{"id":"W4413312092","doi":"10.1073/pnas.2506316122","title":"Sparse autoencoders uncover biologically interpretable features in protein language model representations","year":2025,"lang":"en","type":"article","venue":"Proceedings of the National Academy of Sciences","topic":"Machine Learning in Bioinformatics","field":"Biochemistry, Genetics and Molecular Biology","cited_by":25,"is_retracted":false,"has_abstract":true,"ca_institutions":"Toronto Metropolitan University","funders":"National Institute of General Medical Sciences; National Cancer Institute; Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology; National Institutes of Health; Massachusetts Institute of Technology","keywords":"Computer science; Artificial intelligence; Natural language processing; Pattern recognition (psychology)","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005932061,0.00006104749,0.00008023609,0.000100279,0.00006085427,0.00001436406,0.0005014774,0.00008582924,0.000003460843],"category_scores_gemma":[0.00105287,0.00004025407,0.00004351229,0.0003597408,0.0003291265,0.0000195027,0.000197488,0.0001260608,2.667073e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001228381,"about_ca_system_score_gemma":0.00004829631,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001084546,"about_ca_topic_score_gemma":6.814482e-7,"domain_scores_codex":[0.9992814,0.000006970071,0.0002141868,0.0001472868,0.0002551557,0.00009501402],"domain_scores_gemma":[0.9996976,0.00001907805,0.0001730853,0.000011316,0.00008795793,0.00001091247],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.00002512449,0.00003101429,0.009507165,0.00005193671,0.00001154244,2.617388e-9,0.0001939266,0.02548548,0.9389358,0.02411386,0.001499435,0.0001447823],"study_design_scores_gemma":[0.0003325432,0.00007264909,0.04302323,0.0001920485,0.000008836798,0.000002902306,0.0005949383,0.1109069,0.8120158,0.03246585,0.0002498682,0.0001344981],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9605126,0.00008679647,0.00008834035,0.001684892,0.000008439686,0.0002055045,0.00001014422,0.000005423725,0.0373978],"genre_scores_gemma":[0.9869943,0.000009083942,0.01104641,0.0003964252,0.00001034873,0.00001272307,6.946573e-7,0.00000152779,0.001528551],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.12692,"threshold_uncertainty_score":0.1641513,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01540583249850305,"score_gpt":0.3230372084270323,"score_spread":0.3076313759285293,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}