p. 22

4.4 ASSAYFORMER architecture

• AssayFormer is a three-layer, two-head transformer that combines screen descriptions with labeled history genes, produces per-gene scores, and selects the top 100 untested candidates.A frozen text encoder represents the screen description, while the pooled description token yields a screen latent used by a bilinear gene-scoring head.Source: §4.4 · pp. 22–23

4.5 Training procedure

• AssayFormer training initializes gene embeddings with probit BPMF, uses simulated-history supervised learning, and applies GRPO on stochastic full-trajectory rollouts with context-delta rewards.The reward encourages the policy to exploit revealed context and outperform a frozen context-free policy over sequential campaigns.Source: §4.5 · pp. 23–24

4.6 LLM handoff strategy

• AssayLoop uses an LLM for the first three rounds to provide a biological warm start, then switches to the reinforcement-learned AssayFormer policy for feedback-driven adaptive selection.The handoff point k = 3 is selected on validation data.Source: §4.6 · p. 24

4.7 ASSAYLLM

• AssayLLM post-trains a 27B-parameter Qwen backbone with supervised acquisition trajectories and GRPO for either all rounds or warm-start control before AssayFormer.Rewards jointly reflect campaign enrichment, hits found during LLM rounds, distinct valid observations, complete batches, and handoff readiness.Source: §4.7 · pp. 24–25

4.8 Gene-gene influence analysis; 4.9 Baselines

• Gene-gene influence is estimated by contrasting target scores before and after a probe hit across randomized backgrounds, with directional and screen-specific variants; baselines include adaptive, transfer-based, LLM, and agentic methods.Comparisons include Probability-of-Hit, BioBO, BPMF, Screen-kNN, MAML, prior-hit ranking, multiple LLM families, LLMNN, ICBR-EF, and a coding agent.Source: §4.8 · pp. 25–26, §4.9 · p. 26

A Dataset Composition

• The benchmark retains 1,349 training screens and selects 20 validation and 20 test screens from AssayBench using its temporal split.The source benchmark contains 1,920 CRISPR screens.Source: A Dataset Composition · p. 27
• Screen selection enforces genome-wide coverage, sufficient hit signal, an LLM-signal floor, phenotype-balanced allocation, and within-phenotype description diversity.Eligible libraries contain 18,000–22,000 genes, at least 50 hits with hit rate ≤15%, and Gemini-3-Pro AnDCG@100 ≥0.05; phenotype allocation has a soft cap of six screens.Source: A Dataset Composition · p. 27
• Because selection requires detectable LLM predictive signal, ASSAYBENCH-LOOP evaluates adaptive discovery on signal-rich screens rather than an unbiased sample of genome-wide CRISPR screens.Supplementary Tables 1 and 2 summarize phenotype distributions and the 20 test screens.Source: A Dataset Composition · p. 27, Table 1 · p. 27, Table 2 · p. 28
• The study distinguishes a 23,724-gene training vocabulary, a 21,147-gene fixed acquisition universe, and screen-specific measured libraries.The acquisition universe contains genes appearing in at least two screens.Source: A Dataset Composition · p. 27

B Ablations

• ASSAYFORMER ablations remove or vary gene embedding initialization, the reinforcement-learning reward, and conditioning on the screen description c_s.These are the three ablated components of ASSAYFORMER.Source: B Ablations · p. 29

B.1 Embedding initialization

• Embedding-source performance varies substantially: after supervised fine-tuning, EF ranges from 1.06 for GenePT to 3.83 for BPMF across six initialization strategies.Compared initializations include BPMF, GenePT, SVD, MF, MF-Sphere, and K562 Perturb-seq embeddings.Source: §B.1 · p. 29
• BPMF embeddings achieve the best reported performance, reaching EF 3.83 after supervised fine-tuning and 4.83 after reinforcement learning.RL improves every initialization, while BPMF’s advantage over alternatives widens after RL; SVD reaches EF 3.72 after SFT and 4.40 after RL.Source: §B.1 · p. 29
• BPMF embeddings primarily organize genes by shared screen phenotype, and their geometry remains nearly unchanged during training with Pearson r = 0.999.Downstream performance does not track recovery of canonical relationships in STRING, CORUM, SIGNOR, or Reactome, suggesting adaptive discovery uses structure distinct from standard pathway annotations.Source: §B.1 · p. 29

B.2 Training strategies

• Context-delta reward outperforms terminal enrichment-factor optimization because it isolates the benefit of conditioning on accumulated experimental history.The terminal objective uses T = 10 acquisition rounds but provides a weaker learning signal for history-dependent adaptation.Source: §B.2 · pp. 29–30
• Removing the screen description has relatively minor performance effect, while enabling the LLM warm start to supply biological context before feedback-driven adaptation.AssayFormer’s initial acquisitions rely largely on cross-screen gene-hit patterns learned from training data rather than screen-description text.Source: §B.2 · p. 30

C Complete Results

• Supplementary Table 4 compares baselines and ablations using EF, nAUC, hit fraction, mean shortfall, Vendi ratio, pathway overlap, and pathway-group coverage metrics.EP-B, EP-S, and EP-D measure effective Reactome level-2 pathway groups at batch, screen, and dataset scopes and are not directly comparable.Source: Table 4 · p. 31

D.1 Gene embeddings initialization

• The gene-embedding initializer fits a latent-utility matrix-completion model to observed screen–gene pairs, using truncated-normal and conjugate Gibbs updates.Unobserved pairs are excluded rather than treated as zero labels; the model uses 1,349 training screens, 23,724 genes, and 64% observed pairs.Source: §D.1 · p. 32
• After 2,000 iterations with burn-in and thinning, ten-dimensional posterior-mean gene factors initialize covered genes without per-gene bias, while absent genes use zero factors and marginal-frequency bias.For covered genes, ranking signal comes from the inferred screen latent and the bilinear gene representation.Source: §D.1 · p. 32

D.2 GRPO

• GRPO normalizes remaining-return advantages within each group and updates Plackett–Luce policy probabilities with REINFORCE plus a 0.01 entropy bonus.The policy acquires from and is evaluated on the full gene universe, including genes without hit labels in the screen.Source: §D.2 · p. 32

D.3 Training settings

• AssayFormer training uses 40 SFT epochs followed by 100 GRPO epochs across four B200 GPUs, generating 1,078,400 traces and selecting the final checkpoint by validation handoff performance.GRPO uses group size 8, context-delta reward, policy updates every 25 epochs, and rollouts over 21,147 genes appearing in at least two screens.Source: §D.3 · p. 32

D.4 ASSAYLLM

• AssayLLM uses the Qwen3.6-27B backbone with SDPA attention, bf16 parameters, and DeepSpeed ZeRO-3; sequence packing is omitted because SDPA is used throughout.ZeRO-2 exceeds host memory for this model size, and FlashAttention-2 is numerically unstable on the stated accelerators.Source: §D.4 · pp. 32–33

D.4.1 SFT

• AssayFormer SFT trains on multi-turn screen trajectories containing ranked 100-gene outputs and incremental feedback, masking loss on assistant tokens only.Malformed assistant gene lists are dropped, affecting roughly 3% of rows; optimization uses full-parameter fine-tuning across 16 GPUs for 3 epochs with global batch size 32 and learning rate 1×10−6.Source: §D.4.1 · p. 33

D.4.2 GRPO

• GRPO rollout samples 16 responses per screen prompt at temperature 1.0 and top-p 0.95, then optimizes group-normalized advantages with a clipped surrogate objective.The implementation uses global batch size 48, actor learning rate 1×10−6 with 3% warmup, ZeRO-3/bf16, checkpointing, and SDPA attention.Source: §D.4.2 · p. 33

D.5 Screen-kNN implementation

• Screen-kNN combines an empirical gene-level hit-rate prior with similarity-weighted evidence from historical screens, increasingly emphasizing observed screen history as feedback accumulates.The acquisition score uses an exponential similarity kernel with τ = 0.1 and interpolates using λ_t = |G_t−1|/(|G_t−1| + κ), with κ = 100.Source: §D.5 · pp. 33–34

E.1 LLM Prompt Examples

• The example LLM prompt requires exactly 100 ranked HGNC human gene symbols, excludes previously sampled genes, and can provide either labeled hit feedback or unlabeled sampling history.It specifies the screen context, hit definition, FDR-based ranking criterion, comma-separated output format, and a no-prose response.Source: §E.1 · pp. 35–37
• 12.1% cumulative hits were obtained after validating 99 genes in the active-learning prompt example, with 12 hits and 87 non-hits recorded.The example concerns an SK-N-DZ CRISPRa screen testing resistance to RSL3-induced ferroptosis, where hits increase viability and satisfy FDR < 0.05.Source: §E.1 · p. 36

E.2 Qualitative analysis of ASSAYLOOP on two biologically distinct screens

• ASSAYLOOP is qualitatively evaluated against ASSAYFORMER and Gemini-3.1-Pro on two biologically distinct screens.The screens cover NF-κB/TNF signaling and AAV transgene silencing.Source: Figure 1 · p. 38
• The comparison visualizes gene composition across acquisition rounds and the hits recovered from each proposed batch.Each screen-model pair contains an upper acquisition-batch histogram and a lower hit-composition histogram.Source: Figure 1 · p. 38
• The figure distinguishes ASSAYFORMER, Gemini-3.1-Pro, and the ASSAYLOOP handoff through separate model rows.ASSAYFORMER appears in the top row, the handoff in the middle row, and Gemini-3.1-Pro in the bottom row.Source: Figure 1 · p. 38
• In the AAV transgene silencing screen, ASSAYFORMER proposes 100 genes for the first batch, with DNA repair as the largest category.Pathway bars are colored according to Reactome gene pathways.Source: Figure 1 · p. 38
• The qualitative analysis uses pathway-level composition to show how proposed acquisitions and resulting hits differ across models and screens.The visual reports pathway composition rather than only aggregate batch sizes.Source: Figure 1 · p. 38

F Gene Embedding Analysis

• The gene-embedding appendix provides additional analysis of ASSAYFORMER’s embeddings and complements the ablation reported in Section B.1.This subsection is distinct from the two-screen qualitative comparison.Source: F Gene Embedding Analysis · p. 39

F.1 Embedding drift during training

• 0.999 Pearson r and median RBO above 0.95 show that BPMF pairwise geometry and gene neighborhoods are almost perfectly preserved through training.The results indicate that ASSAYFORMER learns its acquisition policy on top of an effective fixed structure rather than reorganizing BPMF embeddings.Source: §F.1 · p. 39, Figure 2 · p. 40
• Embedding drift varies sharply by source: random embeddings nearly reorganize, SVD, MF, and MF-Sphere substantially change neighborhoods, and GenePT remains relatively stable despite poor downstream performance.Reinforcement-learning fine-tuning produces much smaller changes than supervised fine-tuning across embedding types.Source: §F.1 · p. 39, Figure 2 · p. 40
• BPMF’s effectiveness reflects both stability and alignment with adaptive hit discovery, unlike embeddings that either reorganize without matching its geometry or remain stable but task-misaligned.This phenotype-derived representation may complement existing gene embeddings and motivates evaluation in other downstream tasks.Source: §F.1 · p. 39, §F.3 · p. 43

F.2 Embedding initialization and biological plausibility

• GenePT and K562 best recover canonical STRING, CORUM, SIGNOR, and Reactome relationships but provide the weakest ASSAYFORMER initializations, whereas matrix-factorization embeddings perform best.This negative association suggests that canonical biological plausibility does not predict usefulness for adaptive hit discovery.Source: §F.2 · p. 39, Figure 3 · p. 41

F.3 Structure of the BPMF embedding space

• BPMF embeddings cluster by common-essential, context-essential, selectively essential, immune-response, and apoptosis phenotypes, with five unresolved DNA-damage-response clusters.These clusters were obtained with HDBSCAN using a minimum cluster size of 150 in the original embedding space.Source: Figure 4 · p. 42, §F.3 · p. 43
• CORUM and Reactome annotations show minimal spatial clustering in BPMF, except for partial Reactome organization within the common-essential region.Genes from the same complexes or pathways are generally distributed across the embedding space rather than forming compact groups.Source: Figure 4 · p. 42, §F.3 · p. 43
• BPMF embeddings organize primarily by shared screen phenotype, with proximity reflecting historical perturbation profiles rather than canonical pathway or complex membership.HDBSCAN identifies common-essential, context-essential, selectively essential, immune-response, apoptosis, and five DNA-damage-response clusters, while canonical annotations show little spatial grouping.Source: Figure 4 · p. 42, §F.3 · p. 43

Reactome composition

• LLMs generally acquire similar Reactome pathway distributions, emphasizing RNA and protein metabolism and under-representing bulk metabolism and developmental biology.The comparison aggregates requested genes across 20 test screens and contrasts pathway shares with uniform random sampling.Source: Figure 5 · p. 44

Model comparison

• Differences among LLMs are smaller than their differences from other baseline methodologies such as kNN.The figure reports model-level pathway composition as a complement to the method-level comparison in Figure 5A.Source: G Biological Diversity Analysis · p. 43, Figure 5 · p. 44

Coverage and output size

• Each LLM covers a narrow range of effective pathway counts, while Kimi-K2.6 and Opus often provide fewer than 100 requested genes.The lower gene counts may reflect biological safety training and result in fewer unique genes.Source: Figure 5 · p. 44

Hit enrichment factor

• The primary hit enrichment factor (EF) measures hits found relative to random expectation, normalizing by in-library, real out-of-library, and unfilled acquisitions while penalizing hallucinated genes.This differs from the classical normalization by total acquisitions and avoids penalizing valid genes outside the screen library.Source: H Metrics used in ASSAYBENCH-LOOP · pp. 44–45

Normalized area under the cumulative-hits curve

• Normalized AUC (nAUC) captures hit-ordering quality across the full acquisition trajectory, using effective queried budget and an oracle normalization; 1.0 means all hits are acquired first.Unfilled slots do not advance the curve, so under-supplying policies end at a smaller effective budget rather than being carried flat.Source: H Metrics used in ASSAYBENCH-LOOP · pp. 45–46

Recall and validity

• Fraction of hits (FH) measures whole-trajectory recall, while shortfall (SF) measures the fraction of acquisitions outside the screen library, including real out-of-library and hallucinated genes.FH counts only in-library hits, and FH=1.0 indicates that all screen hits were found within budget.Source: H Metrics used in ASSAYBENCH-LOOP · p. 46

Percentage of essential genes

• The percentage of essential genes (%ess) diagnoses concentration among broadly essential recovered hits using 1,827 common essential genes from DepMap.It is the fraction of acquired hits that are common essentials, which may not reflect screen-specific biology.Source: H Metrics used in ASSAYBENCH-LOOP · p. 46

Effective Pathways

• Effective Pathways (EP) measures biological breadth as the effective number of Reactome level-2 pathway groups, with rarefied scopes of 30, 200, and 6000 annotated genes.Higher EP indicates broader pathway coverage; Monte Carlo rarefaction and one-group-per-gene assignment improve comparability and prevent annotation multiplicity from inflating diversity.Source: H Metrics used in ASSAYBENCH-LOOP · p. 47

Batch diversity metrics

• Vendi ratio (VR) and pathway overlap (PO) assess batch diversity: higher VR indicates more diverse embeddings, whereas PO=1.00 matches random selection, PO>1 indicates narrower neighborhoods, and PO<1 broader batches.VR uses GenePT embeddings and Vendi score [14], while PO uses normalized pairwise GO biological-process Jaccard overlap; the metrics are complementary.Source: H Metrics used in ASSAYBENCH-LOOP · p. 47

I ASSAYFORMER Reveals Directional Gene-Gene Relationships

• ASSAYFORMER’s influence analysis reveals structured, context-dependent relationships among functionally related genes, while influences among canonical cancer drivers are relatively weak.The analysis measures acquisition-score changes after adding a probe gene as a hit across 10 random background contexts; Supplementary Fig. 7 covers 31 representative genes.Source: I ASSAYFORMER Reveals Directional… · p. 48, Figure 7 · p. 48
• The learned gene relationships are strongly directional: influence-matrix/transposed-matrix correlation is r = 0.16, and 44% of reciprocal pairs have opposite signs.For example, MDM2 increases PFDN4’s acquisition score by +0.33, whereas PFDN4 decreases MDM2’s by −0.46.Source: I ASSAYFORMER Reveals Directional… · p. 49
• After removing database-annotated pairs, high-magnitude influences still identify relationships consistent with co-essentiality, synthetic lethality, epistatic masking, and lineage-specific exclusion.Examples include increased spliceosome dependence after MYC hits, increased nucleolar-stress dependence after MDM2 hits, and reduced mitotic or OXPHOS dependence after PIK3CA or SMAD4 hits.Source: I ASSAYFORMER Reveals Directional… · p. 49, Figure 8 · p. 49

J Model Scaling

• Increasing training data improves ASSAYFORMER performance, whereas scaling from 0.28M to 10.97M parameters produces no meaningful improvement, making training screens the apparent bottleneck.The comparison spans 150 models, including the 4.44M-parameter default model; Supplementary Fig. 9 reports data but not meaningful model-size scaling.Source: J Model Scaling · p. 49, Figure 9 · p. 50

K Revisiting whether LLMs can learn from lab-in-the-loop feedback

• LLM performance consistently declines when previous-round hit labels are removed, showing that multiple LLM families use experimental feedback for subsequent acquisitions.The evaluation preserves previously acquired genes while omitting hit labels, directly measuring the benefit of outcome information.Source: K Revisiting whether LLMs can learn… · p. 51, Figure 10 · p. 51
• Despite benefiting from feedback, standalone LLMs are consistently outperformed by ASSAYFORMER, whose feedback-conditioned adaptation is substantially larger.This finding contrasts with prior reports from Gupta et al. and aligns with Wainrib et al. [51].Source: K Revisiting whether LLMs can learn… · pp. 50–51
• ASSAYLOOP combines an LLM warm start with later ASSAYFORMER acquisition, achieving strong label-blind performance and better in-context learning than standalone LLMs.The result supports explicitly training acquisition policies for feedback-conditioned sequential hit discovery.Source: K Revisiting whether LLMs can learn… · p. 51