eshmun-thinking

Thinking-aware instruction-tuning dataset for protein language models: chain-of-thought <think>...</think> reasoning traces, built programmatically from UniProt/SCOP knowledge graphs (not distilled from an external LLM or self-sampled), covering protein annotation (sequence → description) and generation (description → sequence) in both directions.

Hypothesis under test: does inserting an explicit reasoning step grounded in real sequence features (motifs, domains, GO function/process/localization, EC numbers, SCOP fold/superfamily/ family, PPI partners) before the final answer improve instruction-tuned protein models over direct instruction tuning (InstructProtein/ProLLaMA-style)?

Configs

8 configs, each with train/validation/test splits (sequence-identity-based via MMseqs2: validation <70% identity to train, test <30% identity to train):

config	task family	direction	examples
task_a_annotation	family/function/process/localization/catalytic activity (human)	sequence → property	42,714
task_a_generation	same, generation direction	property → sequence	132,817
all_organism_annotation	same task family, all reviewed SwissProt organisms	sequence → property	1,802,616
all_organism_generation	same, generation direction	property → sequence	3,075,604
ppi_annotation	protein-protein interaction (1:3 positive:negative)	pair → yes/no	104,376
ppi_generation	protein-protein interaction	partner + context → sequence	52,188
scop_annotation	SCOP fold/superfamily/family	sequence → property	38,457
scop_generation	SCOP fold/superfamily/family	property → sequence	36,945

Each row has instruction, reasoning (the <think>...</think> block), answer, plus task-specific identifier columns (entry, entry_a/entry_b, target_relation, target_value, label).

Protein sequence encoding

Sequences are wrapped as <protein>...</protein> with a Ƥ character joining every residue ("Ƥ".join(sequence)), e.g. <protein>MƤSƤLƤGƤLƤL...</protein>. This matches the tokenization convention of the InstructProtein-based pilot this dataset currently trains (khairi/Eshmun-Thinking-Pilot), whose tokenizer has no protein-specific vocabulary and would otherwise BPE-merge multiple residues into one token, losing residue-level granularity. Strip Ƥ to recover the raw sequence.

Reasoning trace subjects

Annotation direction: reasoning cites real, known facts about the specific queried protein (or, for SCOP, the other two structural levels of the same solved entry) — subject is literally protein (or protein_a/protein_b for PPI, since both are given inputs).

Generation direction: the model is given only one stated property as input (e.g. member_of = "Cytochrome P450 family"), so reasoning cannot cite other instance-specific facts about a not-yet-generated sequence without asserting things the model can't know and inverting the real causal order (sequence determines function, not the reverse). Instead it cites population-level statistics about the target category, computed from train-split entries only, and the triple subject reflects that — a short (one/two-word) category label, not protein:

target_relation	subject
catalyzes	catalytic_activity
has_function	molecular_function
involved_in	biological_process
located_in	cellular_component
member_of	family
scop_fold	fold
scop_superfamily	superfamily
scop_family	scop_family

Example (task_a_generation, target catalyzes = 1.-.-.-):

instruction: I want a protein sequence with the catalytic activity described by EC number 1.-.-.-.
reasoning: <think>
(catalytic_activity, has_length, [>400] amino acids)
(catalytic_activity, involved_in, glycolipid biosynthetic process)
(catalytic_activity, located_in, mitochondrion)
...
</think>
answer: <protein>MƤLƤPƤNƤTƤGƤRƤL...</protein>