Erudio Bio — Technical, Business & Strategic Roadmap

Centered on bioTCAD · authored from the CTO / AI vantage

Prepared by Sunghee Yun (Co-Founder & CTO) · draft for internal alignment

Framing — one company, two engines, one thesis

Erudio Bio is a single company with two revenue engines and one underlying thesis.

• VSA (Versatile Smart Assay) — a dynamic-force-spectroscopy (DFS) measurement platform. Today it is a diagnostics product (multi-cancer biomarker screening); structurally it is a per-molecule, kinetics-resolved measurement instrument.
• bioTCAD — measurement-anchored physics + AI simulation for molecular design. Today a validated proof-of-concept; structurally the design engine that consumes measurement and produces optimized molecules.

The thesis that unifies them: biology’s bottleneck is not compute or model architecture — it is the absence of high-quality, kinetics-resolved, per-molecule measurement to anchor prediction. Generic force fields and public-data-trained ML both drift from reality on the specific molecule in front of you. Erudio closes that gap by generating the measurement (VSA) and using it to calibrate the simulation (bioTCAD) — a loop no competitor has vertically integrated.

Strategic posture: diagnostics pays the bills; bioTCAD is the prize. VSA revenue (Korea, near-term, non-dilutive-friendly) funds and de-risks the runway while bioTCAD matures into the high-multiple asset.

Mandate & the value we deliver

Overarching mandate: Make biology measurable and predictable — supply the measurement-anchored ground truth that converts AI from a correlational guesser into a reliable engine for molecular design and diagnosis. In one line: the measurement layer for AI-era biology.

Value, by customer:

Customer	What they need	What Erudio delivers
Government — DoW / biodefense	Rapid, onshore, auditable medical-countermeasure design against emerging threats	Measurement-anchored, physics-grounded candidate design + validation that black-box generative models can’t audit; first-in-kind binding data; a domestic capability
Government — FDA	Credible, calibrated in-silico evidence to safely compress pre-IND timelines	bioTCAD as a qualifiable NAM delivering confidence-bounded predictions of the molecular determinants (binding, selectivity, MABEL inputs) that govern the first human dose
Pharmaceutical companies	Better lead candidates, faster, with quantified confidence	bioTCAD as a drug-discovery platform that compresses and de-risks the lead-to-optimization phase — higher-quality candidates, fewer synthesize-and-assay cycles
Hospitals / screening centers	Affordable, accurate, convenient early cancer screening	VSA multi-cancer panel at ~⅓ the cost of separate assays, lower false-positive rate (cross-reactivity filtering), single 10 µL sample

The through-line: to each customer we sell the same core capability — trustworthy per-molecule measurement, and prediction anchored to it — packaged as a diagnostic (hospitals), a design platform (pharma), regulatory-grade evidence (FDA), or a national capability (DoW).

How we deliver it — the bioTCAD + VSA platform

VSA — the measurement engine

Dynamic force spectroscopy measures single-molecule bond rupture, yielding unbinding force, on/off kinetics, and specificity from 50–100 interactions per 10 µL sample, with cross-reactivity filtering. Two faces:

• As product: the diagnostics revenue engine (Section 3.4).
• As infrastructure: a source of kinetics-resolved, ground-truth binding labels — exactly the data public databases lack and AI models starve for.

bioTCAD — the design engine

The Phase I seven-stage pipeline (structure prediction → MD relaxation → CD structural measurement → force-field customization → complex simulation → contact-map analysis → measurement-anchored binding-energy extraction), validated on the Vanderbilt cyclic-peptide set: ≥88% bind/no-bind classification vs 37.5% unmodified MD and 25% for the original method. The core proven capability: measurement corrects the force field for the specific molecule, turning generic MD into a molecule-specific predictor.

The integrated loop — the moat

   generative / AI proposal
            │
            ▼
      bioTCAD simulation  ──►  candidate ranking  ──►  synthesize top-k
            ▲                                                │
            │                                                ▼
   force-field calibration  ◄──  VSA / CD / SPR measurement (ground truth)
            │
            ▼
   first-in-kind labeled dataset  ──►  trains every downstream AI model

VSA’s DFS output feeds bioTCAD’s affinity anchoring (via a Bell–Evans interpretation) directly — so the measurement that calibrates the simulation is made in-house, not bought from third parties. This is the vertical integration competitors cannot replicate, and (see §5.3) it is also the fix for the Phase I affinity-anchor reliability problem.

Where the AI lives (my domain)

bioTCAD is already a multi-layer AI system, not a simulator with AI bolted on. Mapped to the five-zone landscape:

• Zone 1 — AI already latent: DL structure prediction (AlphaFold3 / RoseTTAFold3), ML force-field potentials (ANI/NequIP/MACE) inside the calibration loop, simulation-based inference for Bayesian calibration, the contact-map classifier, the labeled-data factory, and the ranking engine.
• Zone 2 — AI feeding in: generative proposal (RFdiffusion / ProteinMPNN), inverse-problem net for CD interpretation, active-learning oracle, GNN force-field selector.
• Zone 3 — AI derived out: per-molecule foundation model, affinity surrogate, parameter recommender, thermodynamic-decomposition model.

The strategic asymmetry: every calibration run produces proprietary, measurement-anchored training data (Zone 1 data factory → Zone 3 models). The engine and the data asset compound each other. That is the durable advantage, and it is fundamentally an AI advantage.

Customer-sector strategies

Pharma — bioTCAD as a lead-optimization platform

Positioning: a drug-discovery platform that compresses and de-risks the lead-to-optimization phase — reducing the time to a high-quality development candidate and improving the quality (affinity, selectivity, developability) of the molecules that advance.

Why now: AI-first discovery (Recursion, Owkin, Tempus) and pharma R&D both hit the same wall — models trained on noisy, heterogeneous public data can’t reliably rank true binders vs look-alikes, and can’t extrapolate to novel chemotypes (the OOD problem, §5.1). bioTCAD’s measurement-anchored physics is the missing high-fidelity filter.

Engagement ladder (low → high commitment):

1. Data licensing — proprietary measurement-anchored binding datasets ($200–500K / project).
2. SaaS — bioTCAD cloud analysis / candidate ranking ($50–200K / year).
3. Co-development — milestone-based partnerships on specific targets/series.

TAM anchor: AI drug discovery ~$5B (2024), growing 25–35%/yr; the measurement-anchored training-data layer beneath it is a distinct, defensible slice.

FDA — credibility and standard-setting

Play: convert the RFI comment letter into a durable regulatory position — get measurement-anchored, model-based candidate analysis recognized as a qualifiable NAM under existing V&V-40 / context-of-use / fit-for-purpose machinery, and use pre-IND engagement as the venue.

My contribution: the uncertainty-quantification argument is the regulatory linchpin — context-of-use, model-risk, and fit-for-purpose validation all depend on the model knowing when it is reliable, which is exactly what measurement-anchored (via SBI, §5.4) confidence bounds provide and pure ML lacks OOD. This is what makes bioTCAD admissible, not just accurate.

Prize: if Erudio helps define the data-quality and calibration standards, competitors must meet Erudio-shaped standards — a regulatory moat.

DoW / biodefense — non-dilutive capability

Play: position bioTCAD + VSA as a domestic, auditable medical-countermeasure design-and-validation capability, adjacent to LLNL’s GUIDE (hybrid physics + ML for biodefense) and the JPEO-CBRND / BARDA / DTRA / ARPA-H funding vectors.

Differentiators for this audience:

• Auditability — physics-grounded, measurement-anchored predictions are inspectable in a way black-box generative models are not (a dual-use / trust advantage).
• Data-provider angle — GUIDE openly states the binding data it needs doesn’t exist; VSA generates exactly that. Erudio can be a data/validation partner, not only a bidder.
• Onshore — aligns with the national-competitiveness rationale driving current policy.

Funding character: non-dilutive, milestone-based — extends runway without dilution while diagnostics revenue scales.

Hospitals / screening centers — the VSA diagnostics engine (Korea-first)

Model (B2B2C): hospitals and health-screening centers adopt the VSA benchtop reader and run the validated 6-marker multi-cancer panel (AFP, CA19-9, CA125, CEA, FT4, TSH) on patient serum.

Revenue (three streams):

1. Hardware — VSA reader systems (70%+ gross margin).
2. Consumables — microfluidic chips + reagent kits (recurring).
3. Per-sample fee — ₩150,000–300,000 / test.

Value to the hospital: ~⅓ the cost of six separate assays, fewer false positives (cross-reactivity filtering), single-sample convenience — into a Korean market of 15M+ annual health screenings.

Anchor partners: SNU Bundang Hospital (IRB in progress; 15M+ patient-data access on approval), Keimyung Dongsan (multi-site validation), Shanghai General (JDA). Regulatory via MFDS with C&R Research (device 1st-grade notification + separate cartridge approval).

Strategic role: near-term cash flow + a real-world data stream that (longer-term) feeds the same measurement-anchored data asset.

Roadmap & milestones

Quarterly through 2027, then annual. Four workstreams: bioTCAD/AI, VSA diagnostics, Regulatory & Government, Corporate/Funding. (Dates are targets; ✦ marks the AI-differentiated deliverables.)

Q3 2026 (current)

• bioTCAD/AI: Close Phase I; scope Phase II. Begin CMAP fitting for D1V1 (the peptide LJ alone couldn’t solve). Stand up the calibration-library data pipeline that every Zone-3 model depends on. ✦ Prototype the inverse-CD net to remove the manual CD↔DSSP comparison.
• VSA: MFDS 품목질의 → 품목허가 preparation with C&R Research; advance SNU Bundang IRB.
• Regulatory/Gov: File the RFI comment letter (docket window). Open GUIDE / JPEO-CBRND / BARDA and Fisher/OPHPR pilot conversations.
• Corporate: Run the lead-investor / Gates $5M-follow-on process (A2G and others).

Q4 2026

• bioTCAD/AI: ✦ Solve D1V1 via residue-restricted CMAP → two validated per-molecule calibrations, not one. First prospective design attempt (design a novel candidate, not just classify existing ones). ✦ Affinity-surrogate v0 and parameter-recommender data schema.
• VSA: Submit MFDS filing.
• Regulatory/Gov: Pre-IND / NAM-qualification pathway engagement; biodefense proposal(s) in progress.
• Corporate: Close lead investor / Series A first tranche; trigger Gates follow-on.

Q1 2027

• bioTCAD/AI: ✦ Parameter-Recommender Network v1 — learn the sequence→calibration-parameter prior so new molecules initialize from learned priors instead of generic defaults (the core generalization lever, §5.1). Extend validated calibration to ~5–10 peptide systems. ✦ Integrate VSA-DFS as the affinity anchor, replacing third-party SPR (fixes the Phase I anchor-reliability gap, §5.3).
• Pharma: First paid data-licensing pilot.
• VSA: MFDS review; first hospital pilot site prep.

Q2 2027

• bioTCAD/AI: ✦ Active-learning oracle in the loop (Bayesian selection of the next molecule to measure) → fewer wet-lab cycles per optimized candidate. Scope small-molecule bioTCAD (ADMET/solubility anchoring). SaaS alpha.
• Pharma: Second pilot; SaaS early-access.
• VSA: MFDS approval (target); first commercial screening-center deployment.

Q3 2027

• bioTCAD/AI: ✦ Antibody/CDR bioTCAD (SPR/BLI + structural anchoring of CDR loops). ✦ Affinity surrogate enabling million-scale virtual screens; begin fine-tuning a per-molecule foundation model on the accumulated calibration corpus.
• VSA: Korean commercialization ramp; expand panel menu.
• Corporate: Series B narrative + metrics assembled.

Q4 2027

• bioTCAD/AI: ✦ Closed-loop autonomous prototype — LLM agent orchestrating bioTCAD + synthesis procurement + VSA measurement (propose → simulate → synthesize if threshold cleared → measure → recalibrate). ✦ Selectivity-panel bioTCAD as a regulatory-grade off-target safety product.
• Regulatory/Gov: First NAM-qualification milestone / formal FDA interaction on context-of-use.

2028

• bioTCAD/AI: General platform across ≥2 modalities (peptide + small-molecule and/or antibody). SaaS GA; multiple pharma partnerships. ✦ Per-molecule foundation model + universal calibration prior (compounding ROI on every past run). ✦ Regulatory AI co-pilot (bioTCAD output → submission language) as a standalone offering.
• VSA: Japan / SE-Asia expansion; China via Shanghai General.
• Corporate: Series B.

2029

• bioTCAD/AI: Autonomous design foundry productized; causal/counterfactual layer for mechanistic design; a named biodefense countermeasure program in execution.
• VSA: Broadened diagnostic menu (companion Dx, infectious disease); regional scale.

2030

• Measurement-anchored foundation models positioned as an industry reference; bioTCAD established as the calibrated-evidence layer beneath AI drug discovery; diagnostics a durable cash engine funding the platform.

bioTCAD technical roadmap — the AI deep-dive

This is where the platform is won or lost, and where my expertise is decisive. Phase I proved the mechanism on one molecule (D4V2). The job now is to turn a bespoke, manual, single-molecule result into an automated, general, self-improving platform. Five problems, in priority order.

From D4V2-specific to general — the generalization problem

Where we are: the customized force field is anchored to one peptide. D4V2 succeeded; D1V1 required a different strategy (CMAP). Each new molecule currently implies a bespoke calibration, and some fail. This is, in ML terms, the same out-of-distribution issue at the heart of my FDA argument — a calibration fit to D4V2 is by construction anchored to D4V2, and the open question is whether the approach transfers.

The path (learned priors, not per-molecule from scratch):

1. Parameter Recommender Network — train a sequence/topology → calibration-parameter model on the growing calibration library (LJ ε-scaling, the a/b correction coefficients). New molecules initialize from a learned prior instead of generic defaults. This is the direct analog of TCAD process-parameter libraries transferred across device geometries — Kee’s world, formalized in ML.
2. Universal Calibration Prior — the asymptote of (1): a model that maps molecular topology → measurement-anchored parameters, so calibration converges with fewer experiments as the library grows. Every past run appreciates in value — a data asset with compounding ROID.
3. GNN force-field selector — learn CHARMM-vs-AMBER (and beyond) per scaffold, removing a human judgment call and an iteration cycle.
4. Cross-molecule transfer learning — treat each calibration as a point in parameter space; learn the manifold; interpolate/extrapolate to novel scaffolds with quantified confidence.

Success metric: calibration cost per new molecule (wet-lab cycles + compute + human hours) falling toward zero as the library grows; validated accuracy holding across a held-out set of molecules never calibrated before (the real OOD test).

Automating the manual pipeline

Where we are: Phase I was heavily manual — hand-iterated LJ fitting, manual CD↔DSSP comparison, human force-field selection, manual measurement-method choice, bespoke analysis scripts. This is the single biggest throughput ceiling.

The automation stack (each replaces a manual step):

• Force-field fitting → Bayesian optimization / active learning. Replace hand-tuning of LJ (and CMAP) with an optimizer that proposes the next parameter set to try, and — via the active-learning oracle — the next molecule to measure. Turns a manual art into a self-directing search.
• CD interpretation → inverse-problem neural net. Map raw CD spectra → secondary-structure fractions with calibrated uncertainty, eliminating the manual comparison and feeding the loop faster.
• Force-field selection → GNN selector (as §5.1).
• Screening → affinity surrogate. A model predicting Kd from contact maps in milliseconds (vs hours of MD) lets us screen libraries orders of magnitude larger than full simulation allows — full MD reserved for the shortlist.
• Orchestration → agentic pipeline. An LLM agent wires the stages together (the Zone-5 autonomous foundry), moving from “a scientist runs each stage” to “the system runs the loop, a scientist supervises.”

Success metric: human-hours per calibrated molecule; end-to-end wall-clock from candidate proposal to calibrated prediction.

Fixing the affinity anchor — and turning a weakness into the moat

Where we are: Phase I’s affinity side is the soft spot — third-party SPR disagreed with published BLI by 2–3 orders of magnitude, and the MM/GBSA scaling (Eq. 1) was fit from two points with zero validation. Ironically, this is the very SPR-unreliability we argue against in the FDA letter.

The path:

1. Replace the anchor with VSA-DFS. Bring affinity measurement in-house via VSA’s single-molecule force spectroscopy (Bell–Evans), removing dependence on the third-party platform we critique and giving a kinetics-resolved anchor richer than an equilibrium Kd. This closes the loop and resolves the letter-vs-report tension in one move.
2. Validate the scaling model properly. Grow the (calibrated ΔH/ΔS, measured Kd) set far beyond two points; replace the arbitrary single-condition anchor with a principled multi-measurement fit; consider simulation-based inference so the correction carries calibrated uncertainty instead of a point estimate.
3. Thermodynamic-decomposition model. Learn enthalpy- vs entropy-dominated signatures across the library → predict which correction strategy a new scaffold needs (encoding med-chem intuition).

Success metric: anchor reproducibility (in-house VSA vs external); scaling-model accuracy on held-out molecules; uncertainty calibration.

Uncertainty quantification — the regulatory and scientific backbone

Why it’s load-bearing: UQ is simultaneously (a) the FDA admissibility hook (context-of-use / model-risk / fit-for-purpose all assume the model knows when it’s reliable), (b) the fuel for active learning (measure where uncertainty is highest), and (c) honest science (know when we’re extrapolating).

The path: promote bioTCAD from point-estimate calibration to a Bayesian framework via simulation-based inference (already named in the provisional). Distinguish epistemic vs aleatoric uncertainty; produce confidence bounds that widen as molecules move away from measured anchors. Physics-grounded bounds, not statistical resemblance to a training set — the property pure ML lacks OOD.

Success metric: calibration curves (predicted confidence vs realized accuracy); demonstrable widening of bounds on OOD inputs.

The arc — classification → design → autonomy

The maturity ladder that ties the above together:

1. Retrospective classification (Phase I, done) — classify 8 known peptides at ≥88%.
2. Prospective design (2026–27) — design new candidates and prospectively validate them (the first true test of the platform).
3. General platform (2027–28) — learned priors + automation make calibration cheap across molecule classes; SaaS + data licensing productize it.
4. Autonomous foundry (2028–29) — the closed-loop agent (bioTCAD + synthesis + VSA) runs discovery with human supervision; the per-molecule foundation model and universal prior make each cycle cheaper than the last.

Modality scope expands along the ladder: peptides → small molecules (ADMET anchoring) → antibodies/nanobodies (CDR loops) → nucleic acids, plus the selectivity-panel and proteolytic-stability extensions — each an addressable market covered by the existing provisional claims.

Key risks & open questions

• Generalization risk (the central technical bet). Does per-molecule anchoring transfer via learned priors, or does every molecule need bespoke work? §5.1 is the make-or-break; the held-out-molecule test is the honest arbiter.
• Affinity-anchor validity. The VSA-DFS anchor (§5.3) must be demonstrably more reproducible than the SPR it replaces, or the scaling model stays fragile.
• Data-flywheel throughput. The Zone-3 models need volume; if calibration stays slow/manual (§5.2), the data asset accumulates too slowly to compound.
• Regulatory timing. NAM qualification and MFDS approval are external-dependency milestones; treat their dates as targets, not commitments.
• Focus vs breadth. The five-zone landscape is deliberately expansive; execution requires sequencing (this roadmap’s ordering) rather than pursuing all vectors at once.

The through-line

Erudio’s defensibility is not the simulator and not any single model — it is the measurement-anchored data loop: VSA generates ground truth no one else has, bioTCAD turns it into calibrated prediction, and every run enriches a proprietary corpus that makes the next prediction better and cheaper. Diagnostics funds the journey; the FDA and DoW work builds credibility and non-dilutive runway; pharma is where the platform monetizes at scale.

That loop is fundamentally an AI and software asset — which is precisely where my mandate as CTO sits, and where the next two years of value creation will be decided.