Programs
Discovery and data
Computational discovery and data infrastructure: structure-based small-molecule design, privacy-preserving synthetic multi-omics data and rare-disease diagnosis from variants and phenotypes.
Projects
Structure-Based AI Small-Molecule Discovery
Link public protein structures and small-molecule activity records to assess how well computational scores transfer to unseen molecules.
Early Boost for Core Health R&D Technology
Measure how much statistical structure and predictive information synthetic multiomics and clinical data retain.
Rare Disease Diagnostic Research
Combine public variant evidence with a phenotype ontology and evaluate the ranks and supporting evidence for candidate genes.
Target, evidence, next check — kept apart
A target is what the project set out to reach. Evidence is what has been measured, in which scope. The next check is what would change the reading.
| Project | Internal target | Evidence so far | Next check |
|---|---|---|---|
| 5-1 · AI drug discovery | Binding free-energy error of 1 kcal/mol or less | Common synthetic checks | Select records with measured binding values, specify assay conditions and lock the split by scaffold, target family and date. |
| 8-1 · Synthetic data | At least 95% of downstream predictive performance retained | Synthetic harness (synth8) · Common synthetic checks | Run the same harness on the fixed TCGA-BRCA institution split with the baseline generator, then train the main generator and require it to beat the baseline on every endpoint at the same privacy budget. |
| 19-1 · Rare disease diagnosis | At least 90% of causal genes in the top five | Published-case evaluation · Public-record audit (acmg19) · Synthetic replay (rare19) · Common synthetic checks | Read the literature for every expert-pathogenic missense variant instead of a sample, so that functional, segregation and case findings enter the ledger. Then re-measure sensitivity on genes outside the current 40, with the condition that specificity does not fall. |
Evaluation data
Official D3R Grand Challenge 4 Free Energy Sets
Direct thermodynamic measurements with independent scaffold groups
TCGA-BRCA PanCancer Atlas RNA, copy number and progression data
Monarch Phenopacket Store with the December 2022 Human Phenotype Ontology
NCBI ClinVar GRCh38 VCF archive
Paired patient-level VCF and HPO with adjudicated labels
Boundary cases recorded before evaluation
- Evaluation leakage across molecular scaffolds and homologous target families
- Inconsistent assay conditions and units across binding measurements
- The research specification also sets limits of 0.4 structural similarity, 2.0 Å RMSD over 100 ns and five retrosynthetic steps
- Original rows or repeated records for the same person appearing across splits
- Differential privacy budget and proximity of synthetic rows to source records
- An internal target of zero biologically impossible variable combinations
- Gene ranks in synthetic cases containing known variants
- Ranking and evidence changes after missing or miscoded phenotypes
- Reference build mismatches, conflicting classifications and missing literature references
What the evidence covers
The computational benchmark covers scores and chemical metrics. Patent novelty, synthesis, toxicity and therapeutic effects each require their own evidence.
TCGA, GEO and MIMIC-IV are separate sources. Validation follows each source's endpoint, access terms and verified linkage.
Evaluation can overstate performance when inputs already include the variant annotation or gene–disease association used as the reference answer.