What the platform computes, and how it says so.

On demand

Boltz-2: co-fold a drug with a target, and read the pose back.

Boltz-2 is an open, MIT-licensed structure model that folds a protein and a ligand together and predicts their binding affinity. Curasynth runs it as a service beside the data. From a drug or target monograph, Predict binding names the partner, returns a cost estimate, and runs on confirmation under a dollar ceiling you set; the result lands as a new row against the pair.

What comes back is more than a number. The predicted potency and binder probability sit beside the measured value where one exists; the pose is read for the model's own doubt — interface predicted aligned error, per-residue confidence — and for what a chemist would check: contacts, buried area, clashes, and whether the ligand sits in the known pocket. The complex renders in the structure viewer, colored by confidence and labeled predicted.

The same service screens a drug's structural neighbors at its primary target: each Morgan-fingerprint analog scored for predicted potency with its similarity to the reference and the reference's own measured potency beside it.

16 pairs co-folded to date · 6 analog screens · measured K_i shown beside every prediction

• Ranking

  A composite that asks about this disease

  Ranking targets by how strongly they are implicated in a disease answers only half the question, and outside cancer it answers almost none of it. The composite folds in eight parts: the association evidence under your own weights, tractability, chemical matter (measured occupancy where it exists, a predicted binder probability discounted where it does not), dependency read in the disease's own cell lines, tumour expression, clinical precedent — the phase a drug on this target reached for this disease — a mechanistic route from target to disease, and collateral lethality. The three parts that only mean something in cancer weigh zero elsewhere and render as a dash rather than a zero, and the rest scale up. Every part, its weight and its reason ride on the row.

  Type 2 diabetes returns PPARG, GLP1R, ACE, DPP4, SLC5A2 — how the disease is actually treated
• Pharmacology

  Receptor occupancy and target engagement

  For every drug with binding data, its targets are solved together for occupancy at a free-ligand exposure — the coupled answer beside the independent L/(L+K_d) for each — across a ladder from 1 nM to 1 µM. The binding constant is the geometric median over uncensored K_i/K_d measurements, with functional IC50/EC50 used only where no binding constant exists and marked as such; the measured range becomes the occupancy interval. A profile can be compared to a reference drug's by RMS distance, and scored for liability burden over the standard in-vitro safety-pharmacology panel with the primary target excluded.

  261,772 occupancy rows · 30,835 liability screens · 43-target panel (Bowes et al., 2012)
• Safety

  Pharmacovigilance with shrinkage

  Twelve quarters of FDA adverse-event reports, deduplicated to 4.2 million cases, are counted into one 2×2 table per drug and MedDRA term, with strata, onset days and the serious share. Each cell is scored for disproportionality: proportional reporting ratio, reporting odds ratio, and the empirical-Bayes geometric mean with its interval, plus a sex-by-age adjusted odds ratio and the onset pattern. A signal is at least three co-reports with an EBGM lower bound of two or a PRR lower limit above one — a hypothesis for review, never an incidence, because spontaneous reports have no denominator.

  1,913,561 2×2 tables · 864,709 scored signals · 2023 Q3 – 2026 Q2
• Clinical

  Arm-level and pooled trial effects

  Where a registered study has posted results, its arms are compared from the reported event and participant counts to give an effect with its interval; trials of the same comparison in the same disease are pooled. The trials landscape counts how many studies carry a computed arm effect so absence reads as absence.

  230,987 arm effects · 19,236 pooled comparisons
• Genetics

  Dependency, copy number and collateral lethality

  CRISPR dependency from DepMap is rolled up per target with the pattern a decision needs — selectively essential or broadly dispensable — and measured in three-dimensional organoid screens beside the two-dimensional ones. Co-dependency ranks each gene's top partners; co-deletion and deletion frequency come from TCGA PanCancer copy number; together they yield collateral-lethality candidates — a passenger deleted alongside a tumor suppressor whose paralog the tumor now depends on — with the mechanism recorded. Published synthetic-lethal pairs are served beside them.

  874,159 co-deletions · 447,650 co-dependencies · 39,275 collateral-lethality candidates · 250,480 synthetic-lethal rows
• Binding

  Consensus potency over censored measurements

  Measurements from ChEMBL, BindingDB and the PDSP K_i database are collapsed by one rule — duplicates across sources removed first, so every median is over distinct measurements — into a consensus per drug–target pair: the median in p-affinity space, the best value in nanomolar terms, the count, the sources, whether any value was censored, and which source won precedence. A parent-level rollup joins the salts.

  461,919 measurements → 145,727 pairs · 149,778 by parent
• Chemistry

  Standardization, similarity and properties

  Every ingredient is standardized before it is given an identity — an all-salt input is refused rather than registered under its counter-ion — which is what makes a drug's identifier an assertion about a molecule. Morgan fingerprints give each single-ingredient drug its twenty nearest neighbors with the targets they share measured; descriptors, a CNS multi-parameter score and a blood–brain-barrier screen are computed for every drug.

  11,758 drugs with properties · 71,610 neighbor rows
• Structure

  Served structures and their annotation

  For each target a render decision is made once: an experimental co-crystal from the PDB where one exists with a ligand of interest, otherwise the predicted structure. Domains, enzyme class and structural family are annotated per entry; each solved structure links to its primary citation. Drugs render from a stored 3-D conformer.

  79,269 target structures · 127,476 domain rows · 116,960 structure–paper links
• Literature

  Citation intent and stance

  Every reference in every held paper is resolved to a work; each citing sentence is classified for intent and stance, so the lineage of a finding can be read as confirmation, use, or dispute rather than a count. Author and organization identity is resolved through ORCID, OpenAlex and ROR, with the resolution method stated on the page.

  301,416,662 references resolved · 208,113,080 intent markers · 9,256,585 authors with metrics

Next, with technical partners

The methods we would rather call than build.

The registry that resolves a method to a backend was built so that a partner's model can be called from the same cited workflow with the same provenance rules. The next slots are pocket-aware docking where a co-crystal exists and blind docking where it does not; ADMET endpoints fine-tuned on labeled data; and a calibrated target-validity model to replace hand-set constants. If you build one of these, we should talk.

Continue