Biocomputation, AI and Q Capital.
We audited our own pipeline: the thermodynamic channel adds no value in missense (AUC +0.000 over the evolutionary signal, which already reaches ~0.90), but it classifies synonymous variants (AUC 0.63–0.71), where the protein language model is blind. Correcting your own record as the product.
read →The synonymous CNN reduces to two interpretable physical observables — the signed stacking ΔΔG (Turner) + the G>A bias — that match it (AUC 0.68) and generalize across genes. Glass-box, deployable as pip install ef-synonymous + a free web tool.
read →Formal comparison across 7 genes with a stratified meta-analysis. EF reaches ΔAUC of −0.024 to −0.040 against massive predictors — it sits below all three overall, and beats CADD in the sodium channel SCN1A (0.913 vs 0.832). Whether the signal is orthogonal and additive is a hypothesis yet to be tested.
read →A multichannel 1D CNN that combines the mRNA stacking free energy with evolutionary signal from ESMC-300M. 8 genes in the preprint, intra-gene AUC up to 0.99; 42-gene panel with results unpublished. Zero-shot transferability across protein families.
read →The stacking ΔG profile adds a dimension of its own next to secondary structure (78 % of its variance is neither composition nor base pairing). A dual optimizer: a computational proof of concept reaching more favourable ΔG and MFE values than BNT162b2 and mRNA-1273, without experimental validation.
read →Synonymous variants don't change the protein but do change the mRNA's thermodynamic profile. A CNN predictor with pure thermodynamic signal: AUC 0.683 ± 0.018, not exceeding SpliceAI. Conceptual framework Λ = S ⊗ Φ.
read →A model of the firm where the unit of capital is neither the work hour nor the physical asset, but the token: the inference compute that, governed and captured, accrues as proprietary, defensible knowledge.
read →A Kanban board where several agents collaborate under human supervision: fail-closed approval broker, hash-chain audit, RBAC and a multi-LLM adapter. Born as the testbed for the lab's pipelines.
read →An adaptation of the Business Model Canvas into nine blocks to force the expensive decisions — segment, autonomy, KPIs, integrations, costs, risks — before writing a single line of code.
read →