Explore the research

Biology

Can AI predict how biological systems behave and turn its designs into working results?

33%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Useful novelty25%
Structural accuracy45%
Research efficiency30%
Generalization25%
Functional prediction30%
Design success40%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Design success

40%estimated progress

What we track

Experimentally functional designs / all synthesized AI proposals, with explicit activity, stability, and specificity requirements.

Why this estimate?

Experimentally successful binders demonstrate strong bounded design capability across targets. People still choose targets, run acceptance assays and establish broader biological utility.

Advancements

AlphaProteo binds 7 target proteins, with an 88% wet-lab success rate on BHRF1September 5, 2024. Google DeepMind reported that AlphaProteo designed binders for seven target proteins with 3 to 300 times better binding affinity than the best existing methods, including an 88% experimental success rate for candidates against the viral protein BHRF1. It failed to produce successful binders against an eighth target, TNF-alpha.
BindCraft designs protein binders with 46% average success2025 publication. BindCraft reports target-specific experimental binder success rates of 10–100%, averaging 46.3%, with affinities predominantly in the nanomolar range.
Designed serine hydrolases reach kcat/KM of 2.2x10^5 with sub-angstrom structural match2025. Lauko, Pellock et al., Science 388, eadu2454, report computationally designed serine hydrolases with catalytic efficiencies (kcat/KM) up to 2.2 x 10^5 M-1 s-1, whose crystal structures match the design models with C-alpha RMSDs below 1 angstrom.
RFantibody designs antibodies that bind C. difficile toxin at mid-nanomolar affinityNovember 5, 2025. The UW Institute for Protein Design reported RFdiffusion-designed antibodies against bacterial toxins, influenza and coronavirus proteins and cancer targets; anti-Clostridium difficile toxin designs showed mid-nanomolar affinity and retained potency as full IgG. Of five designed antibodies imaged with their targets, four matched the computational predictions closely.

Read the dated findings and sources

Status: Measured experimental design success. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2025 publication: BindCraft reports target-specific experimental binder success rates of 10–100%, averaging 46.3%, with affinities predominantly in the nanomolar range. BindCraft, Nature.

Interpretation and tracking value

The reported average is not a pooled success fraction for every generated or synthesized design, and binding alone does not establish all activity, stability and specificity requirements.

Next comparable measurement

Record all synthesized designs per target; require prespecified binding, function, stability and off-target thresholds and independent assay replication. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Functional prediction

30%estimated progress

What we track

Prospective prediction error and calibration for measured binding, enzyme activity, or cellular response. Use new experiments rather than only retrospective databases.

Why this estimate?

Variant-effect benchmarks demonstrate useful prediction across modalities, but prospective experimental calibration and autonomous improvement of the scientific process remain unestablished.

Advancements

AlphaMissense classifies 89% of 71 million possible missense variantsSeptember 19, 2023. Google DeepMind's AlphaMissense, published in Science, classified all 71 million possible human missense variants, calling 57% likely benign and 32% likely pathogenic and leaving 11% uncertain, at a threshold giving 90% precision on a database of known disease variants.
Boltz-2 reaches 0.62 Pearson correlation on the held-out FEP+ affinity benchmarkbioRxiv posting dated 2025-06-18. Boltz-2, from MIT's Jameel Clinic and collaborators, reports an average Pearson correlation of 0.62 on the standard FEP+ (OpenFE) affinity benchmark with targets held out of training, and outperformed all methods submitted to the CASP16 affinity challenge across 140 complexes.
AlphaGenome leads 25 of 26 variant-effect evaluationsJanuary 28, 2026. AlphaGenome matches or exceeds external models in 25 of 26 variant-effect evaluations.

Read the dated findings and sources

Status: Strong retrospective benchmark; prospective calibration incomplete. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 28, 2026: AlphaGenome matches or exceeds external models in 25 of 26 variant-effect evaluations. AlphaGenome, Nature.

Interpretation and tracking value

This is a count of benchmark comparisons, not 96% prediction accuracy or prospective experimental validation. It does not yield a common error or calibration baseline for binding, enzymes and cells.

Next comparable measurement

Register predictions before new assays; report effect-size errors, calibration and failures separately by endpoint and experimental batch. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Generalization

25%estimated progress

What we track

Performance on held-out protein families, cell types, and experimental conditions, reported separately.

Why this estimate?

Genomic holdouts provide some transfer evidence, but do not establish reliable performance across unfamiliar protein families, cell types and conditions.

Advancements

Evo 2 predicts variant effects across all domains of life with no finetuningPosted February 21, 2025. Evo 2, trained on 9.3 trillion DNA base pairs spanning all domains of life, is released at 7B and 40B parameters with a 1 million token context at single-nucleotide resolution, and predicts functional effects of genetic variants, including noncoding pathogenic mutations and clinically significant BRCA1 variants, from sequence alone without task-specific finetuning.
State model predicts perturbation responses in cell contexts unseen in trainingPosted July 10, 2025. Arc Institute's State model, trained on gene expression from over 100 million perturbed cells with a cell embedding trained on 167 million observational cells, identified strong perturbations in novel cellular contexts where no perturbations were seen during training, and improved discrimination of effects on large datasets by more than 30%.
CASP16 finds deep learning no better than traditional methods for RNA structures2025 Nov 3. The CASP16 overview reports that for RNA structures deep learning methods were not superior to traditional approaches, that most predictions were inaccurate without a structural template, and that the best-performing group used a hybrid of physics-based and deep learning methods. Antibody and nanobody-antigen complexes also remained difficult.
AlphaGenome is tested on unseen genome regionsJanuary 28, 2026. AlphaGenome was evaluated on unseen genome intervals and across regulatory modalities using human and mouse data.
Virtual Cell Challenge 2026 sets a zero-shot task on six unseen cell linesAugust 20, 2026. Arc Institute opened the 2026 Virtual Cell Challenge with no challenge-specific training set, requiring zero-shot prediction of CRISPRi knockdown responses in six cell lines never seen perturbed, three for the leaderboard and three held back for final testing, scored on an aggregate of six metrics. The 2025 edition drew more than 5,000 registrants from 114 countries and over 1,200 submitting teams.

Read the dated findings and sources

Status: Held-out genomic evidence; broader transfer unmeasured. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 28, 2026: AlphaGenome was evaluated on unseen genome intervals and across regulatory modalities using human and mouse data. AlphaGenome, Nature.

Interpretation and tracking value

Genomic holdouts do not establish transfer to unfamiliar protein families, cell types or experimental conditions. Report these dimensions separately.

Next comparable measurement

Freeze family, cell-type and condition holdouts; report absolute performance and drop from in-distribution results with uncertainty. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Research efficiency

30%estimated progress

What we track

Experimental cycles, human hours, elapsed time, and total cost per independently validated functional result.

Why this estimate?

Automated characterization speeds a defined workflow; complete cost per independently validated function and human-free laboratory management remain unmeasured.

Advancements

AI co-scientist proposes AML drug repurposing candidates confirmed in vitroSubmitted February 26, 2025. The Co-Scientist paper describes a Gemini-based multi-agent system whose hypotheses improve with more test-time compute, validated in three biomedical areas; in one case it identified drug repurposing candidates and combination therapies for acute myeloid leukemia that were confirmed in in vitro experiments.
Robin multi-agent system proposes and validates ripasudil for dry AMD19 May 2025. FutureHouse's Robin, a semi-autonomous multi-agent system, generated the hypothesis of enhancing retinal pigment epithelium phagocytosis for dry age-related macular degeneration and identified ripasudil, a clinically used ROCK inhibitor not previously proposed for the disease, which was then validated; a Robin-proposed follow-up RNA-seq experiment pointed to ABCA1 upregulation. All hypotheses, experimental plans, analyses and main-text figures were produced by Robin.
AI co-scientist reproduces a years-long gene-transfer finding in days9 September 2025. Imperial College London reported that Google's AI co-scientist, given the team's starting question about how cf-PICIs spread between bacterial species, independently proposed hypotheses matching the team's experimentally proven mechanism in a matter of days, where the human work had taken years. Two companion papers appeared in Cell (DOIs 10.1016/j.cell.2025.08.018 and .019).
SAPP platform characterizes hundreds of protein designs a dayAugust 20, 2026. The SAPP study reports hundreds of protein designs characterized per day, at least tenfold throughput improvement, and a roughly 48-hour process with about six hours of bench work.

Read the dated findings and sources

Status: Measured workflow throughput proxy. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

August 20, 2026: The SAPP study reports hundreds of protein designs characterized per day, at least tenfold throughput improvement, and a roughly 48-hour process with about six hours of bench work. Protein design validation platform, Nature Communications.

Interpretation and tracking value

Characterization throughput is not cost or time per independently validated biological function. Gene synthesis, failed designs and downstream functional assays must be counted.

Next comparable measurement

Measure total time, human hours and dollars from proposed sequence to externally reproduced function; retain failures and all assay stages. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Structural accuracy

45%estimated progress

What we track

Accuracy against subsequently released experimental structures using preregistered structural and interface metrics; split by molecular class and similarity to training data.

Why this estimate?

Joint molecular structure prediction shows broad bounded capability and comparison against experimental structures. New-target selection and faithful prospective validation still rely on human-designed workflows.

Advancements

AlphaFold 3 predicts protein complexes with nucleic acids and small moleculesMay 8, 2024. AlphaFold 3 reports improved joint prediction of protein complexes with nucleic acids, small molecules and other partners.
Google DeepMind releases AlphaFold 3 model code and weights for academic useNovember 11, 2024. An update on Google's AlphaFold 3 announcement, dated November 11, 2024, states that AlphaFold 3 model code and weights were released for academic use. The same page reports AlphaFold 3 is 50% more accurate than the best traditional methods on the PoseBusters protein-ligand benchmark.
CASP16 assessment: AlphaFold 3 baseline outscores best ligand-pose predictorPublished online October 4, 2025. In the CASP16 pharmaceutical assessment, 30 groups submitted predictions for 229 protein-ligand pose targets and 140 affinity targets across five protein systems. An AlphaFold 3 baseline reached a mean LDDT-PLI of 0.8, above the best CASP16 predictor's 0.69, while the best affinity correlation was a Kendall's tau of 0.42 against a theoretical maximum of about 0.73.
CASP16 overview reports AlphaFold 3 gives only a modest gain over AlphaFold 22025 Nov 3. The CASP16 overview in Proteins concludes that AlphaFold 3 yields a modest improvement over AlphaFold 2, with monomer accuracy essentially unchanged between 2022 and 2024 because agreement approaches the limit set by experimental uncertainty. The clearest gain was in protein-ligand poses, where AlphaFold 3 outperformed all participating groups with a median RMSD just over 1 angstrom.

Read the dated findings and sources

Status: Experimental-structure benchmark available. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

May 8, 2024: AlphaFold 3 reports improved joint prediction of protein complexes with nucleic acids, small molecules and other partners. AlphaFold 3, Nature.

Interpretation and tracking value

There is no single meaningful accuracy percentage across these molecular classes. This review does not extract a fresh preregistered post-release structural score; historical benchmark results remain distinct from that target.

Next comparable measurement

Evaluate subsequently released structures using fixed interface and structural metrics, stratified by molecular class and training similarity. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Useful novelty

25%estimated progress

What we track

Designs or mechanisms judged substantively different from known sequences and functions; sequence difference alone does not establish a new function.

Why this estimate?

Designed functional binders show constructive capability, while genuinely new biological function and autonomous novelty assessment are not yet demonstrated by the cited examples.

Advancements

ESM3 generates esmGFP, 58% identical to the closest natural fluorescent protein6.25.24. EvolutionaryScale's 98-billion-parameter ESM3 generated esmGFP, which is about 58% similar in sequence to the closest known natural fluorescent protein and differs by 96 mutations across 229 residues. Two rounds of 96 generated proteins were tested experimentally; esmGFP showed brightness similar to natural GFPs.
BindCraft produces working protein binders from scratch2025 publication. BindCraft demonstrates de novo designed functional protein binders.
Genome language models produce 16 viable bacteriophages with evolutionary noveltyPosted September 17, 2025. A bioRxiv preprint from Arc Institute and Stanford reports generative design of complete bacteriophage genomes using genome language models, with experimental testing of AI-generated genomes yielding 16 viable phages showing substantial evolutionary novelty relative to the phiX174 template.

Read the dated findings and sources

Status: Functional design examples; novelty-qualified count unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2025 publication: BindCraft demonstrates de novo designed functional protein binders. BindCraft, Nature.

Interpretation and tracking value

A new sequence or new binding interface need not constitute a new biological function. The paper's success rate cannot be reused as a substantially new-function rate.

Next comparable measurement

Have independent reviewers classify prior art and function before counting; require experimental function beyond sequence or structural difference. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note