A closed loop for drug discovery
Drug discovery is a chain of decisions. A structural hypothesis must become a binding hypothesis; a binding hypothesis must become a ranked set of molecules or biologics; and the highest-value candidates must be tested, interpreted, and redesigned. When these steps live in separate tools, scientists spend time reconciling incompatible outputs, estimating confidence by hand, and deciding which experiments are worth running next.
AnewDDE is an agentic Drug Discovery Engine for this decision loop. It connects four capabilities: AnewFold for biomolecular structure modelling and pocket analysis, AnewAffinity for fast and inspectable binding-affinity ranking, AnewDesign for agent-assisted biologics design with wet-lab feedback, and AnewMind for ADMET prediction and long-horizon pharmaceutical reasoning. The goal is not a single prediction. It is a repeatable cycle of hypothesis → quantitative prioritisation → candidate generation → experimental validation → next-round design.
Structure prediction for difficult and novel systems
Reliable structure hypotheses are the foundation of the loop. AnewFold models antibody–antigen complexes, protein–ligand interactions, ligand-free pockets, and molecular-glue ternary assemblies. Its advantage is most visible when the target or interaction mode is far from familiar training examples.
Across five seeds, AnewFold reaches top-1 success rates of 62.3% on AF3-AB, 60.8% on PXMeter-AB, and 76.2% on FoldBench-AB. On the similarity-filtered FoldBench protein–ligand benchmark, it reaches a 77.4% top-1 joint success rate. On a post-cutoff molecular-glue benchmark, it reaches 71.8%, compared with 45.0% for Protenix-v1 and 29.8% for Protenix-v2.
In ligand-free pocket identification, AnewFold reaches an overall AUPRC of 0.751, including 0.663 for cryptic pockets and 0.755 for non-cryptic pockets.
From affinity estimates to design decisions
A pose is a starting point. Design teams also need to know which transformation is likely to improve binding and which prediction deserves a second look. AnewAffinity combines learnable models with FEP-compatible, phase- and window-specific confidence estimates, so a result can be ranked quickly and inspected when the chemistry is consequential.
On 199 ligands from eight retrospective JACS systems, AnewAffinity achieves a ligand-count-weighted Pearson R² of 0.553 and Spearman correlation of 0.724, compared with 0.486 and 0.620 for the local Boltz-2 evaluation. Pairwise ΔΔG RMSE is 1.057 kcal/mol for AnewAffinity and 1.016 kcal/mol for Boltz-2; AnewAffinity’s most significant advantage lies in preserving binding-affinity trends and accurately capturing ligand affinity rankings. AnewAffinity evaluates one ligand-pair edge in approximately 1.5 seconds on a single GPU, making it suitable for high-throughput screening and early design-cycle decisions.
The output is not only a scalar. For a held-out BACE transformation, AnewAffinity predicts ΔΔG = +1.12 kcal/mol, close to the experimental +1.33 kcal/mol, while exposing the phase and alchemical windows that support the estimate. Ambiguous transformations can then be escalated to explicit FEP, structural re-analysis, or experiment.
This operating point connects accuracy with throughput: broad chemical series can be ranked early, while the evidence behind a high-value or uncertain edge remains available for review.
Closing the loop for biologics design
For antibodies and nanobodies, the challenge is not only generating a plausible sequence. A candidate must bind to the target epitope, remain developable, express well enough to test, and improve when experimental evidence arrives. AnewDesign turns these constraints into an agent-assisted lab-in-the-loop workflow.
The workflow starts from a target and epitope, generates sequence–structure-aware candidates, triages them for interface plausibility and developability, and sends a compact panel to wet-lab testing. Experimental affinity and developability results then update the next round of design priorities. In the proof-of-concept nanobody campaign, the agent generated 3,092 candidates, formed 287 filtered clusters, and recommended 50 sequences for the first wet-lab round. 7 of 50 designs showed SPR-derived KD values below 400 nM. An in silico maturation round proposed 100 additional sequences. Across the 150-clone campaign, 16 clones reached single-digit-nanomolar KD.
The practical benefit is a smaller, more informative experimental search space. Instead of treating the laboratory as a final checkpoint, AnewDesign uses measured affinity and developability as feedback for the next design round.
Reasoning for ADMET prediction and pharmaceutical R&D
A drug candidate is judged across more than potency. ADMET, developability, selectivity, assay validity, and project strategy all shape whether a molecule should move forward. AnewMind is a scientific language model post-trained at the hundred-billion-parameter scale to connect these kinds of evidence to pharmaceutical decisions.
In zero-shot evaluations across four modalities, AnewMind Preview ranked in the top five in every aggregate setting: small-molecule classification (Elo-AUROC 0.771), small-molecule regression (Elo-Spearman 0.520), cyclic-peptide PAMPA permeability (Elo-AUROC 0.737, ranked first), and antibody developability (Elo-Spearman 0.247).
PharmBench is an internal benchmark developed by pharmaceutical experts that tests the longer chain of reasoning. It contains 50 R&D scenarios and 200 questions built from experimental evidence, covering project decisions, ADMET, SAR interpretation, molecular design and synthesis, chemical reasoning, structure-based design, and biology and assay validity. AnewMind demonstrated consistent strengths across most capability dimensions and cases, with overall performance competitive with leading frontier models worldwide.
The intended role is decision support: explain what the evidence shows, identify the next uncertainty, and propose a testable next step.
Advancing closed-loop drug discovery
AnewDDE brings together the steps that are usually separated: model a difficult interaction, identify a designable site, rank candidate transformations, generate biologics, read experimental feedback, and reason about what should happen next. Its modules are evaluated separately, but their value grows when they are used as one decision loop.
The result is a foundation for agentic drug discovery across modalities. AnewFold supplies structural hypotheses for difficult targets. AnewDesign turns those hypotheses into experimentally testable biologics. AnewAffinity makes affinity ranking fast enough to guide broad screening. AnewMind connects molecular evidence to ADMET and project-level reasoning. Together, they reduce decision friction between a model output and the next experiment.
Our ambition is to turn deeper molecular interaction understanding and iterative experimental learning into a faster path to new high-quality drug candidates.