---
license: mit
tags:
  - protein
  - small-molecule
  - affinity-prediction
  - hit-to-lead
  - qsar
  - boltz
  - structure-based
  - capo
metrics:
  - pearsonr
  - spearmanr
model-index:
  - name: capo-hit-to-lead-ft-binding-boltz-20260706-1111-cd08-best
    results:
      - task:
          type: regression
          name: Hit-to-lead affinity ranking (FEP benchmark, held-out batch)
        metrics:
          - type: pearsonr
            value: 0.7396
            name: All-targets mean log-Pearson r (15 valid targets)
          - type: spearmanr
            value: 0.7000
            name: All-targets mean Spearman r (15 valid targets)
---

# CAPO · Hit-to-lead affinity head on frozen Boltz-2

A **per-target-calibrated QSAR affinity head** that turns [Boltz-2](https://github.com/jwohlwend/boltz)
co-folding outputs into a **calibrated potency ranking** for hit-to-lead. Boltz-2
weights are **frozen** — this is "fine-tuning to the assay" by calibration on
measured `batch_1` potencies, not gradient training. That is precisely why 549
cached co-folds from a prior run could be reused at 0 GPU-hours.

Produced by **CAPO** (Compute-Aware Automated Protein-Model Optimization), run
`binding-boltz-20260706-1111-cd08`. This repo contains the **best** checkpoint only.

## What this is (and is not)

- **Is:** a light regression head (`affinity_head.pkl`) + per-target calibration
  that predicts log₁₀-potency from features derived from a Boltz-2 co-fold of a
  protein–ligand complex.
- **Is not:** a standalone folding/affinity model. It **requires Boltz-2** to
  generate structural features for any new ligand, and no Boltz-2 weights are
  included or modified.

## Features consumed by the head

Per protein–ligand complex (`affinity_head_manifest.json`):

- **Ligand:** ECFP4 fingerprint (radius 2, 1024 bits) + RDKit descriptors
  (MolWt, MolLogP, TPSA, NumHDonors, NumHAcceptors, NumRotatableBonds,
  NumAromaticRings, FractionCSP3).
- **Boltz-2 signals:** affinity_pred_value, affinity_probability_binary, ptm,
  iptm, complex_plddt, complex_pde.
- **Contacts:** contacts_4A, contacts_5A, contact_residues, buried_frac.

## Task & data

Hit-to-lead affinity ranking on a FEP-style benchmark spanning **29 drug targets**
(bace, cdk2, tyk2, shp2, mcl1, eg5, p38, …). Given `batch_1` measured potencies per
target, rank a held-out `batch_2`. **644 complexes** were co-folded (549 reused +
96 new). The intended headline was to show `classical-ML < raw Boltz-2 << CAPO`
on the regime that is hard for **both** baselines.

## Evaluation

All-target mean over the **15 targets** that yielded valid ranking metrics
(`results/analysis/summary.json`):

| Metric (log₁₀ µM) | This model (CAPO) | Raw Boltz-2 | Classical ML |
|---|---:|---:|---:|
| log-Pearson r | **0.7396** | 0.6227 | 0.6416 |
| Spearman r | **0.7000** | 0.5936 | 0.6162 |
| RMSE | **0.5588** | 1.0611 | — |

Calibration improves ordering and **nearly halves RMSE vs raw Boltz-2** — its main
win is fixing magnitude. The `classical < Boltz < CAPO` ordering holds on the
all-target log-Pearson average.

## ⚠️ Important caveats

- **Pre-registered headline is unresolved.** The "hard-for-both" filter
  (raw-Boltz Spearman < 0.30 **and** classical LOO < 0.20, `loo_min_n = 8`) selected
  **zero targets** (`headline_hard_for_both_targets: []`). The primary claim was
  never testable on this benchmark — thresholds were too strict.
- **14 / 29 targets returned NaN** ranking metrics from sparse `batch_2` / too-few
  ligands; only 15 targets contribute to the average above.
- Sparse ligand counts can drive negative per-target correlations (e.g. hsp90
  Spearman −0.80). Trust the pooled average over any single sparse target.

## Files

- `affinity_head.pkl` — the calibrated QSAR affinity head (~23 MB).
- `affinity_head_manifest.json` — feature spec, seed, per-target detail, versions
  (Python 3.10.12).
- `config.json` — `{strategy: "custom_frozen_boltz2_calibrate_qsar", model_id: "custom", n_targets: 29}`.
- `metrics.json` — full pre-registration table + thresholds + per-target results.

## How to use

```python
import pickle
from huggingface_hub import hf_hub_download

REPO = "theoschiff-biie/capo-hit-to-lead-ft-binding-boltz-20260706-1111-cd08-best"
head = pickle.load(open(hf_hub_download(REPO, "affinity_head.pkl"), "rb"))

# 1. Co-fold each protein–ligand pair with Boltz-2 (frozen).
# 2. Build the feature vector per the ECFP4 + RDKit + Boltz-2 + contact spec in
#    affinity_head_manifest.json (feature order must match training).
# 3. head.predict(X) -> calibrated log10-potency; rank within a target to lead-rank.
```

Reproduce feature extraction with the CAPO run's `src/` pipeline; the manifest is
the authoritative feature contract.

## Provenance

- **CAPO run:** `binding-boltz-20260706-1111-cd08` (`RUN_REPORT.md` in the run dir).
- **Model selection:** bypassed — user-specified custom frozen-Boltz-2 pipeline.
- **Cost gate:** projected $10.61 vs $75 cap on 1× A100-SXM4 @ $1.99/h.
- **Cost:** infra $3.35 (1h 57m); total run $22.09. Reuse of 549 co-folds at 0 GPU-h.
- **Push:** `checkpoints/best/` only (private). `checkpoints/last/` retained locally,
  not pushed.

*License follows the Boltz-2 base model (MIT).*