Define the next decade of drug discovery.

Examples by category

Illustrative nominations only; they do not represent selected challenges or define prize conditions.

Target Identification & Validation

Can we assign a confidence score to a causal gene-disease link, such that the score predicts downstream clinical success?

Full example

Why Now?

Minikel et al. (2024) provide target-indication pairs that could be used to track the strength of causal-gene assignment. Meanwhile, the emergence of population-scale sequencing provides us with allelic series, direction of effect, and human knockout evidence gene by gene. Third, by integrating methods such as fine-mapping, locus-to-gene scoring, trans-acting protein quantitative trait loci (trans-pQTLs), enhancer-to-gene features and sequence-to-function models such as AlphaGenome (Avsec et al., 2026), we are starting to move the field of causal gene assignment from a qualitative judgement call into a probabilistic, quantifiable process.

If Solved, What Changes?

Target selection becomes a quantifiable process. Companies and teams could quantify efficacy, identify uncertainty, and prioritize high-value targets before major investment.

Success Criteria

Successfully answering this question means that before we have clinical outcomes, we can reliably predict which gene is causal for a given disease and which direction of effect is required to achieve efficacy. Additionally, instead of indicating a lack of certainty, a low score should predict a low probability of success in a clinical trial. Moreover, this calibration between score and clinical outcome should hold within strata of disease area, modality, orphan status, sponsor characteristics, and calendar year, even where raw success rates differ across those strata.

References

• Minikel et al. (2024), Refining the impact of genetic evidence on clinical success
• Avsec et al. (2026), Advancing regulatory variant effect prediction with AlphaGenome

Binder Generation

Can we design high-affinity de novo binders to cryptic or conformational epitopes within a small, fixed experimental budget?

Full example

Why Now?

Structure prediction algorithms have largely solved static single-chain folding, and the field’s attention is shifting to ensembles and dynamics; complexes with no close structural homologs, such as antibody-antigen interfaces, remain hard. Recent advances in AI de novo binder design algorithms, including RFdiffusion and BindCraft, along with preprints such as BoltzGen, PXDesign, SeedProteo, Chai-2, and Origin-1, have shown that de novo binder design against rigid and convex epitopes can reliably reach nanomolar-range affinity. We hypothesize that a dominant remaining failure mode is binding sites poorly captured by deposited structures, such as cryptic or conformational pockets.

If Solved, What Changes?

The set of epitopes inaccessible to designed protein-based therapeutics greatly shrinks. Cryptic and conformational sites on G protein-coupled receptors (GPCRs), transporters, intrinsically disordered regions, and even transient protein-protein interfaces become tractable by design.

Success Criteria

On a blind target set with no known designed binders and no complex structure sharing more than 60% sequence identity with any training example, where target sites are identified as cryptic or conformational a priori from apo-structure ensembles or molecular dynamics (MD)-predicted pocket-opening events (not from a known binder), the design method must yield experimentally validated sub-nanomolar binders. Additionally, this should be achieved with a small budget, such as fewer than 24 designs per target, to demonstrate that the algorithm has learned to generate high-affinity binders rather than relying on extensive experimental screening.

References

• Watson et al. (2023), De novo design of protein structure and function with RFdiffusion
• Pacesa et al. (2025), One-shot design of functional protein binders with BindCraft
• Stark et al. (2025), BoltzGen: Toward Universal Binder Design
• Protenix Team et al. (2025), PXDesign: Fast, Modular, and Accurate De Novo Design of Protein Binders
• Qu et al. (2025), SeedProteo: Accurate De Novo All-Atom Design of Protein Binders
• Chai Discovery Team et al. (2025), Zero-shot antibody design in a 24-well plate
• Levine et al. (2026), Origin-1: a generative AI platform for de novo antibody design against novel epitopes

Preclinical-to-Human Translation

Can we predict human safety and immunogenicity from preclinical data well enough to determine which molecules advance into first-in-human trials?

Full example

Why Now?

The U.S. Food and Drug Administration’s (FDA) 2025 Roadmap to Reducing Animal Testing in Preclinical Safety Studies, followed by a December 2025 draft guidance on monoclonal antibodies, encourages evaluation of human-relevant alternatives to animal testing. Sponsors have also progressed more compounds and biologics through early clinical development, accumulating paired preclinical and human outcome data, though most of it remains proprietary and unpublished. The heightened incentive to develop alternative models, combined with this growing (if largely private) body of outcome data, creates an opportunity to systematically develop models that predict clinical outcomes and aid preclinical-to-human translation.

If Solved, What Changes?

If we can reliably predict human safety and immunogenicity from preclinical data, we can better eliminate molecules that pose high risks to the human body. This will help teams prioritize candidates and design follow-up studies before first-in-human trials. If successful, such predictive models could reduce Phase I attrition driven by safety and pharmacokinetic (PK) failures.

Success Criteria

On a blinded set of molecules with known but withheld clinical outcomes, the model must predict Phase I safety failures with a sufficiently high negative predictive value to support go/no-go decisions, and predict the human efficacious dose within a pre-specified fold-error. For immunogenicity, because clinical anti-drug antibody (ADA) incidence depends heavily on assay method, sampling schedule, and drug tolerance and is not directly comparable across molecules, predicted scores must correlate with observed ADA incidence within a dataset assayed under a consistent protocol, or reliably separate molecules into low/medium/high immunogenicity-risk tiers.

References

• FDA (2025), Roadmap to Reducing Animal Testing in Preclinical Safety Studies
• FDA (2025), Monoclonal Antibodies: Streamlined Nonclinical Safety Studies (draft guidance)
• FDA (2025), FDA Announces Plan to Phase Out Animal Testing Requirement for Monoclonal Antibodies and Other Drugs
• FDA (2014), Immunogenicity Assessment for Therapeutic Protein Products

Function

Can we specify the desired functional phenotype upfront, such as agonism, antagonism, partial agonism, or biased signaling, and reliably design molecules that achieve that specific functional profile?

Full example

Why Now?

While binding is becoming increasingly tractable, predicting and designing function remains a challenge. This is because efficacy is not a single pose. It requires consideration of a conformational ensemble. One recent campaign recovered agonists by screening thousands of designs rather than by prediction (Muratspahić et al., 2026). At the same time, the emergence of high-throughput functional assays creates an opportunity to use functional assay outputs as training data (Avet et al., 2022), making function-driven design a realistic possibility in the near future.

If Solved, What Changes?

Design stops being a two-step process of making binders and then screening for the ones that perform the intended function. Functional phenotype becomes a design objective rather than a post-design screening step.

Success Criteria

A successful solution means that a design method can accurately predict a molecule's functional phenotype before experimental validation, and that the predicted function is subsequently confirmed in vitro and in vivo. Concretely, among molecules that are experimentally tested after being predicted as agonists, antagonists, partial agonists, or biased signaling molecules, we should expect a > 10% hit rate (fraction confirming the intended phenotype); for predicted agonists and partial agonists, a Spearman correlation of > 0.6 between predicted and measured E_max; for predicted antagonists, an equivalent correlation on pIC₅₀ or pA₂; for predicted biased ligands, a correlation on bias factor for an explicitly specified pair of signaling pathways; and the ability to replicate these successes on unseen target families.

References

• Muratspahić et al. (2026), De novo design of miniproteins targeting GPCRs
• Avet et al. (2022), Effector membrane translocation biosensors reveal G protein and βarrestin coupling profiles of 100 therapeutically relevant GPCRs

Therapeutic Window & Conditional Logic

Can we design conditionally active therapeutics with a predictable in vivo therapeutic window?

Full example

Why Now?

Masked therapeutics have now progressed beyond a single clinical precedent, with multiple programs spanning masked antibody-drug conjugates (ADCs), checkpoint antibodies, cytokines, and protease-activated T-cell engagers. Several early masked-therapeutic programs, including praluzatamab ravtansine and pacmilimab, were discontinued or yielded mixed clinical results rather than advancing outright. That track record, rather than undercutting the case, underscores the need for a predictive understanding of the in vivo therapeutic window before committing a conditionally active construct to the clinic.

If Solved, What Changes?

If we can successfully develop a predictive understanding of the in vivo therapeutic window, teams can prioritize designs that are more likely to succeed in a conditional activation setting. Moreover, the druggable target space will be expanded; those that were deemed inaccessible due to safety concerns can be reconsidered.

Success Criteria

Success will be defined by (1) demonstrating a predefined improvement in the in vivo therapeutic window of conditionally active constructs, relative to experimental controls, and (2) establishing an in vitro readout that predicts the in vivo therapeutic window against a predefined accuracy benchmark in a blinded, diverse set of molecules. It should be noted that human translation will require separately defined evidence.

References

• Johnson et al. (2021), Phase I, First-in-Human Study of the Probody Therapeutic CX-2029 in Adults with Advanced Solid Tumor Malignancies
• NCT03543813, PROCLAIM-CX-2029, Phase 1/2 masked anti-CD71 ADC, completed
• NCT04596150, praluzatamab ravtansine with or without pacmilimab, Phase 2, completed

Systems-Level Design

Can we predict emergent functions of multispecific therapeutics across unseen combinations of binding arms?

Full example

Why Now?

For one, co-folding models such as AlphaFold3 and the Boltz series can now predict multi-chain assemblies with greater accuracy than before, though precision on antibody-antigen interfaces and T-cell engager ternary geometries remains limited. Second, high-throughput functional assays are starting to make multi-arm functional data available, though public datasets pairing multiple binding arms with functional readouts for multispecific constructs remain scarce. Third, the mechanistic theory to embed as inductive bias is an active research area: Bellout et al. (2026) extend ternary binding models to membrane-confined systems with finite copy number, using bispecific T-cell engagers (BiTEs) such as blinatumomab as a worked case.

If Solved, What Changes?

Teams could optimize interacting functional requirements before synthesis. For example, a model might help identify constructs that maintain tumor-cell killing while limiting excessive cytokine release, rather than optimizing isolated arm affinities and discovering the combined behavior only after testing.

Success Criteria

While the endpoint value is important, successfully answering this question means that the model accurately predicts the interaction term, defined as the signed departure from a null prediction based on each arm's affinity and monospecific activity. Performance is demonstrated on emergent endpoints with no monospecific counterpart, including antigen density-dependent selectivity, cis- versus trans-engagement, and decoupling of correlated outputs such as cytotoxicity and cytokine release.

References

• Bellout et al. (2026), A Generalization of the Ternary Binding Model to Membrane-Confined Systems with Finite Copy Number