ML-Readiness Scoring

The ML-readiness score evaluates how suitable a molecule is for machine learning workflows by assessing structural quality, physicochemical properties, complexity, and representation quality across four dimensions totaling 100 points.

Score Dimensions

Dimension	Max Points	What It Measures
Structural Quality	20	Structural soundness for ML pipelines
Property Profile	35	Physicochemical property desirability
Complexity & Feasibility	25	Synthetic tractability and complexity
Representation Quality	20	Numerical representability for ML models

1. Structural Quality (20 points)

Binary pass/fail checks on fundamental structural soundness. Each item either passes (full points) or fails (0 points).

Item	Points	Pass Condition
Single component	5	Exactly 1 fragment — no mixtures, salts, or solvents
Standard organic elements	5	No metal atoms present (no organometallics)
No radicals	3	No atoms with unpaired electrons
Reasonable charge	3	Net formal charge between −2 and +2
No dummy atoms	4	No R-groups or attachment points (atomic number ≠ 0)
Total	20	Sum of passed items

Non-scored caveats (reported as warnings):

• Isotope labels detected (deuterium, ¹³C, tritium)
• Trivial molecules (≤ 3 heavy atoms)

2. Property Profile (35 points)

Desirability-scored physicochemical properties measuring how well a molecule fits within typical ML training set distributions. Each property is scored using a trapezoidal desirability function.

Property	Ideal Range	Max Points
Molecular Weight	200–500 Da	6
LogP (Wildman-Crippen)	0.5–4.0	6
TPSA	40–120 A²	5
H-Bond Donors	0–3	4
H-Bond Acceptors	2–8	4
Rotatable Bonds	1–8	5
Aromatic Rings	1–3	5
Total		35

Desirability Function

Each property is scored using a trapezoidal desirability function:

If min ≤ value ≤ max:  d = 1.0  (ideal range → full score)
If value < min:        d = max(0, 1.0 − (min − value) / range)
If value > max:        d = max(0, 1.0 − (value − max) / range)

where range = max − min
points = d × max_points

Example: A molecule with MW = 600 Da:

• range = 500 − 200 = 300
• d = max(0, 1.0 − (600 − 500) / 300) = 0.667
• points = 0.667 × 6 = 4.0

Property Ranges

These ranges reflect the most common distributions in drug-like compound datasets used for ML training. Molecules outside these ranges aren't necessarily bad — they just fall outside the typical training distribution.

3. Complexity & Feasibility (25 points)

Assesses synthetic tractability and structural complexity, which affect practical utility in ML-driven drug discovery campaigns.

Component	Max Points	Calculation
QED	8	QED.qed(mol) × 8
SA Score	8	Inverse mapping (see below)
Fsp3	4	desirability(Fsp3, 0.2, 0.6) × 4
Stereocenters	5	Complexity-adjusted scoring (see below)
Total	25

SA Score → Points

Synthetic Accessibility Score (1–10) is inversely mapped to points:

SA Score	Points	Interpretation
≤ 3	8.0	Easy to synthesize
3–5	8.0 − (SA − 3) × 2.0	Moderate complexity
5–7	4.0 − (SA − 5) × 2.0	Difficult synthesis
> 7	0.0	Very difficult

Stereocenter Scoring

Stereocenters	Base Score	Notes
0–4	5.0	Manageable complexity
5–8	5.0 − (n − 4) × 0.75	Decreasing linearly
> 8	0.0	Too complex

Penalty: If more than 50% of stereocenters are undefined, the base score is halved (×0.5).

4. Representation Quality (20 points)

Measures how well the molecule can be numerically represented for ML models — the core requirement for any ML application.

Component	Max Points	What It Tests
Descriptor completeness	5	Fraction of 451 RDKit descriptors computed successfully
Fingerprint generation	5	Weighted success across 7 fingerprint types
Fingerprint informativeness	5	Ideal bit density between 1% and 30%
Conformer generation	5	3D coordinate generation via ETKDGv3
Total	20

Descriptors Tested (451 total)

Descriptor Set	Count	Method
Standard RDKit	217	Descriptors.CalcMolDescriptors()
AUTOCORR2D	192	rdMolDescriptors.CalcAUTOCORR2D()
MQN	42	rdMolDescriptors.MQNs_()

Score: round(5.0 × (successful / 451), 2)

Fingerprint Types & Weights

Fingerprint	Bits	Weight	Description
Morgan (radius=2)	2048	8	Circular fingerprints (ECFP4-like)
Morgan Features	2048	8	Feature-based Morgan
MACCS Keys	167	8	166-bit MACCS structural keys
Atom Pair	2048	4	Atom pair descriptors
Topological Torsion	2048	4	Topological torsion descriptors
RDKit FP	2048	4	Daylight-like path fingerprints
Avalon	512	4	Avalon toolkit fingerprints
Total weight		40

Score: round(5.0 × (sum of successful weights / 40), 2)

Fingerprint Informativeness

Measures whether fingerprints have useful information content (not too sparse, not too dense):

Avg Bit Density	Score	Interpretation
1–30%	5.0	Ideal information content
< 1%	Proportional	Too sparse (molecule too simple)
30–45%	Decreasing	Too dense (losing discriminative power)
> 45%	0.0	Not informative

Conformer Generation

Method	Points
ETKDGv3 success (seed=42, maxIter=500)	5
Random coordinate fallback	3
Complete failure	0

Overall Score & Tiers

Total = Structural Quality + Property Profile + Complexity & Feasibility + Representation Quality

Score	Tier	Interpretation
85–100	Excellent	Suitable for most ML workflows without modification
70–84	Good	Minor limitations; generally suitable with standard preprocessing
50–69	Moderate	Usable but may need careful feature selection or preprocessing
30–49	Limited	Significant challenges; consider alternatives or specialized models
0–29	Poor	Not recommended for standard ML pipelines

The UI displays an interpretation banner with:

• Overall score badge
• Tier-specific guidance text
• Per-dimension health tags showing completion percentage

API Usage

curl -X POST http://localhost:8001/api/v1/score \
  -H "Content-Type: application/json" \
  -d '{
    "molecule": "CC(=O)Oc1ccccc1C(=O)O",
    "include": ["ml_readiness"]
  }'

Response:

{
  "ml_readiness": {
    "score": 88,
    "dimensions": [
      {
        "name": "Structural Quality",
        "score": 20.0,
        "max_score": 20,
        "items": [
          {"name": "Single component", "score": 5.0, "max_score": 5, "passed": true},
          {"name": "Standard organic elements", "score": 5.0, "max_score": 5, "passed": true},
          {"name": "No radicals", "score": 3.0, "max_score": 3, "passed": true},
          {"name": "Reasonable charge", "score": 3.0, "max_score": 3, "passed": true},
          {"name": "No dummy atoms", "score": 4.0, "max_score": 4, "passed": true}
        ]
      },
      {
        "name": "Property Profile",
        "score": 30.5,
        "max_score": 35,
        "items": [
          {"name": "MW", "score": 6.0, "max_score": 6, "passed": true},
          {"name": "LogP", "score": 6.0, "max_score": 6, "passed": true}
        ]
      },
      {
        "name": "Complexity & Feasibility",
        "score": 21.0,
        "max_score": 25
      },
      {
        "name": "Representation Quality",
        "score": 16.5,
        "max_score": 20
      }
    ],
    "interpretation": "Good ML candidate with minor limitations...",
    "caveats": []
  }
}

Use Cases

Dataset Curation for ML

Filter molecules before creating training datasets:

ml_readiness_score >= 80 AND validation_score >= 90

This ensures:

• Structurally sound for ML (no mixtures, metals, radicals)
• Properties within typical training distributions
• All required descriptors and fingerprints generate successfully
• 3D conformer can be generated

Model Applicability Domain

Use ML-readiness to define applicability domain:

• Train models only on molecules with score >= 80
• Flag predictions on molecules with score < 80 as uncertain
• Exclude molecules with score < 60 from predictions

Dimension-Specific Analysis

Use individual dimension scores to diagnose issues:

Low Dimension	Likely Issue	Action
Structural Quality	Mixtures, metals, radicals	Clean up structure first
Property Profile	Unusual MW, LogP, TPSA	May be outside typical drug-like space
Complexity & Feasibility	Hard to synthesize, many stereocenters	Consider if practical for your use case
Representation Quality	Descriptor failures, no 3D	May need specialized featurization

Limitations

Does not test:

• Descriptor quality or relevance to specific models
• Model-specific feature requirements
• Chemical space coverage of your training set
• Experimental measurability

Assumes:

• Standard RDKit descriptors are sufficient
• Common fingerprint types are appropriate
• Property ranges derived from typical drug-like datasets

Custom Requirements

ML-readiness tests standard descriptors and fingerprints. If your model uses custom features (e.g., graph neural network features), you'll need additional validation.

References

1. Bickerton, G. R. et al. (2012). Quantifying the chemical beauty of drugs. Nature Chemistry, 4(2), 90–98.
2. Ertl, P. & Schuffenhauer, A. (2009). Estimation of synthetic accessibility score. Journal of Cheminformatics, 1(1), 8.
3. Lovering, F. et al. (2009). Escape from flatland. Journal of Medicinal Chemistry, 52(21), 6752–6756.
4. Rogers, D. & Hahn, M. (2010). Extended-connectivity fingerprints. Journal of Chemical Information and Modeling, 50(5), 742–754.

Next Steps

• Scoring Overview — All scoring systems
• Batch Processing — Score large datasets
• API Reference — Full scoring API