VHH prioritisation for CAR screening
This case examines why an immune-selected VHH does not automatically work inside a CAR, and how functional selection, two-dimensional binding-region grouping and expert judgement prioritise candidates for CAR experiments.
The scientific mismatch
Immune selection rewards antigen recognition and clonal expansion.
CAR-T requires a candidate that still works after entering the intended receptor and cellular context.
Abundance can describe immune history. A planar binding-region model can prioritise hypotheses, but it cannot directly predict CAR compatibility.
One chain, two kinds of evidence
Wet experiments decide whether a candidate is eligible. Dry analysis explains and prioritises candidates that already passed. They are hierarchical, not interchangeable weights.
Hard gate
CAR activation selection
Positive target-density conditions retain functional VHH-CARs; negative conditions remove non-specific activation.
Interpretation
NGS identification
Sequences identify the VHHs that passed the functional selection. Read count is not the primary rank.
Interpretation
Diversity control
CDR3 and sequence families prevent near-neighbours from consuming the downstream budget.
Interpretation
Structure and PDB QC
Docking, interface quality and model QC ask whether the proposed complex and interface are technically plausible.
Interpretation
Binding-region grouping
Predicted antigen contact patterns group candidates; regions already supported by CAR-T function define the priority search space.
Interpretation
Semi-dynamic verification
Short GROMACS MD and MM/PBSA examine relaxed binding energy, residue contributions and antigen-side hotspots.
Existing platform module · /gromacsWet validation
Protein and CAR-T assays
Expression, affinity, cell binding and ultimately CAR-T function determine the outcome.
The current epitope model is two-dimensional
The method across two targets
FcRH5
Evidence now
- 1,042 candidates and interface evidence are preserved in the platform batch.
- The earlier platform shortlist contained 17 recommended candidates across 9 CDR3 communities. The 10 Sep update reports 11 candidates submitted for expression.
- The final panel covers two binding-region groups: 8 control-matched candidates and 3 in another group, all with AI screening totals above 70.
- Control-matched group (Cevo_Final_sorted): 0740 (96.4), 0669 (92.2), 0254 (88.0), 0354 (78.4), 0727 (78.4), 1025 (78.2), 0655 (74.8), 0493 (72.2). IDs use the FcRH5 prefix.
- Other group (New_Final): 0929 (82.4), 0055 (76.4), 0748 (72.2). These match the workbook rows above 70; expression submission is scientist-reported.
- The best-abundance clone in the control-like region sat outside the overall top 200 by abundance.
- The control calibration reference is 50; the actual selection cutoff for this panel is above 70. Protein characterisation and subsequent integrated scoring are pending.
What it means
The decision has produced an 8 + 3 expression panel prioritised by binding-region group and AI score. The workbook calculates AI Total_Score as 40% Stat_Score + 60% MD (Score). Experimental superiority and integrated scoring require the forthcoming characterisation results.
Still open
Expression and characterisation results for all 11 candidates, followed by integrated scoring and CAR-T validation. The earlier 17-candidate shortlist and final 11-candidate panel are separate milestones; their candidate-level reconciliation remains to be documented.
CD70
Evidence now
- The final library contained about 35 distinct CDR3s, a materially narrower diversity base than FcRH5.
- The first dry/wet pass tested 24 candidates and reported one preferred clone among the computational top five.
- Its predicted contact region overlapped the control region by about 55%.
What it means
This is an early feasibility signal that the workflow transfers to a different antigen architecture. It is not yet a performance benchmark.
Still open
The decisive CAR-T functional result is pending.
An operational epitope vocabulary
The method separates structural definitions from decision logic: predicted contacts define groups, prior CAR-T function prioritises groups, and candidate-level evidence drives selection within a group.
- Contact region
- approximately 4.5 Å neighbourhood
- Main epitope residue
- residue pair ≤ 3.5 Å and antigen-residue BSA ≥ 30 Ų
- Epitope group
- candidates sharing a similar predicted antigen contact region; the unit for groupwise comparison
- Priority region group
- a binding-region group supported by prior CAR-T function; a strong search prior, not proof of an exact residue-level epitope
- Cross-pose design mode
- when orientation is the objective, specify the MOA-critical antigen region and ask AI to design toward that constraint; this is a separate generative task, not an extension of the planar rank
What the score means, and how thresholds are governed
The 25-point AI/physicochemical block and 75-point protein-characterisation block normalise heterogeneous indicators into usable ranges. They are not literal claims that the evidence is worth 25% versus 75%.
Rank shows order; the score also shows the margin under the same standard. A top-ranked candidate at 80 and a top-ranked candidate at 65 do not carry the same evidence strength.
For FcRH5, 50 is the control calibration reference. The final expression panel uses an AI screening total above 70.
Total_Score = 0.4 × Stat_Score + 0.6 × MD (Score), as recorded in interface_info.xlsx. This dry-lab score is distinct from the later integrated score incorporating protein characterisation.
The order of evidence and the distinction between wet gates and dry interpretation.
The antigen region, control definition, experimental-entry threshold and other project parameters encode the mechanism-of-action hypothesis.
For FcRH5, prefer F at VHH IMGT position 42. Seek higher tyrosine content across CDR3, but exclude any contiguous YYY motif. These are project-level design parameters, not universal VHH rules.
Each experiment returns its operator, conditions, raw measurements and expert interpretation. Parameter corrections or additions are versioned and traceable rather than silently learned.
The joint contribution
Define the MOA-critical region, controls, project parameters, functional selection logic and experimental interpretation.
Xie Meng / project scientistsExecute the workflow consistently; link every experiment to its operator, conditions, raw data, expert judgement, candidate lineage and parameter version.
IASO Bio AIReview returned evidence to correct or add project parameters, while testing which rules transfer across targets and which remain target-specific.
JointThe claim worth testing
The platform succeeds when its workflow and evidence help a candidate continue through development toward market—not when it merely completes a run, produces a shortlist or assigns a high score.
Established in the working method
- CAR activation is an upstream wet-lab hard gate.
- NGS abundance is not the core candidate rank.
- Predicted binding regions group the passing pool; CAR-T-supported groups are prioritised and candidates are compared within groups.
- The current epitope model is planar because two-dimensional evidence is presently more abundant and reliable than cross-pose data.
- Cross-pose questions move into region-constrained AI design rather than being inferred from the planar rank.
- Experimental data and the responsible scientist's interpretation both return to the project as versioned evidence.
- Thresholds are expert-governed and project-specific.
Evidence needed next
- Complete the CD70 CAR-T functional readout.
- Compare FcRH5 control-like candidates with the control in the same wet-lab context.
- Preserve all candidate-level outcomes together with experimental conditions, operator and expert interpretation—not only the selected success.
- Only then quantify threshold calibration and the incremental value of each dry proxy.