Best Practices for Using MolecularFingerprints.jl

To get the most out of MolecularFingerprints.jl, keep the following engineering principles in mind:

1. Leverage Type Specialization

Julia's compiler specializes code based on types. By defining your calculator once (e.g., calc = ECFP{2048}(3)), you allow the compiler to optimize the internal loops for that specific bit-length. Avoid redefining the calculator inside loops.

2. Threading Performance

The batch fingerprint function uses Threads.@threads. To ensure you are actually using multiple cores, check your environment:

using Base.Threads
println(nthreads()) # Should be > 1

If it returns 1, start Julia with julia -t auto.

3. Bit-Vector vs. Dense Vector

Note that fingerprints are returned as BitVector objects. This is memory-efficient (1 bit per element) and allows for high-speed logical operations (AND, OR, XOR) which are critical for calculating Tanimoto similarities.

4. Canonicalization

While the library handles SMILES parsing, remember that different SMILES strings can represent the same molecular graph (e.g., C1=CC=CC=C1 vs c1ccccc1). For consistency in machine learning tasks, ensure your input data is canonicalized.