イオンチャネル開閉のリアルタイム マルコフモデリングへの深層学習アプローチ Oikonomou et al. (2024), Communications Chemistry 7:280
1. 書誌情報
原題
A deep learning approach to real-time Markov modeling of ion channel gating
邦題(要約)
イオンチャネル開閉のリアルタイム マルコフモデリングへの深層学習アプローチ
著者
Efthymios Oikonomou, Yannick Juli, Rajkumar Reddy Kolan, Linda Kern, Thomas Gruber, Christian Alzheimer, Patrick Krauss, Andreas Maier, Tobias Huth (corresponding)
所属
Friedrich-Alexander-Universität Erlangen-Nürnberg (Institute of Physiology and Pathophysiology / Pattern Recognition Lab / NHR@FAU)
67 ms は推論のみで、idealization (HOHD) + 2D-histogram 作成の時間は含まれない。実機への組み込みでは前処理が律速になる可能性。「office computer の background で走らせられる」と著者は楽観的だが、サンプリング 100 kHz × 10 s = 1 M sample の HOHD を online で回す実装は本論文では未検証。Discussion で「recording software に統合する必要がある」と素直に認めているのは好印象。
Fig 1. NN アーキテクチャ(A: topology, B: rates, C: Channel-Increase module, D: RAE vs MAPE)
PDF Fig 1 caption: "Illustration of the neural network architectures used for Markov modeling. In this study, we used modified versions of the Inception-Res-Net-V2 architecture for (A) topology discrimination and (B) rate estimation. (C) The original Reduction-B module was substituted with a module that increases the filter dimension without pooling. (D) Rate constant prediction was evaluated using the RAE error score (Eq. 3). In comparison to the mean absolute percentage error (MAPE), the RAE score is symmetrical with respect to the ground truth."
PDF Fig 2 caption: "Flow chart of the proposed algorithm. The orange path shows the flow of the experimentally recorded data. It is sequentially fed to the topology estimation NN and to the NN for the estimation of the transition rates. The blue paths indicate training of the topology estimation NN and rates estimation NN with simulated training datasets I and II. Dataset I contains 2D-histograms simulated with a collection of models encompassing different topologies. Dataset II consists of a collection of simulated datasets where each set encompasses only one specific topology with a range of rates kij. The green path shows the two stages of estimating the kinetic model."
(b) 図中要素の解説(サブパネル無し)
オレンジパス(推論ライン)
実験 patch-clamp 時系列 → idealization (HOHD) → 2D-histogram → Topology NN → 該当 Rates NN → HMM 出力。1 ステップずつ独立に置き換え可能な設計。
青パス I(topology NN の訓練)
Dataset I = 18 トポロジー混合の 10 M+ サンプル → topology NN を分類器として学習。一回きり。
青パス II(rates NN の訓練)
Dataset II = トポロジー別(COCOC, CCCOO, ...)の rate-randomized サンプル → 各トポロジーに対応する rates NN を回帰器として学習。トポロジー数だけ NN を用意。
PDF Fig 3 caption: "Topology estimation of Markov models using neural networks. (A) Shows all possible linear five-state topologies that were all encompassed in the training dataset. They are grouped according to the number of open/closed states and their interconductance rank (number of independent C-O links). (B) The accuracy related to the size of the training dataset is displayed. (C, D) The confusion matrices were obtained by testing a single NN that has been trained with 10⁷ 2D-histograms. (C) The recall (diagonal) and FNR (off-diagonal). (D) The precision (diagonal) and FDR (off-diagonal)."
PDF Fig 4 caption: "Transition rates estimation of COCOC and CCCOO models. The regression architecture (Fig. 1A) was trained using datasets No. 2 and 3. (A, B) illustrate the predictions for the overall best-predicted rates k₅₄, k₃₄, and (C, D) for the worst-predicted rates k₂₁, k₂₁ according to the overall RAE score for COCOC and CCCOO. Each test dataset contains 10,000 samples. The orange short-dashed line and the red dashed line indicate error scores (RAE) equal to 0.6 and 1.0."
PDF Fig 5 caption: "Analysis of the transition rates estimation for COCOC and CCCOO models. (A, B) Cumulative distributions of error scores (RAE) for each kij, with paired rates (k12-k21 etc.) visualized as hatched areas. Pink dotted line: randomly predicted rates baseline. (C, D) Five models selected at 100/75/50/25/0 percentile, each re-simulated 1000 times. Box-and-whiskers show median, 25/75/10/90 percentiles. Orange dots: 2D-Fit comparison (4 time series, 64 ensemble runs each). (E, F) Ranked RAE vs number of detected events. Red line: 1025-window geometric average. Dashed blue in (F): 100 M sample training (dataset 4)."
Fig 5E — COCOC, ranked RAE vs detected event count 🟢
左軸: 順位付き RAE (blue line)、右軸: 該当 2D-histogram のイベント数 (orange dots) と 1025-window 幾何平均 (red line)。COCOC ではイベント数と RAE に強い相関なし(情報飽和)。【画像視覚情報】画像で確認: 青い RAE 曲線が 緩やかな S 字状(sigmoid 様)でランクに従って上昇するが、右軸の橙のイベント数ドットは RAE のランク上昇に連動せずほぼ水平に散らばる。赤い 1025-window 幾何平均線も水平に近い。 = イベント数が「RAE を決める主要因」ではない(COCOC は情報飽和済み)。
Fig 5F — CCCOO, ranked RAE vs detected event count 🟢
同様。CCCOO では明確な逆相関(イベント数が多いほど RAE 小 = 予測良)。C-O 接続が 1 本だけなのでイベント数が直接効く。dashed blue = 100 M サンプル学習(dataset 4)で RAE がさらに改善 — データ量を増やせば良くなる典型例。【画像視覚情報】画像で確認: Fig 5E と対比すると 橙ドットが明確に右下がりトレンドを示し、赤い幾何平均線も 右下がりの傾斜 を見せる。さらに、青の dashed line(100 M 学習)は solid line(10 M 学習)より明確に下方にシフト = データ量を 10 倍にすると RAE が体系的に低下。Fig 5E と Fig 5F の対比が「情報飽和済 vs 未飽和」を 1 画面で対照する設計。
PDF Fig 6 caption: "Evaluation of the predicted transition rates of time series with a low signal-to-noise ratio. Datasets No. 2 (SNR=5) and No. 5 (SNR=2) of COCOC topology. (A-C) Excerpts of time series for best-predicted (A), median (B), worst-predicted (C) models, with corresponding 2D-histograms (2D_GT, 2D_Pr) and 2D_Diff. 100 simulations of each predicted model. Volume deviations V_D(G, H_1..H_100) and V_R(H_1..H_100) shown on histograms. (D-F) Distribution of 100 re-predicted rates kij; orange dots = ground truth. (G) Cumulative RAE for SNR=5 (solid) and SNR=2 (dashed); pink dashed = random baseline."
PDF Fig 7 caption: "Transition rates estimation for COCOC models including fast gating rates. k₁₂ to k₄₃ in 0.1–10 ks⁻¹ range; k₄₅, k₅₄ in 10 ks⁻¹–1 Ms⁻¹ (above 10 kHz corner frequency). Regression NN trained on dataset No. 6. (A-C) Time series excerpts for best/median/worst models with 2D-histograms (2D_GT, 2D_Pr, 2D_Diff) and current distributions. 100 re-simulations. (D-G) Scatter plots on test dataset. (D, E) Slow rates k₃₂ and k₃₄ (worst overall). (F, G) Fast rates k₄₅ and k₅₄. Orange/red dashed: RAE = 0.6 / 1.0."
(b) サブパネルごとの解説
Fig 7A — Best fast-gating model 🟢
Fig 6A と同一構造。Fast gating 領域では フリッカリング(高速点滅)が時系列に現れる。\( \bar V_D \approx \bar V_R \) で良予測。電流ヒストグラムは fast gating によって O level が「歪む」(apparent level reduction、ローパスによる平均化)。【画像視覚情報】画像で確認: GT 時系列の特定箇所に 「縦線が密集した黒い帯」が高頻度に現れる = フリッカリング(高速点滅)の視覚的シグネチャ。これが fast gating の指紋。Prediction (Best) も同じ位置に同様の帯を再現。電流ヒストグラムは O level のピーク位置が予想より下にシフトし、ピーク高さも低い(apparent level reduction が明示)が、GT/Pred のヒストグラム形状はほぼ一致。
PDF Fig 8 caption: "Transition rates estimation for data generated with the patch-clamp setup. (A) Experimentally derived step response of patch-clamp setup deviates from simulated 4-pole low-pass Bessel filter. (B) Power spectra of simulated and recorded noise. Blue: bath resistor (10 MΩ), orange: patch resistor (10 GΩ), brown: real cell-attached; red arrows indicate stray noise. Simulated: cyan (Bessel-filtered white), green/red/lime (power-spectrum-based). (C, D) Eight NNs trained with combinations of step response × noise type (Table 1 datasets 7-10 for SNR=4-6; 11-14 for SNR=8-10). Box-whisker on 100 semi-synthetic time series at SNR≈6 (C) and SNR≈8 (D). Kruskal-Wallis + Dunn's test (**p<0.01, ***p<0.001). Orange stripe: NN test set baseline; green stripe: random rates."
Richter-Laskowska et al. (2021) "Application of Machine-Learning Methods to Recognize mitoBK Channels from Different Cell Types Based on the Experimental Patch-Clamp Results" Int. J. Mol. Sci. 22(2):840.
PMID: 33467711 /
DOI: 10.3390/ijms22020840
Almanjahie et al. (2019) "Moving average filtering with deconvolution (MAD) for hidden Markov model with filtering and correlated noise" Eur. Biophys. J. 48(4):383-393.
PMID: 31028435 /
DOI: 10.1007/s00249-019-01368-1
Langthaler et al. (2022) "Ion Channel Modeling beyond State of the Art: A Comparison with a System Theory-Based Model of the Shaker-Related Voltage-Gated Potassium Channel Kv1.1" Cells 11(2):239.
PMID: 35053355 /
DOI: 10.3390/cells11020239