From possibilityto hypothesis,to design,to validation.
AI-Powered Drug Discovery
With AI and molecular simulation
- Candidate design
- Prioritization
- Validation planning
Better decisions before the next experiment.
NEO Drug Discovery connects molecular design, evidence review and experimental planning in one computational discovery platform.
From protein degraders to RNA-directed approaches and small molecules, we help research teams decide what to test—and understand why.
Scope of our work
- Current capabilities
- Protein degraders · RNA/ASO · Small moleculesComputational design and candidate prioritization
- Collaboration scope
- ADC payloads & linkers · PeptidesComputational comparisons that rely on a partner's assets and laboratories
- Exploratory
- Neoantigen prioritization · Quantum computingMethods under review; no results are claimed
Platform§ 01From a hypothesis to a testable plan
Explore broadly.
Test what matters first.
We bring structure, molecular physics, AI and developability into the same decision process. The output is a prioritized hypothesis with a clear rationale and a plan for testing it—not a claim that the molecule already works.
- 01 / 04
Frame the question
Start with biological evidence, available structures or RNA context. Define the desired mechanism and what an experiment would have to show.
- 02 / 04
Design & compare
Compare suitable molecular approaches using structural analysis, simulation and property triage. Rank candidates against the question, not a score in isolation.
- 03 / 04
Challenge the ranking
Check reference cases, applicability and practical constraints. Preserve uncertainty, negative findings and the reasons a promising design may fail.
- 04 / 04
Prepare the handoff
Connect candidate rationale to proposed material specifications, readouts and decision criteria. Experimental results can then guide the next design cycle.
INSIDE THE PLATFORM
Different methods. A traceable decision.
Our technical architecture brings together 13 computational and experimental-interface modules. These are grouped below by scientific purpose; the methods needed are selected for each project.
How we decide
Rules we apply in projects—not reported results.
- When docking and simulation disagree
- We do not pick the more favorable score. We first check whether each method reproduces the reference ligand and whether the pose holds in molecular dynamics; if they still disagree, the conflict is recorded as uncertainty and reflected in the ranking.
- When a candidate is excluded
- Candidates outside a method's applicability domain, with high synthesis or property risk, or supported only by a model that failed on reference cases are removed from the ranking—and the reason is recorded.
- When confidence is low
- We propose the simplest measurement that could change the decision most. If the binding hypothesis is the uncertain part, for example, binding is confirmed by SPR or ITC before cellular degradation and mechanism experiments.
Structure & confidenceKnow what the starting model can support.
Use experimental or predicted structures with their provenance and confidence. AlphaFold-derived structures and complementary prediction approaches inform pocket assessment; agreement is not a substitute for structural validation.
- AlphaFold / UniProt
- RoseTTAFold2
- Pocket analysis
Molecular physics & dynamicsExamine conformations, energetics and model limits.
Molecular dynamics, quantum-chemical parameterization, enhanced sampling, pathway and committor analyses address different questions. Each requires suitable controls, convergence assessment and a defined applicability range.
- ORCA / Psi4
- GROMACS / PLUMED
- PathCV / committor analysis
- OpenFE / RBFE
Pose generation & comparisonCompare proposals rather than trust one score.
Geometry- and diffusion-based pose proposals can be compared with docking and energy-based rescoring. These are complementary computational stages—not independent experimental confirmations of binding.
- EquiBind
- DiffDock
- AutoDock Vina
- MM-GBSA
AI prediction & molecule proposalsBalance binding hypotheses with developability.
Graph-based affinity and ADMET models support prioritization alongside de novo proposal methods, including reinforcement learning. Predictions are interpreted against relevant controls, domain limits and downstream experimental requirements.
- GNN
- ADMET
- De novo design / RL
Specialist electronic-state calculationsUse advanced methods where the question warrants them.
TD-DFT and EOM-CCSD are specialist tools for suitable electronic and photophysical questions. They are not a universal final check of drug affinity, efficacy or safety.
- TD-DFT
- EOM-CCSD
Experimental interface & feedbackSpecify what must be measured next.
Plan binding or thermodynamic measurements such as SPR and ITC with experimental partners, then link suitable cellular readouts to the original hypothesis. Feeding results back into the next design cycle is the intended workflow—not a claim of completed experimental validation.
- SPR / ITC planning
- Cellular assay planning
- Design feedback
Platform briefing
What our AI platform does—and what it does not
A walkthrough of our computational research workflow, candidate prioritization and the boundary between predictions and experimental evidence.
02 PLATFORM CAPABILITIES
One research question.
More than one way forward.
A target does not arrive with a preferred modality. We compare biological rationale, molecular feasibility and the evidence needed to choose the next approach.
Computational design
Design for proximity. Test for degradation.
Connect target recognition, linker design and E3-ligase recruitment in one design workflow. Structural plausibility helps prioritize a candidate; it does not establish protein degradation.
What we evaluate
- Target-binding features and linker attachment geometry
- Linker conformations and ternary-complex plausibility
- Physicochemical properties, permeability risks and design constraints
- Computational output
- Prioritized design hypotheses with structural rationale, known limitations and a proposed synthesis and assay brief.
- What experiments must establish
- Cellular degradation, concentration response, selectivity and mechanism-dependent activity.
Computational design
Start with RNA evidence. Match the mechanism.
Evaluate RNA and splicing hypotheses before choosing an antisense design. Splice modulation and RNA knockdown require different mechanisms, chemistry and experimental readouts.
What we evaluate
- RNA evidence, transcript context and strand-concordance checks
- Accessibility and transcriptome/genome off-target screening
- Mechanism–chemistry matching and suitable experimental controls
- Computational output
- Prioritized sequence-design hypotheses, a quality-control record and an experiment brief. Sequences and undisclosed targets remain confidential.
- What experiments must establish
- The intended RNA or isoform change, off-target effects and functional rescue—not just sequence compatibility.
Computational prioritization
Use more than a docking score.
Combine structural analysis, pose generation, molecular simulation and property prediction to compare candidates within an explicitly defined evaluation domain.
What we evaluate
- Pocket assessment, reference ligands and redocking controls
- Docking, molecular dynamics and calibrated free-energy comparisons
- ADMET flags, synthetic accessibility and candidate-generation hypotheses
- Computational output
- A ranked candidate set with method provenance, reference comparisons and uncertainty—not a claim of absolute binding affinity.
- What experiments must establish
- Binding, biochemical activity, cellular effects and experimental ADME or safety measurements.
Proposed collaboration
Complement biologics with small-molecule insight.
Explore payload and linker prioritization for antibody–drug conjugates, and property triage for peptides or macrocycles. Antibody assets and experimental development remain with the partner.
What we evaluate
- Payload binding hypotheses and physicochemical risk flags
- Linker properties, attachment feasibility and stability hypotheses
- Permeability and efflux triage relevant to payloads and larger molecules
- Computational output
- A bounded computational comparison for a partner-defined set of payloads, linkers or peptides. Scope and transferability are agreed before a pilot.
- What experiments must establish
- Conjugation, drug-to-antibody ratio, aggregation, payload release, cellular activity and any bystander effect. We do not offer de novo antibody design.
Exploratory prioritization
Prioritize immune hypotheses. Do not assume response.
Use class I peptide–MHC prediction to organize neoantigen candidates for follow-up. A binding or presentation prediction is only one part of an immune-response hypothesis.
What we evaluate
- Peptide–MHC class I compatibility and candidate ranking
- Reference comparisons and model applicability
- Evidence requirements for subsequent experimental selection
- Computational output
- A candidate-ranking rationale and a clear statement of what the prediction does—and does not—support.
- What experiments must establish
- Actual presentation, T-cell recognition and functional immune activity. Presentation is not proof of immunogenicity.
Research focus§ 03Our current public program
PRMT5-targeted protein degraders
We design molecules
to induce PRMT5 degradation.
Our current public program explores targeted degradation of PRMT5, an enzyme involved in protein arginine methylation and an active target in cancer research.
MTAP-deleted cancer biology provides one context for studying PRMT5 dependency. That rationale does not, by itself, establish degrader selectivity or a therapeutic window.
Current stage
Design-stage research
We are developing structure-based design hypotheses and reviewing developability.
Next validation goal
Questions to answer with experimental partners.
Cellular degradation
Does PRMT5 protein actually decrease in cells?
Concentration response
Does the extent of degradation depend on concentration?
Mechanism dependence
Does the observed reduction in PRMT5 depend on the E3 ligase and the proteasome?
Selectivity
Is PRMT5 reduced selectively relative to other proteins?
How targeted protein degradation works
General principle · four-step diagram
- 01
Recognition
The degrader recognizes and binds the target protein to be degraded.
- 02
Proximity
The degrader holds the target protein close to an E3 ubiquitin ligase.
- 03
Tagging
Ubiquitin is attached to the nearby target in a chain that serves as a degradation signal.
- 04
Degradation
The tagged protein is broken down by the cell's proteasome, and the degrader can generally go on to act on another target molecule.
04 BUILDABILITY & EVIDENCE
A plausible design
is only the beginning.
Before recommending the next experiment, we ask whether a hypothesis is biologically grounded, technically buildable and interpretable if it fails.
Biological rationale
Is the mechanism supported by relevant structural, RNA or disease-context evidence?
Modality fit
Should the question be approached through a small molecule, a degrader or an RNA-directed strategy?
Design feasibility
Are binding features, molecular geometry or sequence constraints compatible with the proposed mechanism?
Developability
Which synthesis, permeability, stability or delivery risks could prevent a plausible design from working?
Validation & IP review
Which experiment could change the decision, and which prior-art questions require specialist review?
A useful recommendation can be any of these
- AdvanceTest a defined hypothesis.
- ReviseResolve a specific uncertainty.
- StopRecord why the path does not hold.
The reasoning should be reproducible.
- Inputs & provenance
- Keep the source data, assumptions and evaluation scope with the result.
- Controls & calibration
- Use reference cases and report where a method does not transfer reliably.
- Versioned handoff
- Preserve configurations and checksums so a reviewed analysis can be traced.
- Negative findings
- Keep failed hypotheses and corrections, not just favorable rankings.
LIMITATIONS
Shared limitations
These four points apply to everything on this page.
Computation is not experiment
Computational rankings, scores and agreement between AI models do not establish binding, degradation or efficacy. We claim no efficacy, clinical entry or independently validated performance.
Films and diagrams are simplified
The teaser film and concept diagrams simplify general principles. They do not depict the structure or design of any specific candidate, or experimental results.
What we disclose
PRMT5 is the only program we disclose individually. Other targets, candidate structures and sequences, and unpublished results are discussed only under a confidentiality agreement.
Laboratory and legal work are separate
Synthesis and measurement are carried out by experimental partners; this does not imply in-house laboratory facilities or an existing partnership. Patentability and freedom to operate require independent specialist review.
Partnering§ 05A defined question. Complementary expertise.
Computational depth.
Experimental partnership.
We work at the computational front end of discovery. Research teams, biotechnology companies and experimental providers can pair that work with their own assets and validation capabilities.
NEO Drug Discovery
What we bring
A scoped comparison of partner-defined candidates, including baseline controls, risk flags and a traceable recommendation.
The question we ask together
“Can this analysis improve one experimental decision?”
Research and discovery teams
What we discuss
- A defined candidate set and reference cases
- Agreed outputs and evaluation criteria
- A joint review before expanding the scope
NEO Drug Discovery
What we bring
Design rationale, proposed controls and assay questions that connect a computational result to a measurable outcome.
The question we ask together
“Which experiment would confirm—or overturn—the hypothesis?”
Experimental and validation partners
What we discuss
- Synthesis and material quality
- Binding, cellular and developability readouts
- Advance, revise or stop criteria and feedback
NEO Drug Discovery
What we bring
Computational discovery capabilities and our publicly disclosed PRMT5 research direction, with other non-public work discussed only under CDA.
The question we ask together
“How should the evidence determine the next development milestone?”
Co-development and business partners
What we discuss
- Joint discovery or staged co-optimization
- ADC payload/linker and peptide pilot opportunities
- Ownership, confidentiality and evidence-based milestones
FROM DESIGN TO EXPERIMENT
A clear division of work. A shared learning cycle.
We focus on the computational design and analysis layer. Experimental partners provide synthesis and measurement. Responsibilities, controls and decision criteria are defined before a project starts.
Design
NEO Drug Discovery
Hypotheses, candidate priorities and an assay brief.
Make
Experimental partner
Synthesis, material preparation and quality control.
Test
Experimental partner
Binding, cellular and property measurements suited to the question.
Analyze
Joint review
Compare prediction with observation; advance, revise or stop.
PILOT EXAMPLE
Start with one answerable question.
An illustrative structure for a computational pilot. It begins with an email and a first conversation; if unpublished material is needed, a confidentiality agreement comes first.
What the partner provides
- The research question and target information
- The candidate set to compare (structures or sequences)
- Reference cases with known activity or inactivity
- Existing experimental results, if any
Agreed before we start
- One question, and the experiment its answer will decide
- Methods, controls and evaluation criteria
- Report format and how the joint review works
- The decision criteria for the end of the pilot
Duration
Set together in the proposal, once the scope is clear.
Stages
- Agree the scope
Fix the question, candidate set, reference cases and deliverables in writing.
- Reproduce reference cases
Check first, on known cases, that the methods work for this target.
- Compare candidates
Compare candidates across methods, recording uncertainty and reasons for exclusion.
- Joint review and report
Review the report together and decide on the next experiment and the collaboration.
Pilot report: contents
- Decision brief
The research question, scope and recommendation.
- Candidate rationale
Priorities, controls, assumptions and risk flags.
- Experimental brief
Draft protocols, material specifications, readouts and decision criteria.
- Reproducibility record
Input provenance, methods and versioned outputs.
An outline, not actual results. Deliverables are scoped with each partner.
Decisions at the end of the pilot
- Expand
Reference cases were reproduced and the ranking is clear enough to choose the next experiment.
We discuss the experimental handoff and the next stage of collaboration.
- Adjust
Reference cases were only partly reproduced, or the differences between candidates are uncertain.
We change the question, methods or candidate set and set the scope again.
- Stop
Reference cases could not be reproduced, so the question cannot be answered.
We document the reasons and the records, and close the pilot.
Team§ 06The people leading the company
Leadership
CEO
Jihae Lim
KICPA
- Former CFO & COO, LVMH Perfumes & Cosmetics
- Former Senior Director, Baker Tilly
- Former Auditor & Consultant, Deloitte
CTO
Sanghun Lee
MD, PhD
- Professor, University of Ulsan College of Medicine
- Director, Regional Cancer Center, Ulsan University Hospital
- Former visiting scholar, UC Irvine School of Medicine
Contact§ 07Before the first conversation
Good partnerships
start with clear questions.
Official inquiry email
Research and business partnering
Tell us, at a level you can share publicly, what you hope to achieve and which areas interest you.
Inquiries we handle
- Research collaboration
- Computational pilot scope
- Co-development and business partnering
contact@neodrug.ai
Your mail app opens with our address and a subject line; nothing is sent until you press Send. This site does not collect or store inquiries, or confirm whether they are sent or received.
If no mail app opens, copy the address and write from the email service you use.
Please do not send unpublished research data or sensitive personal information.
Before you write
What should I include?
Your partnering goals, research interests, the capabilities you can offer and the scope you would like to discuss, all at a level you can share publicly. Specific candidate structures or unpublished experimental data are not needed for a first conversation.
Should I send unpublished data?
No. Please do not send sensitive research data or personal information. Non-public materials are discussed only after mutual review and a confidentiality agreement (CDA) are in place.
How is my inquiry delivered?
Our official inquiry address is contact@neodrug.ai. The button opens your mail app, and the message is sent only when you press Send there. This site does not confirm whether messages are sent or received.
Do you provide in-house laboratory or antibody development services?
Our role is computational design, prioritization and analysis. Synthesis, laboratory measurements and antibody development require appropriately qualified experimental partners and a separately agreed scope.
How are pilot duration and cost set?
After a first conversation about the question and the size of the candidate set, scope, deliverables and timing are set in a written proposal. We do not commit to a duration before the scope is known.