# AIGA DMTA Report

## Summary

- Patients: 21
- Compounds: 2
- Wet-lab observations: 42
- Epitope state distribution: {'KRKR_high': 16, 'low_broad_negative': 4, 'multi_epitope_high': 1}

## Modeling Concept

Initial structure-based prediction prioritized compounds by IFN-gamma pocket binding. Wet-lab validation showed that the best structural hit was not necessarily the strongest biological rescue. Therefore, the prediction objective was revised from compound-only docking score to epitope-state-conditioned pharmacologic rescue.

Untested NTM antibiotics were introduced into the DMTA workflow as structure-prior candidates rather than validated hits. These compounds expand the candidate space for the next rescue cycle, but their predicted effects remain provisional until corrected by patient-level binding inhibition and pSTAT1 rescue assays.

Structure-prior ranking is used to decide which drugs should enter the next wet-lab cycle. It should not be interpreted as direct evidence of pharmacologic rescue.

Current feature formula:

`Response ~ compound + TLFL_ratio + KRKR_ratio + compound x epitope interaction`

## Model Performance

- binding_inhibition_index: elasticnet; metrics: {'ridge': {'MAE': 0.05867612124484664, 'RMSE': 0.0702088468412962, 'Spearman': 0.3784944493963212, 'MAE_bootstrap_95CI': [0.04696153526322822, 0.07140531967571936]}, 'elasticnet': {'MAE': 0.057356678387391793, 'RMSE': 0.06874875622206912, 'Spearman': 0.3742808524430759, 'MAE_bootstrap_95CI': [0.04647398175055708, 0.0694352857653467]}, 'random_forest': {'MAE': 0.0633969073390467, 'RMSE': 0.07470516729681606, 'Spearman': 0.32112470626367395, 'MAE_bootstrap_95CI': [0.05232135199585477, 0.07541503246479621]}}
- pSTAT1_rescue_index: random_forest; metrics: {'ridge': {'MAE': 0.06763670326208973, 'RMSE': 0.0807308810059446, 'Spearman': -0.08961038961038961, 'MAE_bootstrap_95CI': [0.05086182771871492, 0.08923456508615368]}, 'elasticnet': {'MAE': 0.06647177777713856, 'RMSE': 0.07903906839418101, 'Spearman': -0.18311688311688312, 'MAE_bootstrap_95CI': [0.050882608563924465, 0.08722811332555558]}, 'random_forest': {'MAE': 0.06550765496346687, 'RMSE': 0.08131648136316173, 'Spearman': 0.033766233766233764, 'MAE_bootstrap_95CI': [0.047287506838624344, 0.08904398678461199]}}
- responder: logistic_regression; metrics: {'ROC_AUC': None, 'confusion_matrix': None}

## Structure-prior Model Comparison

- structure_only / ridge / binding_inhibition_index: MAE 0.056, RMSE 0.068, Spearman 0.23345283094928276
- structure_only / random_forest / binding_inhibition_index: MAE 0.056, RMSE 0.068, Spearman 0.23345283094928276
- structure_only / ridge / pSTAT1_rescue_index: MAE 0.065, RMSE 0.079, Spearman -1.0
- structure_only / random_forest / pSTAT1_rescue_index: MAE 0.065, RMSE 0.079, Spearman -0.9883116883116884
- epitope_only / ridge / binding_inhibition_index: MAE 0.063, RMSE 0.079, Spearman -0.038603476954022165
- epitope_only / random_forest / binding_inhibition_index: MAE 0.070, RMSE 0.085, Spearman -0.15559983182687645
- epitope_only / ridge / pSTAT1_rescue_index: MAE 0.069, RMSE 0.084, Spearman -0.22337662337662337
- epitope_only / random_forest / pSTAT1_rescue_index: MAE 0.058, RMSE 0.073, Spearman 0.3038961038961039
- epitope_conditioned_drug / ridge / binding_inhibition_index: MAE 0.053, RMSE 0.068, Spearman 0.49209950571266514
- epitope_conditioned_drug / random_forest / binding_inhibition_index: MAE 0.065, RMSE 0.078, Spearman 0.2363665829349324
- epitope_conditioned_drug / ridge / pSTAT1_rescue_index: MAE 0.070, RMSE 0.094, Spearman -0.21818181818181817
- epitope_conditioned_drug / random_forest / pSTAT1_rescue_index: MAE 0.061, RMSE 0.079, Spearman 0.1948051948051948

## Key Features Associated With Rescue

- numeric__TLFL_ratio: importance 0.278
- numeric__KRKR_ratio: importance 0.256
- numeric__TLFL_ratio_x_compound_C004: importance 0.235
- numeric__KRKR_ratio_x_compound_C004: importance 0.230
- numeric__TLFL_ratio_x_compound_C001: importance 0.000
- numeric__KRKR_ratio_x_compound_C001: importance 0.000
- categorical__compound_id_C004: importance 0.000

## Docking Paradox Summary

The workflow keeps docking features in the model but conditions them on patient epitope state.
This allows compounds such as posaconazole to retain strong structural evidence while still being
penalized if wet-lab pSTAT1 rescue is weak for KRKR-high or spreading-like profiles. It also allows
eravacycline-like candidates to be prioritized when biological rescue is stronger than docking alone
would suggest.

## Recommended Next-Round Candidates

- Eravacycline: Acquisition selected Eravacycline for low_broad_negative at 0.1 uM: predicted rescue 0.171, binding 0.121, uncertainty 0.032; 21 pSTAT1 rescue rows available; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Trimethoprim: Acquisition selected Trimethoprim for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.085, uncertainty 0.080; no wet-lab rows for this compound; acquisition is exploratory; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Amikacin: Acquisition selected Amikacin for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.085, uncertainty 0.080; no wet-lab rows for this compound; acquisition is exploratory; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Imipenem: Acquisition selected Imipenem for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.085, uncertainty 0.080; no wet-lab rows for this compound; acquisition is exploratory; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Sulfamethoxazole: Acquisition selected Sulfamethoxazole for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.085, uncertainty 0.080; no wet-lab rows for this compound; acquisition is exploratory; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Azithromycin: Acquisition selected Azithromycin for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.085, uncertainty 0.080; no wet-lab rows for this compound; acquisition is exploratory; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Posaconazole: Acquisition selected Posaconazole for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.062, uncertainty 0.077; binding-only evidence; pSTAT1 rescue should be tested; backend=botorch_gp; binding_uncertainty=botorch_gp.
- Tigecycline: Acquisition selected Tigecycline for low_broad_negative at 0.1 uM: predicted rescue 0.141, binding 0.085, uncertainty 0.080; no wet-lab rows for this compound; acquisition is exploratory; backend=botorch_gp; binding_uncertainty=botorch_gp.

## Structure-prior Candidates for Wet-lab Entry

- Tigecycline: eravacycline-like structural candidate; score 0.56; best state low_broad_negative; uncertainty high
- Doxycycline: tetracycline-class contrast candidate; score 0.25; best state low_broad_negative; uncertainty high
- Moxifloxacin: quinolone-class exploratory candidate; score 0.33; best state low_broad_negative; uncertainty high
- Trimethoprim: low-docking structural control; score 0.00; best state low_broad_negative; uncertainty high

These structure-prior candidates are not treated as training labels. Wet-lab measured
binding inhibition and pSTAT1 rescue must be added to wetlab_results.csv before these
compounds can update model performance metrics.

## Generated Plate Map

C:\Users\cmu\MD_run\AIGA_DMTA_Model\outputs\plate_maps\next_round_plate_map.xlsx