Framing

Learning to Predict Biology (and what to measure next)

AssayBench & AssayLoop

Learning to Predict Biology
(and what to measure next)

Carl Edwards

Chen Institute Symposium for AI Advancing Science and Society 2026
Lifelong Learning Agents workshop · November 2026

Functional assay examples from the AssayBench explainer

Lab-in-the-loop

ChooseThe next experiments

→

MeasureWhat cells do

→

LearnFrom the results

Better predictions. Better experiments.

Framing

Can we “solve” biology?

I think it’s good to pick research directions based on things we want to exist in the world or big goals.

In that spirit, let’s say we have a (fairly ridiculous) goal of “solving” biology. What does that mean?

Framing

What would “solving biology” mean?

For me: predict the functional behavior of a biological system with high precision, and use those predictions to control it.

If we had a perfect predictive system, we could interrogate it for understanding as much as we wanted.

This is the idea behind the “virtual cell” paradigm.

* understanding biology as a solution would be great, but it’s unclear if that’s even possible in a non-black-box manner

Framing

Biology is like language, not physics

Physics has rules.
Biology is an evolved system. It exists because it works.
Biology is chaotic and filled with exceptions. So is human language.

Viruses protect us from bacteria

Mucus + phages↓Fewer bacteria reach cells

Viruses in the body’s microbiome can help regulate bacterial populations.

Phages in mucus · PNAS, 2013

Genes outside chromosomes

Chromosome+ecDNA

Circular extrachromosomal DNA can carry amplified cancer-driving genes.

Extrachromosomal DNA · Nature, 2017

New genes from noncoding RNA

Noncoding
RNA→New DNA
gene→Defense
protein

Some bacteria assemble new protein-coding genes from noncoding RNA to defend against viruses.

DRT2 gene synthesis · Science, 2024

Rule-based language systems fell short

Hand-written rules struggled with context and exceptions.

→

What worked?

Flexible models trained on large amounts of data.

Framing

So what does this tell us to do?

To “solve” biology, we need a system complex enough to learn its “black box” behavior from data.

A sufficiently flexible model

Learn the complexity of the biological system.

A lot of data

Observe how that system behaves.

Our next big problem: we need a lot of data.

Framing

We need a lot of data

What is the best way to get that data?

Experimental design and hypothesis generation

How can we use the data most effectively?

Algorithms and evaluation

Performance versus training screens on a logarithmic axis; plotted values are GRPO recovery AUC

Scaling laws give us a rough idea of how collecting more data will improve performance.

Algorithmic improvements improve the scaling law, and data moves us towards better performance.

Framing

Choose experiments that maximize value

Possible experiments

Fixed budget→Choose a subset

Experiments to run

Best experiments = arg max expected valuewithin the available budget

Improve the model
Information that improves a modelMore discoveries
Hits found in an experimental screen

Framing

Active learning: let results guide the next experiment

Choose

Select valuable experiments

→

Measure

+−+

Observe the outcomes

→

Update

Revise predictions and value

Use each result to make the next choice more valuable.

This is the “lab-in-the-loop” paradigm!

AssayBench

An Assay-Level Virtual Cell Benchmark
for LLMs and Agents

Edward De Brouwer Equal contribution

Carl Edwards Equal contribution

Alexander Wu Equal contribution

Jenna CollierGraham HeimbergXiner LiMeena SubramaniamEhsan HajiramezanaliDavid RichmondJan-Christian HütterSara MostafaviGabriele Scalia

Genentech · Equal contribution

arXiv:2605.10876

Let’s put this idea into practice with CRISPR screens

CRISPR screens

A question for every gene

CRISPR screens

What is a CRISPR screen?

AssayBench

CRISPR screens ask what the cell did

AssayBench

Why CRISPR screens make a useful benchmark

Predict which perturbations change what a cell does.

Functional endpoints

Drug discovery needs functional outcomes: which targets change drug sensitivity, survival or infection?

Beyond gene-expression prediction

Perturb-seq measures expression changes. Functional screens ask which perturbations change cellular behavior.

A bounded experimental space

A defined gene library gives us a finite set of candidates to rank and evaluate.

Measured experimental outcomes

Evaluate priorities against real screens, with temporal splits and assay-adjusted ranking metrics.

AssayBench

AssayBench collects 1,920 public CRISPR screens

Five phenotype classes from the supplied PPTX

One common task:
rank genes for a specified screen.

AssayBench

Processed and curated screens

1,971

CRISPR screens

1,584

Pass quality control

1,568

Replicates merged

1,920

Benchmark entries

Individual screen records

Screen A · replicate 1Screen A · replicate 2Screen B · replicate 1

→

Check quality

✓ usable measurements
✓ consistent gene identifiers

→

Combine replicates

Screen AScreen B

This is the part where most of the work goes.

AssayBench

Train on the past, test on the future

Temporal cohorts and phenotype distributions from PPTX slide 16

AssayBench-Loop

Predicting Biology Out-of-distribution

Training is dominated by fitness screens; the test set spans a broader mix of phenotypes.

Test hit sets have little overlap with their nearest training screen.

Genes that are frequent hits are enriched for common-essential genes.

Hit rates vary widely across screens

AssayBench

Held-out screens are genuinely new

AssayBench

A long tail of hits

AssayBench

Predicting hits from an assay description

AssayBench

Scoring a ranking

AssayBench

But nDCG is not comparable across screens

AssayBench

AnDCG

AssayBench

AnDCG makes rankings comparable across screens

Reward useful genes near the top

Discount relevance by rank, then normalize by the best possible ordering.

Account for the random baseline

Dense-hit screens are easier to score on by chance.

Adjusted nDCG formula from supplied PPTX slide 18

0 Random expectation

1 Perfect ranking

AssayBench

Generalist LLMs provide a strong starting point

Test set · publication-year split · AnDCG@100

LLM RRF Ensemble

0.163

Gemini 3 Pro

0.157

GPT-5.4

0.147

Gene-frequency baseline

0.122

GPT-OSS 120B

0.121

Biomni

0.108

Embedding kNN

0.065

Gene-relevance predictor

0.057

C2S (Gemma-2B)

0.036

Selected methods · higher is betterOracle kNN: 0.292 (privileged-information reference)

AssayBench

Replicate screens reveal substantial room to improve

Technical replicate reference versus Gemini 3 Pro across phenotype classes

A useful empirical reference

How well does another measurement of the same screen predict the outcome?

Models remain below that reference.

AssayBench

Performance depends on the biology being tested

AnDCG by phenotype category from PPTX slide 22

AssayBench

Well-studied biology is easier to predict

Regression coefficients for Gemini 3 Pro performance by publication year, citations, and phenotype

AssayBench

AssayBench: performance across model sizes

Model-size and optimization results from supplied PPTX

AssayLoop

From zero-shot to lab-in-the-loop

AssayBench

Use prior knowledge to predict experiments zero-shot

AssayLoop

Use prior knowledge and new results to inform future experiments

Can we learn how to collect better data
from the data we already have?

AssayLoop

Learning to choose the next experiment

Carl Edwards Equal contribution

Edward De Brouwer Equal contribution

Xiner Li

Anne BitonPaula CoelhoNamkyeong LeeEhsan HajiramezanaliAviv RegevSara MostafaviGabriele Scalia

Equal contribution

AssayLoop

Discovering hits under a budget

AssayLoop

Spend the budget in rounds

AssayLoop

Choose, observe, adapt

AssayLoop

AssayBench-Loop: train on the past, test on the future

AssayLoop

Measuring hit recovery

AssayLoop

How much better than random is hit discovery?

Illustrative recovery curve: at 5% library coverage, a strategy recovers 20% of hits versus 5% expected at random

Enrichment factor (EF)

Hits foundHits expected at random

Example: 20,000 genes · 200 hits
Test 1,000 genes; find 40 hits.

Random expectation:
1,000 × (200 / 20,000) = 10 hits

EF = 40 / 10 = 4

Four times the hits expected at random.

AssayLoop

Learning an acquisition policy across screens

Independent, LLM-based and amortized active hit discovery

AssayLoop

Two ways to choose the next experiments

Bayesian optimization · BioBO
Fit a model within the current screen.

LLM · Gemini 3.1 Pro
Use biological prior knowledge.

Mean recovery across 20 held-out screens

AssayLoop

Strong priors, weak adaptation

AssayLoop

Learn how to choose experiments from historical screens

Historical screens

Screen A

Screen B

Screen C

Known experimental outcomes

Train across screens→

AssayFormerA learned acquisition policy

Learn which experiment to choose
from the outcomes seen so far.

↓Reuse the policy on a new screen

Choose a batch

Apply the learned policy

→

Observe results

Run the experiments

→

Update the history

Use the new evidence

Learn across screens. Adapt within each screen.

AssayLoop

AssayFormer learns from experimental history

AssayLoop

AssayFormer conditions on the experimental history

AssayFormer architecture and acquisition of the highest-scoring untested genes

AssayLoop

Training uses simulated screening campaigns

BPMF initialization, supervised fine-tuning, and reinforcement learning

Initialize gene representationsLearn recurring hit patterns across historical screens.
Learn from partial historiesPredict unobserved hit labels from simulated experimental context.
Fine-tune with reinforcement learningReward in-context learning ability by simulating full experimental campaigns.

AssayLoop

Reinforcement learning rewards the use of feedback

GRPO rollouts, sampling without replacement, and a worked context-delta reward example

Gene selection

We select untested genes without replacement, using the same candidate pool for both models.

Return for round 7

G₇ = r₇ + r₈ + r₉ + r₁₀

The round-7 update uses rewards from rounds 7–10.

Compare this return across eight rollouts from the same screen.

AssayLoop

A strong prior

AssayLoop

AssayFormer improves with feedback

AssayLoop

Combining priors and feedback

AssayLoop

LLM warm start, then AssayFormer

LLM acquisition for rounds one to three, followed by AssayFormer

Gemini 3.1 ProStrong early recovery; gains taper

AssayFormerStarts lower; benefits from feedback

AssayLoop (Gemini)LLM warm start → AssayFormer

AssayLoop

27.9% of hits at about 5% library coverage

5.66×hit enrichment over random

Policy	Mean EF
GPT-5.6 Sol	4.77
AssayFormer	5.03
AssayLoop	5.66

20 held-out screens. Ten rounds of 100 requested genes.

AssayLoop improves on either component alone in this evaluation.

AssayLoop

AssayFormer gains more from experimental feedback

LLMs benefit from feedback, but the gain is modest.

AssayLoop

Performance scales with data, but not model size

Data and model scaling: recovery versus acquisition budget

AssayLoop

Generalization to an excluded phenotype category

Extended-cohort and leave-one-phenotype-out enrichment results

334-screen extended test cohort

AssayFormer training	Mean EF
All phenotype categories	4.44
Focal phenotype excluded	3.65

Exclusion changes both phenotype coverage and training-set size.

AssayLoop

Gene representations improve acquisition performance

Impact of gene token initialization after SFT and RL

AssayLoop

Feedback changes which genes enter the next batch

Figure 2E: feedback-guided gene prioritization, batch composition, and changes in unassayed gene ranks

AssayLoop

Improved recovery can retain biological breadth

Same illustrative NF-κB screen	Gemini	AssayFormer	AssayLoop
Hits recovered out of 169	50	73	82

AssayLoop

Gene selection preferences

Publication counts and historical hit probabilities across acquisition rounds

AssayLoop

Distributional bias in found hits

Wasserstein distance ratios for the recovered hit distributions

AssayLoop

Preferential recovery of common-essential hits

Recovery of common-essential versus other hits for four methods

AssayLoop

Model influences can nominate biological hypotheses

Context-dependent gene influence examples in three test screens

How does observing gene A as a hit change the acquisition score of gene B?

AssayLoop

Post-training improves an LLM acquisition policy

AssayLLM · Qwen3.6-27B

Supervised fine-tuningLearn from GLM-5.1 acquisition trajectories.
Reinforcement learningOptimize complete screening campaigns.
Handoff trainingOptimize the history passed to AssayFormer.

58% higher mean EF
than the base model

AssayLLM base, SFT, GRPO, and handoff results

AssayLoop

What AssayLoop establishes, and what remains open

Evidence so far

Historical screens can train transferable acquisition policies.
LLM priors and learned feedback use complement each other.
Task-specific post-training improves LLM policies.

Next steps

Prospective evaluation in a live experimental loop.
Better using heterogeneous, multi-scale data
Better integration with automated data collection.
Agents vs latent models?

Outlook

Biology needs more data

More data

Collect more useful experiments.

Better algorithms

Learn more from the data we have.

There is a lot we can’t measure yet.
Where is spatiotemporal transcriptomics data!?

There is even more we don’t know yet we should measure.

We need algorithms that learn from new experiments
to discover what to measure next.

Learning algorithms and experimental systems
have to improve together.

Outlook

A more useful question than “can we solve biology?”

Can each round of experiments make
the next round more informative?

AssayBench
Measure what models can predict.

AssayLoop
Learn how to act on experimental feedback.

AssayBench & AssayLoop

Thanks to my coauthors

Jenna CollierGraham HeimbergMeena SubramaniamEhsan HajiramezanaliDavid RichmondJan-Christian HütterAnne BitonPaula CoelhoNamkyeong LeeSara MostafaviAviv Regev

Discussion