Human ADME/PK is lost in translation and prediction from in silico to in vitro to in vivo

Human ADME/PK is lost in translation and prediction from in silico to in vitro to in vivo | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Human ADME/PK is lost in translation and prediction from in silico to in vitro to in vivo View ORCID Profile Urban Fagerholm , Sven Hellberg doi: https://doi.org/10.1101/2025.02.17.638712 Urban Fagerholm 1 Prosilico AB, Lännavägen 7, SE-141 45 Huddinge , Sweden Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Urban Fagerholm For correspondence: urban.fagerholm{at}prosilico.com Sven Hellberg 1 Prosilico AB, Lännavägen 7, SE-141 45 Huddinge , Sweden Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Preview PDF ABSTRACT Background Measurements and predictions of aqueous solubility (S), apparent cell permeability (P app ), unbound microsomal intrinsic clearance (CL int,u ), unbound fraction in plasma (f u ), log D, Lipinski’s Rule of 5 (Ro5) and BBB+/BBB- (blood-brain-barrier) are commonly used in early drug discovery to evaluate whether compounds are likely to have adequate ADME/PK in humans. The main objective was to evaluate the validity and proposed thresholds for commonly used in silico to in vitro (with particular interest in the ADMET Predictor software) and in vitro to in vivo human ADME/PK prediction methods, and physicochemical estimates and rules of thumb. A secondary aim was to compare validity and thresholds to that for the prediction software ANDROMEDA. Methods Data were collected from literature and own studies. Main measures of validity were Q 2 (true predictive accuracy; in silico models), R 2 (correlation coefficient; in vitro models), Q 2 • R 2 (for translation from in silico to in vivo via in vitro ), skewness, range, % correct predicted class and clinical relevance. Results and Discussion Poor accuracies (Q 2 • R 2 =0.05, 0.05, 0.36 and 0.45; ∼0 at low to moderate, decision-critical levels) and class predictions, limited ranges (not covering low to moderate estimates), systematic errors (often considerable overprediction at low values), poor clinical relevance and inadequate thresholds were found. Predictive accuracy was mainly lost in the translation from in vitro to in vivo . Log D and Ro5 are poor predictors of oral bioavailability and half-life. Ro5 produced 63 % false negatives for prediction of poor/good oral absorption. The overall mean Q 2 for ANDROMEDA was 3 times higher (0.57 vs 0.2). Advantages with ANDROMEDA compared to in vitro data-based in silico models include wider application domain (high resolution at low values), extrahepatic elimination models, minimal skewness, clinical relevance, balanced thresholds, and compound- and parameter-unique confidence intervals. ANDROMEDA successfully predicted the human clinical ADME/PK for all small drugs marketed in 2021, while the in silico to in vitro to in vivo approach was out of reach for all. Conclusion The validity of investigated methodologies (including ADMET Predictor) and thresholds were overall very low. Unless both predicted S, permeability and f u are high and CL int is moderate (an overall criterium not met for investigated modern small drugs) one is more or less lost in the translation, and will jeopardize compound selection and optimization. This clearly shows the need for better and thoroughly validated prediction models and software. Marked improvements in accuracy, range, balance and clinical relevance were achieved with ANDROMEDA, which predicts human clinical ADME/PK directly from chemical structures and has undergone extensive validation. INTRODUCTION In vitro data, physicochemical parameter estimates and rules of thumb, and classification systems, are commonly used in early drug discovery to predict and evaluate whether compounds are likely to have adequate or unfavorable ADME/PK properties in humans. Examples of in vitro and physicochemical parameters include aqueous solubility (S), apparent Caco-2 and MDCK permeability (P app ), unbound microsomal intrinsic clearance (CL int,u ), unbound fraction in plasma (f u ), log P and D (typically at pH 7.4), molecular weight (MW), and number of hydrogen bonding acceptors (HBA) and donors (HBD). Commercially available software for ADME/PK predictions (mainly in silico to in vitro ) include ADMET Predictor by Simulations Plus Inc. and QikProp by Schrödinger. Physicochemical rules of thumb include proposed ideal (for ADME/PK) log D and log P of 0-3 (or 2-3) and 1-5 [ 1 – 3 ], respectively, S>10-200 µM for acceptable in vivo solubility/dissolution [ 4 , 5 ], low microsomal CL int,u (high metabolic stability; 500 g/mol, log P>5, >10 HBA and >5 HBD) for acceptable gastrointestinal uptake and oral bioavailability [ 8 ]. Classification systems include BCS (Biopharmaceutics Classification System; 4 classes depending on high and low S and permeability) and BBB+/BBB- (BBB=blood-brain barrier). In the BCS, compounds in Classes II and III have, or are predicted to have, solubility/dissolution rate-limited and permeability-limited gastrointestinal uptake in humans in vivo , respectively, whereas compounds of Class I possess, or are predicted to have, good/sufficient solubility/dissolution and permeability, and Class IV-compounds have, or are predicted to have, gastrointestinal uptake limited by both solubility/dissolution and permeability [ 9 ]. BCS-classing is commonly done based on measured S at various pH-levels) and P app -estimates, and criteria/thresholds for these. According to BBB+/BBB-, compounds belonging to BBB+ (proposed to have high log equilibrium brain/blood concentration ratio (log BB; various thresholds suggested)) are believed to easily cross the BBB and enter the brain rapidly (sufficient brain uptake and distribution is essential for centrally acting compounds), whereas BBB- compounds are anticipated/predicted to not cross the BBB or slowly and poorly cross the BBB and enter the brain (favorable if no central side effects are wanted). In silico prediction models have been developed for, for example, log D, log S, log P app , log human effective permeability (P eff ), fraction absorbed in vivo (f abs ), microsomal log CL int,u , lin and log f u and log BB (BBB+/BBB-) [ 10 – 17 ], which enables ADME/PK predictions for compound-selection and -optimization, and decision making even before compound synthesis. Sufficient validity (good accuracy, no or minor systematic trends, wide range and high clinical relevance), reliability (high reproducibility), and adequate proposed thresholds are required for in silico and in vitro prediction methods and adequate compound selection. To our knowledge, an evaluation of this for commonly used in silico and in vitro methods, physicochemical parameters (including proposed rules of thumb) and classifications has not been done before. Based on the costs and time spent and failures in drug discovery and development, and wish/requirement to replace, reduce and refine animal experiments (3R) and to assure safety in clinical trials, this is surprising. Lack of concern by users, widespread perception that it is an area with sufficiently accurate methodology, unwillingness by model and software developers to disclose poor results and clinical relevance, and lack of interest and demands from regulatory agencies, are among possible explanations. Another approach (than in vitro data-based in silico prediction models and physicochemical parameter estimates) is in silico ADME/PK-models based on human clinical parameters and estimates. Our team (Prosilico AB, Sweden) has developed and validated such a prediction, simulation and optimization software, ANDROMEDA ( https://prosilico.com/andromeda ). ANDROMEDA is built based on a unique, large data bank, unique algorithms, and conformal prediction and machine learning methods. It produces valid levels of confidence and reaches an average Q 2 (correlation between measured and predicted estimates in cross-validations) of 0.52 for its numerical models and parameters within the major domain of MW 150-750 g/mol. Its in silico models for permeability-based fraction absorbed (f abs,pe ; range 0-100 %), fraction dissolved in vivo (f diss ; range 1-100 %), log CL int (range 1-200,000 mL/min) in vivo in man, and in vitro log f u (f u -range 0.02-100 %), have minimal skewness and Q 2 -values of 0.7-0.8, 0.55, 0.55 and 0.6-0.7 (0.5-0.55 for lin-lin scale), respectively [ 18 – 21 ]. How well other established in silico -to- in vitro - to- in vivo prediction approaches (as above) perform compared to ANDROMEDA (directly from in silico to in vivo in man, without using in vitro data) has not been fully investigated and shown. The main objective of this study was to evaluate the validity and proposed thresholds for commonly used in silico to in vitro and in vitro to in vivo methods, and physicochemical estimates and rules of thumb, for prediction of human clinical ADME/PK. Results obtained with the commercially available software ADMET Predictor by Simulations Plus Inc. were of particular interest. A secondary aim was to compare the validity and thresholds to that for ANDROMEDA. METHODS Sources of Data The following references were used: log D [10, Prosilico data on file], log S and f diss [10, 13, 14, 22-25, Prosilico data on file], log Caco-2 P app [ 10 , 15 , 16 , 23 , 25 , 26 – 30 ], log MDCK P app [ 31 ], log P eff (ADMET Predictor predicts P eff and not P app ) [ 10 , 31 , 32 ], log microsomal CL int,u [6, 7, 10, 30, 31-35, Prosilico data on file], f u (13, 20), Ro5 [ 36 ], log BB and BBB+/BBB- [ 37 – 39 ], and in vivo f abs [ 40 ]. In silico to in vitro ADME/PK results obtained using ADMET Predictor by Simulations Plus Inc. were taken from references [ 10 ], [ 14 ] and [ 17 ]. Data Analysis Measures of validity were Q 2 (accuracy; correlation between measured and predicted estimates in cross-validations, for test or hold-out sets), mean absolute error (MAE), root mean square error (RMSE), skewness (Q 2 and intercepts on axis for observed/measured values), range (minimum to maximum observed and predicted values), % correct predicted class and clinical relevance (including analysis of similarities and differences between in vitro and in vivo parameters and conditions, as well as levels of accuracy). The predictive performance of in vitro methods is often demonstrated as R 2 -values (correlation coefficient for measured vs predicted estimates; retrospective/non-prospective), where values <limit of quantification (LOQ) have been excluded. Proposed thresholds and optimal ranges and limits of quantification for in vitro and physicochemical parameters were compared to corresponding apparent in vivo values. The translation from in silico to humans in vivo (Q 2 ), via in vitro (R 2 ), was done by multiplying the both estimates (Q 2 • R 2 ) for each parameter (log S to in vivo f diss , log P eff , via log P app , to in vivo f abs , and in vitro to in vivo CL int ). Results were compared to those obtained with ANDROMEDA for corresponding parameters. RESULTS In silico to in vitro Predictions Log D For a test set of 287 compounds with log D ranging from −3 to 4 (no data in the interesting and challenging region log D>4), ADMET Predictor produced Q 2 , MAE and RMSE for in silico vs in vitro log D of 0.79, 0.59 and 0.76, respectively [ 10 ]. Ca 3 % of compounds had 1.5- to 2-fold prediction errors. Solubility (S) The Q 2 , MAE and RMSE for in silico vs in vitro log S with ADMET Predictor (for a test set of 691 compounds; no or few with log D>4) were 0.20, 1.06 and 1.23, respectively [ 10 ]. Using a 4-class system (low/poor, moderate, high, very high), S-classification was correctly predicted for 56 % of compounds. Ca 95 % of low S compounds according to measurements (can be interpreted as undesirable/unacceptable S; 90 µM), whereas only ca 5 % of compounds with very high S were predicted to have low to high S ( Figure 1 ). Ca 8 % of compounds with predicted very high S (desirable/acceptable S) had low measured S (undesirable/unacceptable S) and every other compound predicted low S did not have low measured S. Ca 67 % of compounds with very high predicted S also had very high measured S. Download figure Open in new tab Figure 1. Percentage correct in silico predicted class (low and very high) for in vitro S, P app & P eff and CL int,u for ADMET Predictor) [ 10 ]. Chevillard et al. [ 14 ] also used ADMET Predictor, and reached a Q 2 of 0.32 and a RMSE of 1.25 for a test set from the Solubility Challenge (n=26; lowest S ca 0.1 µM). They also tested 10 other methods/software, including QikProp, and found Q 2 - and RMSE-values of 0.18-0.57 (0.34 for QikProp) and 1.09-1.66 (1.24 for QikProp) [ 13 ]. Corresponding results for another dataset (n=62) were 0.13-0.62 (0.56 for ADMET Predictor and 0.45 for QikProp) and 0.51-0.95 (0.72 for ADMET Predictor and 0.86 for QikProp). In a study by Falcón-Cano et al., Q 2 of 0.4-0.8 and MAE of 0.6-0.8 were reached [ 14 ]. The Q 2 for in silico to in vitro log S was estimated to average ca 0.4 ( Figure 2 ). Download figure Open in new tab Figure 2. Q 2 (true predictive accuracy; in silico models), R 2 (correlation coefficient; in vitro models) and Q 2 • R 2 (for translation from in silico to in vivo via in vitro ) for S to in vivo f diss , log P eff to in vivo f abs and microsomal CL int,u to in vivo CL int , Q 2 for in silico prediction of f u and F (both ADMET Predictor), and predicted log D vs F. *Values for log D and ADMET Predictor were similar (ca 0.1) and combined. Note: LOQ is a limitation for in vitro S, microsomal CL int,u and f u , and therefore, many compounds with low to moderate estimates are missing and in silico prediction models for these show overprediction trends in these regions. Q 2 , R 2 and Q 2 • R 2 for in vitro based models are much lower at low values . Apparent Caco-2 permeability (P app ) and effective human permeability (P eff ) The % correct permeability classing and Q 2 for in silico P eff vs in vitro log P app were 61 % and 0.61 (n=516) for a local model built by [ 10 ], and 47 % and unknown (roughly estimated to 40-50% based on the difference in correctly predicted permeability class) for ADMET Predictor [ 10 ], respectively. For a smaller set of commercially available compounds (n=26) the Q 2 was 0.78 [ 10 ]. In cannot be ruled out that this set, or parts of it, was included in the training set for the model, and that these results, therefore, were exaggerated. Winiwarter et al. also reached Q 2 =0.60 [ 16 ]. Q 2 -values from 7 other studies were 0.16-0.74 [ 15 ]. The overall Q 2 from the 10 studies was ca 0.6 ( Figure 2 ). The impact of test compounds also present in training sets is unknown. The RMSE was estimated to 0.6-0.8, which corresponds to ca ±15-20 %, ±20 % and ±10 % variability of in vivo f abs at P app of ca 0.1, 1 and 10 • 10 -6 cm/s ( in vivo f abs ca 30, 60 and 90 %, respectively [ 23 ]) [ 15 ]). At P app corresponding to incomplete in vivo f abs (<80-90 % f abs ), there was virtually no correlation between measured and in silico predicted P app (R 2 <0.1) [ 15 ]. For example, there were compounds with in vitro and in silico predicted f abs of ca 80-90 % vs ca 10-15 % (7-fold difference) and ca 5-10 % vs 60 % (8-fold difference), respectively [ 15 ]. In 53 % of cases, incorrect in vitro P app -class was predicted using ADMET Predictor [ 10 ]. Ca 30 % of low/poor P eff compounds (2 • 10 -6 cm/s; desirable/acceptable permeability), and none of the compounds with predicted very high P eff (>3 • 10 -4 cm/s; desirable/acceptable permeability) had low measured P app (<2 • 10 -6 cm/s; undesirable/unacceptable permeability) ( Figure 1 ). Ca 70 % of compounds with predicted low P eff had low measured P app , ca 13 % of compounds with predicted low P eff had very high measured P app (>10 • 10 -6 cm/s), and ca 67 % of compounds with predicted very high P eff also had very high measured P app . Human microsomal unbound intrinsic clearance (CL int,u ) ADMET Predictor produced a Q 2 of 0.50 for in silico vs in vitro log CL int,u for a test set consisting of 1199 compounds [ 10 ]. 37 % of predicted classifications were correct. Ca 68 % of very high CL int,u compounds (undesirable/unacceptable CL int,u ; >80 µL/min/mg) were predicted to have low to high CL int,u (desirable/acceptable CL int,u ), ca 18 % of compounds with predicted low CL int,u (desirable/acceptable CL int,u ; <10 µL/min/mg) had very high measured CL int,u (undesirable/ unacceptable CL int,u ), and ca 26 % of compounds with predicted very high CL int,u (undesirable/unacceptable CL int,u ) had low to high measured CL int,u (desirable/acceptable CL int,u ) ( Figure 1 ). In studies by Shah et al. [ 6 ] (2 classes; low (ca <40-50 µL/min/mg) and high microsomal CL int,u ) and Podlewska and Kafel [ 7 ] (3 classes; low (ca 40-60 µL/min/mg)), 80 and 70 % correct class predictions were achieved, respectively. ADMET Predictor was also used to predict in vitro microsomal log CL int,u for a set of 75 commercial reference compounds, and a Q 2 of 0.14 was reached [ 10 ]. The average Q 2 for ADMET Predictor for the two test sets was 0.32 ( Figure 2 ). Unbound fraction in plasma (f u ) Yun et al. used 3 different in silico methods, ADMET Predictor and models by Watanabe et al. and Ingle et al., to predict the in vitro f u of 818 compounds with f u ranging from 1 to 100 % (those with f u <1 %, which represent just over 10 % of all compounds and are extra challenging to measure and predict, were excluded) [ 17 ]. For models by Watanabe et al and Ingle et al, no test set compounds were used in the model training sets. This was, however, not assured for ADMET Predictor (implies risk of exaggerated outcome). Q 2 -values (for linear predicted vs measured f u ) for ADMET Predictor and models by Watanabe et al. and Ingle et al. were 0.52, 0.46 and 0.37, respectively [ 17 ]. Corresponding MAE- values were 13, 14 and 16 %, respectively. Clear trends were found. There was general overprediction at low (1 to 50-fold individual underpredictions), and at 90- 100 % f u there was on average ca 2-fold underprediction. In vitro to in vivo Predictions Solubility (S) to in vivo dissolution (f diss ) An investigation of the relationships between log S and in vivo f diss and f abs showed that these two measurements correlate poorly: R 2 =0.13 for log S vs in vivo f diss (n=82) and R 2 =0.00 for log S vs in vivo f abs (n=452) ( Figure 3 ), respectively [ 22 , 25 ]). The R 2 -estimates for S and dose-corrected log S vs in vivo f abs were also zero. Download figure Open in new tab Figure 3. The correlation between measured log aqueous solubility (S) and fraction absorbed in vivo (f abs ) for 452 compounds with log S ranging from −4.0 to 6.0 and f abs ranging from ca 0 to 100 %. For 70 compounds with S<5 mg/L (ca <10-20 µM) and 17 compounds with S<1 mg/L (ca <2-4 µM) there were weak correlations between log S and in vivo f abs (R 2 =0.07 and 0.02, respectively) [ 25 ]. Complete or near complete f abs has been demonstrated for many compounds with S∼2-4 µM, and moderate uptake has been shown for compounds with S∼0.2-0.4 µM [ 23 , 25 ]. Out of 129 compounds of the Solubility Challenge data set [ 24 ], 27 have S in the 1-100 µM- range, and only one of these have an in vivo f abs clearly limited/determined by solubility/dissolution [ 25 ]. The level of S corresponding to general solubility/dissolution-limited uptake in vivo in man (overall median <1 µM; individual <1-100 µM; approximated average S at f diss =80 % [ 25 ]) is low compared to proposed thresholds proposed (<10-200 µM [ 4 , 5 ]; 10- and >2-fold average and median ratio between maximum and minimum reported estimates, respectively) [ 25 ]. For example, there are 8400- and 4500-fold differences for highest and lowest reported S for dipyramidole and diclofenac (low to very high S for both), respectively [ 23 ]. For some low solubility drugs (at least 30 found), S has not been possible to quantify [ 41 ], including the well absorbed (high f abs ) amlodipine, aripiprazole, bicalutamide, conivaptan, lomustine, posaconazole, venlafaxine and zolpidem, and poor f diss -compounds atavaquone, artemether, lumefantrine and paricalcitol. Thus, it has not been possible to predict their f abs using S-data and establish in vitro-in vivo relationships along the whole S-scale. Almost every other compound with a non-quantifiable S, in vivo dissolution is or is predicted to be complete or near complete, and there at least over a handful of drugs with S>90 µM (very high in vitro S) are known to have incomplete in vivo dissolution. In an investigation, 77, 84 and 100 % of compounds (n=73) predicted to be in BCS classes I (high P app - high S), III (low P app - high S) and IV (low P app - low S) based on in vitro data also belonged to these classes based on their measured in vivo f abs , respectively [ 25 ]. For compounds that were predicted to belong to BCS class II (high P app - low S), however, only 31 % belonged to in vivo BCS class II (69 % had an in vivo f a of 90-100 %; BCS class I). For compounds that have BCS-classification from different sources (n=140), 64 % show contradictory classifications. Permeability to fraction absorbed (f abs ) The R 2 between log Caco-2 P app and in vivo f abs ranges between ca 0.2 and 0.8 (average ca 0.6; Figure 2 ), depending on compound choice, f abs -range and laboratory [ 26 – 30 ]. Limitations with in vitro permeability assays include greater uncertainty and quantification problems for low permeability compounds, low recovery (for example, for highly lipophilic compounds) and variability among labs. Skolnik et al. [ 26 ] found that 1/8 and 50 % of compounds were subject to <30 % (44 % of poorly soluble compounds) and 100 compounds that needed to excluded from evaluation of log Caco-2 P app vs in vivo f abs -relationships for such reasons [ 41 ]. According to an evaluation and summary by Pham-Tee et al. [ 23 ], Caco-2 values corresponding to low f abs (<30- 40 %) are ≤0.1 • 10 -6 cm/s, and this is considerably lower than the threshold for low P app (<2 • 10 - 6 cm/s; corresponding f abs ca <65 %) set by Sohlenius-Sternbeck and Terelius who evaluated ADMET Predictor [ 10 ]. The method for estimation of in vivo P eff has difficulties to accurately quantify for compounds with f abs <ca 70-90 % (similar as the f abs of the average drug) and P eff ca <1 • 10 -4 cm/s [ 31 , 32 ], which limits it for applicability to predict complete (ca 90-100 %) or incomplete uptake (ca <70-90 %). In an evaluation of MDCK cell P app -based predictions of P eff and human in vivo f abs (similar to the ADMET Predictor approach) (n=46; all except one with apparent permeability limited uptake), a R 2 of 0.43 and an intercept of 39 % f abs were reached, respectively [ 31 ]. A considerable portion of the compounds (nearly 40 %) of test compounds were, however, already used in the training set, and therefore, the accuracy was most likely exaggerated. Intrinsic clearance (CL int ) Our previous investigation showed that the R 2 for predicted and observed in vivo log CL int for human hepatocyte CL int,u -based predictions was 0.3-0.4 [ 21 , 25 ]. Stringer et al. [ 33 ] found a R 2 that appeared to be lower than that, both for hepatocyte and (in particular) microsome CL int,u - based predictions. Using a test set of 73 drugs, we found that the R 2 between predicted and observed in vivo log CL int for human microsome CL int,u -based predictions was 0.18 ( Figure 2 ) [ 30 ]. In these microsomal experiments, drugs such as naloxone, lidocaine, glipizide and ondansetron with moderate to moderately high in vivo CL int ( in vivo CL int corresponds to in vitro CL int,u ) were found to have in vitro CL int,u <LOQ. The maximum prediction error was 350-fold (overprediction). In vivo CL int threshold for proposed low CL int ranged between 15 and 1,900 mL/min. The average and maximum intra-laboratory variability were 6.7- and 58-fold [ 10 , 30 , 34 ], respectively. Every other compound had individual microsome CL int,u -measurements in more than one class (at least two measurements were done for each compound), and 10 % of them had individual CL int,u - measurements spanning over 2-3 classes (one compound, alfentanil, had low and very high CL int,u on different occasions). Thus, the reproducibility of the microsome assay is limited. For every other compound in vitro and in vivo CL int -classes differed [ 10 , 30 ]. Ca 30 % of compounds with low measured microsomal CL int,u had low in vivo CL int , and 67 % of compounds with low in vivo CL int had low in vitro CL int,u . which shows an overprediction trend. By adding human microsome CL int,u for 7 OATP-substrate drugs (with available in vivo CL int data) from [ 35 ] to the set of 73 compounds, the R 2 decreased to 0.11 [ 30 ]. The in vivo CL int for OATP-substrates was on average more than 10-fold higher than predicted from microsome CL int,u [ 35 ]. Due to LOQ-limitations, a substantial fraction of compounds is normally excluded from such correlations and evaluations [ 25 , 33 , 41 ]. Stringer et al. [ 33 ] found that none of their selected test compounds with in vivo CL int 7,000 mL/min had quantifiable microsomal in vitro CL int,u , respectively [ 33 ]. Thus, there is an approximately 50 % chance that a compound with in vitro microsomal CL int <LOQ has high to very high in vivo CL int . The estimated median in vivo CL int and our proposed limit for low CL int for drugs are 850 and 150 mL/min, respectively. The corresponding limit for low CL int that was set by Sohlenius-Sternbeck and Terelius (ca 2,000 mL/min) [ 10 ] is about 3- and 13-fold higher than these, respectively. Corresponding thresholds for very high CL int are ca 20,000 and 50,000 mL/min (2.5- fold lower in vitro [ 10 ]), respectively. Thus, limits proposed for microsome CL int,u are relatively high at low CL int and low at high CL int . For metabolically unstable compounds, in vitro t ½ could be 100 µL/min/mg), which implies higher uncertainty of measured and predicted CL int . Among top 5 % CL H -drugs, less than 1/3 belong to those with highest CL int , showing that also f u (and CL int • f u ) are important parameters for metabolic stability in vivo . For those with the highest estimated CL H , none belong to the very high CL int -class. Peak CL H is found at CL int =24,000 mL/min, which is about 1/10 of the maximum estimated in vivo CL int . The mean fraction excreted renally (f e ) at in vivo CL int of 150 (our proposed threshold for low CL int ), 850 (median CL int for drugs), 2,000 (approximate threshold for low CL int according to [ 10 ]), 5,000 (approximate threshold for very high CL int according to [ 10 ]) and 50,000 (our proposed threshold for very high CL int ) mL/min are 54, 27, 19 (range 0-85 %), 13 (range 0-70 %) and 2 %, respectively. Interlaboratory variability for unbound microsomal fraction (f u,mic ), which is used to convert CL int to CL int,u , is also a source of uncertainty for CL int -predictions. For example, ratios between highest and lowest reported f u,mic -estimates for amitriptyline and imipramine are 5- to 6-fold. In silico prediction models for f u,mic reach R 2 of ca 0.6, with 1.5 average-fold error and some cases with 10- to 100-fold prediction errors [ 31 ]. This sets the upper accuracy limit for CL int,u - predictions. Blood-brain barrier permeability and brain uptake An evaluation of BBB permeability and uptake demonstrated that this barrier is highly permeable, sufficient to absorb compounds with MW>1000 g/mol, log D270 Å 2 well [ 37 ]. The in vitro BBB P app is higher than in Caco-2 cells [ 38 ], on average 11-fold and maximally 34-fold higher. Brain uptake index-data from rats in vivo showed rapid and extensive brain uptake of high permeability compounds during a very short experimental period and some absorption of low permeability substances [ 37 ]. For example, antipyrine, caffeine, nicotine and propranolol were absorbed to ∼70-100% within 5-15 s, and the uptake of hydrocortisone and sucrose was 1.4 % [ 37 ]. The BBB passage times for different molecules (from the low permeability compound sucrose to the highly permeable propranolol) in capillary endothelial BBB cells were approximated to ∼0.1 to ∼4 s [ 37 ]. Corresponding in vivo estimates obtained in rats in vivo were ∼0.3 s to ∼12 min (0.2 s, 7 s and 50 min for caffeine, morphine, and inulin (MW∼5000 g/mol), respectively). In humans in vivo , many highly permeable substances, including anaesthetics and nicotine, have a very rapid on-set of CNS effects (in the order of seconds) following injection or inhalation. Morphine, with moderate passive BBB permeability, significant BBB efflux (by MDR-1; ratio between unbound brain and plasma concentrations (K p,uu,brain ) in rats = 0.15) and low binding capacity to brain tissue has an on-set time of 5-10 min. The transport through the gut-wall, rather than the passage across the BBB, appears to be rate-limiting. Remifentanil has a blood-brain equilibration t ½ of 1 min and a rapid onset of action (general anesthesia) in humans. K p,uu,brain -estimates obtained in rats and humans differ. A 0.01 correlation between the two was found [ 42 ]. Among reasons to the poor relationship include the relatively high expression of mdr-1 and low expression of bcrp in rats (compared to human MDR-1 and BCRP), 15 % difference in MDR-1/mdr-1 homology between humans and rats, and overall higher plasma f u in in rats [ 42 ]. One compound with high K p,uu,brain in man (2.8) is effluxed and predicted to have high passive permeability, but has no apparent efflux at the rat BBB. A MDR-1 substrate with high K p,uu,brain in rat (2.4) has very low K p,uu,brain in man (0.15) and is predicted to have high passive permeability. A high log BB is (according to the BBB+BBB- concept) indicative of high BBB permeability and CNS-activity potential. There are, however, CNS-active compounds with very low log BB in rats (for example, log BB<2 (BB<0.01)) [ 37 ]. In silico and Physicochemical Properties to in vivo Predictions Log D to oral bioavailability (F) and half-life (t ½ ) Figures 4 and 5 show the correlations between measured log D and F (n=212; R 2 =0.23; corresponding R 2 for predicted log D vs F=0.11) and log t ½ (n=423; R 2 =0.065; corresponding R 2 for predicted log D vs log t ½ =0.035), respectively. Note that F- and/or log D-values for many highly lipophilic compounds are not available. Download figure Open in new tab Figure 4. The correlation between measured log D and oral bioavailability (F) for 212 compounds with log D ranging from −3.2 to 4.8 and F ranging from ca 0 to ca 100 %. For predicted log D vs F the R 2 was 0.11. Download figure Open in new tab Figure 5. The correlation between measured log D and log half-life (t ½ ) for 423 compounds with log D ranging from −3.9 to 5.4 and t ½ ranging from 4 min to 35 days. For predicted log D vs log t½ the R 2 was 0.035. According to the fitted relationship between log D and F, the optimal log D for achieving the highest possible F is ca 0.5. At log D 0-1 there are, however, compounds with very low F (70 % F has been reached for compounds with log D from −2.5 to 4.7. At log D 0-1 estimated t ½ ranges from 25 min to 35 days. Rule of 5 (Ro5) and fraction absorbed (f abs ) Among a collection of 129 Ro5 violent compounds (2 or more of the rule criteria violated) we found 59 with available f abs -values. The mean f abs for these was 42 % (compared to ca 80 % for compounds with Ro5 acceptance) [ 36 ]. Only 37 % of these had poor f abs (65 %. Out of 77 selected compounds in the study that do not violate more than 1 Ro5 criteria 3 (4 %) were found to be false negatives (having f abs <10 %). Suenderhauf et al. built in silico models for the prediction (rather fitting, since test set was apparently missing) of f abs , and reached a R 2 of 0.6 and a RMSE of 26 % (n=458) [ 40 ]. The intercept for observed vs predicted f abs deviated, however, markedly from zero, and was as high as ca 50 % f abs . Ca 20 % of compounds with f abs <10 % were predicted to have f abs 50 %. 10 compounds with predicted 90-98% f abs are poorly absorbed in vivo (including 2 with f abs <5 %), showing risk of selecting compounds with unacceptable uptake properties. With class models (only f abs 80 %), the Receiver Operating Characteristics (ROC) averaged ca 0.7 [ 40 ]. Numerical models by other groups reached R 2 (no forward-looking predictions) and RMSE of 0.7-0.9 and 14-35 %, respectively. With ADMET Predictor (where f abs is predicted via P eff , and many of the test set compounds probably were present in the training set), a R 2 of 0.1 and an intercept of roughly 40 % f abs (overprediction trend at low f abs ) were reached (n=several hundred; data on file). This R 2 is of same size as found for in silico predictions of F with ADMET Predictor (R 2 ca 0.15 for compounds with complete absorption, and likely not more than half of that if including incompletely absorbed compounds) with this software [ 19 ]. In silico to in vitro to in vivo Predictions and Thresholds Solubility to in vivo dissolution (f diss ) The Q 2 for in silico vs in vitro log S (including ADMET Predictor) and R 2 for log S vs in vivo f diss were ca 0.4 (range 0.13-0.8) and 0.13, respectively ( Figure 2 ). The corresponding in silico-in vitro Q 2 - in vitro-in vivo R 2 product is 0.05. The Q 2 • R 2 applies within the moderately low to very high S range. Sufficient information for determination of Q 2 • R 2 for compounds with low S (<1- 10 µM) and incomplete f diss is lacking. An overprediction trend at low S (ca 95 % of low S compounds were in silico predicted to have moderate to very high in vitro S with ADMET Predictor [ 10 ]), lack of many S-values due to LOQ-limitation, and a relatively high set in vitro threshold for poor/low S (apparently, roughly at least 15-fold higher than found for in vivo f diss ; 10 vs 0.5-1 µM) causes uncertainty for prediction of in vivo f diss from S. For compounds with predicted very high S, ca 67 % have very high measured S, and ca 8 % have poor S (<10 µM) (according to ADMET Predictor results by [ 10 ]), and at the same time, 10 µM often consistent with complete in vivo f diss . Permeability to fraction absorbed (f abs ) The overall Q 2 for in silico vs in vitro log permeability and R 2 for log P app vs in vivo f diss were both ca 0.6 (range ca 0.2-0.8) ( Figure 2 ). The corresponding in silico-in vitro Q 2 - in vitro-in vivo R 2 product is 0.36. The Q 2 • R 2 applies within the moderately high to very high permeability and f abs range (>80-90 % f abs ). At low and moderate permeability and f abs (<80-90 %), close to zero Q 2 (and Q 2 • R 2 ) was found. Note: It is possible that compounds without in vitro P app due to low experimental recovery were excluded . Corresponding in silico vs in vitro Q 2 and Q 2 • R 2 for ADMET Predictor, where P eff is an intermediate between P app and f abs , was lower, ca 0.45 and ca 0.25 (roughly estimated), respectively [ 10 ]. With a similar approach (MDCK P app to in vivo f abs via P eff ), 0.43 and ca 0.2, respectively [ 31 ]. Using ADMET Predictor, ca 70 % of compounds with predicted (proposed) low P eff (<1 • 10 -4 cm/s; corresponds to ca <90 % f abs [ 32 ]) had measured P app <2 • 10 -6 cm/s (which corresponds to ca <65 % according to [ 23 ]). P eff - and P app -thresholds for low in vivo permeability are <0.1-0.3 • 10 -4 cm/s (<ca 30-60 % f abs [ 32 ] and <0.1 • 10 -6 cm/s (f abs <ca 30-40 % [ 23 ]) rather than the 3- to 10- and 20-fold higher limits proposed (by [ 10 ]), respectively. Ca 13 % of compounds with predicted P eff 10 • 10 -6 cm/s), and ca 67 % of compounds with predicted very high P eff also had very high measured P app . None of the compounds with predicted very high P eff (>3 • 10 -4 cm/s) had measured P app <2 • 10 -6 cm/s). Intrinsic clearance (CL int ) The Q 2 for in silico vs in vitro log CL int using ADMET Predictor and R 2 for microsomal in vitro CL int vs in vivo log CL int were 0.32 (mean of 0.50 and 0.14) and 0.15 (based on R 2 values with and without OATP-substrates from different studies, excluding compounds with CL int,u LOQ). Metabolic stability in vivo (e.g. CL H ) is also determined by f u . Based on prediction accuracies for both CL int (0.15 for in vitro to in vivo ; 0.05 for in silico to in vitro to in vivo ) and f u (0.45 for in silico to in vitro ), CL H is also likely to be poorly predicted. The corresponding in vivo CL int for the proposed low microsomal CL int (<15 µL/min/mg) is approximately 3 times higher than for the typical marketed drug (2,000 vs 850 mL/min), and considered moderately low to moderate. What was considered very high and undesirable for microsome CL int is ca 2.5-fold lower than the proposed threshold for the in vivo situation (ca 20,000 vs 50,000 mL/min). Ca 18 % of compounds with in silico predicted proposed low CL int had proposed very high and undesirable measured in vitro CL int , and the percentage in the opposite direction was ca 26. In contrast, there was an underprediction trend at low CL int for in vitro to in vivo as only ca 30 % of compounds with low microsomal CL int had low in vivo CL int . DISCUSSION In this investigation it is apparent that commonly used early discovery ADME/PK numerical and classification models (including the ADMET Predictor software) and rules of thumb have poor validity (poor accuracy, limited range, systematic errors, very low/no clinical relevance) and inadequate thresholds for poor/acceptable values. Poor reliability of in silico to in vitro to in vivo and in silico to in vivo predictions was shown by low accuracy (Q 2 • R 2 ) for log S to f diss (0.05), log permeability to f abs (0.36; 0.25 for the ADMET Predictor approach that includes the intermediate P eff -step), log CL int,u to log CL int (0.05), log CL H (no number estimated, but apparently, very low), log D vs F (0.11) and log t ½ (0.035). At low to moderate, often decision-critical, levels, Q 2 • R 2 -estimates were ∼0. The low accuracies and limited ranges for CL int and f abs , and skewness and limited ranges for S- and f u -models, indicate that CL H , CL, F and t ½ must also be poorly predicted. Note: The assumption that Q 2 • R 2 correctly reflects the true in silico to in vivo predictive power might not be accurate. True values might be somewhat lower or higher . The average Q 2 for in silico to in vitro log S, log P app and log CL int was estimated to 0.44, which was higher than the corresponding mean R 2 for in vitro to in vivo (0.26). Thus, most of the predictive accuracy from in silico to in vivo was lost in the translation between in vitro and in vivo . Furthermore, application/prediction range is also lost with in vitro data. For example, low S and f u and low to moderate (sometimes up to high) microsomal CL int is often not quantifiable. Thus, unless predicted S, permeability and f u are high and CL int is moderate one is more or less lost in the translation and predictions and will jeopardize compound selection and optimization. Apparently, none of 28 modern small drugs (based on drugs and data from 2021) falls into this character [ 18 ]. Eighteen % of them did not meet any of these criteria, and nearly half of them only reached maximum one criterium ( Figure 6 ) [ 18 ]. The typical (based on median observed or predicted values for small drugs marketed in 2021 [ 18 ]) modern small drugs has a f diss of 95 % (60 % with in vivo dissolution limitation), f abs of 70 % (83 % with in vivo f abs <90 %), f u of 4 % (52 % with f u <5 %) and CL int of 2,500 mL/min (45 % with very high in vitro CL int,u according to previously established limit) [ 18 ]. It is therefore highly probable that at least one of in silico predictions of S, P app , CL int,u and f u for modern small drug candidates will be very poor (uncertain, overpredicted and/or underpredicted). This will likely lead to erroneous interpretations and decisions and shows the need for other better tools. The findings and/or predictions that significant renal excretion, gut-wall extraction and/or bile excretion are involved for 30-55 % of the new drugs show additional obstacles and needs [ 18 ]. Download figure Open in new tab Figure 6. Estimated coverage (prediction ranges for solubility/dissolution, permeability/absorption, metabolic instability and unbound fraction in plasma) of in silico to in vitro to in vivo (for example, ADMET Predictor) and ANDROMEDA ( in silico to in vivo ) prediction methodologies for 28 modern small drugs (marketed in 2021). Due to poor to moderately high R 2 and Q 2 , there were also quite high percentages of incorrect predicted class, 43 to 63 % incorrect for in silico to in vitro log S (in particular for low S compounds), f abs (in particular for high P app compounds) and CL int ( Figure 1 ). Thresholds for poor S (low f diss ), poor permeability (low f abs ) and undesirably high CL int are lower (at least ca 15-fold for S), lower (ca 10-fold for P eff and ca 20-fold for P app ) and higher (ca 2.5-fold for CL int ) according to our in vivo data based proposals. An adjustment to these is assumed to lead to better decision-making (stop/go). Higher accuracy was demonstrated for in silico predictions of in vivo f abs (0.1-0.6) and in vitro f u (0.4-0.5). However, it could not be ruled out that the presence of training set compounds in test sets had exaggerated the predictive power, and both models were shown to not be applicable for compounds with lower estimates. Lack of correlation between predicted and observed values and marked skewness at low permeability and f abs (large overprediction trend at low f abs ) limits the applicability of numeric in silico f abs -models, whereas lack of highly bound (f u <1 %; including many compounds with f u <LOQ) in evaluations and marked skewness (large overprediction trend at f u <25 % and underprediction trend at high f u ) limits the applicability in silico the investigated f u -prediction models in [ 17 ]. Since many modern small drugs have f u of maximally 2 % (for ca 30 % of small drugs; [ 18 ]) and/or limited permeability and f abs (for nearly every other small drug [ 18 ]), this is believed to have a major negative impact on the applicability of these models. Limited range and skewness were also evident for S and CL int,u at the lower values. The portion of non-quantifiable compounds was particularly high for CL int,u (up to every other compound). A belief that 10 µM generally corresponds to solubility/dissolution problems in vivo and that an in vitro CL int,u of 15 µL/min/mg is low and reflects good metabolic stability leads to an underestimation of dissolution potential and overestimation of metabolic stability in vivo . This might cause incorrect decision-making, compound selection and optimization strategies. The clinical relevance and validity are also poor for aqueous S and microsomal CL int,u . The R 2 s for log aqueous S vs S in simulated (FaSSIF) and real (HIF) intestinal fluids and in vivo f diss are estimated to 0.31, 0.15 and 0.13, respectively (data on file). The low R 2 s for log S reflect the differences between media regarding content, dynamics and pHs. There are also marked differences between incubated microsomes and hepatocytes in the liver in vivo . Microsomes lack an outer cell membrane and its transporter proteins and lack some binding components. They show reduced enzymatic activity (especially for conjugation) and there is considerably smaller ratio between cells/cell components (microsomes in vitro and hepatocytes in vivo ) and fluid (incubation fluid in vitro and blood in vivo ). There is also uncertainty how to compensate for binding capacity (f u,mic is used for conversion of microsome CL int to CL int,u , but its clinical relevance is questioned) and detaching rates from binding sites. The Caco-2 cell model appears closer to the in vivo situation, which is also indicated by a relatively high in vitro-in vivo correlation. Due to the importance of f u for CL H (a low f u compensates for a high CL int in vivo ) and the observation that it is generally not the compounds with highest CL int that have the highest CL H , it is our proposal that CL int • f u (or CL H ) should replace CL int as a measurement and classification of metabolic instability and that in silico models for f u -predictions include compounds with very low values. The importance of f u , steady-state volume of distribution (V ss ) and excretion for overall CL and t ½ is apparent. Some drugs with very good metabolic stability (very low in vivo CL int ; probably <LOQ for microsomes) have short t ½ (<2 h) - captopril, carboplatin, inogatran, adefovir, amoxicillin, aztreonam, cefuroxime, melagatran and metformin, to name a few. On the contrary, some drugs with very high in vivo CL int have long t ½ (days to weeks). These include clomipramine, itraconazole and amiodarone. Other drugs worth mentioning are atenolol (low in vivo CL int ; zero, low and moderate microsomal CL int,u reported [ 30 ]), diclofenac (high in vivo CL int ; very high microsomal CL int,u ; low to high hepatocyte CL int,u (57-fold difference between lowest and highest reported values) [ 30 , 43 , 44 ]), gemfibrozil (moderate in vivo CL int ; moderate microsomal CL int,u ; low to high hepatocyte CL int (138-fold difference between lowest and highest reported values) [ 30 , 43 , 44 ]), and telmisartan (very high in vivo CL int ; zero and low microsomal CL int,u (>200-fold underprediction) [ 35 ]). In vivo in man, telmisartan is a very high CL int , very low f u drug, but it appears likely that in silico models would predict limited microsomal CL int,u and much higher f u than estimated in vitro (using ADMET Predictor or models by Watanabe et al. and Ingle et al.). Earlier proposed ideal log D for ADME/PK (log D of 0-3 or 2-3) got support by the peak F found at log D=0.5 ( Figure 4 ). However, the correlation between log D and F is weak, many compounds within the range have poor F, and outside the proposed optimum zone there are several compounds with good F. An even weaker correlation was found for log D and log t ½ ( Figure 5 ). Thus, this rule of thumb is not particularly reliable for these two secondary ADME/PK parameters. The evaluation of Ro5 for classification of poor and adequate f abs (10 % f abs ) showed that this rule is skewed and produces a large portion of false negatives and a minor portion of false positives. 63 % of drugs classified to have poor absorption showed moderate to complete f abs in vivo . Thus, the application of log D-, MW- and HB-dependent Ro5 implies a substantial risk to opt out compounds with sufficiently good oral absorption potential in vivo in man. The BBB+/BBB- concept (based on rodent log BB-data and thresholds) does not seem valid and useful either. This is because the BBB appears highly permeable, at least for small drugs, log BB is highly dependent on the binding to brain and blood components, species differences in brain uptake capacity and efflux exist, and there are many CNS-active compounds with very low log BB (belonging to proposed BBB- class) [ 37 ]. K p,uu,brain is a more relevant parameter for assessing and describing uptake capacity across the BBB [ 42 ]. However, an investigation showed no correlation between rat and human K p,uu,brain -values [ 42 ]. Whether this was mainly due to experimental factors or differences in efflux transporter expression and activities is unknown. The BBB permeability and efflux (mainly MDR-1 and BCRP) and brain binding reference system developed by us, included in the ANDROMEDA prediction software, is an alternative [ 45 ]. ANDROMEDA is an alternative to the in silico to in vitro to in vivo approach. This software is mainly based on models built and trained on human clinical data and predicts directly from chemical structure to in vivo in man (without the need and use of in vitro data for S, P app and CL int,u ). Advantages compared to in vitro data-based in silico models (including ADMET Predictor) include wider application domain (high resolution at low values), higher accuracy, more parameters (including renal and biliary CL and gut-wall extraction), low/minimal skewness, high clinical relevance, lower variability/uncertainty of data used for model training and clinically more relevant thresholds. Q 2 -values (corresponding Q 2 • R 2 - values for in silico-in vitro-in vivo within parentheses) for f abs , f diss , log CL int and f u are 0.7-0.8 (0.36), 0.55 (0.05), 0.55 (0.05) and 0.5-0.55 (0.4-0.5), respectively. ADMET Predictor reached a Q 2 of 0.15 for F for compounds without absorption limitations (assumed to be <0.15 including compounds with incomplete absorption). The mean Q 2 for F with ANDROMEDA, reached in 4 different validation and benchmarking studies, was 0.49 [ 19 ]. The overall mean values for these 5 parameters are 0.57 (0.2) (3 times higher for ANDROMEDA). Corresponding prediction and application ranges with ANDROMEDA are 0-100 % (ca 0-100 %), 1-100 % (higher than 1 %, but unknown, to 100 %), 1-200,000 mL/min (roughly 50-5,000 mL/min to ca 80,000 mL/min), 0.02-100 % (ca 25-100 %) and 0-100 %, respectively. Other advantages with ANDROMEDA include additional essential parameters, such as renal and biliary CL, gut-wall extraction ratio, blood-plasma concentration ratio, V ss , t ½ , CYPID and efflux transporter ID, and unique 70 % confidence intervals predicted for each compound and parameter estimate. The latter implies that measures of confidence are produced. ANDROMEDA was very successful in predicting the human clinical ADME/PK for the small drugs released on the market in 2021 and a challenging set of compounds proposed for benchmarking and validation of prediction methods ( Figure 6 ) [ 18 ]. It predicted f diss , f abs , CL int and f u for all these drugs, and the average and maximum individual prediction error for all investigated parameters were 2.6- and 16-fold, respectively [ 18 ]. With traditional models and thresholds there is an imminent risk to incorrectly opt out compounds and select compounds that (later) will show poor ADME/PK-characteristics. That is a matter of costs, time, uncertainties and risks. These risks are considerably lower with ANDROMEDA. CONCLUSION The validity of investigated methodologies (including ADMET Predictor) and thresholds were overall very low, and much of the predictive power was lost by including and predicting via in vitro data. Unless both predicted S, permeability and f u are high and CL int is moderate (an overall criterium not met for any of the investigated modern small drugs) one is more or less lost in the translation and predictions, and this will likely lead to erroneous interpretations and decisions and jeopardizing of compound selection and optimization. It is surprising that these findings have not been shown before. This clearly shows the need for better and thoroughly validated prediction models and software. Marked improvements in accuracy, range, balance and clinical relevance were achieved with ANDROMEDA, which predicts more than 30 human clinical ADME/PK parameters directly from chemical structures and has undergone extensive validation and benchmarking. ANDROMEDA successfully predicted the human clinical ADME/PK for drugs marketed in 2021, and for which the in silico to in vitro to in vivo approach seems out of reach. REFERENCES 1. ↵ Leeson PD , Empfield JR. 2010 . Reducing the risk of drug attrition associated with physicochemical properties . Annual Reports in Med. Chem . 2. Hu , DX. Why LogD matters / ΔLogD Cheat Sheet . https://drughunter.com/resource/logdsheet . 3. ↵ Lim S. 2022 . Using Log P and Log D to assess drug bioavailability . https://ftloscience.com/log-p-log-d-drug-bioavailability/ . 4. ↵ Thomas VH , Bhattachar S , Hitchingham L , Zocharski P , Naath M , Surendran N , Stoner CL , El- Kattan A . 2006 . The road map to oral bioavailability: an industrial perspective . Expert Opin. Drug Metab. Toxicol . 2 : 591 – 608 . OpenUrl CrossRef PubMed 5. ↵ Hill AP , Young RJ . 2010 . Getting physical in drug discovery: A contemporary perspective on solubility and hydrophobicity . Drug Disc. Today 15 : 648 – 655 . OpenUrl CrossRef PubMed 6. ↵ Shah P , Siramshetty VB , Mathé E , Xu X . 2024 . Developing robust human liver microsomal stability prediction models: Leveraging inter-species correlation with rat data . Pharmaceutics 16 : 1257 . OpenUrl CrossRef PubMed 7. ↵ Podlewska S , Kafel R . 2018 . MetStabOn - Online platform for metabolic stability predictions. Intnl. J Mol . Sci . 8. ↵ Lipinski CA , Lombardo F , Dominy BW & Feeney PJ . ( 1997 ) Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings . Adv. Drug Deliv. Rev . 23 , 3 – 25 . OpenUrl CrossRef Web of Science 9. ↵ Amidon GL , Lennernäs H , Shah VP , Crison JR . 1995 . A theoretical basis for a biopharmaceutic drug classification: the correlation of in vitro drug product dissolution and in vivo bioavailability . Pharm. Res . 12 ( 3 ): 413 – 20 . OpenUrl CrossRef PubMed Web of Science 10. ↵ Sohlenius-Sternbeck A-K , Terelius Y . 2022 . Evaluation of ADMET Predictor in early discovery drug metabolism . Drug Metab. Dispos . 50 : 95 – 104 . OpenUrl Abstract / FREE Full Text 11. ADMET Predictor software by Simulations Plus . https://www.simulations-plus.com/software/admetpredictor/ . 12. QikProp ADME prediction software by Schrödinger . https://www.schrodinger.com/platform/products/qikprop/ . 13. ↵ Falcón-Cano G , Molina C , Cabrera-Pérez MA . 2020 . ADME prediction with KNIME: In silico aqueous solubility consensus model based on supervised recursive random forest approaches . ADMET & DMPK 8 ( 3 ): 251 – 273 . OpenUrl PubMed 14. ↵ Chevillard F , Lagorce D , Reynès C , Villoutreix BO , Vayer P , Miteva MA. 2012 . In silico prediction of aqueous solubility: A multimodel protocol based on chemical similarity. Mol . Pharmaceutics 9 : 3127 – 3135 . OpenUrl 15. ↵ Paixão P , Gouveia LF , Morais JAG. 2010 . Prediction of the in vitro permeability determined in Caco-2 cells by using artificial neural networks . Eur. J Pharmaceut. Sci . 41 : 107 – 117 . OpenUrl CrossRef PubMed 16. ↵ Winiwarter S , Ahlberg E , Watson E , et al. 2018 . In silico ADME in drug design - Enhancing the impact . ADMET DMPK 6 : 15 – 33 . OpenUrl CrossRef 17. ↵ Yun YE , Tornero-Velez R , Purucker ST , Chang DT , Edginton AN . 2021 . Evaluation of quantitative structure property relationship algorithms for predicting plasma protein binding in humans . Comput. Toxicol . 17 : 100142 . OpenUrl CrossRef PubMed 18. ↵ Fagerholm U , Hellberg S , Alvarsson J , Spjuth O . 2023 In silico prediction of human clinical pharmacokinetics with ANDROMEDA by Prosilico: Predictions for an established benchmarking data set, a modern small drug data set, and a comparison with laboratory methods . Altern. Lab. Anim . 51 , 39 – 54 . OpenUrl CrossRef PubMed 19. ↵ Fagerholm U , Hellberg S , Spjuth O . 2021 . Advances in predictions of oral bioavailability of candidate drugs in man with new machine learning methodology . Molecules 26 , 2572 . OpenUrl CrossRef PubMed 20. Fagerholm U , Spjuth O , Hellberg S . Comparison between lab variability and in silico prediction errors for the unbound fraction of drugs in human plasma . Xenobiot 2021 ; 51 : 1095 – 1100 . OpenUrl CrossRef 21. ↵ Fagerholm U , Spjuth O , Hellberg S . 2022 . The impact of reference data selection for the prediction accuracy of intrinsic hepatic metabolic clearance . J. Pharm. Sci . 111 , 2645 – 2649 . OpenUrl CrossRef PubMed 22. ↵ Newby D , Freitas AA , Ghafourian T . 2015 . Decision trees to characterise the roles of permeability and solubility on the prediction of oral absorption . Eur. J. Med. Chem ., 90 , 751 – 765 . OpenUrl CrossRef PubMed 23. ↵ Pham-The H , Garrigues T , Bermejo M , González-Álvarez I , Cruz Monteagudo M , Ángel Cabrera-Pérez M . 2013 . Provisional classification and in silico study of biopharmaceutical system based on Caco-2 cell permeability and dose number . Mol. Pharmaceut . 10 , 2445 – 2461 . OpenUrl CrossRef 24. ↵ Llinàs A , Avdeef A . 2019 . Solubility challenge revisited after ten years, with multilab shake-flask data, using tight (SD ∼ 0.17 log) and loos (SD ∼ 0.62 log) test sets . J Chem. Inf. Mod . 59 : 6 . OpenUrl 25. ↵ Fagerholm U. An analysis of laboratory variability and thresholds for human in vitro ADME/PK methods . bioRxiv , Sept. 2022 . 26. ↵ Skolnik S , Lin X , Wang J , Chen X-H , He T , Zhang B. 2010 . Towards prediction of in vivo intestinal absorption using a 96-well Caco-2 assay . J Pharm. Sci . 99 : 3246 – 3265 . OpenUrl CrossRef PubMed 27. Lin X , Skolnik S , Chen X , Wang J . 2011 . Attenuation of intestinal absorption by major efflux transporters: quantitative tools and strategies using a Caco-2 model . Drug Met. Disp . 39 : 265 – 274 . OpenUrl Abstract / FREE Full Text 28. Thomas S , Brightman F , Gill H , Lee S , Pufong B . 2005 . Simulation modelling of human intestinal absorption using Caco-2 permeability and kinetic solubility data for early drug discovery . J Pharm Sci . 48 : 604 – 613 . OpenUrl 29. Matsson P , Bergström CAS , Nagahara N , Tavelin S , Norinder U , Artursson , P . 2005 . Exploring the role of different drug transport routes in permeability screening . J Med. Chem . 48 : 604 – 613 . OpenUrl CrossRef PubMed 30. ↵ Fagerholm U , Lundquist P. Unpublished data. Prediction of in vivo CL int and fraction absorbed using microsomal CL int and Caco-2 P app -data . 31. ↵ Broccatelli F , Veeravalli V , Cashion D , Baylon JL , Lombardo F , Jia L . 2024 . Application of mechanistic multiparameter optimization and large-scale in vitro to in vivo pharmacokinetics correlations to small-molecule therapeutic projects . Mol. Pharm . 2 : 4312 – 4232 . OpenUrl 32. ↵ Sjögren E , Westergren J , Grant I , Hanisch G , Lindfors L , Lennernäs H , Abrahamsson B , Tannergren C. 2013 . In silico predictions of gastrointestinal drug absorption in pharmaceutical product development: Application of the mechanistic absorption model GI-Sim . Eur. J Pharmaceut. Sci . 49 : 679 – 698 . OpenUrl CrossRef PubMed 33. ↵ Stringer R , Nicklin PL , Houston JB . 2008 . Reliability of human cryopreserved hepatocytes and liver microsomes as in vitro systems to predict metabolic clearance . Xenobiot . 38 : 1313 – 1329 . OpenUrl CrossRef 34. ↵ Sohlenius-Sternbeck A-K , Afzelius L , Prusis P , Neelissen J , Hoogstraate J , Johansson J , Floby E , Bengtsson A , Gissberg O , Sternbeck J , Petersson C . 2010 . Evaluation of the human prediction of clearance from hepatocyte and microsome intrinsic clearance for 52 drug compounds . Xenobiot . 40 : 637 – 649 . OpenUrl CrossRef 35. ↵ Izumi S , Nozaki Y , Komori T , Takenaka O , Maeda K , Kusuhara H , Sugiyama Y . 2017 . Pharmacokinetics, pharmacodynamics and drug transport and metabolism comparison of the predictability of human hepatic clearance for organic anion transporting polypeptide substrate drugs between different in vitro-in vivo extrapolation approaches . J Pharm. Sci . 106 : 2678 – 2687 . OpenUrl CrossRef PubMed 36. ↵ Fagerholm U , Hellberg S , Alvarsson J , Ekmefjord M , Spjuth O. August 2024 . Comparing Lipinski’s Rule of 5 and machine learning based prediction of fraction absorbed for assessing oral absorption in humans . bioRxiv . 37. ↵ Fagerholm U . 2007 . The highly permeable blood-brain barrier: an evaluation of current opinions about brain uptake capacity . Drug Disc. Today , 12 ( 23-24 ): 1076 – 82 . OpenUrl CrossRef PubMed Web of Science 38. ↵ Lundquist S , et al. 2002 . Prediction of drug transport through the blood-brain barrier in vivo : a comparison between in vitro cell models . Pharm. Res . 19 , 976 – 81 . OpenUrl CrossRef PubMed Web of Science 39. ↵ Vilar S , Chakrabarti M , Costanzi S. 2011 . Prediction of passive blood-brain partitioning: straightforward and effective classification models based on in silico derived physicochemical descriptors . J Mol. Graph. Model . 3 ;28: 899 – 903 . OpenUrl 40. ↵ Suenderhauf , C , Hammann F , Maunz A , Helma C , Huwyler J . 2011 . Combinatorial QSAR modeling of human intestinal absorption . Mol. Pharmaceut . 8 ; 213 – 224 . OpenUrl CrossRef 41. ↵ Fagerholm U , Hellberg S , Alvarsson J , Spjuth O. Nov . 2023 . ANDROMEDA by Prosilico software successfully predicts human clinical pharmacokinetics of 300 drugs out of reach for in vitro methods . bioRxiv . 42. ↵ Fagerholm U . Evaluation of the reliability and applicability of human unbound brain-to-plasma concentration ratios . bioRxiv , Nov 2022 . 43. ↵ Bowman CM , Benet LZ . 2019 . Interlaboratory variability in human hepatocyte intrinsic clearance values and trends with physicochemical properties . Pharm Res . 36 : 113 . OpenUrl CrossRef PubMed 44. ↵ Fagerholm U. In silico predictions of the hepatic metabolic clearance in humans for 10 drugs with highly variable in vitro pharmacokinetics . bioRxiv , April 2023 . 45. ↵ Fagerholm U , Hellberg S , Alvarsson J , Spjuth O. Prediction and classification of the uptake and disposition of antidepressants and new CNS-active drugs in the human brain . bioRxiv , Sept. 2022 . View the discussion thread. Back to top Previous Next Posted February 22, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Human ADME/PK is lost in translation and prediction from in silico to in vitro to in vivo Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Human ADME/PK is lost in translation and prediction from in silico to in vitro to in vivo Urban Fagerholm , Sven Hellberg bioRxiv 2025.02.17.638712; doi: https://doi.org/10.1101/2025.02.17.638712 Share This Article: Copy Citation Tools Human ADME/PK is lost in translation and prediction from in silico to in vitro to in vivo Urban Fagerholm , Sven Hellberg bioRxiv 2025.02.17.638712; doi: https://doi.org/10.1101/2025.02.17.638712 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Pharmacology and Toxicology Subject Areas All Articles Animal Behavior and Cognition (7637) Biochemistry (17705) Bioengineering (13899) Bioinformatics (41970) Biophysics (21463) Cancer Biology (18605) Cell Biology (25526) Clinical Trials (138) Developmental Biology (13385) Ecology (19911) Epidemiology (2067) Evolutionary Biology (24329) Genetics (15615) Genomics (22514) Immunology (17743) Microbiology (40424) Molecular Biology (17194) Neuroscience (88650) Paleontology (667) Pathology (2835) Pharmacology and Toxicology (4827) Physiology (7648) Plant Biology (15160) Scientific Communication and Education (2046) Synthetic Biology (4302) Systems Biology (9825) Zoology (2271)